Profit Maximization Techniques for Operating Chemical Plants [1. ed.] 1119532159, 9781119532156

Citation preview

Profit Maximization Techniques for Operating Chemical Plants

Profit Maximization Techniques for Operating Chemical Plants Sandip Kumar Lahiri National Institute Of Technology, Durgapur, India

This edition first published 2020 © 2020 John Wiley & Sons Ltd All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted, in any form or by any means, electronic, mechanical, photocopying, recording or otherwise, except as permitted by law. Advice on how to obtain permission to reuse material from this title is available at http://www.wiley.com/go/permissions. The right of Sandip Kumar Lahiri to be identified as the author of this work has been asserted in accordance with law. Registered Offices John Wiley & Sons, Inc., 111 River Street, Hoboken, NJ 07030, USA John Wiley & Sons Ltd, The Atrium, Southern Gate, Chichester, West Sussex, PO19 8SQ, UK Editorial Office John Wiley & Sons Ltd, The Atrium, Southern Gate, Chichester, West Sussex, PO19 8SQ, UK For details of our global editorial offices, customer services, and more information about Wiley products visit us at www .wiley.com. Wiley also publishes its books in a variety of electronic formats and by print-on-demand. Some content that appears in standard print versions of this book may not be available in other formats. Limit of Liability/Disclaimer of Warranty In view of ongoing research, equipment modifications, changes in governmental regulations, and the constant flow of information relating to the use of experimental reagents, equipment, and devices, the reader is urged to review and evaluate the information provided in the package insert or instructions for each chemical, piece of equipment, reagent, or device for, among other things, any changes in the instructions or indication of usage and for added warnings and precautions. While the publisher and authors have used their best efforts in preparing this work, they make no representations or warranties with respect to the accuracy or completeness of the contents of this work and specifically disclaim all warranties, including without limitation any implied warranties of merchantability or fitness for a particular purpose. No warranty may be created or extended by sales representatives, written sales materials or promotional statements for this work. The fact that an organization, website, or product is referred to in this work as a citation and/or potential source of further information does not mean that the publisher and authors endorse the information or services the organization, website, or product may provide or recommendations it may make. This work is sold with the understanding that the publisher is not engaged in rendering professional services. The advice and strategies contained herein may not be suitable for your situation. You should consult with a specialist where appropriate. Further, readers should be aware that websites listed in this work may have changed or disappeared between when this work was written and when it is read. Neither the publisher nor authors shall be liable for any loss of profit or any other commercial damages, including but not limited to special, incidental, consequential, or other damages. Library of Congress Cataloging-in-Publication Data Names: Lahiri, Sandip Kumar, 1970- author. Title: Profit maximization techniques for operating chemical plants / Dr Sandip Kumar Lahiri. Description: First edition. | Hoboken, NJ : John Wiley & Sons, Inc., 2020. | Includes bibliographical references and index. Identifiers: LCCN 2019058766 (print) | LCCN 2019058767 (ebook) | ISBN 9781119532156 (hardback) | ISBN 9781119532217 (adobe pdf ) | ISBN 9781119532170 (epub) Subjects: LCSH: Chemical engineering–Cost effectiveness. | Engineering economy. | Profit. Classification: LCC TP155.2.C67 L34 2020 (print) | LCC TP155.2.C67 (ebook) | DDC 660.068/1–dc23 LC record available at https://lccn.loc.gov/2019058766 LC ebook record available at https://lccn.loc.gov/2019058767 Cover Design: Wiley Cover Images: © David Burton/Getty Images, © Lightspring/Shutterstock Set in 10/12pt WarnockPro by SPi Global, Chennai, India Printed and bound by CPI Group (UK) Ltd, Croydon, CR0 4YY 10 9 8 7 6 5 4 3 2 1

Dedicated to my Parents, wife Jinia and two lovely children Suchetona and Srijon

vii

Contents Figure List xix Table List xxv Preface xxvii

1.1 1.2 1.3 1.4 1.5 1.6 1.7

1 Introduction 1 Who is This Book Written for? 3 What is Profit Maximization and Sweating of Assets All About? 4 Need for Profit Maximization in Today’s Competitive Market 7 Data Rich but Information Poor Status of Today’s Process Industries 8 Emergence of Knowledge-Based Industries 9 How Knowledge and Data Can Be Used to Maximize Profit 9 References 10

2

Big Picture of the Modern Chemical Industry 11

2.1 2.2 2.3

New Era of the Chemical Industry 11 Transition from a Conventional to an Intelligent Chemical Industry 11 How Will Digital Affect the Chemical Industry and Where Can the Biggest Impact Be Expected? 12 Attaining a New Level of Functional Excellence 12 Manufacturing 13 Supply Chain 14 Sales and Marketing 14 Research and Development 15 Using Advanced Analytics to Boost Productivity and Profitability in Chemical Manufacturing 15 Decreasing Downtime Through Analytics 16 Increase Profits with Less Resources 17 Optimizing the Whole Production Process 18 Achieving Business Impact with Data 19 Data’s Exponential Growing Importance in Value Creation 19 Different Links in the Value Chain 20 The Insights Value Chain – Definitions and Considerations 21 From Dull Data to Critical Business Insights: The Upstream Processes 22

1

2.3.1 2.3.1.1 2.3.1.2 2.3.1.3 2.3.1.4 2.4 2.4.1 2.4.2 2.4.3 2.5 2.5.1 2.5.2 2.5.2.1 2.6

Concept of Profit Maximization

viii

Contents

2.6.1 2.6.2 2.7 2.7.1 2.7.2 2.7.3

Generating and Collecting Relevant Data 22 Data Refinement is a Two-Step Iteration 23 From Valuable Data Analytics Results to Achieving Business Impact: The Downstream Activities 25 Turning Insights into Action 25 Developing Data Culture 25 Mastering Tasks Concerning Technology and Infrastructure as Well as Organization and Governance 25 References 26

3

Profit Maximization Project (PMP) Implementation Steps 27

3.1 3.1.1 3.1.2 3.1.3 3.1.4 3.1.5

Implementing a Profit Maximization Project (PMP) 27 Step 1: Mapping the Whole Plant in Monetary Terms 27 Step 2: Assessment of Current Plant Conditions 27 Step 3: Assessment of the Base Control Layer of the Plant 28 Step 4: Assessment of Loss from the Plant 29 Step 5: Identification of Improvement Opportunity in Plant and Functional Design of PMP Applications 29 Step 6: Develop an Advance Process Monitoring Framework by Applying the Latest Data Analytics Tools 30 Step 7: Develop a Real-Time Fault Diagnosis System 30 Step 8: Perform a Maximum Capacity Test Run 30 Step 9: Develop and Implement Real-Time APC 31 Step 10: Develop a Data-Driven Offline Process Model for Critical Process Equipment 31 Step 11: Optimizing Process Operation with a Developed Model 32 Step 12: Modeling and Optimization of Industrial Reactors 32 Step 13: Maximize Throughput of All Running Distillation Columns 33 Step 14: Apply New Design Methodology for Process Equipment 33 References 34

3.1.6 3.1.7 3.1.8 3.1.9 3.1.10 3.1.11 3.1.12 3.1.13 3.1.14

4

4.1 4.2 4.3 4.4 4.4.1 4.4.2 4.4.3 4.5 4.6 4.7 4.7.1 4.7.2 4.7.3

35 Introduction 35 How is Operating Profit Defined in CPI? 36 Different Ways to Maximize Operating Profit 36 Process Cost Intensity 37 Definition of Process Cost Intensity 37 Concept of Cost Equivalent (CE) 39 Cost Intensity for a Total Site 39 Mapping the Whole Process in Monetary Terms and Gain Insights 40 Case Study of a Glycol Plant 40 Steps to Map the Whole Plant in Monetary Terms and Gain Insights 43 Step 1: Visualize the Plant as a Black Box 43 Step 2: Data Collection from a Data Historian and Preparation of Cost Data 46 Step 3: Calculation of Profit Margin 46 Strategy for Profit Maximization

Contents

4.7.4 4.7.5 4.7.6 4.7.7 4.7.8

Step 4: Gain Insights from Plant Cost and Profit Data 48 Step 5: Generation of Production Cost and a Profit Margin Table for One Full Year 51 Step 6: Plot Production Cost and Profit Margin for One Full Year and Gain Insights 51 Step 7: Calculation of Relative Standard Deviations of each Parameter in order to Understand the Cause of Variability 52 Step 8: Cost Benchmarking 53 Reference 54

5

Key Performance Indicators and Targets 55

5.1 5.2 5.2.1 5.2.2 5.2.3 5.2.4 5.2.5 5.2.6 5.2.7 5.3 5.3.1 5.3.2 5.3.3 5.4

Introduction 55 Key Indicators Represent Operation Opportunities 56 Reaction Optimization 56 Heat Exchanger Operation Optimization 58 Furnace Operation 58 Rotating Equipment Operation 59 Minimizing Steam Letdown Flows 59 Turndown Operation 59 Housekeeping Aspects 59 Define Key Indicators 60 Process Analysis and Economics Analysis 61 Understand the Constraints 61 Identify Qualitatively Potential Area of Opportunities 65 Case Study of Ethylene Glycol Plant to Identify the Key Performance Indicator 66 Methodology 66 Ethylene Oxide Reaction Section 67 Understand the Process 67 Understanding the Economics of the Process 68 Factors that can Change the Production Cost and Overall Profit Generated from this Section 69 How is Production Cost Related to Process Parameters from the Standpoint of the Cause and Effect Relationship? 69 Constraints 69 Key Parameter Identifications 70 Cycle Water System 71 Main Purpose 71 Economics of the Process 71 Factors that can Change the Production Cost of this Section 72 Constraints 72 Key Performance Parameters 72 Carbon Dioxide Removal Section 73 Main Purpose 73 Economics 73 Factors that can Change the Production Cost of this Section 73

5.4.1 5.4.2 5.4.2.1 5.4.2.2 5.4.2.3 5.4.2.4 5.4.2.5 5.4.2.6 5.4.3 5.4.3.1 5.4.3.2 5.4.3.3 5.4.3.4 5.4.3.5 5.4.4 5.4.4.1 5.4.4.2 5.4.4.3

ix

x

Contents

5.4.4.4 5.4.4.5 5.4.5 5.4.5.1 5.4.5.2 5.4.5.3 5.4.5.4 5.4.6 5.4.6.1 5.4.6.2 5.4.6.3 5.5 5.6 5.7 5.7.1 5.7.2 5.8

Constraints 74 Key Performance Parameters 74 EG Reaction and Evaporation Section 74 Main Purpose 74 Economics 75 Factors that can Change the Production Cost of this Section 76 Key Performance Parameters 76 EG Purification Section 76 Main Purpose 76 Economics 77 Key Performance Parameters 77 Purpose to Develop Key Indicators 77 Set up Targets for Key Indicators 78 Cost and Profit Dashboard 78 Development of Cost and Profit Dashboard to Monitor the Process Performance in Money Terms 78 Connecting Key Performance Indicators in APC 79 It is Crucial to Change the Viewpoints in Terms of Cost or Profit 80 References 80

6

Assessment of Current Plant Status 83

6.1 6.1.1 6.1.2 6.2 6.3 6.4 6.5 6.6

Introduction 83 Data Extraction from a Data Historian 83 Calculate the Economic Performance of the Section 84 Monitoring Variations of Economic Process Parameters 90 Determination of the Effect of Atmosphere on the Plant Profitability 90 Capacity Variations 91 Assessment of Plant Reliability 91 Assessment of Profit Suckers and Identification of Equipment for Modeling and Optimization 91 Assessment of Process Parameters Having a High Impact on Profit 93 Comparison of Current Plant Performance Against Its Design 93 Assessment of Regulatory Control System Performance 94 Basic Assessment Procedure 96 Assessment of Advance Process Control System Performance 97 Assessment of Various Profit Improvement Opportunities 97 References 98

6.7 6.8 6.9 6.9.1 6.10 6.11

7

7.1 7.2 7.3 7.3.1 7.3.2 7.3.3

99 Introduction 99 Problems to Develop a Phenomenological Model for Industrial Processes 100 Types of Process Model 101 First Principle-Based Model 101 Data-Driven Models 101 Grey Model 101

Process Modeling by the Artificial Neural Network

Contents

7.3.4 7.4 7.5 7.5.1 7.5.2 7.5.3 7.5.4 7.5.5 7.6 7.6.1 7.6.2 7.6.2.1 7.6.2.2 7.6.2.3 7.6.3 7.6.4 7.6.5 7.6.6 7.7 7.8 7.8.1 7.8.2 7.8.3 7.8.4 7.8.5 7.8.6 7.8.7 7.9

Hybrid Model 101 Emergence of Artificial Neural Networks as One of the Promising Data-Driven Modeling Techniques 106 ANN-Based Modeling 106 How Does ANN Work? 106 Network Architecture 107 Back-Propagation Algorithm (BPA) 107 Training 108 Generalizability 110 Model Development Methodology 110 Data Collection and Data Inspection 110 Data Pre-processing and Data Conditioning 110 Outlier Detection and Replacement 112 Univariate Approach to Detect Outliers 112 Multivariate Approach to Detect Outliers 112 Selection of Relevant Input–Output Variables 113 Align Data 113 Model Parameter Selection, Training, and Validation 113 Model Acceptance and Model Tuning 115 Application of ANN Modeling Techniques in the Chemical Process Industry 115 Case Study: Application of the ANN Modeling Technique to Develop an Industrial Ethylene Oxide Reactor Model 116 Origin of the Present Case Study 116 Problem Definition of the Present Case Study 117 Developing the ANN-Based Reactor Model 119 Identifying Input and Output Parameters 119 Data Collection 120 Neural Regression 121 Results and Discussions 122 Matlab Code to Generate the Best ANN Model 124 References 125

8

Optimization of Industrial Processes and Process Equipment 131

8.1 8.2 8.3 8.3.1 8.3.2 8.4 8.4.1 8.4.2 8.4.3 8.5

Meaning of Optimization in an Industrial Context 131 How Can Optimization Increase Profit? 132 Types of Optimization 133 Steady-State Optimization 133 Dynamic Optimization 133 Different Methods of Optimization 134 Classical Method 134 Gradient-Based Methods of Optimization 134 Non-traditional Optimization Techniques 135 Brief Historical Perspective of Heuristic-based Non-traditional Optimization Techniques 136 Genetic Algorithm 138 What is Genetic Algorithm? 138

8.6 8.6.1

xi

xii

Contents

8.6.2 8.6.3 8.6.3.1 8.6.3.2 8.6.3.3 8.6.3.4 8.6.3.5 8.6.4 8.6.5 8.6.5.1 8.6.5.2 8.6.5.3 8.6.6 8.7 8.7.1 8.7.2 8.7.3 8.7.4 8.7.5 8.8 8.8.1 8.8.2 8.8.3 8.9 8.9.1 8.10

Foundation of Genetic Algorithms 138 Five Phases of Genetic Algorithms 140 Initial Population 140 Fitness Function 140 Selection 140 Crossover 140 Termination 141 The Problem Definition 141 Calculation Steps of GA 141 Step 1: Generating Initial Population by Creating Binary Coding 141 Step 2: Evaluation of Fitness 142 Step 3: Selecting the Next Generation’s Population 142 Advantages of GA Against Classical Optimization Techniques 144 Differential Evolution 145 What is Differential Evolution (DE)? 145 Working Principle of DE 145 Calculation Steps Performed in DE 145 Choice of DE Key Parameters (NP, F, and CR) 145 Stepwise Calculation Procedure for DE implementation 146 Simulated Annealing 149 What is Simulated Annealing? 149 Procedure 149 Algorithm 150 Case Study: Application of the Genetic Algorithm Technique to Optimize the Industrial Ethylene Oxide Reactor 151 Conclusion of the Case Study 152 Strategy to Utilize Data-Driven Modeling and Optimization Techniques to Solve Various Industrial Problems and Increase Profit 153 References 155

9

Process Monitoring 159

9.1 9.2 9.3 9.4 9.5 9.6 9.6.1 9.6.2 9.6.3 9.7 9.8

Need for Advance Process Monitoring 159 Current Approaches to Process Monitoring and Diagnosis 160 Development of an Online Intelligent Monitoring System 161 Development of KPI-Based Process Monitoring 161 Development of a Cause and Effect-Based Monitoring System 163 Development of Potential Opportunity-Based Dash Board 163 Development of Loss and Waste Monitoring Systems 164 Development of a Cost-Based Monitoring System 165 Development of a Constraints-Based Monitoring System 166 Development of Business Intelligent Dashboards 166 Development of Process Monitoring System Based on Principal Component Analysis 167 What is a Principal Component Analysis? 168 Why Do We Need to Rotate the Data? 169 How Do We Generate Principal Components? 170 Steps to Calculating the Principal Components 170

9.8.1 9.8.2 9.8.3 9.8.4

Contents

9.9 9.9.1

10

10.1 10.2 10.3 10.3.1 10.4 10.5 10.6 10.6.1 10.6.2 10.7 10.7.1

Case Study for Operational State Identification and Monitoring Using PCA 171 Case Study 1: Monitoring a Reciprocating Reclaim Compressor 171 References 174 Fault Diagnosis 177

Challenges to the Chemical Industry 177 What is Fault Diagnosis? 178 Benefit of a Fault Diagnosis System 179 Characteristic of an Automated Fault Diagnosis System 180 Decreasing Downtime Through a Fault Diagnosis Type Data Analytics 180 User Perspective to Make an Effective Fault Diagnosis System 181 How Are Fault Diagnosis Systems Made? 183 Principal Component-Based Approach 184 Artificial Neural Network-Based Approach 184 A Case Study to Build a Robust Fault Diagnosis System 185 Challenges to a Build Fault Diagnosis of an Ethylene Oxide Reactor System 187 10.7.2 PCA-Based Fault Diagnosis of an EO Reactor System 187 10.7.3 Acquiring Historic Process Data Sets to Build a PCA Model 188 10.7.4 Criteria of Selection of Input Parameters for PCA 189 10.7.5 How PCA Input Data is Captured in Real Time 191 10.7.6 Building the Model 192 10.7.6.1 Calculations of the Principal Components 192 10.7.6.2 Calculations of Hotelling’s T 2 192 10.7.6.3 Calculations of the Residual 193 10.7.7 Creation of a PCA Plot for Training Data 193 10.7.8 Creation of Hotelling’s T 2 Plot for the Training Data 194 10.7.9 Creation of a Residual Plot for the Training Data 194 10.7.10 Creation of an Abnormal Zone in the PCA Plot 194 10.7.11 Implementing the PCA Model in Real Time 194 10.7.12 Detecting Whether the Plant is Running Normally or Abnormally on a Real-Time Basis 195 10.7.13 Use of a PCA Plot During Corrective Action in Real Time 197 10.7.14 Validity of a PCA Model 198 10.7.14.1 Time-Varying Characteristic of an EO Catalyst 198 10.7.14.2 Capturing the Efficiency of the PCA Model Using the Residual Plot 199 10.7.15 Quantitive Decision Criteria Implemented for Retraining of an Ethylene Oxide (EO) Reactor PCA Model 200 10.7.16 How Retraining is Practically Executed 200 10.8 Building an ANN Model for Fault Diagnosis of an EO Reactor 200 10.8.1 Acquiring Historic Process Data Sets to Build an ANN Model 200 10.8.2 Identification of Input and Output Parameters 201 10.8.3 Building of an ANN-Based EO Reactor Model 201 10.8.3.1 Complexity of EO Reactor Modeling 201 10.8.3.2 Model Building 202 10.8.4 Prediction Performance of an ANN Model 203

xiii

xiv

Contents

10.8.5 10.8.6 10.8.7 10.9 10.10

Utilization of an ANN Model for Fault Detection 203 How Do PCA Input Data Relate to ANN Input/Output Data? 204 Retraining of an ANN Model 206 Integrated Robust Fault Diagnosis System 206 Advantages of a Fault Diagnosis System 208 References 208

11

Optimization of an Existing Distillation Column

11.1 11.1.1 11.2 11.3 11.4 11.5 11.5.1 11.5.2 11.5.3 11.5.4 11.6 11.6.1 11.6.2 11.6.3 11.6.4 11.6.5 11.6.6 11.7 11.7.1 11.7.2 11.7.3 11.7.4 11.8 11.9 11.9.1 11.9.1.1 11.9.1.2 11.9.1.3 11.9.1.4 11.9.1.5 11.9.1.6 11.9.2 11.9.3 11.9.4

209 Strategy to Optimize the Running Distillation Column 209 Strategy 209 Increase the Capacity of a Running Distillation Column 210 Capacity Diagram 211 Capacity Limitations of Distillation Columns 212 Vapour Handling Limitations 214 Flow Regimes – Spray and Froth 214 Entrainment 215 Tray Flooding 215 Ultimate Capacity 217 Liquid Handling Limitations 217 Downcomer Flood 217 Downcomer Residence Time 217 Downcomer Froth Back-Up% 219 Downcomer Inlet Velocity 220 Weir liquid loading 221 Downcomer Sizing Criteria 221 Other Limitations and Considerations 221 Weeping 221 Dumping 222 Tray Turndown 222 Foaming 223 Understanding the Stable Operation Zone 223 Case Study to Develop a Capacity Diagram 224 Calculation of Capacity Limits 224 Spray Limit 224 Vapor Flooding Limit 226 Downcomer Backup Limit 226 Maximum Liquid Loading Limit 227 Minimum Liquid Loading Limit 227 Minimum Vapor Loading Limit 228 Plotting a Capacity Diagram 228 Insights from the Capacity Diagram 229 How Can the Capacity Diagram Be Used for Profit Maximization? 229 References 230

12

New Design Methodology 231

12.1 12.2

Need for New Design Methodology 231 Case Study of the New Design Methodology for a Distillation Column 231

Contents

12.2.1 12.2.2 12.3 12.4 12.5 12.6 12.7 12.8 12.9 12.9.1 12.10 12.11 12.12 12.13 12.14

Traditional Way to Design a Distillation Column 231 Background of the Distillation Column Design 232 New Intelligent Methodology for Designing a Distillation Column 234 Problem Description of the Case Study 237 Solution Procedure Using the New Design Methodology 237 Calculations of the Total Cost 238 Search Optimization Variables 239 Operational and Hydraulic Constraints 239 Particle Swarm Optimization 241 PSO Algorithm 241 Simulation and PSO Implementation 242 Results and Analysis 243 Advantages of PSO 245 Advantages of New Methodology over the Traditional Approach 246 Conclusion 248 Nomenclature 248 References 250 Appendix 12.1 251

13

Genetic Programing for Modeling of Industrial Reactors

13.1 13.2 13.3 13.4 13.5 13.6 13.7 13.7.1 13.7.2 13.7.3 13.7.4 13.7.5 13.7.6 13.7.7 13.8 13.8.1 13.8.2 13.8.3 13.9 13.9.1 13.9.2 13.9.3 13.9.4

259 Potential Impact of Reactor Optimization on Overall Profit 259 Poor Knowledge of Reaction Kinetics of Industrial Reactors 259 ANN as a Tool for Reactor Kinetic Modeling 260 Conventional Methods for Evaluating Kinetics 260 What is Genetic Programming? 261 Background of Genetic Programming 262 Genetic Programming at a Glance 263 Preparatory Steps of Genetic Programming 264 Executional Steps of Genetic Programming 264 Creating an Individual 267 Fitness Test 268 The Genetic Operations 269 User Decisions 271 Computing Resources 272 Example Genetic Programming Run 272 Preparatory Steps 273 Step-by-Step Sample Run 274 Selection, Crossover, and Mutation 275 Case Studies 277 Case Study 1 277 Case Study 2 278 Case Study 3 279 Case Study 4 280 References 281

14

Maximum Capacity Test Run and Debottlenecking Study

14.1 14.2

283 Introduction 283 Understanding Different Safety Margins in Process Equipment 283

xv

xvi

Contents

14.3 14.4 14.5 14.6 14.7 14.7.1 14.7.2 14.7.3 14.8 14.8.1 14.8.2 14.8.3 14.8.4 14.8.5 14.8.6 14.8.7 14.8.8 14.8.9 14.8.10 14.8.11 14.8.12 14.8.13 14.9 14.9.1 14.9.2 14.9.3 14.10 14.10.1 14.10.2 14.10.3

Strategies to Exploit the Safety Margin 284 Capacity Expansion versus Efficiency Reduction 285 Maximum Capacity Test Run: What is it All About? 286 Objective of a Maximum Capacity Test Run 287 Bottlenecks of Different Process Equipment 288 Functional Bottleneck 288 Reliability Bottleneck 288 Safety Interlock Bottleneck 290 Key Steps to Carry Out a Maximum Capacity Test Run in a Commercial Running Plant 291 Planning 291 Discussion with Technical People 296 Risk and Opportunity 296 Dos and Don’ts 297 Simulations 298 Preparations 299 Management of Change 299 Execution 300 Data Collections 300 Critical Observations 302 Report Preparations 303 Detailed Simulations and Assembly of All Observations 303 Final Report Preparation 304 Scope and Phases of a Detailed Improvement Study 304 Improvement Scoping Study 305 Detail Feasibility Study 305 Retrofit Design Phase 305 Scope and Limitations of MCTR 306 Scope 306 Two Big Benefits of Doing MCTR 306 Limitations of MCTR 306

15

Loss Assessment 309

15.1 15.2 15.3 15.4 15.4.1 15.4.2 15.4.3 15.4.4 15.4.5 15.4.6 15.4.7

Different Losses from the System 309 Strategy to Reduce the Losses and Wastages 309 Money Loss Audit 310 Product or Utility Losses 312 Loss in the Drain 312 Loss Due to Vent and Flaring 313 Utility Loss 314 Heat Loss Assessment for the Fired Heater 314 Heat Loss Assessment for the Distillation Column 315 Heat Loss Assessment for Steam Leakage 316 Heat Loss Assessment for Condensate Loss 317

16

Advance Process Control 319

16.1

What is Advance Process Control? 319

Contents

16.2 16.3 16.4 16.5 16.6 16.7 16.8 16.9 16.9.1 16.9.2 16.9.3 16.9.4 16.9.5 16.9.6 16.9.7 16.10 16.10.1 16.10.2 16.10.3 16.10.4 16.10.5 16.10.6 16.10.7 16.10.8

Why is APC Necessary to Improve Profit? 320 Why APC is Preferred over Normal PID Regulatory Control 322 Position of APC in the Control Hierarchy 324 Which are the Plants where Implementations of APC were Proven Very Profitable? 327 How do Implementations of APC Increase Profit? 328 How does APC Extract Benefits? 330 Application of APC in Oil Refinery, Petrochemical, Fertilizer and Chemical Plants and Related Benefits 334 Steps to Execute an APC Project 336 Step 1: Preliminary Cost –Benefit Analysis 336 Step 2: Assessment of Base Control Loops 337 Step 3: Functional Design of the Controller 337 Step 4: Conduct the Plant Step Test 338 Step 5: Generate a Process Model 338 Step 6: Commission the Online Controller 338 Step 7: Online APC Controller Tuning 339 How Can an Effective Functional Design Be Done? 339 Step 1: Define Process Control Objectives 340 Step 2: Identification of Process Constraints 342 Step 3: Define Controller Scope 343 Step 4: Variable Selection 344 Step 5: Rectify Regulatory Control Issues 346 Step 6: Explore the Scope of Inclusions of Inferential Calculations 347 Step 7: Evaluate Potential Optimization Opportunity 347 Step 8: Define LP or QP Objective Function 348 References 349

17

150 Ways and Best Practices to Improve Profit in Running Chemical Plant 351

17.1

Best Practices Followed in Leading Process Industries Around the World 351 Best Practices Followed in a Steam and Condensate System 351 Best Practices Followed in Furnaces and Boilers 355 Best Practices Followed in Pumps, Fans, and Compressor 357 Best Practices Followed in Illumination Optimization 359 Best Practices in Operational Improvement 359 Best Practices Followed in Air and Nitrogen Header 360 Best Practices Followed in Cooling Tower and Cooling Water 361 Best Practices Followed in Water Conservation 362 Best Practices Followed in Distillation Column and Heat Exchanger 363 Best Practices in Process Improvement 364 Best Practices in Flare Gas Reduction 365 Best Practices in Product or Energy Loss Reduction 365 Best Practices to Monitor Process Control System Performance 366 Best Practices to Enhance Plant Reliability 367

17.2 17.3 17.4 17.5 17.6 17.7 17.8 17.9 17.10 17.11 17.12 17.13 17.14 17.15

xvii

xviii

Contents

17.16 17.17 17.18 17.19

Best Practices to Enhance Human Resource 368 Best Practices to Enhance Safety, Health, and the Environment 368 Best Practices to Use New Generation Digital Technology 369 Best Practices to Focus a Detailed Study and R&D Effort 370 Index 373

xix

Figure List Figure 1.1

Various constraints or limits of chemical processes 5

Figure 1.2

Optimum operating point versus operator comfort zone

Figure 2.1

Developing stages of the chemical industry 12

Figure 2.2

Three major ways digital transformation will impact the chemical industry 13

Figure 2.3

Three major impact areas where advance analytic tools will help to increase profit 16

Figure 2.4

Different components of the insights value chain

Figure 2.5

Overview of the insights value chain upstream processes (A–B) and downstream activities (D–E) 21

Figure 2.6

Data science is an iterative process that leverages both human domain expertise and advanced AI-based machine learning techniques 24

Figure 3.1

Different steps in profit maximization project (PMP) implementation 28

Figure 4.1

Different ways to maximize the operating profit of chemical plants 36

Figure 4.2

Schematic diagram of a glycol plant 41

Figure 4.3

Steps to map the whole plant in monetary terms and to gain insights 43

Figure 4.4

Representing the whole plant as a black box

Figure 4.5

Mapping the whole plant in monetary terms 47

Figure 4.6

Break-up of the total cost of production

Figure 4.7

Cost of raw material 50

Figure 4.8

Cost of different utilities (USD/h)

Figure 4.9

Cost of different chemicals (USD/h) 51

Figure 4.10

Variations of profit margin (USD/h) throughout the year 51

Figure 4.11

Variations of profit margin (USD/MT of product) throughout the year 52

Figure 4.12

Variations of production cost (USD/MT) throughout the year

Figure 4.13

Variations of MEG production (MT/h) throughout the year 52

6

21

44

49

50

52

xx

Figure List

Figure 5.1

Five-step process of a key parameter identification 60

Figure 5.2

Queries normally asked to perform a process analysis and economic analysis of a whole plant 61

Figure 5.3

Major six categories of limitations in a plant to increase profit

Figure 5.4

Some examples of process limitations 63

Figure 5.5

Some examples of equipment limitations 64

Figure 5.6 Figure 5.7

Examples of instrument limitations 64 Guideline questionnaires to initiate the discussion with plant people 65

Figure 5.8 Figure 6.1

Various causes of catalyst selectivity increase 70 Comparison of daily actual profit (sorted) versus best achieved profit in US$/h terms for one year of operation 86

Figure 6.2 Figure 6.3

Daily opportunity loss in million US$ for one year of operation 86 Cumulative opportunity loss in million US$ for one year of operation 86

Figure 7.1 Figure 7.2

Advantage and disadvantage of the first principle-based model 102 Advantages and disadvantages of data-driven models 103

Figure 7.3

Advantages and disadvantages of the grey modeling technique 104

Figure 7.4

Advantages and disadvantages of the hybrid modeling technique 105

Figure 7.5

Typical pseudo code of a back-propagation algorithm 109

Figure 7.6

Architecture of a feed-forward network with one hidden layer 109

Figure 7.7 Figure 7.8

Steps followed in data collection and data inspection 111 Task performed in the data pre-processing and data conditioning step 112

Figure 7.9 Figure 7.10

Two main univariate approaches to detect outliers 112 Guidelines for selection of the relevant input output variables 114

Figure 7.11

Relation between catalyst selectivity and promoter concentration in a commercial ethylene oxide reactor for the latest generation high selectivity catalyst 118

Figure 7.12

Actual selectivity versus ANN model predicted selectivity 122

Figure 7.13

Prediction error percent between actual selectivity and predicted selectivity 122

Figure 7.14

Plot of actual selectivity versus predicted selectivity for testing and training data 123

Figure 7.15

ANN model performance for testing and training data 123

Figure 7.16

Different ANN algorithms developed by different scientists in the last 30 years 124

Figure 7.17

Different activation functions used in an ANN 124

Figure 8.1

Different minimum values of a function depending on different starting points 135

62

Figure List

Figure 8.2

Principle features possessed by a genetic algorithm

Figure 8.3

Foundation of the genetic algorithm 139

Figure 8.4

Five main phases of a genetic algorithm 140

Figure 8.5

Mechanism of crossover 143

Figure 8.6

Calculations steps performed in DE 146

Figure 8.7

Schematic diagram of DE 147

Figure 8.8

Calculation sequence of a simulated annealing algorithm 151

Figure 9.1

Cause and effect relationship of a steam increase in the distillation column 164

Figure 9.2

KPI-based process monitoring 166

Figure 9.3

Projection of a three-dimensional object on a two-dimensional plane 168

Figure 9.4

Projection of a three-dimensional object on a two-dimensional principal component plane 169

Figure 9.5

Projection of data towards a maximum variance plane 169

Figure 9.6

Steps to calculating the principal components

Figure 9.7

Normal and abnormal operating zones are clearly different when plotted on the first three principal component planes 172

Figure 9.8

Trends of the first principal component

Figure 9.9

Variance explained by the first few principal components

Figure 9.10

Front end to detect abnormality in the reciprocating compressor

Figure 9.11

Normal and abnormal data projected onto the first two and first three principal component planes 174

Figure 10.1

New business challenges versus improve performance 178

Figure 10.2

Pyramid of a process monitoring system 178

Figure 10.3

Fault diagnosis system 179

Figure 10.4

Characteristics of an automated real–time process monitoring system 180

Figure 10.5

Concerns when building an effective fault diagnosis system 182

Figure 10.6

Different requirements of different stakeholders from fault diagnosis software 182

Figure 10.7

Summary of user perspective and challenges to build an effective fault diagnosis software 183

Figure 10.8

Principal component plot

Figure 10.9

Schematic of an ethylene oxide reactor and its associated unit 186

139

171

173 173

184

Figure 10.10 EO reactor process parameters along with a schematic 187 Figure 10.11 Various challenges to develop an EO reactor fault diagnosis 188 Figure 10.12 Chloride versus catalyst selectivity plot 2

174

189

Figure 10.13 PCA scores plot, T plot, and residual plot

193

xxi

xxii

Figure List

Figure 10.14 Interface between a data historian and a dedicated PC loaded with PCA and ANN software 195 Figure 10.15 Contribution plots of 15 variables 197 Figure 10.16 Dynamic movement of the reactor status from the normal zone to the overchloride zone 198 Figure 10.17 Steps to build a PCA-based fault diagnosis system 199 Figure 10.18 Actual versus ANN model predicted selectivity and equivalent ethylene oxide (EOE) 205 Figure 10.19 Integrated robust fault diagnosis system 207 Figure 11.1 Effect of tower loading on the tray efficiency valve versus sieve tray 210 Figure 11.2 Capacity diagram or feasible operating window diagram 211 Figure 11.3 Vapor liquid flow pattern on the tray 213 Figure 11.4 Froth regime versus spray regime operation 214 Figure 11.5 Jet flooding and its impact on entrainment and tray efficiency 216 Figure 11.6 Downcomer choking 218 Figure 11.7 Vapor recycle increases the vapor load 218 Figure 11.8 Downcomer filling 219 Figure 11.9 Effect of weeping on efficiency 222 Figure 11.10 Operational guide for deriving the operating window 225 Figure 11.11 Capacity diagram of the case study 228 Figure 12.1 Operating limits of a distillation column tray 239 Figure 12.2 Various constraints need to be satisfied during a distillation column design 240 Figure 12.3 Various downcomer-related constraints need to be satisfied during distillation column design 240 Figure 12.4 Various process constraints need to be satisfied during distillation column design 241 Figure 13.1 Some chemical engineering applications of genetic programming 262 Figure 13.2 Five major preparatory steps for the basic version of genetic programming that the human user is required to specify 264 Figure 13.3 Flow chart of genetic programming 266 Figure 13.4 A typical individual that returns 5(x + 7) 267 Figure 13.5 Two-offspring crossover genetic operation 270 Figure 13.6 Example of sub-tree mutation 271 Figure 13.7 Initial population of four randomly created individuals of generation 0 275 Figure 13.8 Fitness of the evolved functions from generation 0 275 Figure 13.9 Population of generation 1 (after one reproduction, one mutation, and one two-offspring crossover operations) 276

Figure List

Figure 14.1 Figure 14.2 Figure 14.3 Figure 14.4

Different ways to increase plant throughput 284 Schematic diagram of strategy 2 of the maximum capacity test run 293 Schematic diagram of strategy 3 of the maximum capacity test run 294 Schematic diagram of strategy 4 of the maximum capacity test run 295

Figure 16.2

Different low grade heat recovery options 311 Flow scheme of a simple cracking furnace using an advance process controller 321 Hierarchy of the plant-wide control framework 325

Figure 16.3 Figure 16.4 Figure 16.5

Features of potential plants for APC implementation 328 Capital investment versus benefits for different levels of controls Typical benefits of APC 329

Figure 16.6

APC stabilization effect can increase plant capacity closer to its maximum limit 331

Figure 16.7

Reduced variability allows operation closer to constraints by shifting the set point 331 Operating zone limited by multiple constraints 332

Figure 15.1 Figure 16.1

Figure 16.8

328

Figure 16.9 Typical intangible benefits of APC 334 Figure 16.10 Typical payback period of APC 335 Figure 16.11 Typical benefits of APC implementation in CPI 335 Figure 16.12 Advance control implementations by one of the major APC vendors 336 Figure 16.13 Spread of APC application across the whole spectrum of the chemical process industries 336 Figure 16.14 Different steps in the APC implementation project 337 Figure 16.15 Steps in the functional design stage 339

xxiii

xxv

Table List Table 2.1 Table 4.1 Table 4.2 Table 6.1 Table 6.2 Table 6.3 Table 6.4 Table 7.1 Table 8.1 Table 8.2 Table 8.3 Table 8.4 Table 9.1 Table 10.1 Table 10.2 Table 10.3 Table 10.4 Table 11.1 Table 11.2 Table 12.1 Table 12.2 Table 12.3 Table 12.4 Table 12.5 Table 12.6 Table 13.1

Comparisons between smart and conventional chemical industries 13 Representing the whole plant as a black box with consumption and cost data 45 Summary of profit margin and cost intensity 49 Table to calculate production cost, cost intensity, profit, and profit/MT of product 85 Table to relate production cost, cost intensity, with key parameters 88 Plant reliability assessment 92 Typical performance of control loops in industry 95 Input and output variables for the ANN model 120 Initial Population of x1 and x2 and Their Fitness 148 Mutation and Crossover 148 New Generation Populations 149 Optimum Value of Input Variables Corresponding to the Maximum Value of Selectivity 152 Performance Parameter for Major Process Equipment 162 Input Parameters of a PCA-based EO Reactor Model 191 Input and output parameters of an ANN-based EO reactor model 202 Prediction performance of an ANN model 204 Comparison of PCA and ANN input data 206 Conditions of the Most Constrained Tray 224 Tower and Plate Dimensions 225 Simulation results 235 Simulation results for different feed tray locations 235 Optimization variables with their upper and lower limits 236 Different constraints and their limits 236 Optimal column geometry using improved PSACO methods 244 Value of constraints corresponding to the optimum solution 245 Examples of primitives used in GP functions and terminal sets 273

xxvi

Table List

Table 13.2

Best model generated by the GP algorithm and corresponding RMS error 278

Table 13.3

Best model generated by the GP algorithm and the corresponding RMS error 279

Table 15.1

List of Process Coolers (Water Cooler and Fin Fan Air Cooler) along with Their Duty and Money Lost 311 Calculation of Money Loss 312

Table 15.2 Table 15.3

Table to Estimate the Money Lost from an Entire Plant Due to the Drain 312

Table 15.4

Table to Estimate the Money Lost from an Entire Plant Due to Vent and Flaring 313 Typical benefits of APC implementation in refinery 335

Table 16.1

xxvii

Preface In chemical process industries there is an ongoing need to reduce the cost of production and increase the profit margin. Due to cut-throat competition at the global level, the major chemical process industries are now competing to optimize raw material and utility consumption, to increase equipment and process performance, to reduce emissions, and to minimize pollution. Profit maximization is the buzzword of today’s chemical process industries. Profit maximization in running chemical plants itself is a huge challenge, which needs to be addressed by holistic vision and procedures. However, there are no dedicated books available to discuss basic concepts, provide practical methods, and explain industrial application procedures. This book is written to fill this gap with the following people in mind: practicing process or chemical engineers, production engineers, supervisors, senior technicians working in chemical, petrochemical, pharmaceuticals, paper and pulp, oil and gas companies, and petroleum refinery across the globe. This book will also become very useful for large numbers of managers, general managers, top-level senior executives, and senior technical service consultants, whose main jobs include strategic planning and implementation of various optimization projects to increase profit in chemical process industries. Undergraduate and postgraduate chemical engineering students and business students who want to pursue careers in the chemical field will also greatly benefit from this book. The book is aimed at providing practical tools to people who face challenges and wish to find opportunities for improving profit in running chemical plants. It aims to convey concepts, theories, and methods in a straightforward and practical manner. This book provides engineers in all practical aspects of a profit maximization project in running plants, as well as expert guidance on how to derive maximum benefits. The book will present the core of a systematic approach covering profit optimization strategy, solution methodology, supporting structure, and assessment methods. In short, it will describe what it takes to make sizable reductions in operating costs for process plants and how to sustain profit improvement benefits. Short on theory and long on step-by-step information, it covers everything plant process engineers and technical managers need to know about identifying, building, deploying, and managing profit improvement applications in their companies. Readers are able to take away methods and techniques for identifying, analysis, optimization, engineering design, and monitoring that are required to identify, assess, implement, and sustain profit improvement opportunities.

xxviii

Preface

The main feature of this book, which differentiates it from other available books on the market, is its practical content, which helps the reader to understand all the steps of profit maximization project implementation in an actual commercial plant. The key features of this book that differentiate it from other available chemical engineering books are summarized below: • The reader can develop a thorough understanding of steps for building a profit maximization application in running a chemical plant. All practical considerations to identify, build, and deploy a profit improvement project in the commercial running of the plant form the essence of this book. • The benefits of this effective approach include identification of large profit improvement projects by applying assessment methods, capturing hidden opportunities in process operation by the use of advance monitoring and fault diagnosis, increasing plant capacity by a systematic way of performing a test run and debottlenecking study, optimizing process performance through various online conventional and stochastic optimization procedures, pushing the plant operation towards multiple constraints by advance process control, and maintaining continuous improvement by using regular review and performance matrices.

Overview of Contents The chapter contents are described below. Concept of Profit Maximization

The first chapter contains the foundation of the profit maximization project in running process industries. Sweating of assets and deriving maximum benefit from assets forms the essence of profit maximization. After implementation of data historian software in the last decade, todays chemical process industries (CPI) are very data rich but unfortunately remain information poor. No effective platform is still available to utilize this large amount of data. This chapter explains the emergence of knowledge-based industries and only CPIs employing knowledge to drive the business are likely to survive in the future. This essentially means generating an effective platform that can generate knowledge from available business data and use this knowledge to develop a unified framework to support faster business decisions to respond to external market uncertainties. This chapter gives an overview of how to build a framework where advanced computational knowledge and experience-based heuristics are applied to utilize this wealth of data to maximize profit. In simple terms, profit maximization means maximization of dollar ($)/h generation from the plant while subject to constraints that all process and safety constraints need to be honored and all equipment limitations should not be violated. The need for profit maximization in today’s competitive market is explained in this chapter. Big Picture of the Modern Chemical Industry

Currently the chemical industry is slowly entering into a new era called the data analytics and artificial intelligence stage, commonly known as industry 4.0. Disruptive technologies like artificial intelligence, machine learning, big data analytics, and the internet of

Preface

things (IoT) have already entered the chemical process industries and have started to change the rules governing the chemical business. Chapter 2 explains how the transition from a conventional to an intelligent chemical industry is slowly taking place. Their influence is starting to see benefits in a significant improvement in production efficiency, energy utilization, optimization of the entire manufacturing process, integration of the supply chain, new product development, product delivery speed, etc. As of now, it is quite clear that digital will have a significant impact on many areas of the chemical industry, with the gains in manufacturing performance potentially among the largest. This chapter gives an overview of how digital will affect the chemical industry and where the biggest impact can be expected. There are three major areas where applications of an advanced analytic tool can give an enormous profit increase, namely predictive maintenance; yield, energy, and throughput analytics; and value-maximization modeling. This chapter gives insights about how to achieve a business impact using data and introduces the concept of how valuable data analytics and upstream and downstream activities can result in achieving a business impact. Profit Maximization Project (PMP) Implementation Steps

Chapter 3 describes different steps for implementing a profit maximization project. It introduces 14 major broad ideas or steps for profit maximization in running commercial plants. These ideas are described in detail in subsequent chapters throughout the book. These generic steps are holistic and can be applied in any process industry, starting from refinery, petrochemical, chemical plants, metals, pharmaceuticals, paper and pulp industries, etc. It starts with mapping the whole plant in monetary terms (US$/h) instead of flow terms. This gives an idea of where to focus maximization of the profit and what low hanging fruits are needed that can be easily translated to increase profit without much investment. Practical guidelines to build a profit maximization framework, easily implementable solutions, numerous examples, and case studies from industries give a completely new computational approach to solve process industry problems and are the hallmark of this book. Strategy of Profit Maximization

A strategy of profit maximization is the essence of Chapter 4. This chapter describes different ways to maximize the operating profit. The concept of process cost intensity and how to calculate it are introduced in this chapter. The procedure for mapping the whole process in monetary terms and gain insights is described by way of an ethylene glycol plant case study. This chapter describes in detail eight key steps in mapping current process conditions against different process constraints and limits. The first three major steps are (i) define plant business and economic objectives, (ii) identify various process and safety limitations, and (iii) critically identify the profit scope. Key parameter identification steps for economics, operations, and constraints of the plant are discussed in detail. How to evaluate and exploit potential optimization opportunity is discussed with industrial case studies. Key Performance Indicators and Targets

Knowing what key operating parameters to monitor and defining the targets and limits for these parameters is an important step for profit optimization. We also need to

xxix

xxx

Preface

know the economic values of closing gaps between actual and targeted performances to create incentives for improvement. This chapter deals with how to identify the key performance parameters in running the plant and the whole process is explained with a real-life commercial plant case study. It provides a methodology to identify qualitatively potential areas of opportunities. The system of key indicators is the cornerstone of a sustainable profit management system. Assessment of Current Plant Status

An assessment of current plant status and know where you are is the first major step in building a profit maximization project. This chapter deals with the holistic approach to assess the current plant status. How to assess the performance of the base regulatory control layer and the advance process control layer of running a plant is discussed in detail in this chapter. A performance assessment of the major process equipment and an evaluation of the economic performance of the plant against a benchmark are two key focus areas discussed in this chapter. An assessment of profit suckers and identification of equipment for modeling and optimization and an assessment of process parameters having a high impact on profit are two takeaways in this chapter. Readers are enlightened with an assessment of various profit improvement opportunities. Process Modeling by an Artificial Neural Network

Chapter 7 emphases the need for data-driven black box and grey box modeling techniques where building of a first principle-based model is infeasible or time consuming due to the complexity of the industrial equipment. How an artificial neural network (ANN) can be utilized as an effective tool of black box modeling in an industrial context is discussed in this chapter with various real-life applications. A step-by-step procedure to build an ANN-based modeling platform to utilize a large amount of process data is explained in detail with example calculations. The new horizon of modeling process performance parameters like selectivity, yield, and efficiency and how these models can be utilized to increase profit is explained here. Different examples and case studies of ANN models already applied in diverse fields of process industries are illustrated to give the reader a feel for large scope and potential of applications of the ANN in industry. Optimization of Industrial Processes and Process Equipment

Due to cut-throat competition in business, companies now want to reduce their operating costs by optimizing all of their available resources, be it man, machine, money, or methodology. Optimization is an important tool, which can be utilized to strike a proper balance so that profit can be maximized in the long run. Since capital cost is already incurred for a running plant, optimization essentially boils down to minimization of the operating cost for the operating plants. In running a chemical plant, there is a huge scope to optimize the operating parameters, like temperature, pressure, concentration, reflux ratio, etc., which gives either a higher profit through higher production or lower operating costs. There are many ways to optimize the operating conditions of reactors, distillation columns, absorbers, etc., to enhance their profitability. Chapter 8 lays the foundation about how parameter optimization can be utilized to increase

Preface

profit in running the chemical plant. Conventional optimization techniques are initially discussed to enlighten the reader about the scope and huge potential of optimization in the process industry. This chapter introduces new advanced Meta heuristic optimization techniques that can be applied where application of a conventional technique is limited due to the complexity of the industrial context. Different Meta heuristic optimization techniques, like the genetic algorithm (GA), differential evolution (DE), simulated annealing (SA), etc., are described in detail in this chapter. A basic algorithm, step-by-step procedure to develop an optimization technique and different uses of GA, DE, and SA in various fields of process optimization are explained here in order to develop an understanding of this new area. A case study in reactor optimization is illustrated to explain the advantage and ease of implementation of Meta heuristic methods over conventional methods. Process Monitoring

Today’s complex chemical plants need advanced monitoring and control systems to quickly identify the suboptimal operation of process equipment and implement a quick optimization strategy. Running the plant at the highest possible capacity for profit maximization necessitates the development of an intelligent real-time monitoring system. However, due to the large amount of process data, it is a herculean task to monitor each and every piece of process data. Chapter 9 enlightens the readers about an online intelligent monitoring system, KPI-based process monitoring, a cause and effect-based monitoring system, etc. It also gives an idea regarding the development of a potential opportunity-based dashboard, loss and waste monitoring systems, a cost-based monitoring system, a constraints-based monitoring system, and how all these can be integrated into business intelligent dashboards. In this chapter, a new advanced computational technique, namely principal component analysis (PCA), is discussed to visualize data. The advantage of such an online monitoring system is to visualize the plant condition from a higher level but with a lower dimension space. A step by step procedure to build a PCA-based advance monitoring system is explained in detail, with examples and industrial case studies. Fault Diagnosis

Chemical industries recently discovered that a large amount of profit becomes eroded due to unplanned shutdowns of the plant. Due to spurious trips of equipment much potential profit is lost. One major ingredients of profit maximization is to increase plant reliability and running hours. Plant shutdown can be avoided by building a robust fault diagnosis system that will detect and alert the operator about any potential event that can lead to plant disturbance and eventually plant shutdown before it starts happening. How a robust fault diagnosis system can be made by PCA and ANN that can be implemented in industry is discussed in detail in this chapter with industrial case studies. Different aspects of enhancement of plant reliability by an advance monitoring and fault diagnosis system is the main focus of the chapter. Optimization of the Existing Distillation Column

Often distillation columns cause a bottleneck to increase plant capacity. It is very important to understand the operation and capacity limits of distillation columns in

xxxi

xxxii

Preface

commercial plants. Chapter 11 enlightens the reader about how to evaluate a feasible operating window by using a capacity diagram. Calculations based on the capacity diagram and the effect of different design and operating variables on the capacity diagram are explained in detail with example calculations. This chapter enlightens the reader about operating profile assessment, tower rating assessment, tower efficiency assessment, and hydraulic performance evaluations of running distillation columns. It also provides practical guidelines regarding what to look for in distillation column optimization in an industrial context and explains the whole concept with real-life case studies.

New Design Methodology

Due to intense competition among chemical industries across the globe, it is now absolutely necessary to minimize the cost of equipment during the design phase. Equipment costs consist of the initial capital cost of the equipment and the operating costs of the equipment. Due to the availability of a faster computer, it is now feasible to design one million different design configurations for any equipment. It is important to choose the lowest cost equipment among those one million options, but one that also obeys all of the constraints of operation, safety, maintainability, etc. Hence, to survive in today’s cut-throat competition, it is necessary to put the minimization of equipment cost as the main design target and an optimization algorithm is required to search all feasible design configurations to arrive at a minimum cost design quickly. This gives rise to a new design methodology of process equipment. Earlier traditional design methodology, where cost is not considered as a design target during the design phase, no longer produces a competitive design. In this chapter, a new design methodology of a plate-type distillation column is considered as a case study to show the essence of the new design methodology. This chapter evolves a strategy to optimize various tray geometric parameters, like tray diameter, hole diameter, fractional whole area, downcomer width, etc., and also decides on the optimum feed tray location based on the overall cost minimization concept by particle swarm optimization techniques.

Genetic Programing for Modeling of Industrial Reactors

Industrial reactors are the most potential candidates used to increase profit, yet they are the most neglected in the optimization project in industry. This is due to fear of process engineers to change reaction parameters beyond their usual boundaries because of poor knowledge of reaction kinetics. Conventional methods for evaluating complex industrial reaction kinetics have their own limitations. Chapter 13 introduces a completely new advanced computational technique, namely geometric programing (GP), to model industrial reaction kinetics. Being a new computational technique, the main advantage of GP is that process engineers do not have to assume any form of kinetic equation beforehand; it will be generated on its own from available industrial reactor data. The theoretical basis of GP with its various features, an algorithm of GP, and different case studies are discussed in detail to enlighten the reader about this new technique. How a generated kinetic model can be used online and offline to increase profit from an industrial reactor is described in detail through case studies.

Preface

Maximum Capacity Test Run and Debottlenecking Study

All over the world, chemical plants are running 100–140% of their installed nameplate capacity. Profit maximization by capacity enhancement is a very common route in process industries. This high capacity running more than their design capacity is possible due to inherent safety margins available in process equipment. This chapter helps readers to understanding different safety margins available in process equipment and explains the strategy to exploit those margins. Chapter 14 explains in a systematic way how to increase plant capacity without affecting safety and reliability. Plant real bottlenecks and the potential opportunity to increase capacity can be evaluated by actually performing a demo test run of the plant with the highest possible capacity for one week, which is commonly known as a maximum capacity test run. Fourteen key steps to carry out a maximum capacity test run in commercial running plants is the key focus area of this chapter. All the steps of the maximum capacity test run are explained with real-life examples and industrial case studies. The next step of the maximum capacity test run is to find real bottlenecks of the plant by a debottlenecking (DBN) study. How to carry out a DBN study from the existing data and its different aspects are discussed in this chapter. Loss Assessment

One of the strategies of profit maximization is to reduce the wastage of resources. In other words, this essentially means to minimize all the money drain from the plant. Chapter 15 enlightens the reader about various strategies used to reduce the losses and wastages. A step-by-step procedure for a money loss audit is explained in detail in this chapter. Advance Process Control

Advance process control (APC) already established itself as an effective tool to increase profit by pushing the process at its constraints. What is APC, how APC brings benefits, a candidate for APC implementation, and typical benefits obtained from commercial APC applications across the world are the subjects described in Chapter 16 to enlighten a new reader about the potential of APC. A step-by-step procedure to build an effective APC application is discussed in detail. Special emphasis is given to describe the most important step, namely the functional design step. Do’s and Don’ts of APC application buildings are explained so that readers are aware how to develop a robust application. 150 Ways and Best Practices to Improve Profit in Running a Chemical Plant

There are a million ways to increase profit. Leading refineries, petrochemicals, and chemical companies around the world have developed some best practices on their journeys for continuous improvement. These best practices are not confined to some process area and discipline, but are in diverse fields. The author has visited many fortune 500 chemical companies around the world and collected a list of these best practices. Due to brevity, only some of the high-impact best practices are shared here. Leading chemical companies had discovered these best practices as part of

xxxiii

xxxiv

Preface

their learning experience spanning over decades. When they implemented these best practices in their plants, they found huge profit improvements. Chapter 17 gives an overview of 150 best practices followed by leading process industries around the world. All of the best practices discussed here are implemented and followed by some of the leading companies. The practicing process engineers or production engineers working in chemical plants normally have much less exposure or knowledge to implement a site-wide holistic profit maximization project. The available books on the market on chemical engineering and process engineering do not cover the practical aspects of how to increase profit in a running commercial chemical plant. The available books on chemical engineering optimization and energy improvement place emphasis on unnecessary theoretical details, which normally are not required by practicing engineers and those theories have very little relevance for commercial implementation of profit maximization in a running plant. Most of the books cover theoretical aspects and unnecessary detailed calculations mainly for the design of process equipment. Few books address optimization in chemical plants as their focus is on development of a superior theoretical optimization algorithm and completely ignore the real-life application side. Every CEO in chemical plants understands that profit maximization in running a chemical plant is a vast subject and should be tackled holistically. There are various ways and means with varying degrees of usefulness to increase profit in operating plants. Profit can be increased by reducing waste, by optimizing processes, by increasing equipment performance, by improving reaction selectivity/yield/efficiency, by reducing raw material and utility consumption, by pushing the process towards its constraints, by running the plants at the highest possible capacity, by installing low-cost high-efficacy equipment, etc. Most of the available books cover the whole subject only partially. Some books cover energy optimization, some deal with waste minimization, and some introduce low-cost design of equipment. However, no book is available on the market to cover the whole gamut of the subject holistically and in this context this will be the first such book. It provides engineers with all practical aspects of the profit maximization project in running plants, as well as expert guidance on how to derive maximum benefits from running the plants. Clearly, it was not a small effort to write the book, but the absence of such a practical oriented book on the market and its requirement in a large number of process industries spurred me to writing it. I had an opportunity to work with leading petrochemical plants across the globe in the last 28 years was fortunate to see a wide spectrum of profit improvement initiatives taken by various fortune 500 companies. I have tried to incorporate all my learning and global experience in this book. I would like to thank Mr Mansoor Husain of M/s Scientific Design, USA, for teaching and exposing me to the practical field of profit improvement. Finally, I am truly grateful to my family, my wife Jinia and my two lovely children Suchetona and Srijon, for their understanding and generosity of spirit in tolerating my absence during the writing of this book. Dr Sandip Kumar Lahiri January, 2020

1

1 Concept of Profit Maximization 1.1 Introduction There has been a drastic change of business environment in the last 20 years. Shrinking profit margins in chemical process industries (CPIs) due to globalization and an uneven level playing field in international chemical businesses have given rise to cut-throat competition among process industries and has changed the global chemical business scenario (Lahiri, 2017b). Introduction of low cost technology in the market, cheap oil prices, a decline in the growth rate of Chinese and EU economies, the recent discovery of cheap US shale gas, for example, have added new dimensions in the business environment in recent years. Modern process industries are experiencing the following new challenges in their businesses: • Enforcement of stringent pollution control laws. • Pressure from government agencies to change to more energy efficient processes and equipment. • Constant encouragement from government to shift to safer and less pollutant processes and technology. • Decline of sales prices of end products in the international market. • More focus on sustainability and more reliable processes. All these issues are forcing the process industries to look for new technological innovations so that new ways of doing business can be explored. Profit maximization is at the core of every chemical company’s vision and mission nowadays. Making money by safely producing chemicals and selling locally are no longer adequate to survive in today’s business environment. Maximization of profit, continuous improvement of operation, sustainability, and enhanced reliability to reduce production cost are buzzwords in today’s CPI. Industries are slowly shifting their priorities to energy efficiency, real-time process optimization, environment friendliness, and sustainability. Running the plant at their highest feasible capacity by exploiting the margin available in process equipment is no longer a luxury but a necessity. Maximizing the profit margin by reducing waste products, by increasing mass transfer and energy efficiency of equipment, and by pushing the process to their physical limit are the current trends of CPI (Lahiri, 2017b).

Profit Maximization Techniques for Operating Chemical Plants, First Edition. Sandip Kumar Lahiri. © 2020 John Wiley & Sons Ltd. Published 2020 by John Wiley & Sons Ltd.

2

1 Concept of Profit Maximization

Another challenge process industries are facing is the unprecedented fast rate of obsolescence. Chemical process technology and equipment are becoming obsolete at a very fast rate. New generation energy efficient processes, environment friendly low waste generated processes, and higher selectivity catalysts are coming to market every year and pose a survival threat to older plant and technology. Simple distillation columns that have dominated the process industries and refineries for the last few decades are no longer remaining competitive. Reactive distillation columns and low energy intensive membrane technology are slowly replacing the older technology. Apart from profit maximization, business objectives also venture out to the fields of sustainability, reliability, low environment impact, etc. With the advent of faster computers, online optimization, advance process control, real-time plant monitoring, online equipment fault diagnosis, big data-based process parameter data analytics, artificial intelligence (AI), and internet of things (IOT)-based sensors and technology have found their way into process industries. During this turbulent time, people and industries that cannot change their way of doing business quickly gradually become obsolete. Many industries and technologies that were running 25 years ago no longer exist as they could not cope with the changing requirements of modern times. After globalization, chemical industries needed to purchase their raw material and sell their finished products to the international market at the international market price. They have to compete with cheap raw material prices of OPEC countries, cheap manpower costs of China, big capacity of the Middle East plants, etc. Most of the traditional industries could not cope with this uneven competition and gradually became obsolete. Hence, to survive in today’s business environment, industry needs to find new ways of doing business. Therefore, a completely new way of thinking and optimization is required to maximize profit. A complete paradigm shift in mindset is the need of the hour. Profit maximization is one of the prime goals of every management team of process industries across the globe. Profit maximization in running chemical plants is itself a huge challenge that needs to be addressed by holistic vision and procedures. What is the objective of profit maximization in the process industries? It is not simply waste or loss minimization and it is not simply energy and raw material minimization. It is much bigger than that. The ultimate goal of profit maximization is to use every resource available to the process industry in the most efficient manner so that dollar per hour generation is maximized and sustained over years. To achieve this goal, energy, raw materials, equipment, manpower, processes, and environment as well as the mindset of the people must be aligned and be holistically optimized. Most of the time, engineers and managers working in process industries tackle this problem with a fragmented approach. Process engineers try to optimize process parameters, production engineers try to meet a production target rate, energy managers want to minimize energy consumption, maintenance engineers try to increase equipment uptime, reliability engineers try to increase equipment and process an on-stream factor, safety engineers try to reduce accidents and incidents, etc. All of them try to increase profit in their own way without paying attention to what other people are doing. Sometimes, their objectives are contradictory. However, a profit maximization project needs to holistically combine all their effort in the most efficient way so that sustainability and profit maximization are achieved.

1.2 Who is This Book Written for?

The main constraints of a technical work force in various process industries today is that they are too loaded with various types of managerial and technical jobs that they do not have time to see things holistically. Furthermore, the mind sets of plant managers and engineers are not aligned with the larger vision of profit maximization and sustainability. They are not equipped with techniques and methods that can combine all the fragmented approaches and do an overall optimization. No such procedures have been available until now. Therefore, the current challenge of a profit maximization project is: how can we develop a road map so that all of the available intelligent tools and techniques can be utilized effectively to support plant engineers as well as top management? Over the years, engineers working in process industries have believed that reducing loss and waste is the only way to achieve efficiency. Thus, to them, energy efficiency means to identify and rectify steam leaks, faulty steam traps, damage insulation in steam lines, less flaring, cleaning of fouled exchangers, etc. In the last decade they focused on these basic old housekeeping parts of energy efficiency. These are the techniques of past decades. With the advent of faster computers and data historians, every minute many process parameter data are stored, with the availability of offline process simulators (like Aspen, Pro II, etc.) and online advanced process control (APC) and real-time optimization (RTO) applications. With the advance of artificial intelligence-based data analytics a complete new generation of profit maximization tools and techniques are currently available. This book attempts to highlight some of these proven new generation tools, which slowly being introduced in progressive process industries. Adaptions of these new techniques in process industries need a mindset change of engineers and management. The real challenge becomes: how can a mindset change of management be made so that new generation techniques can be quickly applied to generate profit. How can these new generation tools and techniques be learnt and adapted quickly? The challenge boils down to: which methods should be selected, how can they be tailor-made to fit them into specific applications, and how can they be implemented for specific circumstances?

1.2 Who is This Book Written for? However, there are no dedicated effective books available to discuss a basic roadmap to utilize various intelligent tools and techniques, provide practical methods to implement them on the shop floor, and explain industrial application procedures. This book has been written to fill this gap with the following people in mind: practicing process or chemical engineers, production engineers, supervisors, and senior technicians working in chemical, petrochemical, pharmaceuticals, paper and pulp, oil and gas companies and petroleum refineries across the globe. This book will also become particularly useful for large numbers of managers, general managers, top level senior executives, and senior technical service consultants whose main jobs include strategic planning and implementation of various optimization projects to increase profit in chemical process industries. Undergraduate and postgraduate chemical engineering students and business students who want to pursue careers in the chemical field will also greatly benefit from this book. The book is aimed at providing various intelligent computational tools to engineers and managers working in CPI who face challenges and are looking for new ways to increase profit in running chemical plants. This book aims to convey concepts, theories, and methods in a straightforward and practical manner.

3

4

1 Concept of Profit Maximization

This book provides engineers with all practical aspects of profit maximization projects, as well as practical advice on how to derive maximum benefits from running plants. The book will present various intelligent computation techniques covering profit optimization strategy, application methodology, supporting structures, and assessment methods. In short, it will describe completely new ways and techniques to maximize profit for process plants and how to sustain profit improvement benefits. Short on background theory and long on step-by-step application procedures, it covers everything plant process engineers and technical managers need to know about identifying, building, deploying, and managing profit improvement applications in their companies. Readers are able to take away new ways to increase profit in their current plant, background computational tools and techniques for identifying profit improvement opportunities, and analysis, optimization, and monitoring procedures that are required to identify, assess, implement, and sustain profit improvement opportunities.

1.3 What is Profit Maximization and Sweating of Assets All About? In chemical process industries (CPIs), profit maximization is attained in many fragmented ways. Energy managers in a plant try to increase profit by increasing energy efficiency, production engineers increase profit by pushing the plant to its highest possible capacity, control engineers try to optimize the plant in real time by advance process control, maintenance engineers try to maximize the critical single line equipment availability by doing proper preventive and predictive maintenance, reliability engineers try to reduce the failure rate by proper inspection, the human resource (HR) department tries to reduce manpower cost and increase employee’s productivity, safety engineers try to minimize the incidence and accident rate, etc., and many people try to maximize profit in a multi-dimensional fragmented way. However, all of the above approaches are not independent but are deeply interrelated and sometimes conflicting. For example, running equipment beyond its design limit for maximization of plant capacity will definitely increase its failure rate. Therefore, profit maximization is an approach that sees all these conflicting attempts in a holistic way and evaluates the strategy that will maximize profit of the plant in the long run and sustain it. In simple terms, profit maximization means maximize dollar per hour generation from the plant and make sure that this is sustained. In mathematical terms, Maximize Profit generation in $/h terms from the plant Subject to constraints: all process and safety constraints need to be honored and all equipment limitations should not be violated Some common ways to maximize profit are (but not limited to) (Lahiri, 2017a): • Maximize plant throughput while obeying all operational and safety limits imposed by the designer. • Minimize raw material and utility consumption.

1.3 What is Profit Maximization and Sweating of Assets All About?

• Reduce production costs by maximizing process efficiency (like catalyst selectivity, yield). • Increase plant and process equipment reliability while obeying all design and safety limitations so that the profit-making production process can be sustained for longer periods, etc. In still other cases, there is a tradeoff between increased throughput and decreased process efficiency and so process optimization is needed. There are many multi-faceted dimensions of the profit maximization project. The profit maximization project (PMP) involves all the activities to increase profit in the plant. As the scope is vast and multi-disciplinary, this book only addresses some of the ways that are related to chemical engineering/process engineering and suggest some alternative new ways to generate profit, some of which are given in Figure 1.2 below. Critically assess current plant operation and identify and exploit the opportunities. For any process unit, there are various constraints and limits, as shown in Figure 1.1 which stops the plant from continuously increasing capacity (Lahiri, 2017b). The normal operating zone for any process is bounded by these limits. It is a common tendency among panel operators and production engineers to operate the plant at the center of this acceptable operating region, far from any constraints. The reason is simple: the panel operator gets a maximum amount of time to respond to disturbances before it drives the process beyond the acceptable operating region. This center region is the comfort zone of the operator as it gives some flexibility in operation. However, to get maximum profit from the process, it has to push several constraints or limits and usually

Safety limits (e.g., various safety interlocks)

Product quality constraints (e.g., upper limit on product impurities)

Operational constraints (e.g., a compressor surge limit, a tower differential pressure to avoid flooding)

Actuator limits (e.g., a valve is either open or closed)

Equipment limits (e.g., the maximum vessel working pressure or temperature)

Figure 1.1 Various constraints or limits of chemical processes

5

1 Concept of Profit Maximization

Product purity limit Feed supply limit

g tin ra pe O

Pressure limit

Ac ce pt ab le

6

Operator's preferred operating zone

Compressor amp limit

Economic optimum zone

Column flooding limit

Region Furnace max temperature limit

Figure 1.2 Optimum operating point versus operator comfort zone

this most economic operating point lies at the edge of the boundary limit (see Figure 1.2) (Lahiri, 2017a). The most common way to increase profit is to run the plant at maximum possible capacity. This essentially means push the plant to its limit so that multiple pieces of equipment or assets touch their maximum operating limits. This is called “sweating of the asset.” One major target of the profit maximization project is sweating of all assets in the plant. Plant engineers and managers are also considered as valuable human assets of the company and sweating them intellectually is also needed. Running the plant at maximum capacity does not mean to run it at its nameplate capacity, i.e. process flow diagram (PFD) capacity. That is the bare minimum target. All over the world, good companies are running at 125–150% of their nameplate design capacity. Normally they follow three basic steps to increase plant capacity: Step 1: A 10–15% capacity increase over its nameplate design capacity is usually possible by exploiting the design margin usually available in process equipment. Step 2: Another 5–10% capacity increase is possible by a small investment or minor modifications with resources already available in the plant or outside with little capex. Step 3: Rest 20–25% capacity can be done by a major revamp and with big capex. All good plants follow these three steps in order and continuously improve themselves so that with the same plant they can run 25–50% extra capacity. This is one of the surest ways to increase profit. Running the plant lower that its nameplate design capacity is no longer a viable option and all the plants running at a lower capacity will not be able to cope with stiff international competition and eventually will perish over time. Hence the first and most important step in profit maximization is to know the techniques of how to run the plant at its highest possible capacity. This essentially means:

1.4 Need for Profit Maximization in Today’s Competitive Market

• How to know and exploit the design margin available in installed equipment? • How to know the equipment that is a bottleneck for a further capacity increase? • How to carry out a detailed cost benefit analysis for a major revamp project? However, the profit maximization project does not end by maximizing the capacity only and involves all the multi-faceted activities to increase profit in the plant. Following this project an old generation plant can be transformed to a new generation plant. As the scope is vast, this book only addresses some of the proven techniques related to chemical engineering and suggests some alternative innovative ways to generate profit. Some of the ways to increase profits (but are not limited to these) are given below (make a diagram similar to that in Figure 1.2): • Assess existing plant operation and identify and exploit the opportunities to increase profit. • Implement an advance online process monitoring system to monitor equipment and process performance in real time. • Implement a real-time fault diagnosis system to detect any abnormality of equipment/process at its incipient stage and take preventive and corrective action. • Identify and implement a major debottleneck project. • Utilize a process modeling and simulation technique to optimize process parameters to increase profit. • Identify hidden margins available in major distillation columns and push them to their limit. • Utilize different modeling techniques (data driven or kinetic driven) to generate a model of a major reactor and subsequently optimize reactor parameters to increase profit. • Identify the scope of utility savings and waste reduction to increase profit by implementing them. • Install APC and RTO to stabilize and optimize the process in real time to increase profit amid various disturbances.

1.4 Need for Profit Maximization in Today’s Competitive Market Due to globalization today CPIs have to compete with the global competitive market. The raw material costs of refinery and petrochemicals are varying along with the global crude oil price. Crude oil prices are unpredictable and depend on many complex geo-political parameters of OPEC and devolved countries. As raw material cost comprises a major component of the product price, profit margins of refineries and petrochemical companies vary in tune with the crude oil price. The final product price of polymers and downstream chemicals are also varying with the crude oil price but the amplitude of variations is not the same as the crude price. Some of the major challenges faced by CPI today are as follows: • Middle East companies (especially those in Saudi Arabia, Qatar, UAE, etc.) and those of China are building plants of world scale capacity, which therefore minimize their production costs. Eight of 10 of the world’s largest refineries or petrochemical plants are situated in the Middle East countries or China.

7

8

1 Concept of Profit Maximization

• Cheap raw material costs of Middle East companies make them so competitive that companies in other countries find it difficult to compete with them. • Companies have to operate in a global market (both purchase and sales) due to rapid globalization where there are no trade barriers. • Cheap manpower costs of China and SEA companies make them cost effective. • CPI has to fulfill more and more stringent pollution control norms imposed by local pollution control authorities. • Discovery of cheap US shale gas and the building of mega sized refineries and petrochemical factories in the US will change the business map of global competition in the near future as major oil and chemical demands of the US will be satisfied by US companies themselves. This will reduce US demands from China and OPEC countries. • The US may very soon drastically reduce their oil import from OPEC countries and rely on their own oil production. This will change the global oil demand pattern as the US is the major oil importer from Saudi companies. This new oil demand supply equation will affect the crude oil price in an unprecedented way, which may lead to uncertainty in the global crude oil price. • The recent trade war between the US and China will make the polymer and chemical pricing and demand supply scenario uncertain as it evolves. • A faster gross domestic product (GDP) growth rate of the world’s top two biggest countries, China and India, will increase demands for polymer and oil, which may shift the global demand pattern into the Asia region in the near future. Today’s CPI has to operate under all of the above uncertain scenarios. Unless it finds an effective mechanism to optimize all the resources that are under its direct control and maximize its profit, it will not generate enough cash to prepare it to deal with uncertain product prices and uneven global competition. Companies need to be flexible in their product pattern and should have the capability to increase or decrease capacity in tune with global demand. For that companies need to have enough extra cash in hand so that they can use it as a buffer to respond to global business uncertainties. Implementation of an effective profit maximization project is the only way to generate this buffer money and strategically position the company in a better way to sustain their operation amid global business uncertainties.

1.5 Data Rich but Information Poor Status of Today’s Process Industries With the advent of faster computers in CPI, a large amount of process data is collected and stored every minute by data historian software like IP21, the Pie system, Exaquantum, etc. Every second or minute data of all process parameters of whole plants are now available. This large historical data depository is a distinct feature of today’s CPI as compared to older generation plants. These real-time process data are like an untapped gold mine. Many insights and much process knowledge can be generated from these large sets of operating data. However, very little has been done so far. Due to the unavailability of effective process data analytics, the knowledge hidden in such data could not be

1.7 How Knowledge and Data Can Be Used to Maximize Profit

tapped properly. The main concerns of CPIs are how to extract meaningful information from these data. Thus, today’s chemical industry remains data rich but information poor. There is a need to generate an effective framework where knowledge can be extracted from this wealth of data. Advanced AI based big data analytics systems need to be applied to extract knowledge. The capability to meet this challenge is key for business excellence.

1.6 Emergence of Knowledge-Based Industries The speed of technological advancement in the last 20 years makes older technologies obsolete at a speed never before achieved. In today’s cut-throat global competitive environment, companies that follow the old way of doing business gradually become obsolete and die over time. Many chemical companies of the 1980s or 1990snow no longer exist. Companies who could not adopt new technologies and new ways of doing business gradually perish. Only knowledge-based chemical industries survive. Only CPIs employing knowledge to drive their businesses are going to survive in the future. This essentially means generating an effective platform that can generate knowledge from available business data and use this knowledge to develop a unified framework to support faster business decisions to respond to external market uncertainties. Companies who utilize this knowledge to drive their businesses are called knowledge-based chemical industries. It is survival of the fittest scenario and only knowledge-based industries that adapt to a changing business scenario will survive in the future. All other companies, who fail to integrate their knowledge with business, will gradually perish. This gives rise to a new generation of process industries. The emergence of these new generation process industries in this decade is the most important phenomena in CPI.

1.7 How Knowledge and Data Can Be Used to Maximize Profit New ways of doing business are key for survival. Intelligent industries are those who can adapt quickly to this knowledge and innovation era. However, this needs a complete mindset change. How we generate useful knowledge and integrate it with business decisions is the real challenge of today’s CPI. A new look to the old problems is absolutely necessary. A new way to increase equipment reliability, novel methods for process data monitoring, and a new emphasis on real-time optimization are what is now needed. How data and knowledge can be used to maximize profit is the real key driver and all the chapters of this book are dedicated to that. Companies took multi-faceted a completely new advanced approach to deal with this challenge. Some of the common solutions Global good companies have implemented are as follows: • Real-time optimization (RTO) and advanced process control (APC) are implemented for real-time optimization of plant. These tools ensure running the plant with simultaneous multiple constraints. • Implementation of an advanced artificial intelligence (AI) base, online data monitoring, and fault diagnosis detect any abnormality of process equipment at its incipient stage.

9

10

1 Concept of Profit Maximization

• • • • • •

All these plants run 25–50% more capacity than their design capacity. Online equipment reliability monitoring systems are implemented. Risk-based inspection systems are in place. Online supply chain management system. SAP-based reliability centered maintenance practice. A management framework to encourage people participation and to tap their ideas for small improvements in the plant.

References Lahiri, S.K. (2017a). Front matter. In Multivariable Predictive Control (pp. i–xxxiii). https://doi.org/10.1002/9781119243434.fmatter. Lahiri, S.K. (2017b). Introduction of model predictive control. In Multivariable Predictive Control (pp. 1–21). https://doi.org/10.1002/9781119243434.ch1.

11

2 Big Picture of the Modern Chemical Industry 2.1 New Era of the Chemical Industry Since 1746, evolution of the modern chemical industry can be divided into four distinct stages of development: the industrialization era (chemical industry 1.0), the scale and capacity building era (chemical industry 2.0), and the automation and computerization era (chemical industry 3.0). Currently the chemical industry is slowly entering into a new era called data analytics and the artificial intelligence (AI) stage (4.0). Disruptive technologies like artificial intelligence, machine learning, big data analytics, and the internet of things (IoT) have already entered inside the chemical process industries and are already changing the rule of the chemical business (Ji, He, Xu, and Guo, 2016). Their influence is starting to see benefits in a significant improvement in production efficiency, energy utilization, optimization of the entire manufacturing process, integration of the supply chain, new product development, product delivery speed, etc. Figure 2.1 shows the development stages of the chemical industry.

2.2 Transition from a Conventional to an Intelligent Chemical Industry The recent advances of these disruptive digital technologies give birth to a new generation of intelligent chemical industries (Ji et al., 2016). The old method of doing business by the conventional chemical industry are slowly becoming obsolete. Distinct features of a new generation of intelligent chemical industries are given below (but are not limited to these): • A new generation of intelligent chemical industries use data analytics to take informed decisions in every phase of business, be it manufacturing, marketing, or R&D (research and development). These intelligent chemical industries develop a complete infrastructure of digital platforms to collect and analyze data and integrate it with business processes. This is called digital transformation. • They generate knowledge from the available data by using artificial intelligence-based algorithms. This knowledge is used to integrate shareholder value, market demands, and sustainable development. • Manufacturing facilities of these new generation of chemical industries are transformed from island mode to integrated mode. Operation of the supply chain, Profit Maximization Techniques for Operating Chemical Plants, First Edition. Sandip Kumar Lahiri. © 2020 John Wiley & Sons Ltd. Published 2020 by John Wiley & Sons Ltd.

12

2 Big Picture of the Modern Chemical Industry

Chemical industry 4.0 AI and data analytics Intelligent manufacturing

Chemical industry 3.0

Sustainability and green plant.

Automation DCS and PLC

Chemical industry 2.0

Computer process control. 2010s

Scale Rapid scale up of petroleum industry

Chemical industry 1.0

Polymer industry started. 1970s

Industrialization Inorganic industry matured, organic industry started th

Early 20 century

Mid 18th century

Figure 2.1 Developing stages of the chemical industry

manufacturing facility, marketing, and R&D are integrated to leverage a larger optimization scope. • The process control of these process industries is not confined to normal PID (Proportional, Integration and derivative) control but expands to advance process control and real-time optimization covering the production process, entire marketing, and supply chain operation. • With the help of data analytics and artificial intelligence-based algorithms these chemical industries develop a knowledge-based decision-making capability in every aspect of business and make themselves better prepared to handle more stringent environmental requirements and changing customer needs. The comparisons between an intelligent chemical industry and the conventional chemical industry are listed in Table 2.1 (Ji et al., 2016).

2.3 How Will Digital Affect the Chemical Industry and Where Can the Biggest Impact Be Expected? The global management firm M/s Mckinsey has studied and reported that digital transformation is changing the entire chemical business in three major ways, as depicted in Figure 2.2 (Klei et al., 2017). 2.3.1 Attaining a New Level of Functional Excellence

Data analytics and AI-based interpretation is helping efficiency improvement of all core business processes of the chemical industry, including manufacturing, marketing and sales, and R&D. Data-based decision making, called digital in short, provides the means to unlock a new level of productivity enhancement (Klei et al., 2017).

2.3 How Will Digital Affect the Chemical Industry and Where Can the Biggest Impact Be Expected?

Table 2.1 Comparisons between smart and conventional chemical industries Conventional chemical industry

Items

Intelligent chemical industry

Integration mode

Integration for processes

Integration of supply chain network

Optimization goals

Profit optimization on specific conditions

Profits optimization considering market demand, device status, energy conservation and emissions reduction

Optimization patterns

Serial mode conducted offline

Synchronous optimization of decision-making and control adjustment employed online

Technical economic feature

Large-scale

Equilibrium between large-scale and necessary flexibility

Operation mode

Specialized manufacturing

Combination of manufacturing and service

Decision factors

Operational and technical factors

Users’ requirements, products, quality standard, operating condition, resource, system reliability status

Control mode

Discrete control

Advanced process control

Intelligent degree

Low level

Artificial intelligence embedded in the process optimization control

Control platform

Discrete control system

Contemporary integrated process system

Flexibility

Limited flexibility, adaptive scope and function redundancy

More flexible configuration, adaptive to multiple optimization control modes

Data supporting

Local small data

Big data

Algorithm

Traditional statistical analysis

Statistical analysis, data mining, AI and visualization techniques

Functional excellence

Application of digital-enabled approaches to improve companies’ core business processes

Value Chain

Potential for digital to affect demand patterns in end markets, with implications for the chemical industry’s value chains

Change in business model

Digital developments lead to changes in the business models through which chemical companies capture and create value for customers

Figure 2.2 Three major ways digital transformation will impact the chemical industry

2.3.1.1 Manufacturing

Manufacturing operations consume most of the production costs and digital technology can bring the highest impact in this area. This is true for all segments of the chemical industry, from petroleum refinery to petrochemicals to pesticides to specialty chemicals. The global management firm Mckinsey estimate the potential for a threeto five-percentage point improvement in return on sales from employing digital in

13

14

2 Big Picture of the Modern Chemical Industry

production operations (Klei et al., 2017). With the advent of data historians two decades ago, an enormous amount of production-related data has been collected by all of the major chemical industries. However, due to the absence of proper data analytics software, most of these data remains unutilized. AI-based algorithms can extract knowledge from these data and utilize that knowledge to achieve higher efficiency and throughput, lower energy consumption, and more effective maintenance. For many companies, these are low hanging fruits and benefit can be achieved immediately using existing IT (information technology) and process control systems. The contribution to profits can be substantial. Examples are many in all leading chemical industries around the world. A major petrochemical company applied advanced AI-based data analytics to a billion data points that it collected from its naphtha cracker manufacturing plant. With the help of an AI-based stochastic optimization algorithm, this plant optimizes different process parameters that lead to an increase in the ethylene production by 5% without making any capital investments and generated cost savings by reducing energy consumption by 15%. A leading refinery company takes another approach at one of its main plants: it used AI-based advanced analytics to model its production process and make a virtual plant, and then used the model to provide detailed, real-time guidance to DCS (distributed control system) panel operators on how to adjust process parameters to optimize performance. Once it was implemented, profit from this plant increased by over 25% and yields increased by seven percentage points, thus saving on raw materials, while energy consumption fell by 26% (Holger Hürtgen, 2018). Besides this AI-driven analytics-based opportunity, there are other digital-enabled advances that have started creating profit in the manufacturing operations area. Examples include IoT-based steam trap monitoring, IoT-based wireless vibration and temperature monitoring of critical pumps and other single-line rotating equipment, the use of digital sensors to monitor vent gas composition, etc. These advances help to reduce maintenance costs and improve process reliability and safety performance. At the same time, deploying a holistic automated and centralized data analytics and plant performance management system should enable the plant engineers to monitor the plant better and take proper corrective and preventive actions faster. 2.3.1.2 Supply Chain

Digital technology also can bring enormous value to the entire supply chain, including inbound and outbound logistics and warehousing. From past historical data, an intelligent algorithm can significantly improve accuracy of forecasting, which helps to optimize the entire sales and operations planning process (Klei et al., 2017). Digital technology can be used to leverage better scheduling of batch production, shorter lead times, and lower safety stocks with a higher level of flexibility. A digital enabled holistic system can be built to develop integrated “no touch” ordering and scheduling systems. 2.3.1.3 Sales and Marketing

Data analytics and AI-based digital technology can be used for intelligent decision making in sales and marketing. Mckinsey estimate- that digital-enabled initiatives in marketing and sales could improve the industry’s average return on sales (ROS) by two to four percentage points. Digital initiatives in marketing and sales include developing intelligent pricing systems, generating growth opportunities from data, and using algorithms to predict

2.4 Using Advanced Analytics to Boost Productivity and Profitability in Chemical Manufacturing

churn at the individual-customer level and then suggesting countermeasures to the sales force. The impact of these initiatives can be significant. A large polymer company used advanced analytics to reset prices for hundreds of thousands of product-customer combinations in three core countries, based on individual risk and willingness to pay. By developing an AI-based intelligent algorithm, the company was able to achieve price increases of 3 to 5%, compared to 1% increases in previous years. In some other petrochemical companies, the company’s manufacturing unit is connected with the sale and marketing unit by an optimization algorithm and the company’s production plant process parameters, product split in a multi-product plant, and capacity are adjusted by the demand scenario coming from sales and marketing forecasting. 2.3.1.4 Research and Development

Due to plastic pollution, pollution from cars and from various carcinogenic chemicals, the usage patterns and demands of various chemicals across the globe is changing very fast. This poses a challenge to chemical industries who makes those products. One of the ways a research and development department of chemical plants can respond to this challenge is by creating higher-value-added, higher-margin products at a faster pace, in particular in specialty chemicals and crop-protection chemicals (Klei et al., 2017). Through intelligent algorithms, chemical companies will be able to use high-throughput optimization to develop and adjust molecules that offer more value. They will also be able to deploy advanced analytics and machine learning to simulate experiments, to use digital predictive power to systematically optimize formulations for performance and costs, and to data-mine information available from past successful and failed experiments. Not least, they will be able to identify the best possible resource allocation to enhance the performance of R&D teams and the innovation pipeline. Many of these practices are already established in the pharmaceutical industry but were largely unaffordable for chemical companies. With the emergence of inexpensive computing power on a massive scale, this is likely to change.

2.4 Using Advanced Analytics to Boost Productivity and Profitability in Chemical Manufacturing As of now, it is quite clear that digital will have a significant impact on many areas of the chemical industry, with the gains in manufacturing performance potentially among the largest companies (Holger Hürtgen, 2018). Chemical companies have already created the infrastructure to collect and store enormous amounts of process data from hundreds of thousands of sensors, but very few have succeeded so far to take advantage of this data gold mine of potential intelligence. With the availability of cheaper computational power, IoT-based cheap sensors, and intelligent advanced analytics tools, all chemical companies can now use those data to make more profit, extract knowledge from those data, and using machine-learning and visualization platforms to uncover ways to optimize plant operations (Wang, 1999). AI-based machine learning tools can be used to develop insights into what happens in a chemical plant’s complex manufacturing operations; this can help chemical companies solve previously impenetrable problems and reveal those that they never knew existed, such as hidden bottlenecks or unprofitable production lines.

15

16

2 Big Picture of the Modern Chemical Industry

Predictive maintenance Predictive maintenance analyzes the historical performance data of production units and their machinery to forecast when equipment is likely to fail, to limit the time it is out of service, and to identify the root cause of the problem.

Yield, energy, and throughput analytics It can be used to ensure that the individual production units are as efficient as possible when they are operating, helping to increase their yields and throughput or to reduce the amount of energy they consume.

Value-maximization modeling It scrutinizes the thousands of parameters and conditions that have an impact on the total profitability of an integrated supply chain—from raw-materials purchasing through the complex and often interrelated chemical-manufacturing steps to final sales—and then provides intelligence on how best to capitalize on given market conditions.

Figure 2.3 Three major impact areas where advance analytic tools will help to increase profit

There are three major areas where applications of advanced analytics tools can give an enormous profit increase, namely predictive maintenance; yield, energy, and throughput analytics; and value-maximization modeling, as shown in Figure 2.3 (Wang, 1999). 2.4.1 Decreasing Downtime Through Analytics

One of the major profit suckers in chemical industries is a sudden trip of critical single line equipment. Once a plant trips, millions of dollars get lost in terms of less or no production and more time is required to bring back the plant to on-spec production after a disturbance. Besides this, a lot of money is lost in terms of flaring, venting, or draining of costly chemical gas or liquids (Wang, 1999). Big data analytics can be used to develop fault diagnosis software to anticipate the failure of critical equipment at a very early stage and thus give sufficient time to plant engineers to take preventive or corrective actions. Such fault diagnosis systems analyze historical data to generate insights that cannot be observed using conventional techniques. By implementing an intelligent analytics-based fault diagnosis system, companies can determine the circumstances that tend to cause a machine to break. Then a real-time automated system can be developed to monitor all relevant parameters and give early fault signals, so engineers can intervene before breakage happens, or be ready to replace a component when it does, and thus minimize downtime. Companies who has implemented such systems typically reduce machine downtime by 30 to 50% and increases machine life by 20 to 40% (Wang, 1999). Chemical companies are already starting to see substantial gains in this area. One major polymer producer consistently ran into problems with extruders at its largest plant. When one of the shafts of the extruder broke, the plant had to stop production for 3 days while a replacement was installed; these shafts are expensive, besides the cost impact of the production loss. Engineers had done a detailed study to determine the possible root causes of failure; alternative materials in the shaft were also tried out, as well as different process conditions, but none of them solved the problem. A principal component-based fault diagnosis approach changed all this. It combined a detailed analysis of data from hundreds of sensors with the plant engineers’ expert

2.4 Using Advanced Analytics to Boost Productivity and Profitability in Chemical Manufacturing

domain knowledge, and reexamined the process variables and other data sources; it then developed a real-time-based algorithm to predict when a failure was imminent. The problem occurred with only one of the polymer grades, and not with all batches, suggesting the key lay in specific process conditions in the equipment. The team developed a model based on “a hybrid principal component analysis and artificial neural network” algorithm that took into account the specific parameter settings in production, such as extremes of temperature and temperature progression, together with information on the polymer product type and composition. A real-time visual platform was developed, which flagged an early warning to engineers when the plant conditions approached a state that could ultimately lead to shaft failure. When it flags that a failure is imminent, the plant operators undertake a 15-minute cleaning of specific parts of the machinery to prevent the failure from occurring. The improvements to performance that resulted from using the advanced-analytics approach have been substantial. Instead of a 3-day production loss plus a costly extruder shaft replacement, the company was now dealing with just a 15-minute production interruption, and the approach has cut production losses by 60% and maintenance costs by 85%. 2.4.2 Increase Profits with Less Resources

Increasing plant throughput is the best and most effective way to increase plant profit. In plant, every equipment has some extra margin rising from its design safety margin. Advance analytics can scrutinize the past 3–4 years of plant operation data and can estimate how much capacity increase is possible in the plant without investing a single penny. AI-based optimization techniques can then be used to optimize process parameters so that these safety margins can be exploited to increase plant yield and throughput or minimize energy costs. Even small percentage improvements in operational efficiency can significantly enhance earnings before interest, taxes, depreciation, and amortization (EBITDA) performance. The approach does that by balancing yield, energy use, and throughput – while also taking account of varying raw materials costs – to maximize the profitability of each process step (Wang, 1999). One petrochemical company was having capacity limitations at its naphtha cracker furnace that makes ethylene and propylene. The furnace’s unstable production rate and low overall output meant that it represented a serious bottleneck for a high-margin segment of output. An artificial neural network and genetic programing-based furnace model building exercise was carried out. All the temperature, pressure, flow, and composition data related to the furnace had been collected over 3 years of production, comprising 900 000 samples, each with 360 tags – almost 80 million data points. This analysis identified critical process parameters and made it possible to build a data-driven furnace model. The model generated mathematical relations between the furnace throughput and other input parameters that can influence the throughput. By running the models at different input conditions, a deeper understanding of furnace operations is generated. A test run of the furnace confirmed the model’s findings. The plant engineer had long suspected that manipulating some of the levers identified in the model could improve productivity, but they never had the mathematical tools or data to confirm it. Based on its new advanced analytics-based understanding of its process, the company developed

17

18

2 Big Picture of the Modern Chemical Industry

an automated, real-time-based operator guidance platform that advises the operators how to adjust a range of process parameters to get the best performance. The result was an output increase of 10 to 15%, which represented a net profit-contribution increase of around USD 15 million a year. The company estimates that applying the same kind of advanced-analytics approach across all the different manufacturing operations at the site could generate a USD 75 million annual profit gain (Holger Hürtgen, 2018). Advance analytics can also be used to develop a data analytic platform to show a companywide energy usage pattern and guide the operators to change various energy usage options to minimize energy consumption. AI-based techniques are sometimes use for a simple process, such as how to exploit favorable seasonal conditions (say a less cooling water temperature) to improve plant profitability. 2.4.3 Optimizing the Whole Production Process

Whereas predictive maintenance and yield, energy, and throughput analyses are designed to improve the efficiency and profit-making capability of individual pieces of equipment, value-maximization modeling covers the whole plant or whole site and helps to optimize the interaction between those pieces of equipment across processes. This optimization and modeling technique utilizes its inherent analytic capability to show in real time how to maximize the rate of profit generation in complex production systems and supply chains, encompassing every step from purchasing to production to sales. Unlike the limitations of human planners, this advanced-analytics approach typically solves the complex maze comprising as many as 10 000 variables and one million constraints to help producers figure out what to buy and when, what to make and how much, and how they should make it to generate maximum profit in each period (Wang, 1999). The uncertainty in pricing and demands of turbulent chemical markets poses a complex business challenge, which needs to be solved every day to figure out the most optimum buy and sell decision and also how much to produce. The uncertain and frequently changing nature of chemical companies’ businesses and product lines means they must be capable of solving a complex objective function: volatile costs and prices, multiple plants, and products that can be made in various ways from diverse combinations of materials, involving output of different combinations of co-product of varying values, maximizing output of a high-value product, as well as managing by-product flows. The following example from one large, diversified integrated refinery and petrochemical complex shows the kinds of gains to be captured. The company was selling a broad range of petrochemicals and specialty chemicals from the site to a global marketplace through a mixture of spot and long-term contracts. On the other hand, it was buying the raw material, i.e. crude oil, from various countries with varying quality and price. Being a multinational company with a presence in different countries, purchase, sales, and production decisions were made by local offices and pricing was arbitrarily set by different regions and departments. Organizational responsibilities were scattered across multiple business units and corporate functions. Underlying all this was the typical chemical-industry challenge of commodity products underpinning specialties production, while the commodity output brought with it lower-value co-products, multiplying the hurdles to maximizing profitability. Due to the absence of a global optimization algorithm, the company lost a lot of money due to non-optimal decisions that were taken

2.5 Achieving Business Impact with Data

locally. A mixed-integer programming model encompassing the 900 variables explored nonlinear cost curves and the 4000 constraints related to production capacities, transportation, and contracts; the hundreds of steps in production with alternative routes and feedback loops; nonlinear price curves and raw-materials cost structures; and intermediate inventories (Wang, 1999). Using the model, the team solved a global optimization problem and were able to increase profits by USD 20 million a year (Wang, 1999). For example, the company started making an intermediate product on an underused line instead of buying it from a third party. At the same time, the team optimized different process parameters of a furnace, various distillation columns, an absorber, etc., which gave higher yields, thereby reducing raw-material consumption. It identified some extra cushions available in some of its plant to expand capacity by increasing the throughput, and it increased sales revenues by raising the capacity for some product categories. It also maximized the production of some of the products that fetched a higher profit margin. The analytics approach revealed some counterintuitive improvements. The model suggested that eliminating the production of a particular polymer grade would increase profitability overall. The company had been selling this lower-grade polymer to a local customer for a long time, but generated limited returns while incurring high logistical costs. By shifting the raw material, i.e. ethylene, used to make this polymer, to manufacturing another value-added product, the company was able to make more profit. That switch might never have been suggested if the decision had been left to the manager of the polymer business, who previously had the decision rights. These changes enabled the chemical company to boost its earnings before interest and taxes by more than 50%.

2.5 Achieving Business Impact with Data For the last two decades chemical industries have been generating, collecting, and storing huge amounts of operation and maintenance data using various software. These data are like a gold mine and now is the best time to achieve an impact with (your) data. More and more data are available, computing power is ever increasing, and mathematical techniques and the so-called data science are becoming more and more advanced. Yet while data is considered as the “new oil” or the “new gold” these days, several technologyand business-related challenges prevent chemical industries from realizing the potential impact data can make on their business (Holger Hürtgen, 2018). 2.5.1 Data’s Exponential Growing Importance in Value Creation

The following facts regarding data have changed the business outlook in recent times: Rapid increase in data volume: The number of delivered sensors globally has increased sevenfold from 4 billion in 2012 to greater than 30 billion in 2016 (Mckinsey). Data has not only increased exponentially in volume but has also gained tremendous richness and diversity. In the chemical industry, data is not only generated from various flow, temperature, and pressure transmitters but also from cameras and analyzer to vibration monitors, enabling richer insights into process behavior.

19

20

2 Big Picture of the Modern Chemical Industry

Falling IoT sensor price: There was a 50% reduction in IoT sensor price between 2015 and 2020. Cheap computational power: Better processors and graphics processing units increased investment in massive computing clusters, often accessed as cloud services, improvements in storage memory, etc., have recently increased computational power. New data analytic tools: In recent times, many new tools have been coming to market to convert this flood of raw data into insights and eventually into profit. Machine learning and artificial intelligence: These new generation algorithms are rapidly replacing the old method of calculations and emerge as new data analytics. Both data and computational power enable next-generation machine learning methods, such as a deep learning neural network. Value creation: As a consequence, data has become the new oil of the chemical industry – and the best way for companies to generate and access is to digitize everything they do. Digitizing customer feedbacks provides a wealth of information for marketing, sales, and product development, while digitizing manufacturing processes generates data that can be used to optimize operations and improve productivity. The confluence of data, storage, algorithms, and computational power today has set the stage for a wave of creative disruption in the chemical industry. 2.5.2 Different Links in the Value Chain

Data in its raw and most basic form is virtually worthless until we generate knowledge and business insights from it. The biggest challenge to confront the chemical industry today is how to generate business insights from these huge data banks sitting in their server and convert that knowledge to increase profit. Today every leading chemical industry talks about Big Data and Advanced Analytics and even machine learning and artificial intelligence (AI). Today’s leading chemical industry is in a hurry to implement the advance analytics in their business and they focus too much on single technical components of the “insights value chain,” as we call it. However, the value creation of data consists of following five components and companies need to focus on all the components if they want to capture the full value (or any value at all) from relevant (smart) data (Figure 2.4): [ ][ ] Quality Effectiveness of Value creation from data = of data data analytics ] ⎡Adaption⎤ ⎡IT infrastructure⎤ [ ⎢ ⎥ People of × ⎢ to capture and ⎥ ⎢ ⎥ mindset ⎢⎢ business ⎥⎥ ⎣ store data ⎦ ⎣ process ⎦

(2.1)

It is important to understand Equation (2.1), which reveals that the insights value chain is multiplicative, meaning that if one single link in that chain is zero, your impact will be zero. In other words: the entire data ecosystem is only as good as its weakest component. The chemical industry needs to understand this critical concept and should give importance to developing all components and steps of the insights value chain – not focusing on only one piece and forgetting about the others.

es Pr oc es s

pl e Pe o

IT

Da ta

An al yt ic s

2.5 Achieving Business Impact with Data

New data sources

Machine learning

Cloud sourcing

Cultural change

automation

Raw unstructured data

AI

Data visualization

Data engineer

Agile process

Predictive Statistics

Programing language

Organization

Adaptation

Data security

Value captured

The insights value chain is muItipIicative, i. e., you are only as good as the weakest link in the chain

Figure 2.4 Different components of the insights value chain

Figure 2.5 Overview of the insights value chain upstream processes (A–B) and downstream activities (D–E)

The following sections briefly explain the function of each of the insights value chain’s core components (see Figure 2.5) along with its upstream as well as its downstream steps and processes. 2.5.2.1 The Insights Value Chain – Definitions and Considerations (Holger Hürtgen, 2018)

The insights value chain has two foundations, namely a technical foundation and a business foundation. The technical component of the technical foundation consists of data, analytics (algorithms and technical talent), and an IT infrastructure (Hürtgen, 2018). This essentially means that the value creation from data is possible when efficient data scientists and domain experts use smart algorithms to extract meaningful information from high-quality data. In today’s world of Big Data, companies also need an IT infrastructure capable of capturing, storing, and processing large amounts of data fast. Second, the business foundations of the insights value chain consist of the components of people (non-technical talent) and the company’s adaptive processes, both of which are required to turn the knowledge gain from data into (business) action.

21

22

2 Big Picture of the Modern Chemical Industry

Here are some key considerations concerning the components of the insights value chain: Data: The basic building block of this value chain is data and data must be thought of as the entire process of collecting, linking, cleaning, enriching, and augmenting internal information (potentially with additional external data sources). In addition, the security and privacy of the data throughout the process are fundamental. Analytics: The second component of the insights value chain is the data analytics, which can be considered as an IC engine that will utilize the data (new oil) to generate business insights. Analytics describes the set of digital algorithms (e.g. software) deployed to extract knowledge from data as well as the talent (e.g. data engineers and data scientists, domain experts in the chemical industry) capable of deploying the right tools and methods. IT: IT infrastructure is the technical layer enabling the capturing, storing, and processing of data, e.g. data lakes, two-speed IT architecture. People: People are the ultimate drivers who will implement those insights in business actions. People from the front lines of manufacturing and sales are needed to guide and run an advance analytics course that converts data into insights and successfully implements those insights in the business. Today’s chemical companies need to change the old mindset and should develop this critical capability to “translate” analytics- and data-driven insights into business implications and actions. Process: Another crucial challenge in the digital journey is to develop adaptive processes and systems within the company that can deliver these business actions at scale. To develop the ability of seamless implementation, some old operating procedures might need to be adapted, some might need to be fully automated, and others might need to be made more agile. In addition, there is an overarching frame and an underlying governance in which the insights value chain is operating: Strategy and vision are the overarching frames in which the insights value chain is meant to operate. Data analytics should not be “done” for the sake of a data analytics but in fulfillment of the organization’s vision and in support of its overall business strategy. “Think business backwards, not data forward” (Holger Hürtgen, 2018). The operating model is the underlying governance in which the insights value chain lives. Core matters to be addressed include deciding where the analytics unit will sit within the organization and how it will function and interact with BUs (e.g. centralized, decentralized, hybrid).

2.6 From Dull Data to Critical Business Insights: The Upstream Processes The insights value chain’s upstream processes comprise two steps (see Figure 2.5). 2.6.1 Generating and Collecting Relevant Data

Today’s big chemical complexes have at least 10–20 plants and each plant consists of approximately 4000–7000 transmitters or sensors that can collect data every second

2.6 From Dull Data to Critical Business Insights: The Upstream Processes

(Holger Hürtgen, 2018). For instance, a petrochemical complex having 10 individual chemical plants generates 3.127 trillion data in one year. It is very costly (and perhaps impossible) to capture and save every bit of the tera- and petabytes of data that will be generated every second and will create a data overload on the system. Not all the data are relevant to make an impact on the business in the chemical industry. Hence it is important to know which data should be collected and at which frequency so that it can be used to generate business insights and drive the profits up. One easy solution is to follow the use case earlier implemented in any other chemical industry. Defining certain requirements based on particular use cases will help ensure that only relevant data are captured. As shown in the study of Mckinsey, the situation requires a vision and critical thinking to find which data to store in its original granularity and which to aggregate or pre-analyze. In the case of relevance, the more classical “hypothesis-driven” or “use case backwards” approach often delivers better results than the often praised “Big Data, brute force” approach. Data layering is another critical requirement while handling enormous volumes of data of the chemical industry. Instead of overloading the data analytic algorithm with all types of raw data, carefully organizing it into several logical layers and then employing a logic by which to stack these layers can help generate more meaningful data. 2.6.2 Data Refinement is a Two-Step Iteration

Once all relevant raw data of the chemical industry is captured and stored, the next step is the process of translating the enormous amount of unrefined raw data into actionable business insights. This is a very critical step and here lie all the challenges. The two-step process is comprised of enrichment and extraction (Holger Hürtgen, 2018). Step 1: Enriching data with additional information and/or domain knowledge. It is important to understand that a data engineer working alone is not sufficient to do this translation. A domain expert who runs the chemical company is absolutely necessary to enrich the data with additional domain knowledge – which is a somewhat more complex process. This essentially means that human expertise and domain knowledge is as important to making data useful as is the power of analytics and algorithms. The blend of data analytics capability along with a domain expert’s knowledge is still the optimal approach to data enrichment, although this may well change one day due to further developments in AI space. Therefore, before we even start with machine learning, we need to involve human experts who use their expertise to explain their hypotheses. The task of a data scientist (or sometimes a data engineer at this stage) is now to translate, i.e. codify, additional information and/or this domain knowledge into variables. Concretely, this means transforming existing data into new variables – often also called feature engineering. Man + machine example: Predictive maintenance in big compressors in the chemical industry. Deep knowledge in the data science field is important in choosing the right machine learning algorithm and in fine tuning models in a way that best predicts machine/component failure in compressors. At the same time, engineering domain

23

24

2 Big Picture of the Modern Chemical Industry

collect raw data Take insights into action

Define target variable

Create smart variables

Pedict future

Test performance on validation sample

split data for training and validation Build predictive models

Figure 2.6 Data science is an iterative process that leverages both human domain expertise and advanced AI-based machine learning techniques

knowledge specific to compressors will make a big difference when interpreting and identifying the root causes of failures. Sometimes data collected from other industries running with similar types of compressors can further optimize these models by providing additional ideas for drivers of failures that can be used in the predictive model to further improve the predictive power. Domain knowledge can also help to interpret the results and derive concrete business-based solutions to prevent failures in the future. Finally, business knowledge is critical in implementing the recommendations, so that processes can be appropriately aligned, e.g. training maintenance engineers to schedule predictive maintenance using the outputs of the predictive maintenance model in their daily work. Step 2: This step involves extracting insights using the machine learning algorithm. The purpose of this step is to select and run the appropriate machine learning algorithm and doing the actual maths and number crunching. The objective is to find the patterns in the data and feature selection. Though a sophisticated AI-based algorithm is capable of finding all the features in the data, involvement of domain experts during this process helps to generate insights and improve the ability to explain the evolved solutions. Creating new features just helps the machine to find patterns more easily and also helps humans to describe and act on these patterns. The purpose of this step is to utilize different machine learning algorithms to identify these patterns. Typically, one can distinguish among descriptive analytics (what happened in the past and why?), predictive analytics (what will happen in the future?), and prescriptive analytics (how can we change the future?). In all these steps, simple but also quite sophisticated methods can be used. More and more advanced techniques in AI and machine learning are being used due to the increased amount of available data and computing power. Figure 2.6 depicts how data science becomes an iterative process

2.7 From Valuable Data Analytics Results to Achieving Business Impact: The Downstream Activities

that leverages both human domain expertise and advanced AI-based machine learning techniques.

2.7 From Valuable Data Analytics Results to Achieving Business Impact: The Downstream Activities The downstream part of the insights value chain is comprised of non-technical components. It involves people, processes, and business understanding that – through a systematic approach – these new data-driven insights can be operationalized via an overall strategy and operating model (Holger Hürtgen, 2018). 2.7.1 Turning Insights into Action

Once we have extracted important insights from the models, the next crucial step starts: turning these insights into action in order to generate a business impact. An example would be when a predictive maintenance model gives you warning as to when a compressor or some asset might break down, but maintenance is still required. It is very crucial to understand that just knowing the probability of a breakdown is not sufficient; prevention, not prediction, is the key to business impact. Turning insights into action thus requires two things: first, to understand the insights coming from the data analytics and to know what to do. Second, even once it becomes clear what action needs to be taken, success will depend on when and how that action is taken. This step is very crucial. Knowledge generation by data analytics software is not sufficient; taking corrective and preventive actions from these insights is the key driver for business impact. 2.7.2 Developing Data Culture

The long-term success of digital transformation requires a company to develop a data culture. This essentially means developing a culture so that all of the business decisions of the company would be based on data analytics and the regular employees of the company would be equipped to implement the data analytics insights into their day-to-day actions. A company’s internal structure and reward systems should be adopted in such a fashion that promotes the data culture. 2.7.3 Mastering Tasks Concerning Technology and Infrastructure as Well as Organization and Governance

In this step, the organization work process, culture, responsibility hierarchy, governance structure, etc., need to be changed in such a way that will facilitate organizations to take action on the insights from advanced analytics and create an impact. An organization needs the right set of easy-to-use tools – e.g. dashboards or recommendation engines – to enable personnel to easily generate business insights and a working environment that facilitates the integration of those insights, e.g. governance that enables and manages the necessary changes within the organization.

25

26

2 Big Picture of the Modern Chemical Industry

References Holger Hürtgen, N.M. (2018). Achieving business impact with data. Retrieved September 25, 2019, from https://www.mckinsey.com/business-functions/mckinsey-analytics/ website: https://www.mckinsey.com/business-functions/mckinsey-analytics/ourinsights/achieving-business-impact-with-data. Ji, X., He, G., Xu, J., and Guo, Y. (2016). Study on the mode of intelligent chemical industry based on cyber-physical system and its implementation. Advances in Engineering Software, 99: 18–26. https://doi.org/10.1016/j.advengsoft.2016.04.010. Klei, A., Moder, M., Stockdale, O., Weihe, U., & Winkler, G. (2017). Digital in chemicals: From technology to impact. Retrieved September 25, 2019, from https://www.mckinsey .com/industries/chemicals/our-insights/ website: https://www.mckinsey.com/ industries/chemicals/our-insights/digital-in-chemicals-from-technology-to-impact/. Wang, X.Z. (1999). Data Mining and Knowledge Discovery – An Overview. https://doi.org/ 10.1007/978-1-4471-0421-6_2.

27

3 Profit Maximization Project (PMP) Implementation Steps 3.1 Implementing a Profit Maximization Project (PMP) Once a process unit is identified as a potential PMP application, then the following project steps are required to implement the project. These project steps are not written in stone and can be changed and modified as per the needs of individual chemical companies. Various global giant chemical companies have implemented these steps to a varying degree and have sometimes modified the sequence of these steps as per their own strength and business requirements. The project steps are summarized in Figure 3.1. In the following sections, each of these steps is discussed. 3.1.1 Step 1: Mapping the Whole Plant in Monetary Terms

The aim of a profit maximization project is to maximize the profit generation in dollar per hour terms and sustain the profit at its peak value. Hence the first step of a PMP project is to calculate how much USD/h profit is generating from the plant in every hour on a real-time basis. As a first step this is done by considering the whole plant as a big black box and mapping it as raw material and utilities as input to the black box and product waste and vent losses as output from the box. The value of each of these inputs and outputs are then calculated as a USD/h term. This gives an overall idea of how much profit is generating from the whole plant. In a second step, a more detailed calculation was done to estimate the USD/h generated or consumed in each major process equipment for the whole plant. Mapping the whole process in USD/h terms makes it easy to view where the actual loss of profit is occurring and to focus on that. This also combines different performance parameters usually used in process industries, like energy efficiency, yield, selectivity, specific consumption of utilities, etc., in a single cost framework and in order to make it easy to visualize the whole process in a single unit of measurements. This will help to access the current operation of the plant and gives an indication of where to focus in order to increase profit. 3.1.2 Step 2: Assessment of Current Plant Conditions

Various process equipment like distillation columns, reactors, furnaces, heat exchangers, pumps and compressors, etc., form the backbone of chemical plants. Before jumping to increase profit, it is very important to know how they are performing currently in the Profit Maximization Techniques for Operating Chemical Plants, First Edition. Sandip Kumar Lahiri. © 2020 John Wiley & Sons Ltd. Published 2020 by John Wiley & Sons Ltd.

28

3 Profit Maximization Project (PMP) Implementation Steps

Mapping the whole plant in monetary terms

Perform a maximum capacity test run

Develop and implement real time APC

Assessment of current plant condition

Develop real time fault diagnosis system

Develop data driven offline process model for critical process equipment

Assessment of base control layer of plant

Develop advance process monitoring framework by applying latest data analytics tools

Optimizing process operation with developed model

Apply new design methodology for process equipment

Assessment of loss from plant

Identification of improvement opportunity in plant and functional design

Modeling and optimization of industrial reactors

Maximize throughput of all running distillation columns

Figure 3.1 Different steps in profit maximization project (PMP) implementation

plant. In this step, the performance of various process equipment is assessed and any performance limitations are critically examined. It also involves assessing the efficiency of different parts of the process, such as yield, catalyst selectivity, quality consistency, etc. It is of the utmost importance to assess the performance of a base process and equipment layer and improve it before trying to build a PMP application over it. In this step an overall assessment of the base equipment layer was performed scientifically and corrective steps were taken to rectify limitations, if any (Lahiri, 2017c). This essentially means identification of any problems in distillation columns, reactors, furnaces, heat exchangers, pumps, and compressors and applies to various techniques like maintenance of rotating equipment, heat exchanger cleaning, etc., to rectify the problems. 3.1.3 Step 3: Assessment of the Base Control Layer of the Plant

The regulatory proportional, integral, and derivative (PID) control loop forms the base layer of a control system. It is of the utmost importance to assess the performance of the base control layer and improve upon it before trying to build a PMP application over it. In this step an overall assessment of the base control layer was performed scientifically and corrective steps were taken to rectify limitations, if any. This essentially means identification of any problems in control valves (like hysteresis, stiction, valve oversize, or undersize phenomena), measuring sensors (like noise, range of instruments, calibration, etc.), PID controller tuning, oscillation in process parameters, etc., and application of various techniques like controller tuning, maintenance of control valves, calibration of instruments, etc., to rectify the problems (Lahiri, 2017a). Enhancement of control performance actually reduces the variations in key economic parameters of a process and then the DCS panel engineer is able to push the process further near to its constraints. Using only this step, a 1–5% increase in profit has been reported in various literatures by various global chemical companies.

3.1 Implementing a Profit Maximization Project (PMP)

3.1.4 Step 4: Assessment of Loss from the Plant

What are the major energy and product losses in a process? This is the first question that people should ask before embarking on a significant effort to improve profit. The answer to this question could lead to identification of major improvement opportunities and help to define the need for a large profit improvement effort. In a chemical process, a valuable product can be lost either with wastewater or vent to flare. In this step, a systematic approach is followed to calculate how much money gets lost in USD/h terms due to waste and vent. Not only product loss, but also energy loss, account for a major erosion of profit in many chemical plants. In a process, energy losses consist of both thermal and mechanical losses (Zhu, 2013). Thermal losses typically originate from column overhead condensers, product rundown coolers, furnace stack, steam leaks, poor insulation of heat exchangers/piping and vessels, and so on. Mechanical losses could also be significant, which usually occurs in rotating equipment, pressure letdown valves, control valves, pump spill back, heat exchangers, pipelines, and so on. Some of the wastewater, vent gas to flare, and thermal and mechanical losses are recoverable with a decent payback of investment, but many others do not. An energy and product loss audit seeks to identify key recoverable losses. The audit is relatively quick and is designed to determine improvement potential. If the energy loss audit identifies large energy or product losses, more detailed energy assessment efforts will be undertaken later if so required. After identifying all the product and energy losses in a chemical complex, small improvement projects can be initiated and implemented to stop or reduce these losses. In this way, by reducing the money drain from plants, profit can be increased. Many companies in the world have been able to increase their profit 1–5% by following this simple but effective step. 3.1.5 Step 5: Identification of Improvement Opportunity in Plant and Functional Design of PMP Applications

Before building a PMP application for a process, the concerned chemical engineer must understand all the relevant aspects of the process, its various limitations, how it makes profit, and what area can be exploited to increase profit. As a starting step, the PMP engineer usually surveys PFDs and P&IDs of the process under study, meets with operations engineers and a specialized work force, and finds all the opportunities and constraints to increase profit in a plant. In this step the PMP engineer usually assesses the current operation, analyzes historical data, understands the various safety and process constraints and equipment limitations, etc. A functional design aims to identify all the existing opportunities to increase the plant profit (Lahiri, 2017c). In this step, the PMP engineer formulates various profit improvement strategies and identifies all potential applications where application of data analytics and modeling and optimization techniques can be applied to improve profit. A preliminary feasibility study is undertaken to identify whether an APC application can be implemented. In this step, an overall idea and forward path is made regarding which PMP application will be used and where to tap into the profit increase opportunity. A functional design step basically provides a map of every opportunity and which methods will be used to exploit a particular opportunity. The success of the profit maximization will greatly depend on how the functional design is formulated in order to tap into the potential margins available in the process. This step requires synergy between expertise and experience of the domain engineer or plant process engineer and the PMP engineer.

29

30

3 Profit Maximization Project (PMP) Implementation Steps

3.1.6 Step 6: Develop an Advance Process Monitoring Framework by Applying the Latest Data Analytics Tools

The most important step in developing a profit management solution to optimize a process is to be able to measure what process performance looks like against a reasonable set of benchmarks. This involves capturing process performance and relevant cost data related to the process and organizing it in a way that allows operations to quickly identify where the big cost consumers are and how well they are doing against a consumption cost target that reflects the current operations. Only then it is possible to do some analysis to determine the cause of deviations from a target and take appropriate remedial action. For this purpose, the concept of key profit indicators is introduced (Zhu, 2013). In this step, a dashboard is prepared for a monitoring purpose where all the key profit indicators along with their target values are shown in real-time. This dashboard enables an operation engineer to quickly identify any deviations of cost performance of the process and take necessary preventive and corrective actions. In this step an advance process monitoring system is developed by using latest data analytics, like clustering, fault tree, a cause and effect diagram, artificial neural network, etc. This monitoring framework helps to visualize the whole process and its performance against targets and quickly highlights any deviations of performance or improvement opportunities. 3.1.7 Step 7: Develop a Real-Time Fault Diagnosis System

Disruption of whole operations due to malfunctions of various process equipment is very common in chemical plants. In the worst case, due to malfunction of compressors, distillation columns, a process instrument, the electrics of a whole plant tripped and a large amount of money was lost. Today’s chemical plants are so complicated and interrelated between various sections that once a plant trips, the whole process becomes destabilized and it takes 1 or 2 days or more to stabilize the process and continue on-spec production. This not only reduces the profit due to a lower production rate but also represents a huge loss due to flaring/draining, production of off-spec production, etc. Early detection of a fault or equipment malfunction can help to take corrective or preventive actions at their incipient stage and thus avoid the financial loss. Fault diagnosis of equipment or a process is an online real-time system, which continuously monitors various equipment-related data (say temperature, pressure, vibrations, etc.) and sends an early alert signal when a fault is detected. This early alert is triggered before the fault actually disrupts the process. This will help the concerned engineer to focus on the particular fault and take preventive action to avoid process disturbance. In most cases where a fault is detected at its incipient stage the operator will be able to avoid trips and reduce the financial loss associated with a plant trip. It is absolutely necessary nowadays to implement a fault detection system in running a plant to increase its on-stream factor, i.e. running hours. 3.1.8 Step 8: Perform a Maximum Capacity Test Run

Capacity expansion is the single most important strategy used to increase profit. Capacity expansion means an increase in the plant throughput (measured by product

3.1 Implementing a Profit Maximization Project (PMP)

flow in MT/h) over and above its name plate capacity. There are various design margins available in various process equipment. A maximum capacity test run is a process that systematically increases plant capacity and exploits these margins. It is possible to increase plant capacity by 5–10% without any major investment and simply utilizing the hidden margins available in installed equipment. The main idea behind a maximum capacity test run is to slowly push the plant capacity until it reaches major equipment or process limitation. This is a very important tool used to increase the plant profit by systematically running the plant at a higher capacity. 3.1.9 Step 9: Develop and Implement Real-Time APC

PID control formed the backbone of a control system and is found in a large majority of CPIs. PID control has acted very efficiently as a base layer control over many decades. However, with the global increase in competition, process industries have been forced to reduce the production cost and need to maximize their profit by continuous operation in the most efficient and economical manner. Most modern chemical processes are multivariable (i.e. multiple inputs influence the same output) and exhibit strong interaction among the variables (Lahiri, 2017b). In a process plant, it is only seldom that one encounters a situation where there is a one-to-one correspondence between manipulated and controlled variables. Given the relations between various interacting variables, constraints, and economic objectives, a multi-variable controller is able to choose from several comfortable combinations of variables to manipulate and drive a process to its optimum limit and at the same time achieve the stated economic objectives. By balancing the actions of several actuators that each affect several process variables, a multi-variable controller tries to maximize the performance of the process at the lowest possible cost. In a distillation column, for example, there can be several tightly coupled temperatures, pressures, and flow rates that must all be coordinated to maximize the quality of the distilled product. Advance process control (APC) is a method of predicting the behavior of a process based on its past behavior and on dynamic models of the process. Based on the predicted behavior, an optimal sequence of actions is calculated. The first step in this sequence is applied to the process. Every execution period a new scenario is predicted and corresponding actions calculated, based on updated information. The real task of APC is to ensure that the operational and economic objectives of the plant are adhered to at all times. This is possible because the computer is infinitely patient, continuously observing the plant and prepared to make many, tiny steps to meet the goals (Lahiri, 2017b). APC has established itself as a very efficient tool to optimize the process dynamically, minimize variations of key parameters, and push the plant to multiple constraints simultaneously and improve the profit margin on a real-time basis. 3.1.10 Step 10: Develop a Data-Driven Offline Process Model for Critical Process Equipment

If data is the new oil in a modern chemical industry, then the data-driven modeling technique is a combustion engine. Industrial chemical processes are complex to understand and difficult to model. However, to increase profit and run the chemical processes at their optimum, availability of a reliable mathematical model is very crucial.

31

32

3 Profit Maximization Project (PMP) Implementation Steps

Chemical engineering has not developed to accurately develop the phenomenological model of complex chemical processes like catalytic reactors, adsorption, etc. As the chemical industry has a huge amount of historical operating data available in its server, a data-driven process modeling technique has emerged as a viable option. How to use various data-driven modeling techniques to increase profit in plants is a key challenge. Data-driven modeling is an important tool used to increase profit by modeling major process equipment and then optimizing various process parameters associated with it so that performance of the equipment is maximized. In this step, big ticket items like a reactor, major distillation columns, and major compressors are identified where developing a model and later optimizing it can have a big impact on plant profitability. 3.1.11 Step 11: Optimizing Process Operation with a Developed Model

Once a process model is developed for a major equipment, the next step is to optimize various process parameters so that efficiency or performance of that equipment is maximized. Due to cut-throat competition in business, companies now want to reduce their operating costs by optimizing all their available resources, be it man, machine, money, or methodology. Optimization is an important tool that can be utilized to strike a proper balance so that profit can be maximized in the long run. Since capital costs are already incurred for a running plant, optimization essentially boils down to minimization of operating costs for the operating plants. In running a chemical plant, there is a huge scope to optimize the operating parameters, like temperature, pressure, concentration, reflux ratio, etc., which give either higher profit through higher production or lower operating costs. There are many ways to optimize the operating conditions of reactors, distillation columns, absorbers, etc., to enhance their profitability. Various recent stochastic optimization techniques, like genetic algorithm, differential evolution, particle swarm optimization, etc., have been used to optimize the developed data-driven model. 3.1.12 Step 12: Modeling and Optimization of Industrial Reactors

A reactor is the only major equipment that can convert raw materials to a value-added product. In chemical plants, the real value addition is done only in the reactor. Downstream separation units like distillation towers can be considered as cost centers because they consume energy to separate different products made in the reactor. All the downstream unit operations are for separation of the product and enrichment. Separation units consume energy and cost, whereas the reactor generates money by adding value to raw materials by converting them to a product. Hence, from a profit point of view, reactors are different from a downstream distillation column or other separation units. In that respect, there is a huge potential impact of reactor optimization on overall plant profitability. The first step of reactor optimization is to know the governing kinetic equations. Industrial reaction kinetics are not known in many cases. A recent AI-based technique, namely genetic programing, can be used to determine the kinetic equations of unknown industrial reactions. An artificial neural network (ANN) or genetic programing can then be used to model the industrial reactors. Once a reliable model is developed, various stochastic-based optimization techniques can be used to

3.1 Implementing a Profit Maximization Project (PMP)

optimize the reactor parameters to increase selectivity, yield, throughput, etc. In this step a modeling and optimization framework is made to derive more profit from the industrial reactor. 3.1.13 Step 13: Maximize Throughput of All Running Distillation Columns

Distillation is the largest separation unit in any refinery, petrochemicals or chemical plants. Though distillation is considered the most efficient separation process among other separation processes, a distillation column consumes a lot of energy in terms of steam in reboilers. Steam costs in various distillation columns constitute a large chunk of operation costs and in most cases are the second largest contributor to the cost component after raw material costs. Not only the operating cost but also large distillation columns and their presence in sheer large numbers in any CPI, contribute heavily to the plant’s initial investment cost. In short, any CPI distillation unit contributes a very large percentage of both capital costs and operating costs. Therefore, any strategy to reduce capital and operating costs of distillation columns have a significant impact on plant profitability and the strategy can be seen as a multiplier, i.e. it can be applied to many distillation columns already present in the plants. There are three main strategies by which more profit can be earned from an existing distillation column. Strategy 1: Increase the feed in the distillation column to produce more products until limited by process constraints like flooding, entrainment, etc. The main constraints in the distillation column should be hydraulically stable so that it can produce an on-spec product consistently at a higher load. Strategy 2: Reduce reflux to reduce steam consumption. However, the constraint is that one must always produce a required purity of the product. For an energy intensive distillation column this is a major strategy and can be considered as a problem of process simulation. Strategy 3: Exploit the variations in product purity or steam flow or feed flow. If there are many variations in product flow or product purity, then the column operation has to be stabilized by improving process control. Producing an ultra-pure product (i.e. a purer product than its market specifications) has no economic benefit in the market. Impurity of the product should be at its allowable limit. Either an efflux rate reduction or feed increase has to be performed in order to reduce extra purity of product. This will increase profit. This problem can be tackled by APC. Since a distillation operation can severely impact plant profitability, special attention is needed to optimize a running distillation column. In this step, various computation tools and modeling techniques are applied in a distillation column to maximize an economic benefit from it. 3.1.14 Step 14: Apply New Design Methodology for Process Equipment

To date, traditional methodology has been followed when designing a new process equipment. In the traditional method, equipment is designed based on its functionality. Cost is not taken as an objective function and minimization of the total cost is never taken as a design target in the traditional designing method. A moderate sized chemical

33

34

3 Profit Maximization Project (PMP) Implementation Steps

plant uses 100 heat exchangers, distillation columns, reactors, etc., and minimization of their total cost, i.e both capital and operating costs, can be considered as a potential area to greatly increase profits. Cost minimization based on new design methodology can be applied during the grassroots design time and huge savings can be obtained. With the advent of commercial steady-state simulators (started around the mid-1980s) process equipment design took a giant leap forward and computer software/simulators marketed by Aspen, Hysis, and Pro-II have been extensively used by designers around the world. With the faster computer, it is now possible to check billions of design options for a single piece of equipment and the lowest cost design can be finally selected. In this step a new stochastic optimization-based methodology is developed for process equipment design. This new methodology uses minimization of the total cost as the design target while obeying all operational, safety constraints and equipment limitations. Various stochastic optimization algorithms are used to search the entire feasible space and find out the most cost-effective design.

References Energy and Process Optimization for the Process Industries (2013). In F.X.X. Zhu (Ed.), Energy and Process Optimization for the Process Industries. https://doi.org/10.1002/ 9781118782507. Lahiri, S.K. (2017a). Assessment of regulatory base control layer in plants. In Multivariable Predictive Control (pp. 77–99). https://doi.org/10.1002/9781119243434.ch6. Lahiri, S.K. (2017b). Introduction of model predictive control. In Multivariable Predictive Control (pp. 1–21). https://doi.org/10.1002/9781119243434.ch1. Lahiri, S.K. (2017c). MPC implementation steps. In Multivariable Predictive Control (pp. 55–62). https://doi.org/10.1002/9781119243434.ch4.

35

4 Strategy for Profit Maximization 4.1 Introduction Profit maximization is usually the top priority of a company’s top management. However, many external and internal factors influence a company’s profit. Most of the external factors like international price fluctuation of raw materials and final products, and price variations of utilities and chemicals are usually beyond the control of the company’s management. However, production costs, energy efficiency, and process and equipment efficiency can be influenced by a company’s operating persons. In such a case, profit maximization essentially boils down to minimization of the cost of production. In some chemical plants, cost optimization work falls into no man’s land. However, all the operating persons starting from plant operators, production engineers, and supervisors feel that they are operating the plant in the most cost-efficient way. The reality is that there is a large scope for cost efficiency improvement. To find out the truth, you may ask a few questions: What key performance index (KPI) is applied to measure the cost efficiency? How is profit tracked and monitored on a real-time basis? What cost indicators are defined for the key equipment? How do you set up targets for these indicators (Zhu, 2013)? The quality of answers to these four questions will indicate whether the plant management only has good intentions but without a proper measure in place. If no profit or cost matrices are used to measure the performance level, no KPIs are applied for major processes or equipment, and no targets are employed for monitoring or identifying improvement, then a profit maximization program is only on the basis of good intent but without foundation. If you cannot measure profit in real time, you cannot control in real time. A good foundation needs to be built first before building a castle on it. It is possible to get people motivated with good intentions or with a big vision of management. However, to sustain that motivation, plant people need to know what to do and what is the right direction to maximize profit (Zhu, 2013). To build a solid foundation for a profit maximization framework two key concepts need to be understood first, namely cost intensity and key cost indicators. The concept of cost intensity sets the basis for measuring economic performance of the process, while the concept of key cost indicators provides guidance for what to do and how. Both cost intensity and key indicators are the cornerstones of an effective and sustainable profit management system.

Profit Maximization Techniques for Operating Chemical Plants, First Edition. Sandip Kumar Lahiri. © 2020 John Wiley & Sons Ltd. Published 2020 by John Wiley & Sons Ltd.

36

4 Strategy for Profit Maximization

4.2 How is Operating Profit Defined in CPI? A company operates to make profit. The best way to summarize a profit maximization project’s objective is “Maximize profit in USD/h and sustain it.” Usually profit is defined as: ∑ ∑ Raw material ratei Profit in USD∕h = Product ratei × Unit sales valuei − ∑ × Unit cost valuei − Utility consumption ratei ∑ × Unit cost valuei − Overheads (4.1)

4.3 Different Ways to Maximize Operating Profit To maximize the profit function given in Equation (4.1), some of the options are given below (refer to Figure 4.1): Maximize each product rate or maximize the total production rate: This is done by running the plant at the highest possible capacity or debottleneck the plant to enable it to produce more. Maximize the unit sales value: The unit sales value is usually determined by the domestic or international market supply–demand scenario and is usually beyond the control of the company’s local management. However, making a premium quality product and establishing a brand image or reputation in the market can lead to a higher sales value. Good customer focus and genuine after-sales service can influence this. Maximize the product rate that gives a higher profit margin: Not all products earn the same profit margins. Some products have much higher profit margins compared to others. Maximizing the product rate that fetches a higher profit margin can be obtained by sacrificing the low margin product. This is easy to say but sometimes difficult to achieve in a plant. How we can identify the influencing process parameters that will help to do that will be discussed in later chapters in this book. Minimize the raw material consumption rate: In the process industry the raw material cost is the biggest share in the production cost. Thus, slightly reducing raw material consumption while keeping the same target production will have a huge impact on profitability. Specific consumption of raw material is defined as the consumption of a particular raw material in MT per MT of product. The aim is to reduce specific

Maximize each product rate

Minimize raw material consumption rate Maximize unit sales value

Maximize product rate which gives higher profit margin

Minimize utility consumption Minimize raw material unit cost

Figure 4.1 Different ways to maximize the operating profit of chemical plants

Minimize utilities unit cost

4.4 Process Cost Intensity

consumption of the raw material. Raw material specific consumption can be reduced by increasing catalyst selectivity, process yield, by reducing waste, by increasing process efficiency, by increasing raw material purity, etc. Judicious optimization of the reactor operation is the surest way to minimize raw material consumption. This will be discussed in detail later in this book. Minimize raw material unit cost: This is done by selecting a good vendors’ supply at a cheap price, a long-term contract with the vendor for price negotiation, exploring the world map (especially China and Middle East companies) to source raw material, reducing raw material transportation costs, blending of cheap raw materials, etc. A large refinery in India is sourcing cheap inferior quality crude from different parts of world and then keeping them in different crude tanks (with, for example, 20 tanks for each different quality of crude) in the plant. Some crudes have a high sulfur content with low wax, some crudes have a low sulfur content but high wax, etc. Due to the inferior quality, which does not conform to the feed specifications of most refineries, this crude can be purchased at a very low price. A blending algorithm can be made to blend crudes of different qualities so that the final quality of crude after blending conforms with the refinery feed specifications. In this way, the Indian refinery is making money before it refines the crude. Minimize utility consumption: This is an area that is relatively easy to achieve in a plant but is also the most neglected area. Some of the following methods are commonly used to minimize utility consumption: • Carry out a detailed energy audit to identify an area where energy or a utility is lost. Reducing the loss and waste of utilities can surely reduce their consumption. • Increase energy efficiency of major energy intensive process equipment. This is done by constantly monitoring their performance parameters and tracking their responsible parameters on a daily basis. • Increase process–process heat transfer wherever possible. This is done by a pinch study of the heat exchanger network and distillation columns across the complex. Minimize utilities unit cost: Every plant has a different utility cost based on plant locations and plant configurations. A petrochemical plant was burning naphtha in their boiler to produce heat as there was no source of fuel gas. The same plant was producing CBFS (carbon black feed stock) as a low value by-product in their plant and sold it to external parties at a very low throwaway price. The plant has made a detailed study to partially use CBFS in their boiler along with naphtha and modified the boiler burners accordingly. CBFS was mixed with naphtha at a predetermined optimum ratio and was then used to fire the boiler. This has reduced their overall fuel cost. Another plant with the same problems built a coal-based steam plant, which produces steam at much lower price than a naphtha-based steam plant. The payback was one year. Therefore, identifying a cheap energy source and putting in capital investment to reduce the per unit utility cost is one of the ways to reduce utility costs.

4.4 Process Cost Intensity 4.4.1 Definition of Process Cost Intensity

The first step to improve process economic performance is to have an operational quantitative definition of the process economic performance that everyone can agree on and relate to and act upon.

37

38

4 Strategy for Profit Maximization

Let us start with the specific question: how do we define performance for a process? People might think of energy efficiency or process efficiency first. Although energy efficiency or process efficiency is a good measure as everyone knows what it is about, it does not relate resources use to the process feed rate and yields, and thus it is hard to connect the concept of energy efficiency or process efficiency to plant managers and engineers. To overcome this shortcoming, the concept of cost intensity is adopted, which connects process resources (energy, material, assets, etc.) use, and production activity. The cost intensity is developed from a similar concept of energy intensity that originated from Zhu (Zhu, 2013). The concept of cost intensity allows them to examine the trends that prevailed during both increasing and decreasing energy prices. By definition, cost intensity (CI) is described by Cost intensity (CI) =

Cost incurred C = Activity A

(4.2)

The total cost incurred to perform an activity (C) is the numerator, while the common measure of activity (A) is the denominator. The physical unit of the production plant can be MT/h of total product or total feed. Thus, the cost intensity of a process can be defined as Cost incurred in USD∕h (4.3) Cost intensity (CI) = Quantity of product in MT∕h Cost intensity thus refers to the cost incurred by the process per ton of product processed. Cost intensity was defined in Equation (4.2) and directly connects all resource (assets, energy, manpower, etc.) use to production as it puts production as the basis (denominator). In this way, resource use is measured on the basis of production, which is in the right direction of thought: a process is meant to produce products supported by resources (Zhu, 2013). Cost intensity thus has a strong correlation with energy efficiency and process efficiency or every resource efficiency that was put in the process to generate a product. Directionally any efficiency improvement in process and equipment can contribute to a reduction in cost intensity. Therefore, we can come to agree that cost intensity is a more general concept for measuring process efficiency indirectly. Before adopting the concept of cost intensity, one legitimate question is: Which one feed rate or product rate should be used as the measure of activity (Zhu, 2013)? For plants with a single most desirable product, the measure of activity should be the product. For plants making multiple products, it is better to use the feed rate as the measure of activity. The explanation is that a process may produce multiple products and some products are more desirable than others in terms of market value. Furthermore, some products require more resources to make than others. Thus, it could be very difficult to differentiate products for resource use. If we simply add all products together for the sum to appear in the denominator in Equation (4.2), we encounter a problem, which is the dissimilarity in products, as discussed. However, if feed is used in the denominator, the dissimilarity problem is nonexistent for cases with a single feed, and the dissimilarity is much less a concern for multiple feed cases than for multiple products because, in general, feeds are more similar in composition than products (Zhu, 2013).

4.4 Process Cost Intensity

The above discussions leads us to define the cost intensity on the feed basis as follows: Cost intensity (CI) =

Cost incurred in USD∕h Quantity of feed processed in MT∕h

(4.4)

4.4.2 Concept of Cost Equivalent (CE)

A process consumes raw materials, different utilities (like steam, fuel, power, water, instrument air, nitrogen, etc.), and chemicals (like caustic, acids, additives, etc.). Most of these utilities are expressed in different units of measurement. However, to represent them on a comparable basis, all of these utilities are converted to a cost term (USD/h). If all activities of the process can be converted into a cost term, then there will be a uniform basis on which to compare them. The loss from the system, like vent gas or waste water or any other waste stream, can also be converted to a USD/h term.

4.4.3 Cost Intensity for a Total Site

The structure of cost intensity indicators can be organized in a hierarchal manner. That is, intensity indicators are developed for processes first and toward a total site. One may question why the concept of cost intensity does not apply to process sections (e.g. reaction section, product fractionation section) of a process. The reason is that there is strong energy and mass integration between sections of a process unit, and thus cost intensity for sections cannot fairly represent section cost performance. Energy and mass transfer across process units could occur, but the chance is much smaller compared with between-process sections. In case of heat transfer between processes, some adjustment must be made to account for it (Zhu, 2013). To calculate the cost intensity index for the whole site, aggregate cost intensity could be defined simply as the ratio of the total cost equivalent divided by total activity: ∑ CEi Isite = ∑i i Fi where CEi is the total cost equivalent for process i. However, there is a problem here with this simple aggregate approach. Although the cost equivalent is additive, feeds (F) are not because processes usually have different feeds with very different compositions. In other words, the problem with the equation is the dissimilarity in feeds, which cannot be added without treatment. To overcome this dissimilarity problem in feeds, we could think of a reference process with cost intensity for each process known in prior. Thus, the total amount of cost incurred could be calculated for the reference process, as the summation of the cost intensity for the reference processes. Let us derive the mathematical expressions along this line of thought. The reference process can be taken as a similar plant or process elsewhere or for simplicity it can be taken as the PFD value of the process.

39

40

4 Strategy for Profit Maximization

When the cost intensity for a reference process is known or can be calculated, applying the equation above gives the cost use for a reference process as CEi,ref = Ii,ref Fi,ref Since the cost equivalent is additive, the total cost incurred for the reference site is ∑ ∑ CEsite,ref = CEi,ref = Ii,ref Fi,ref Then an intensity index for the site of interest can be defined as the ratio of actual and reference cost use: CEsite,actual Isite = CEsite,ref CEsite,actual can be readily calculated from individual raw material and energy users consisting of fuel, power, steam, boiler feed water (BFW), and cooling water across the site battery limit.

4.5 Mapping the Whole Process in Monetary Terms and Gain Insights The aim of any business is to maximize profit on a USD/h basis and either sustain or continually improve upon it while obeying all safety and operational constraints. Hence it is necessary to express each activity on a USD/h basis. The first step of a profit maximization project is the paradigm shift of a mindset. Process engineers in a plant express the process performance by varieties of term-like yield, catalyst selectivity, raw material and utility specific consumption, energy index, etc. They use many units to express these performance parameters; for example, selectivity is expressed as the percentage of desired product as compared to total product, specific consumption of a utility is expressed as the utility consumed per unit product produced (MT/MT), and the energy index is commonly expressed as kcal/MT of the product. Expressing them in a different fashion and in different unit and fragmenting them in different processing units, such as a distillation column, etc., makes it difficult to visualize the process efficiency in a uniform and holistic way. Mapping the whole process in USD/h terms enables engineers to holistically estimate true process efficiency in business terms. This also enables them to use a single framework of cost and judge different efficiencies like energy efficiency, equipment efficiency, and process efficiency in the same unit of measurement. The first step of a PMP is to change the mindset and start to visualize the process in USD/h terms. The significance of the paradigm shift will be gradually cleared later in this chapter. To understand the concept, a reference ethylene glycol process will be first discussed and then the whole process will be mapped in USD/h terms.

4.6 Case Study of a Glycol Plant Figure 4.2 shows a schematic diagram of a glycol plant. Ethylene is introduced into the recycling gas going into the reactor. Oxygen is mixed at the mixing station in a static mixer. Feed gas is sent to the reactor through the feed

ETHYLENE METHANE

OXYGEN

STEAM

K-2 K-1 R-2

BFW V-1

CHLORIDE TK-1

STEAM

C-1

R-1

STEAM

C-2

C-6

C-5

BFW

E-4

V-2 E-2 C-3

E-1

C-4

DRAIN C-10

C-12

C-11

C-13

TK-2

TK-3

TK-3

MEG PRODUCT

DEG PRODUCT

TEG PRODUCT

Figure 4.2 Schematic diagram of a glycol plant

C-7

C-8

C-9

E-3

PEG

E-5

E-6

E-7

42

4 Strategy for Profit Maximization

gas heater (H-2). A tubular reactor (R-1) consists of several thousand tubes having an internal diameter of 20–50 mm and a length of 6–13 m. Tubes are packed with silver catalyst. Along with the formation of EO (reaction 1), combustion of ethylene into carbon dioxide and water also take place (reaction 2) in the reactor. The fraction of ethylene oxide (EO) further oxidizes to form combustion products (reaction 3). O

1. H2C = CH2 + ½ O2 →

2. H2C = CH2 + 3 O2

3.

O H2C

+ CH2

H2C

CH2

ΔHR,298 = –105.08 kJ/mol

→ 2 CO2 + 2 H2O

ΔHR,298 = –1324.10 kJ/mol

5

ΔHR,298 = –1223 kJ/mol

2 O2

→

2 CO2 + 2 H2O

The catalyst must maximize primary reaction 1 and minimize reactions 2 and 3. Selectivity and activity are important parameters for the EO catalyst. The reactor temperature is maintained at 225 ∘ C at the start run and 270∘ C at an end run with an average pressure of 20 barg. Heat generated in the reactor is taken out by boiling water. Steam is produced in the steam generator in V-1 and V-2 by vaporizing 2 to 4% of water by thermosyphon action. The gas mixture flows from the gas–gas exchanger to the EO reactor. The outgoing product gas stream from the EO reactor exchanges heat with the incoming gas steam in the feed gas heater (E-2). Ethylene oxidation is a continuous process with very low conversion of ethylene per pass in the reactor. The reactants are in the gas phase and react over a silver catalyst in a tubular shell side boiling water-cooled reactor. The reaction temperature is 235 ∘ C for the fresh catalyst, which increases as the catalyst ages. Boiling water on the shell side of the multi-tubular reactors removes the heat of reaction. Water circulation through the shell side of the reactors is by thermosyphon action. The steam–water mixture from the reactor shell side is sent to steam drums where make-up boiler feed water is preheated and steam is separated from water and sent to the 21 barg steam header. EO is separated from the product gas by absorption in water in the EO absorber (C-1). Along with EO, small quantities of methane, ethylene, and carbon dioxide are also absorbed. These gases are removed in C-3 and C-6. EO along with water is taken out from the stripping column (C-3). Overhead gases from C-1 are sent to the CO2 absorber C-2, where the gas makes contact with hot activated potassium carbonate solution and CO2 is absorbed to form bicarbonate. The bicarbonate solution is regenerated in regenerator C-4 using steam and is recycled back for reuse. CO2 released in the regenerator is vented to atmosphere. The overhead gas from C-2 is compressed and recycled back to reactor R-1 by cycle gas compressor K-1 and a small stream of recycled gas is purged to control the build-up of nitrogen and argon. In the reabsorber column (C-5), EO is reabsorbed at very low pressures in water. The quantity of other gases reabsorbed at lower pressures is negligible and they leave the top of the reabsorber column to be reclaimed again and added back to cycle gas loop.

4.7 Steps to Map the Whole Plant in Monetary Terms and Gain Insights

Reclaim compressor K-2 is used to compress gases coming out from C-5 along with methane, which is used as ballast. The EO–water mixture from C-5 at 383.15 K is heated to 418.15 K in a glycol feed heater (H-8) and sent to the ethylene glycol (EG) reactor (R-2). The glycol reactor is an adiabatic pipeline reactor designed to produce the maximum amount of mono ethylene glycol (MEG) (reaction 6) with water to an EO ratio of 22:1 in the reactor feed. The residence time in the reactor is about 6 to 10 minutes. The higher amount of water is necessary to reduce subsequent exothermic non-catalytic reactions, which produce diethylene glycol (DEG) (reaction 7), tri-ethylene glycol (TEG) (reaction 8), and higher glycols. The formation of these by-products is unavoidable due to the faster rate of formation than MEG. Reactor effluent consists of 86% water coming out at 463.15 K. The reactions are: 6.

7.

8.

O H2C

CH2 O

H2C

CH2 O

H2C

CH2

+ H2O → HO-CH2-CH2-HO

ΔHR,298 = –95.24 kJ/mol

+ HO-CH2-CH2-HO → HO-CH2-CH2-O-CH2-CH2-OH ΔHR,298 = –129.90 kJ/mol + HO-CH2-CH2-O-CH2-CH2-OH →

HO-CH2-CH2-O-CH2-CH2-O-CH2-CH2-OH

ΔHR,298 = –101.78 kJ/mol

4.7 Steps to Map the Whole Plant in Monetary Terms and Gain Insights The following sections explain various steps used to map the plant in monetary terms and how to extract knowledge from it. Figure 4.3 summarizes the steps. 4.7.1 Step 1: Visualize the Plant as a Black Box

The first step is to visualize the plant as a black box where raw material, utilities, and chemicals are input to the plant and products and waste stream are output. Figure 4.4 and Table 4.1 summarize the idea. Ethylene and oxygen are input as raw material and

Visualize the plant as black box

Data collection from data historian and preparation of cost data

Calculation of profit margin

Gain insights from plant cost and profit data

Cost benchmarking

Calculation of relative standard deviations of each parameters to understand the cause of variability

Plot Production cost and profit margin for full one year and gain insights.

Generation of Production cost and profit margin table for full one year.

Figure 4.3 Steps to map the whole plant in monetary terms and to gain insights

43

Raw materials Ethylene Oxygen

Product MEG DEG TEG

Consumption UOM 51.15 MT/hr MT/hr 56.15

Utilities Consumption UOM 15.42 MMB/hr Fuel Gas 16.60 MWH Electric Power 19.53 Cooling Water KM3/hr BFW 284.63 MT/hr 79.06 MT/hr Steam (HP, MP and LP) 592.97 NM3/hr Nitrogen 500.94 NM3/hr Instrument air Chemicals Consumption UOM Reactor Inhibitor 1.95 Kg/hr Potasium Carbonate Kg/hr 1.73 Vanadium Penta Oxide Kg/hr 0.59 Kg/hr Antifoam 0.13 Kg/hr Caustic Soda 118.59 Kg/hr 63.25 Hydrochloric Acid

Figure 4.4 Representing the whole plant as a black box

Raw Material Utilities

Chemicals

Consumption UOM 79.06 MT/hr 7.61 MT/hr 0.38 MT/hr

Finished Product Plant as black box

Return condensate

Return utilities Consumption UOM 60 Condensate MT/hr

Vent and waste losses Losses Purge Waste water PEG

Consumption UOM MT/hr 0.95 MT/hr 54.55 MT/hr 0.04

4.7 Steps to Map the Whole Plant in Monetary Terms and Gain Insights

Table 4.1 Representing the whole plant as a black box with consumption and cost dataa)

UOM

Consumption production rate

Rate/unit cost (USD)

Total cost (USD/h)

MEG

MT/h

79.06

755.30

59 715.75

DEG

MT/h

7.61

868.11

6609.56

TEG

MT/h

0.38

1442.40

547.39

Total sales value

USD/h

Add return condensate

MT/h

60.00

0.50

30.00

Ethylene

MT/h

51.15

367.21

18 784.28

Oxygen

MT/h

60.32

51.99

3136.52

Description

Production

66 872.70

Raw material

Total cost of raw material

21 920.80

Utilities Fuel gas

MMB

15.42

0.76

11.70

Electric power

MWH

16.60

32.40

537.94

Cooling water

KM3

19.53

17.69

345.36

BFW

MT/h

284.63

0.41

115.27

Steam (HP, MP, and LP)

MT/h

79.06

3.81

300.99

Nitrogen

NM3

592.97

0.05

32.41

Instrument air

NM3

500.94

0.005

Total cost of utilities

2.28 1345.96

Chemicals Reactor inhibitor

kg

1.95

1.62

Potassium carbonate

kg

1.73

1.24

2.13

Vanadium pentaoxide

kg

0.59

22.74

13.49

Antifoam

kg

0.13

5.27

0.68

Caustic soda

kg

118.59

0.15

18.01

Hydrochloric acid

kg

63.25

0.14

8.54

Total cost of chemicals

3.16

46.02

Waste/loss Purge loss

MT/h

0.90

128.53

115.67

Waste water

MT/h

54.55

0.86

47.13

PEG

MT/h

0.04

600.00

Total cost of waste

24.00 186.81

a) List of abbreviations: UOM (unit of measurement), MEG (monoethylene glycol), DEG (diethylene glycol), TEG (triethylene glycol), BFW (boiler feed water), HP (high pressure), MP (medium pressure), LP (low pressure), PEG (polyethlene glycol), MMB (one million British thermal units), MWH (megawatt-hour), KM3 (kilo cubic meter).

45

46

4 Strategy for Profit Maximization

fuel gas, electric power, cooling water, boiler feed water, steam, nitrogen, and instrument air are considered as utilities and treated as second input. Different chemicals used in the glycol production process, such as inhibitor, carbonate, antifoam, caustic, and hydrochloric acids are considered as a third input to the plant. Outputs from the plants are three main products from the plant, such as MEG, DEG, and TEG. A glycol plant generates condensate after condensing the steam in different reboilers and exports this condensate to utility and offsite plants. This return condensate is considered as second output from the plant. During processing, the plant produces some vent gas, waste water, and polyethylene glycol, which are considered as waste products and the third output. 4.7.2 Step 2: Data Collection from a Data Historian and Preparation of Cost Data

For the purpose of profit calculation of a process unit, the important thing is to identify the main cost consumers and give a reasonable estimate for missing data. Going overboard to collect miniature details and chase utmost precision should be avoided at this stage. Doing so may actually waste effort because such fine details are most likely not needed at this stage and will not make a reasonable impact on profit maximization (Zhu, 2013).What values in USD/h terms does a plant generate? How much cost has plant incurred in terms of raw material, utilities, and chemical consumption in USD/h to generate the above value? All the hourly consumption and product rate data as summarized in Table 4.1 should be collected from a data historian for a minimum one-year period. One such data when plant was run at 100% nameplate capacity (approximately) are shown in Figures 4.4 and 4.5 and Table 4.1. It is important to realize that the unit cost of product, raw material, utilities, and chemicals should be taken as the actual cost to the specific plant. Business plant data for the plant under study can be taken. Those cost data should not be taken from any international price list or any external published cost. Cost data should be as realistic as actual figures for the specific plant under study. Waste stream cost can be calculated from the composition of waste streams and how much raw material or utilities are lost to generate that waste. Add any waste treatment cost over and above that to generate unit cost of waste stream. Note that waste stream cannot be discharged as it is basis and it has to undergo some waste treatment (like waste water treatment cost) before it can be discharged outside company’s battery limit. 4.7.3 Step 3: Calculation of Profit Margin

From Table 4.1 the following cost parameters can be calculated by using the following formulas: Cost of production (USD∕h) = Cost of all raw materials + Cost of all utilities + Cost of all chemicals + Cost of treatment for all waste products

Raw materials Ethylene Oxygen

Product MEG DEG TEG

Consumption UOM 18784 USD / hr USD / hr 3137

Utilities Consumption UOM 12 USD / hr Fuel Gas 538 USD / hr Electric Power 345 Cooling Water USD / hr BFW 115 USD / hr 301 USD / hr Steam (HP, MP and LP) 32 USD / hr Nitrogen 2 USD / hr Instrument air Chemicals Consumption UOM Reactor Inhibitor 3 USD / hr Potasium Carbonate USD / hr 2 Vanadium Penta Oxide USD / hr 13 USD / hr Antifoam 1 USD / hr Caustic Soda 18 USD / hr 9 Hydrochloric Acid

Figure 4.5 Mapping the whole plant in monetary terms

Raw Material Utilities

Chemicals

Consumption UOM 59716 USD / hr 6610 USD / hr 547 USD / hr

Finished Product Plant as black box

Return condensate

Return utilities Consumption UOM 30 Condensate USD / hr

Vent and waste losses Losses Purge Waste water PEG

Consumption UOM USD / hr 116 USD / hr 47 USD / hr 24

48

4 Strategy for Profit Maximization

Cost of production (USD∕h) =

∑

Raw material consumption (MT∕h)

i

× Unit cost (USD∕MT) ∑ Utility consumption (MT∕h) × Unit cost (USD∕MT)

+

j

+

∑

Chemical consumption (MT∕h) × Unit cost (USD∕MT)

k

+

∑

Waste generated (MT∕h) × Unit treatment cost (USD∕MT)

l

Sales value (USD∕h) =

∑

Product rate (MT∕h) × Unit sales value (USD∕MT)

i

Profit margin (USD∕h) = Sales value (USD∕h) − Cost of production (USD∕h) Many oil refineries and petrochemical plants call the above profit margin as net payback. Note that Table 4.1 was generated from hourly data. It can be generated also by taking the average hourly data from daily data or monthly data. This table should give a representative true picture of plant hourly profit generation. The following two parameters are also collected in order to gain a clear idea about plant profitability: Profit margin (USD∕h) Total product rate (MT∕h) Cost of production (USD∕h) Cost intensity (USD∕MT) = Total product rate (MT∕h) Profit margin per ton of product (USD∕MT) =

Table 4.2 summarizes all these cost and profit parameters. It is worth mentioning here that this profit margin is usually refereed as the operating profit margin and is completely different from the plant profit margin as calculated by the plant finance department after deducting overhead costs, taxes, fixed costs, etc. Those overhead terms and fixed costs are neglected here. 4.7.4 Step 4: Gain Insights from Plant Cost and Profit Data

It is very important to understand the different cost terms in Table 4.1 and find implications of those on plant profitability. The following insights have come from Table 4.1. Insight 1: For the reference glycol plant under study 94% of production cost is due to raw material and 6% due to utilities, as summarized in Figure 4.6. Chemical contribution is negligible. Therefore, the main focus to reduce production costs should be directed towards raw material consumption reduction as a priority and then towards utility consumption reduction as the second priority. Effort should not be paid to chemical consumption reduction as their contribution is small and does not bear any large impact on profitability.

4.7 Steps to Map the Whole Plant in Monetary Terms and Gain Insights

Table 4.2 Summary of profit margin and cost intensity Description

USD/h

Total sales value

66 873

Production cost Raw material

21 921

Utilities

1346

Chemicals

46

Waste treatment

187

Total production cost

23 500

Profit margin

43 373

Profit margin per MT of product (USD/MT)

498

Cost intensity (USD/MT)

270

Figure 4.6 Break-up of the total cost of production

utilities 6%

chemicals 0%

raw material 94%

49

4 Strategy for Profit Maximization

Insight 2: From Figure 4.7 it is noted that the contribution of ethylene cost is much higher than oxygen cost. Therefore, effort should be paid to reduce ethylene consumption first. Insight 3: Figures 4.8 and 4.9 show the cost of different utilities and chemicals. The same logic effort should be directed to reducing the consumption of power, cooling water, steam, and BFW as that of the descending priority. Instrument air, nitrogen, and fuel gas reduction are of low priority and have a low impact area.

3136.52

18784.28

50

ETHYLENE

OXYGEN

Figure 4.7 Cost of raw material

Instrument air Nitrogen Steam (HP, MP and LP) BFW Cooling Water Electric Power Fuel Gas 0.00

100.00

Figure 4.8 Cost of different utilities (USD/h)

200.00

300.00

400.00

500.00

600.00

4.7 Steps to Map the Whole Plant in Monetary Terms and Gain Insights

7%

19%

5%

Reactor Inhibitor Potasium Carbonate 29%

Vanadium Penta Oxide Antifoam

39%

Caustic Soda 1%

Hydrochloric Acid

Figure 4.9 Cost of different chemicals (USD/h)

Profit Margin (USD/Hr)

55 000 50 000 45 000 40 000 35 000 30 000 25 000 20 000

0

50

100

150

200

250

300

350

400

No of days in a year

Figure 4.10 Variations of profit margin (USD/h) throughout the year

Insight 4: If someone wants to reduce chemical consumption (though a low priority area), then the focus should be on caustic soda, vanadium pentoxide, and hydrochloric acid. No effort should be wasted on reduction of other chemicals as their impact is insignificant on plant profitability.

4.7.5 Step 5: Generation of Production Cost and a Profit Margin Table for One Full Year

As all the consumption data and cost data are collected for one full year, a consolidate data table combining Table 4.1 and Table 4.2 can be built for one whole year. For simplicity, the daily average hourly data for a full day can be taken as representative data for that day. In this way 365 records will be collected for a year.

4.7.6 Step 6: Plot Production Cost and Profit Margin for One Full Year and Gain Insights

Figures 4.10, 4.11, and 4.12 summarize the variability of the profit margin and production costs for one whole year. Figure 4.13 shows the variations of the main MEG product throughout the year.

51

4 Strategy for Profit Maximization

Profit margin (USD/MT)

52

600 550 500 450 400 350 300 0

50

100

150

200

250

300

350

400

300

350

400

300

350

400

No of days in a year

Figure 4.11 Variations of profit margin (USD/MT of product) throughout the year

350.00 300.00 250.00 200.00 150.00 100.00

0

50

100

150

200

250

No of days in a year

Figure 4.12 Variations of production cost (USD/MT) throughout the year

90.00 85.00 80.00 75.00 70.00 65.00 60.00

0

50

100

150

200

250

No of days in a year

Figure 4.13 Variations of MEG production (MT/h) throughout the year

4.7.7 Step 7: Calculation of Relative Standard Deviations of each Parameter in order to Understand the Cause of Variability

The relative standard deviation of any parameter is calculated as follows: Relative standard deviation (%) =

Standard deviation × 100 Average value

4.7 Steps to Map the Whole Plant in Monetary Terms and Gain Insights

In statistics, RSD stands for relative standard deviation and is also known as the coefficient of variance. %RSD (relative standard deviation percentage) is a statistical measurement that describes the spread of data with respect to the mean and the result is expressed as a percentage. The %RSD function is popular with non-statisticians as the interpretation is based on a percentage result and not some abstract value. To calculate the %RSD in Microsoft Excel a short formula must be used: = (STDEV (Data Range)∕AVERAGE (Data Range)) ∗ 100 = (STDEV (A1 ∶ A11)∕AVERAGE (A1 ∶ A11)) ∗ 100 For example, a relative standard deviation of 6% when your average data for a particular process variable is 40 would mean that the vast majority of data falls between 34 and 46. Your result would read 40 +/– 6%. The result is expressed as a percentage, with a low number ( = P, Compute hidden node inputs: ∑ (net (1) p,j ) = w(1,0) j, I xp,i Compute hidden node outputs: ∑ (x(1) p,j ) = S( w(1,0) j, I xp,i ) Compute inputs to the output nodes: ∑ (net (2) p,k ) = w(2,1) k, j x(1) p,j Compute the network outputs: ∑ (Op,k ) = S( w(2,1) k, j x(1) p,j ) Compute the error between Op, k and desired output, Dp, k Modify the weights between hidden and output nodes: ΔW (2,1 ) k,j = η(Dp,k − Op,k ) S, (net (2) p,k ) x(1) p,j

(1)

Modify the weights between input and hidden nodes: ∑ ΔW (1,0 ) j,j = η {(Dp,k − Op,k ) S, (net (2) p,k ) w(2,1) k,j } × S, (net (1) p,j ) x(1) p, j

(2)

End – for End while Figure 7.5 Typical pseudo code of a back-propagation algorithm

Input layer x1

1

W

Hidden layer

H

1.1

nH

1

Σ

f

Output layer yH

1

W

O

1.1

bH1 x2

xR

1 H n 2 Σ

2

R

WHS,R

nO1

Σ

f

1 O n 2 Σ

yH2

bO2

1

1

nHS bHS

1

yO1

f

yO2

f

yOK

bO1

bH2

S

f

f

yHS WOK,S

nOK

K

bOK 1

Figure 7.6 Architecture of a feed-forward network with one hidden layer (Lahiri & Ghanta, 2009a)

109

110

7 Process Modeling by the Artificial Neural Network

weights to try and produce the correct output (within a reasonable error margin). If it succeeds, it has learned the training set and is ready to perform on previously unseen data. If it fails to produce the correct output it re-reads the input and again tries to produce the correct output. The weights are slightly adjusted during each iteration through the training set (known as a training cycle) until the appropriate weights have been established. Depending upon the complexity of the task to be learned, many thousands of training cycles may be needed for the network to correctly identify the training set. Once the output is correct the weights can be used with the same network on unseen data to examine how well it performs. 7.5.5 Generalizability

Neural learning is considered successful only if the system can perform well on test data on which the system has not been trained. This capability of a network is called generalizability. Given a large network, it is possible that repeated training iterations successively improve network performance on training data, e.g. by “memorizing” training samples, but the resulting network may perform poorly on test data (unseen data). This phenomenon is called “over training.” The proposed solution is to constantly monitor the network performance network on the test data. Hecht-Nielsen (1992) proposes that the weight should be adjusted only on the basis of the training set, but the error should be monitored on the test set. The following strategies are applied: training continues as long as the error on the test set continues to decrease and is terminated if the error on the test set increases. Training may thus be halted even if the network performance on the training set continues to improve.

7.6 Model Development Methodology This section describes the typical steps for ANN-based data-driven model development. The presented procedure is generic and can be applied for any data-driven model development. Steps involved in the development of a reliable model are shown in Figure 7.7. 7.6.1 Data Collection and Data Inspection

The data-driven model is made from large number of historical data. The quality of the developed model is as good or as bad as their input data quality. Tasks usually performed in this step are summarized in Figure 7.7. 7.6.2 Data Pre-processing and Data Conditioning (Lahiri, 2017)

Data gathered from process industries normally involve the following critical characteristics: • Data contain spike or a very high value. • Data contain noise due to the process itself or from measuring transmitters. • Data may contain missing value or the value may freeze.

Preparation before data collection: calibration of relevant transmitters and analyzers and accuracy check of collected data should be performed. This step is important to ensure accuracy of collected data.

Setting of data historian: Installation of data historian software (PIMS) and starting of data collection of all relevant tags with time stamp.

Identification of laboratory data needed for soft sensor and setting of laboratory information management system (LIMS) with time stamp.

Evaluation of the steady state part of the data. Normally model are built with data at steady state. Data in unsteady state i.e. when temperature, pressure etc. of the process is changing, the data may not be representative to build soft sensor.

Assessment of the target variable: it has to be checked, if there is enough variation in the output variables and if this can be modelled at all.

Inspection of data to get bird eye view of data structure and identify obvious problems which may be handled at this initial stage (e.g. locked variables having constant value etc.).

Figure 7.7 Steps followed in data collection and data inspection

112

7 Process Modeling by the Artificial Neural Network

handling of missing data data cleaning

outlier detection and replacement

selection of relevant variables i.e. feature selection

Figure 7.8 Task performed in the data pre-processing and data conditioning step

The tasks performed in the data pre-processing and data conditioning step are shown in Figure 7.8. A systematic approach for data pre-processing for data-driven model building is given in Lin (2007). 7.6.2.1 Outlier Detection and Replacement (Lahiri, 2017)

Outliers are commonly defined as observations that are not consistent with the majority of data, including a missing data point and observations that deviate significantly from normal values. Reasons for an outlier in process industry data are many, e.g. malfunction of process equipment or measuring transmitters, non-functioning of the data collection system, etc., can generate outlier process data. Outlier detection and removing them from a data set is very critical for model development because if not detected outliers have a negative effect on the performance of the final soft sensor model (Lahiri, 2017. 7.6.2.2 Univariate Approach to Detect Outliers

Two main univariate approaches to detect outliers are the Hampel identifier and the 3𝜎 edit rule and their calculation method, shown in Figure 7.9. 7.6.2.3 Multivariate Approach to Detect Outliers (Lin, 2007)

1. Principal component analysis (PCA) is a multi-variate statistical method that reduces the data dimensionality by projecting the data matrix to a lower dimensional space with the help of the loading vectors. The loading vectors corresponding to the k largest eigenvalues can capture the variations of the data and thus contain

Hampel identifier

The 3σ edit rule

Median absolute deviation (MAD) is defined as MAD = 1.4826 median [Ixi – x *l] Where x * is the median of the data sequence. The factor 1.4826 is chosen such that expected MAD= standard deviation for normally distributed data. With the limit of xmed ± 3xMAD, the Hempel identifier identifies most outlier successfully.

The 3σ edit rule is a popular univariate approach to detect outliers Ix(i) – xl > 3σ Where x is the mean of the data sequence. This method labels outliers when data points are 3σ or more standard deviation from mean.

Figure 7.9 Two main univariate approaches to detect outliers

7.6 Model Development Methodology

most of the information. The fitness between data and the model can be calculated using the residual matrix and Q statistics, which measure the distance of a sample from the space of the PCA model (Jackson & Mudholkar, 1979). Hotelling’s T2 statistics indicates the distance between a particular data from the multivariate mean of the data; thus, these statistics provide an indication of variability within the normal subspace (Wise, 1991). The combined Q and T2 tests are used to detect outliers. Given the significance level for the Q and T2 statistics, measurements with Q or T2 values over the threshold are classified as outliers. Outliers are located outside of the 99% confidence ellipse. 2. In another approach, data outliers are detected in two phases. In the first phase the outliers are removed using a univariate approach, then a classic PCA is carried out on the new data set. In the second phase, it a multivariate approach is applied and the covariance matrix is calculated. The proposed procedure uses the ellipsoidal multivariate trimming (MVT) approach (Devlin, Gnanadesikan, & Kettenring, 1981). This trimming method iteratively detects bad data based on the squared Mahalanobis distance: di2 = (xi − x∗i )T S∗−1 (xi − x∗i ) where x* is the current robust estimation of the location and S* is the robust estimation of the covariance matrix. Since the data set has been pre-processed with a Hampel identifier, 95% of data with the smallest Mahalanobis distance are retained in the next iteration. Devlin et al. (1981) suggest that the iteration proceeds until the average absolute change in Fisher z transformations of the elements of the correlation matrix between two successive iterations is less than a pre-defined threshold, or the maximum number of iterations is reached. 7.6.3 Selection of Relevant Input–Output Variables

Input selection is a key step to model an input–output relationship during process model building. Some guidelines are summarized in Figure 7.10 (Lahiri, 2017). 7.6.4 Align Data

A different sampling rate in ANN modeling is often applied in multi-variate systems with several operating sample rates. In many industrial chemical processes, normally product quality parameters are measured offline in a laboratory (once in every 8 hours or so) or by an online analyzer with a long dead time (typically 15 minutes). The input variables like temperature or pressure are measured and recorded every second or minute. Therefore it is necessary to align the data in a proper time scale. It is absolutely necessary that laboratory data should be time stamped properly and align with other continuous data in a proper time scale. 7.6.5 Model Parameter Selection, Training, and Validation (Kadlec, Gabrys, & Strandt, 2009; Lin, 2007)

This is the most critical step in ANN model building. As the model is the heart of ANN, a proper and accurate model selection is key to its performance. Model developers need to give the following parameters as user input during the ANN model building phase: the number of nodes in the hidden layer, the type of activation function

113

114

7 Process Modeling by the Artificial Neural Network

Simple model is the best. Minimum number of input for modelling of output is best. More the input, more will be the noise in the predicted output.

Based on process knowledge, find out probable inputs which has a cause and effect relationship with the output.

Try to select intensive property input like temperature, pressure etc. rather than flow etc. As for example, to model an impurity at distillation column overhead product, obvious input would be tower top temperature and pressure, delta P of column, reflux ratio etc. It is advisable to avoid taking input as reflux flow or overhead product flow etc. Input should be independent of each other i.e. there should not be any data co-linearity between the input. As for example, select only one representative tray temperature for a distillation column rather than choosing 3 or 4 tray temperatures which are dependent on each other.

First start with minimum number of inputs and then increase one by one until there is significant performance improvement of the developed model. If prediction performance is not improved significantly by adding one extra input, don’t add that input.

Figure 7.10 Guidelines for selection of the relevant input output variables

in the input layer and output layer and algorithms for weight-up gradation (like the Levenberg Marquardt algorithm for example), etc. There are a lot of options available for model selection but no clearcut guidelines are available to select which model or for what conditions. For most of the cases, the type of model selected by the developer is based on personal choice and expertise. This can be very detrimental to the development of a good model. The best approach is to remain open minded for all the model types. It is good practice to start with a simple model type with a small number of hidden nodes and gradually increase the model complexity as long as significant improvement in the model’s performance can be observed (using, for example, Student’s t-test; see Gosset, 1908). During the model building phase, the performance of the individual model can be judged by unseen validation data (Hastie, Tibshirani, & Friedman, 2001; Weiss & Kulikowski, 1991). The same approach can also be applied to parameter selection of the pre-processing methods, such as, for instance, variable selection. Normally data-driven models need a large amount of data, which is usually available in modern industries. However, in some instances where lab data are used, only a very small amount of data may be available. Additionally, for some industrial processes where few reliable lab data are available, statistical error-estimation techniques like K-fold cross-validation (Kohavi, 1995) can be applied. This method makes an optimal use of the available data by partitioning it in such a way that all of the samples are used for the model performance validation. Another alternative in these circumstances is to apply statistical re-sampling methods like, for example, bagging (Breiman, 1996) and boosting (Freund & Schapire, 1997). In the case of the first method, a set of training data sets is generated by randomly drawing samples (with replacements) from the available data and training one model for each of the random sets. The final model is obtained by averaging over the particular models’ predictions. In contrast to this, in the case of boosting, the probability of each sample to be drawn is not random but related to the prediction error of the model given the data sample. Additionally, in the case of

7.7 Application of ANN Modeling Techniques in the Chemical Process Industry

boosting, the weights of the contributions of the particular models are calculated based on the model’s performance on a validation data set. After finding the optimal model structure and training the model, the trained ANN model performance has to be judged on the new validation data set once again (Weiss & Kulikowski, 1991). The mean squared error (MSE), which measures the average square distance between the predicted and the correct value, is the most popular performance evaluation technique for the model. Another way to make a performance judgement is by using visual representation of the predictions. In these, the four-plot analysis is a useful tool since it provides useful information about the relation between the predictions and the correct values together with analysis of the prediction residuals (Fortuna, 2007). A disadvantage of the visual methods is that they require the assistance of the model developer and a final decision if the model performs adequately, which is up to the subjective judgment of the model developer. There needs to be an evaluation that the developed model has some resemblance to the underlying physics of the process. In Fortuna (2007), the authors of the book stress the necessity for application of process knowledge during the ANN model development phase. 7.6.6 Model Acceptance and Model Tuning

After developing an ANN model, the model is put on test in offline mode to see how the model prediction matches with fresh data currently generated in DCS. If the model prediction closely matches the actual output, then the model can be accepted in the industry. The usual criteria is that the average prediction error should be less than 1% with the R2 value greater than 0.98. It is very common in the industry that the performance of an ANN model deteriorates over time. The reasons are many. The underlying process may change (e.g. catalyst selectivity, the yield may change over time, the heat exchanger may foul, etc.), measuring transmitter data may drift, analyzer readings may change due to recalibration, etc. All of these can cause the performance of the ANN model to deteriorate and have to be compensated for by adapting or re-developing the model. The ANN model needs to be maintained and tuned on a regular basis. In the literature (Lahiri, 2017) researchers tried various adaptive approaches to update the model based on its performance. The neural model is updated every six months with fresh current data when it is found that the present model prediction capability deteriorates over time. Most of these auto model update methods are still limited to research publications and very few have been really applied in an actual industry.

7.7 Application of ANN Modeling Techniques in the Chemical Process Industry Due to the recent focus on digitization, researchers are aiming to build a data-driven model from commercial plant data for complex chemical processes where a phenomenological model is difficult to build. A number of researchers have attempted to develop a data-driven model for various industrial refinery and petrochemical processes. ANN is the most common modeling technique. Lluvia et al. (2013) summarize

115

116

7 Process Modeling by the Artificial Neural Network

the application of ANN in a chemical industry as follows. ANN models have been employed to simulate refinery operations in various processes, such as hydrocracking (Alhajree et al., 2011), hydrodesulphurization reactors (Bellos et al., 2005), crude oil distillation (Liau, Yang, & Tsai, 2004; Motlaghi, Jalali, & Ahmadabadi, 2008), delayed coking (Zahedi, Lohi, & Karami, 2009), and fluid catalytic cracking (Zhao, Chen, & Shangxu, 2000). The models developed by Alhajree et al. (2011), Liau et al. (2004), Motlaghi et al. (2008), and Zhao et al. (2000) were successfully included in optimization approaches, showing good agreement with the data used to obtain the models. López and Mahecha (2009), Liau et al. (2004), and Motlaghi et al. (2008) employed statistical models to simulate crude oil distillation columns and these models were implemented in optimization methodologies. López and Mahecha (2009) used metamodels (for crude oil distillation towers) and rigorous models (for crude oil pre-heat trains) to perform optimization of three crude oil distillation units. The procedure requires the construction and regression of several metamodels for each group of variables (flow rates, product properties, temperatures, etc.) using data generated from rigorous model simulations. Metamodels are reported to be more robust and faster than rigorous models (López & Mahecha, 2009) and are suitable for implementation in a systematic optimization methodology. Liau et al. (2004) and Motlaghi et al. (2008) used ANN models to optimize crude oil distillation columns using real plant measurements. In their methodologies, an ANN model is built using information from the distillation column operating history. The model then constitutes the knowledge database to perform operational optimization. The main advantage of these methodologies is that optimization can be implemented on-line and the ANN model can be updated using new plant measurements. ANN modeling was tried successfully in a commercial ethylene oxide reactor by Lahiri & Khalfe (2008, 2009b, 2010). Differential evolution was applied for optimization on a developed ethylene oxide reactor model (Lahiri and Khalfe, 2010). Other modeling techniques for a commercial ethylene oxide reactor, like support vector regression combined with differential evolution, was also suggested in the literature (Lahiri & Khalfe, 2009b). However, all the techniques developed so far are for old generation catalysts, where sensitivity of promoter concentration on catalyst selectivity is much less. There is no literature that applied such modeling and optimization procedures to a new generation ethylene oxide catalyst.

7.8 Case Study: Application of the ANN Modeling Technique to Develop an Industrial Ethylene Oxide Reactor Model 7.8.1 Origin of the Present Case Study

The business environment has drastically changed in the last twenty years. Globalization, reduced profit margins, and cut-throat competition among process industries has changed the rule of the game. Data becomes the new oil in this century and artificial intelligence-based data analytics emerges as the new combustion engine. Chemical industries have a lot of data that are underutilized and can produce money if properly utilized. As of now, there is no doubt that digital will have a significant impact on many areas of the chemical industry, with the gains in manufacturing performance potentially

7.8 Case Study: Application of the ANN Modeling Technique

among the largest. Chemical manufacturers have already invested in IT systems and infrastructure that generate enormous volumes of data, but many have failed so far to take advantage of this mountain of potential intelligence. With cheaper computational power and better advanced analytics tools now at their disposal, chemical companies can put those data to work, gathering information from multiple sources and using machine-learning and visualization platforms to uncover ways to optimize plant operations. Though many companies are trying to enter this area effective holistic tools are still not available in the market. This proposal aims to create intelligent artificial intelligence-based ANN modeling and optimization tools where industrial complex reactor operation can be modeled and optimized. The present case study concentrates on ethylene oxide reactor operation in an ethylene glycol plant. Ethylene glycol is considered as the second most important petrochemical in the globe as it has a huge application to make polyester for textile industries. Currently, there is no automated algorithm available in any of these plants that can calculate the optimum process parameters of an ethylene oxide reactor in a dynamic plant environment. Engineers and scientists working in a petrochemical plant still use the trial-and-error method and sometimes an experience-based heuristic method to optimize the process parameters in a commercial reactor. In the trial-and-error approach, it is not possible for every operator to optimize the reactor all of the time. Moreover, most of the experienced operators are not trained to use the new generation catalyst available in the market and it takes time for them (usually 1–2 years) to get the knowledge and experience to optimize the reactor heuristically. The cost of non-optimum reactor conditions is huge (running into millions of US dollars) in terms of higher raw material consumption. All these plants are losing a huge amount of money due to the non-availability of tools that can optimize the reactor conditions dynamically on a real-time basis. The present project will attempt to eliminate this gap. 7.8.2 Problem Definition of the Present Case Study

In an ethylene glycol plant, ethylene oxide is produced by reacting ethylene with oxygen at 230–2700 ∘ C temperature and 20 bar pressure in a shell and tube type catalytic reactor, where silver-based catalyst beads are kept as a packed bed in the tube side, and water is circulated in the shell side to remove the heat of reaction. In addition to producing ethylene oxide, carbon dioxide is also produced in an undesirable side reaction (refer to the chemical reaction below) and the whole economics of the process depends on catalyst selectivity: C2 H4 + 1∕2O2 −−−−→ C2 H4 O C2 H4 + 3O2 −−−−→ 2CO2 + 2H2 O (by-product) Catalyst selectivity is defined as Catalyst selectivity =

Moles of EO formed Moles of ethylene consumed

The latest generation high selectivity catalyst coming to market has a selectivity range of 91–85% (start of run to end of run) as compared to a conventional catalyst with a selectivity range of 81–78%. A promoter in ppm level concentration is added in the feed gas to enhance the selectivity. However, in the latest generation of high selectivity

117

7 Process Modeling by the Artificial Neural Network

92 Optimum promoter concentration

90 Selectivity, %

118

88 86 84 82 80 0

1

2 3 4 Promoter concentration in feed gas, ppm

5

6

Figure 7.11 Relation between catalyst selectivity and promoter concentration in a commercial ethylene oxide reactor for the latest generation high selectivity catalyst

catalyst, selectivity is found to be very sensitive to the promoter concentration, which follows a narrow inverted “U”-shaped curve, as shown in Figure 7.11. The optimum promoter concentration will vary with the age of catalyst, reactor temperature, presence of other hydrocarbons in the feed gas, etc., and it is very difficult to calculate and maintain the promoter concentration at its optimum value all the time in a commercial glycol plant. Any other promoter concentration (varying in the ppm range) other than the optimum has a huge negative effect on catalyst selectivity in terms of reduced plant production and impact the economics of the process badly. Commercial glycol plants use a trial-and-error approach to find out the dynamic optimum promoter concentration and to keep on adjusting it continuously. In this trial-and-error approach, huge potential production gains are lost in terms of reduced selectivity and millions of rupees lost for not knowing the exact optimum promoter concentration in the dynamic plant environment. The key challenge faced by commercial glycol plants is to determine and maintain optimum inhibitor concentration to maximize the catalyst selectivity all the time. The objective of this project is to create a modeling and optimization framework where the above job can be performed on a real-time basis. This is an application-oriented project for industry and once developed can be deployed to any commercial ethylene glycol plant on the globe. This application will have a significant economic impact on a glycol plant in terms of potential lost selectivity due to non-optimum operation of a commercial ethylene oxide reactor. To achieve the objective a methodology to develop a credible model of a commercial ethylene oxide reactor from available plant data is studied in the present case study. Modeling will be done using an artificial neural network to explore its capability to model such a complex industrial reactor. Once a reliable model is developed and validated with plant data, a different stochastic optimization algorithm (like a genetic algorithm, differential evolution, particle swarm optimization, etc.) will be utilized to optimize the input space of operating parameter to increase profit. The primary objective of this proposal is to develop a low cost, artificial intelligence (AI)-based modeling and optimization platform that can be used in a commercial

7.8 Case Study: Application of the ANN Modeling Technique

ethylene glycol plant for enhancing its productivity. In order to develop such an application, the present study has the following objectives: To develop a data-driven ANN-based modeling platform to model an ethylene oxide reactor. To develop a real-time optimization framework to find out the optimum value of promoter concentration and other reactor operating parameters like inlet reactant concentration, gas flow, pressure, catalyst bed temperature, etc., in a dynamic environment so that selectivity and plant production can be maximized all the time. 7.8.3 Developing the ANN-Based Reactor Model (Lahiri & Khalfe, 2008, 2009b, 2010)

The strategy of EO reaction process modeling and optimization is the objective of the present case study, which is to model and optimize the industrial EO reactor to maximize catalyst selectivity. This case study will show how to develop the ANN model and, in the next chapter, it will be shown how the developed model can be utilized to optimize different process parameters so that profit from the reactor can be maximized. Unlike academic studies, this project was intended to optimize the actual operation of an industrial operating plant. That is why considerations and real issues to develop EO reactor models are quite complex and the constraints are unique to actual plant operations. 7.8.4 Identifying Input and Output Parameters

The first step in developing an ANN model is to identify all input and output parameters related to the reactor. Since we are interested in developing a model that will predict catalyst selectivity, so selectivity is kept as the output variable. Now all the reactor parameters that can influence catalyst selectivity are shortlisted as input. Identification of all the relevant inputs are not always so obvious for industrial processes. Hence all the domain experts who have worked in the plant for many years are consulted to include all the inputs. If the domain experts are not sure about the influence of one or two particular inputs on catalyst selectivity and they have some reservations to include them in the input list, the best strategy is to include them. The ANN model will automatically delete any particular input that has no influence on the output. This is very crucial to the development of an ANN model. Table 7,1 summarizes all the input and outputs for an ANN model. Out of the number of inputs in the wish list, later we used ANN regression to establish the best set of chosen inputs that describes the reaction behavior. The following criteria guide provides the choice of the set of inputs (Lahiri & Khalfe, 2008, 2009b, 2010): • • • •

The inputs should be as few as possible. Each input should be highly cross-correlated to the output parameter. These inputs should be weakly cross-correlated to each other. The selected input set should give the best output prediction, which is checked by using the statistical analysis (e.g. the percentage error, standard deviation, and cross-correlation coefficient).

119

120

7 Process Modeling by the Artificial Neural Network

Table 7.1 Input and output variables for the ANN model Input variables

Output variables

Reactor inlet oxygen concentration, mole %

x1

Reactor inlet ethylene concentration, mole %

x2

Reactor inlet CO2 concentration, mole %

x3

Reactor inlet chloride concentration, ppm

x4

Cycle gas flow through reactor, MT/h

x5

Cycle gas pressure at reactor inlet, barg

x6

3

Work rate, kg/h/m of catalyst

x7

Catalyst running hours, h

x8

Catalyst selectivity

y

• There should be minimum complexity in the neural network architecture, i.e. a minimum number of hidden layers. While choosing the most expressive inputs, there is a compromise between the number of inputs and prediction. The cross-correlation analysis that signifies the strength of the linear relation between the input and output is then used to find the dependence between them. A number of inputs can be highly cross-correlated to output, but there should not be any strong dependency between these inputs; otherwise, it just adds to the complexity of the structure rather than contributing significantly to improving the quality of the network. One should be careful here. Although the cross-correlation analysis reveals the dependence between inputs and outputs, it also hides non-monotonic relationships. This can result in losing an important input. Therefore, in this study, several sets of inputs were made and tested via rigorous trial-and-error on the ANN. The mentioned criteria were then used to identify the most pertinent set of input groups. Based on the analysis, the eight input variables (in Table 7.1) have been finalized to predict catalyst selectivity.

7.8.5 Data Collection

All the input initially selected and the corresponding output variables are selected for collecting large amounts of operating data. The quality and quantity of the data are very crucial in ANN modeling since neural learning is primarily based on these data. A daily average of actual plant operating data at the steady state was collected for approximately one and a half years. Since the effect of input parameters on catalyst selectivity usually took one full day, the daily average data was preferred over the hourly average. Data were checked and cleaned for obvious inaccuracies and those data were retained when the plant operation was in a steady state and smooth. Finally, 505 records qualified for neural regression. This wide range of database included plant operation data at various capacities, starting from 75% capacity to 110% of design capacity. Also, these data captured plant operations at the whole catalyst life cycle (usually 2 years). Note that silver-based EO catalyst selectivity decreases over time due to clogging of the catalyst active site by carbon or other impurities. Start of the run selectivity of the catalyst is 90%

7.8 Case Study: Application of the ANN Modeling Technique

and after 2 years catalyst selectivity came down to 80% at the end of the run. After that the catalyst is thrown out and fresh catalyst is loaded. 7.8.6 Neural Regression

For modeling purposes, the reaction operating conditions data (see Table 7.1) can be viewed as an example input matrix, X, of size 505 × 8, and the corresponding reaction output data as the example output matrix, Y, of size 505 × 1. For ANN training, each row of X represents an eight-dimensional input vector, x = [x1 , x2, …, x8 ], and the corresponding row of matrix Y denotes the one-dimensional desired (target) output vector y = [y1 ]. Since the magnitude of inputs and outputs greatly differ from each other, they are normalized in 0–1 scales using the following relation: x − xmin xnormal = xmax − xmin To avoid the over-training phenomena described earlier, 80% of the total data set was chosen randomly for training and the remaining 20% was selected for validation and testing. It has been reported that multi-layer ANN models with only one hidden layer are universal approximators. Hence, a three-layer, feedforward neural network (like Figure 7.6) is chosen as a regression model. Since there is no previous idea about the suitability of the particular activation function, all three activation functions (sigmoid, tan hyperbolic, and linear) are chosen in all combinations for both the hidden and output layers. The purpose is to find out which combination gives the lowest error. The number of nodes in the hidden layer is up to the discretion of the network designer and generally depends on problem complexity. With too few nodes, the network may not be powerful enough for a given learning task. With a large number of nodes (and connections), computation is too expensive and time-consuming. In the present study, the optimum number of nodes is calculated by trial-and-error. The ANN model performance is predicted on the basis of the following three performance KPIs, namely prediction error %, standard deviations of error (𝜎) and coefficient of determination (R2): n 1 ∑ || (ypi − yai ) || %error = | | × 100 | n i=1 || yai | √ √ n [ ]2 √∑ 1 | ypi − yai | √ | | 𝜎= × 100 − %error n − 1 || yai || i=1 ∑n i=1 [ya.i − yameani ][ypi − ypmeani ] 2 R =√ √∑ ∑n n 2 2 i=1 [ya.i − yameani ] i=1 [yp.i − ypmeani ] where ypi and yai are the predicted and the actual selectivity for the ith record respectively, and ypmeani and yameani are the mean predicted and the mean actual selectivity for the ith record respectively.

121

7 Process Modeling by the Artificial Neural Network

7.8.7 Results and Discussions

90.00 89.00 88.00 87.00 86.00 85.00 84.00 83.00 82.00 81.00 80.00

Actual Selectivity

predicted selectivity

1 12 23 34 45 56 67 78 89 100 111 122 133 144 155 166 177 188 199 210 221 232 243 254 265 276 287 298 309 320 331 342 353 364 375 386 397 408 419 430 441 452 463 474 485 496

Selectivity, %

Input and output data were fed to an ANN algorithm in the Matlab environment. While the training set was utilized for the error back-propagation-based iterative updating of the network weights, the test set was used for simultaneously monitoring the generalization ability of the MLP model. The MLP architecture comprised eight input (N = 8) and one output (K = 1) nodes. For developing an optimal MLP model, its structural parameter, namely the number of hidden nodes, L, was varied systematically. For choosing an overall optimal network model, the criterion used was the least error % and highest R2 for the test set. The optimal MLP model that satisfied this criterion has 10 hidden nodes, the sigmoidal activation function at the input nodes and the linear activation function at the output nodes. To see the ANN model prediction capability, the selectivity prediction against the actual selectivity is plotted as shown in Figure 7.12 for the whole data set. The prediction error percent between the actual selectivity and predicted selectivity is plotted in Figure 7.13. From these two figures it is seen that the ANN model is capable of learning the inherent relationship between selectivity and eight input parameters. The average prediction error % is calculated as 0.11%. A plot of actual selectivity versus predicted selectivity for testing and training data are shown separately in Figure 7.14. Three performance matrices to evaluate ANN model performance for testing and training data are shown in Figure 7.15. The training set error % and the test set error % values along with the corresponding values of standard deviation of error and R2 are summarized in Figure 7.15. The low and comparable training and test error % values

Number of operating days

Figure 7.12 Actual selectivity versus ANN model predicted selectivity 5 4.5 4 3.5 3 2.5 2 1.5 1 0.5 0

Prediction error%

1 12 23 34 45 56 67 78 89 100 111 122 133 144 155 166 177 188 199 210 221 232 243 254 265 276 287 298 309 320 331 342 353 364 375 386 397 408 419 430 441 452 463 474 485 496 507

PREDICTION ERROR%

122

Figure 7.13 Prediction error percent between actual selectivity and predicted selectivity

7.8 Case Study: Application of the ANN Modeling Technique

TESTING DATA Predicted selectivity, %

87 86.5 86 85.5 85 84.5 84 84.00

84.50

85.00

85.50

86.00

86.50

87.00

Actual Selectivity, %

TRAINING DATA Predicted selectivity, %

90 89.5 89 88.5 88 87.5 87 86.5 86 86.00

86.50

87.00

87.50

88.00

88.50

89.00

89.50

90.00

Actual Selectivity, %

Figure 7.14 Plot of actual selectivity versus predicted selectivity for testing and training data

ANN PERFORMANCE

Training error%

Testing error% Training error Testing error standard standard deviation deviation

Training R2

Testing R2

Figure 7.15 ANN model performance for testing and training data

indicate a good prediction and generalization ability of the trained network model. Good model prediction and generalization performance are also evident from the high and comparable R2 values corresponding to selectivity of the training and test sets. Considering the fact that all the input and output data are from a real plant with their inherent noise, the very low prediction error (0.11% average error) can be considered as an excellent ANN model. Once developed, this ANN model can be used to quantitively predict the effects of all input parameters on the catalyst selectivity.

123

124

7 Process Modeling by the Artificial Neural Network

7.9 Matlab Code to Generate the Best ANN Model The Matlab code is given in Appendix 7.1, which will generate the best ANN model. The code is generic and can be used to model any data-driven ANN model. Industrial data contain spikes, noise, and error. Before attempting to build an ANN model, data cleaning of industrial data is an absolutely necessary step. A data cleaning Matlab program is given in the Appendix, which will automatically clean data based on the principal component analysis of the data. Different ANN algorithms have been developed by various researchers and scientists at various points of time in the last 30 years (refer to Figure 7.16). Also, ANN used various activation functions, as shown in Figure 7.17. There are no guidelines available as to which activation functions and which ANN algorithm would be best for a particular set of data. Also, the number of nodes in hidden

Levenberg-Marquardt algorithm

BFGS quasi-Newton back propagation

Resilient back propagation

Conjugate gradient back propagation with Powell-Beale restarts

Conjugate gradient back propagation with Fletcher-Reeves updates

Conjugate gradient back propagation with Polak-Ribiére updates

One-step secant back propagation

Gradient descent with momentum and adaptive learning rate back propagation

Figure 7.16 Different ANN algorithms developed by different scientists in the last 30 years Figure 7.17 Different activation functions used in an ANN Log-sigmoid transfer function

Hyperbolic tangent sigmoid transfer function

Radial basis transfer function

Triangular basis transfer

References

layers is a free variable and so the optimum number of nodes has to be calculated by trial and error. To relieve users of this complexity to choose the best ANN model, a Matlab code has been developed (shown in Appendix 7.1), which will automatically check all the combinations and give the best ANN model to users. In the process optimization chapter, it will be shown how this ANN model can be used to increase selectivity so that profit from an EO reactor is maximized.

References Alhajree, I., Zahedi, G., Manan, Z.A., & Zadeh, S.M. (2011). Modeling and optimization of an industrial hydrocracker plant. Journal of Petroleum Science and Engineering, 78(September 3–4), 627–636. Bellos, G.D., Kallinikos, L.E., Gounaris, C.E., & Papayannakos, N.G. (2005). Modelling of the performance of industrial HDS reactors using a hybrid neural network approach. Chemical Engineering and Processing: Process Intensification, 44(May 5), 505–515. Breiman, L. (1996). Bagging predictors. Machine Learning, 24(2), 123–140. Devlin, S.J., Gnanadesikan, R., & Kettenring, J.R. (1981). Robust estimation of dispersion matrices and principal components. Journal of the American Statistical Association, 76, 354–362. Dibaba, O.R., Lahiri, S.K., T’Jonck, S., & Dutta, A. (2016). Experimental and artificial neural network modeling of an Upflow Anaerobic Contactor (UAC) for Biogas Production from Vinasse. International Journal of Chemical Reactor Engineering, 14(6), 1241–1254. Fortuna, L. (2007). Soft Sensors for Monitoring and Control of Industrial Processes. Springer. Freund, Y., & Schapire, R.E. (1997). A decision-theoretic generalization of on-line learning and an application to boosting. Journal of Computer and System Sciences, 55(1), 119–139. Gosset, W.S. (1908). The probable error of a mean. Biometrika, 6(1), 1–25. Hastie, T., Tibshirani, R., & Friedman, J. (2001). The Elements of Statistical Learning: Data Mining, Inference, and Prediction. Springer. Hecht-Nielsen, R. (1992). Theory of the backpropagation neural network. In Neural Networks for Perception (pp. 65–93). Academic Press. Jackson, J.E., & Mudholkar, G.S. (1979). Control procedures for residuals associated with principal component analysis. Technometrics, 21(3), 341–349. Kadlec, P., Gabrys, B., & Strandt, S. (2009). Data-driven soft sensors in the process industry. Computers and Chemical Engineering, 33(4), 795–814. Kohavi, R. (1995). A study of cross-validation and bootstrap for accuracy estimation and model selection. In Proceedings of the Fourteenth International Joint Conference on Artificial Intelligence, Vol. 2 (pp. 1137–1145). Lahiri, S.K. (2017). Multivariable Predictive Control: Applications in Industry. John Wiley & Sons. Lahiri, S.K., & Ghanta, K.C. (2008). Development of an artificial neural network correlation for prediction of hold-up of slurry transport in pipelines. Chemical Engineering Science, 63(6), 1497–1509. Lahiri, S.K., & Ghanta K.C. (2009a). Genetic algorithm tuning improves artificial neural network models, Hydrocarbon Processing, 88(1),73–82. Lahiri, S.K., & Ghanta, K.C. (2009b). Artificial neural network model with the parameter tuning assisted by a differential evolution technique: The study of the hold up of the

125

126

7 Process Modeling by the Artificial Neural Network

slurry flow in a pipeline. Chemical Industry and Chemical Engineering Quarterly/ CICEQ, 15(2), 103–117. Lahiri, S.K., & Ghanta, K.C. (2009c). Development of a hybrid artificial neural network and genetic algorithm model for regime identification of slurry transport in pipelines. Chemical Product and Process Modeling, 4(1). Lahiri, S.K., & Ghanta, K.C. (2010). Artificial neural network model with parameter tuning assisted by genetic algorithm technique: Study of critical velocity of slurry flow in pipeline. Asia-Pacific Journal of Chemical Engineering, 5(5), 763–777. Lahiri, S.K. & Khalfe, N. (2008). Process modeling and optimization strategies integrating support vector regression and differential evolution: A study of industrial ethylene oxide 14 reactor. Chemical Product and Process Modeling. 3(1), Article 57. Lahiri, S.K., & Khalfe, N. M. (2009a). Soft sensor development and optimization of the commercial petrochemical plant integrating support vector regression and genetic algorithm. Chemical Industry and Chemical Engineering Quarterly, 15(3). Lahiri, S.K., & Khalfe, N. (2009b). Process modeling and optimization of industrial ethylene oxide reactor by integrating support vector regression and genetic algorithm. The Canadian Journal of Chemical Engineering, 87(1), 118–128, Lahiri, S.K., & Khalfe, N. (2010). Modeling of commercial ethylene oxide reactor: A hybrid approach by artificial neural network and differential evolution. International Journal of Chemical Reactor Engineering, 8(1), Article A4. Lahiri, S.K., Khalfe, N., & Garawi, M. (2008). Process modeling and optimization strategies integrating neural networks and differential evolution. Hydrocarbon Processing, 87(10), 35–50. Lahiri, S.K., Khalfe, N., & Wadhwa, S.K. (2012). Particle swarm optimization technique for the optimal design of shell and tube heat exchangers. Chemical Product and Process Modeling, 7(1), 1934–2659. Liau, L.C.-K., Yang, T.C.-K., & Tsai, M.-T. (2004). Expert system of a crude oil distillation unit for process optimization using neural networks. Expert Systems with Applications, 26(February 2), 247–255. Lin, B., Recke, B., Knudsen, J.K., & Jørgensen, S.B. (2007). A systematic approach for soft sensor development. Computers and Chemical Engineering, 31(5), 419–425. Lluvia, M.O.E., Jobson, M., & Smith, R. (2013), Operational optimization of crude oil distillation systems using artificial neural networks. Computers and Chemical Engineering. López, D.C., & Mahecha, C. (2009). Optimization model of a system of crude oil distillation units with heat integration and metamodeling. Ciencia, Tecnología y Futuro, 3, 159–174. Motlaghi, S., Jalali, F., & Ahmadabadi, M. (2008). An expert system design for a crude oil distillation column with the neural networks model and the process optimization using genetic algorithm framework. Expert Systems with Applications, 35(November 4), 1540–1545. Principe, J.C., Euliano, N.R., & Lefebvre, W.C. (2000). Neural and Adaptive Systems. New York: Wiley. Qin, S.J. (1997). Neural networks for intelligent sensors and control – Practical issues and some solutions. In Neural Systems for Control (pp. 213–234). Academic Press. Weiss, S., & Kulikowski, C. (1991). Computer Systems that Learn: Classification and Prediction Methods from Statistics, Neural Nets, Machine Learning, and Expert Systems. San Francisco, CA, USA: Morgan Kaufmann Publishers Inc.

Appendix 7.1 Matlab Code to Generate the Best ANN Model

Wise, B.M. (1991). Adapting multivariate analysis for monitoring and modeling dynamic systems. PhD thesis. University of Washington. Zahedi, G., Lohi, A., & Karami, Z. (2009). A neural network approach for identification and modeling of delayed coking plant. International Journal of Chemical Reactor Engineering, 7. Zhao, W., Chen, D., & Shangxu, H. (2000). Optimizing operating conditions based on ANN and modified gas. Computers & Chemical Engineering, 24, 61– 65.

Appendix 7.1 Matlab Code to Generate the Best ANN Model %Clear old stuff and create folder named networks if it doesn’t %already exist clear all; fclose all; clc; targetdistance=1000000; if isdir('networks')==0 mkdir('networks'); else delete ('networks\*.mat') disp('network directory removed') end % USER INPUT filename = input ('please give file name with complete address'); %Give the file name as per following format'H:\Matlab working folder\ANN\inputdata.xlsx' sheetname=input ('please give sheet name where the input and output data are residing'); % example sheetname='Input data' xlRange=input ('please give range of all data e.g.A2: D600'); % example xlRange='A2:I512' wholedata = xlsread(filename, sheetname, xlRange); input_number = input ('please give number of inputs'); output_number = input ('please give number of outputs'); %data cleaning %This will call data cleaning function datacleaningusingpca cleandata=datacleaningusingpca(wholedata); % Use of different training algorithm inputs=cleandata(:,1:input_number)'; targets=cleandata(:,input_number+1:input_number+output_number)'; run=1; for zz=1:2 if(zz==1) tr_method = 'trainlm'; % use of Levenberg-Marquardt algorithm elseif(zz==2) tr_method = 'trainbfg'; % use of BFGS quasi-Newton

127

128

7 Process Modeling by the Artificial Neural Network backpropagation elseif(zz==3) tr_method = 'trainrp'; % use of Resilient backpropagation elseif(zz==4) tr_method = 'traincgb'; % use of Conjugate gradient backpropagation with Powell-Beale restarts elseif(zz==5) tr_method = 'traincgf'; % use ofConjugate gradient backpropagation with Fletcher-Reeves updates elseif(zz==6) tr_method = 'traincgp'; % use of Conjugate gradient backpropagation with Polak-RibiØre updates elseif(zz==7) tr_method = 'trainoss'; % use of One-step secant backpropagation elseif(zz==8) tr_method = 'traingdx'; % use of Gradient descent with momentum and adaptive learning rate backpropagation end % Use of different transfer function for k=1:2 if(k==1) trans_func = 'logsig'; % use of Log-sigmoid transfer function elseif(k==2) trans_func = 'tansig'; % use of Hyperbolic tangent sigmoid transfer function elseif(k==3) trans_func = 'radbas'; % use of Radial basis transfer function elseif(k==4) trans_func = 'tribas'; % use of Triangular basis transfer function end for i=1:25 % vary number of hidden layer neurons for j=1:3 % running number of times hiddenLayerSize = i; % number of hidden layer neurons net = fitnet(hiddenLayerSize,tr_method); % create a fitting network net.layers{1}.transferFcn = trans_func; % net.layers{2}.transferFcn = 'logsig'; net.divideParam.trainRatio = 70/100; % use 70% of data for training net.divideParam.valRatio = 15/100; % 15% for validation net.divideParam.testRatio = 15/100; % 15% for testing [net,tr] = train(net,inputs,targets); % train the network outputs = net(inputs(:,tr.testInd)); % simulate 15% test data % outputs2016 = net(inputs2016);

Appendix 7.1 Matlab Code to Generate the Best ANN Model %simulate year 2016 data rmsetest(run)=sqrt(mean((outputstargets(tr.testInd)).ˆ2)); % RMSE for 15% random test data percenterror(run)=mean(abs(((outputstargets(tr.testInd))/targets(tr.testInd))*100)); % rmse2016(i) = sqrt(mean((outputs2016targets2016).ˆ2)); %RMSE for year 2016 test data rsquare(run)=regression(targets(tr.testInd), outputs); distance(run)=sqrt(rmsetest(run)ˆ2+ (1-rsquare(run))ˆ2); % distance = SQRT {RMSEˆ2+(1-R2)ˆ2} % r2016(i)=regression(targets2016, outputs2016); indicator=i*10+j; if(distance(run)Y) flag(i)=1; else flag(i)=0; end end flag1=flag'; xnew=[X flag1]; xnew1=xnew; indices=find(xnew(:,10)==1.0); xnew(indices,:)=[]; cleandata = xnew (:,1:end-1)

131

8 Optimization of Industrial Processes and Process Equipment 8.1 Meaning of Optimization in an Industrial Context Optimization means finding the most profitable balance between two or more competitive events. In operating plant, there are plenty of areas where optimization can be done. For a catalytic reactor, an increase in the throughput of the reactor may reduce catalyst selectivity drastically. It may even reduce the catalyst life. An increase in throughput on the one hand may increase plant profit, while a reduction of selectivity and life of the catalyst on the other hand may reduce profit. Therefore, a balance is required to find the optimum throughput where the plant can maximize the profit. For a distillation column, increasing the reflux will reduce the number of trays, i.e. capital cost, but it will increase the reboiler steam, i.e. increase the operating cost. An economic balance (tradeoff ) between the capital and operating costs must be obtained, so the optimum reflux will be where the total cost is minimum. Nothing is free in this world. If you want to gain in some area, you have to sacrifice something in another area. An optimization calculation will give you an idea of where your gain is more than your sacrifice. Due to cut-throat competition in business, companies are now wanting to reduce their operating costs by optimizing all their available resources, be it man, machine, money, or methodology. Optimization is an important tool that can be utilized to strike a proper balance so that profit can be maximized in the long run. Since capital costs are already incurred for a running plant, optimization essentially boils down to minimization of operating costs for the operating plants. In running a chemical plant, there is huge scope to optimize the operating parameters like temperature, pressure, concentration, reflux ratio, etc., which give either higher profits through higher production or lower operating costs. There are many ways to optimize the operating conditions of reactors, distillation columns, absorbers, etc., to enhance their profitability. In mathematical terms, optimizations mean finding the values of input operating variables of a multivariable objective function so that it minimizes or maximizes the objective function. In short, for most cases, it means finding the values of operating variables that maximize the profit function. Normally in chemical industry there are two types of optimization problems that are encountered. The first one is encountered during the design phase of the equipment. Design optimization essentially means finding the values of free design variables so Profit Maximization Techniques for Operating Chemical Plants, First Edition. Sandip Kumar Lahiri. © 2020 John Wiley & Sons Ltd. Published 2020 by John Wiley & Sons Ltd.

132

8 Optimization of Industrial Processes and Process Equipment

that the total cost of the equipment, i.e. capital cost plus operating cost, is minimized. Suppose a binary distillation column is to be designed to separate 10 MT/h of feed with 99.9% purities in bottom and top products. Optimization essentially means finding the reflux ratio, number of trays, column diameter, etc., so that the total cost comprising capital cost of the column and its ancilliaries and operating cost, i.e. steam cost in the reboiler, is minimized. Mathematically the solution is generally accomplished by means of an economic balance (tradeoff ) between the capital and operating costs. The second scope of applying optimization takes place during the running condition of the plant. Once equipment is designed and is built and running, there is little scope to change anything to reduce the capital cost as all the capital cost has already been encountered. Then the main aim of optimization shifts to maximize capacity of the equipment or minimize the operating cost for a given capacity. In a running plant, optimization essentially means finding out the optimum operating conditions so that the operating profit can be maximized. For an example of a running distillation column, this means finding out the optimum reflux ratio, optimum feed rate, column pressure, column bottom temperature, etc., so that the column throughput is maximized or steam consumption in the reboiler is minimized for a given throughput. This operating optimization is required because the installed column has different safety margins available in it, which should be exploited. In many other cases, the steam cost or energy cost increased significantly from their value during the design phase (say before 20 years) and hence the operating point shifts its position. Husain and Gangiah (1976) reported some of the optimization techniques that are used for chemical engineering applications.

8.2 How Can Optimization Increase Profit? Someplaces where optimization can bring more value are described below. Optimization during the design phase: Most of the traditional techniques of design of process equipment do not include cost during the design phase. For example, traditionally heat exchangers are designed mainly to cater for heat duty requirements. Many free design variables like pipe diameter, pipe length, baffle type, number of baffles, tube pitch, etc., are chosen heuristically and a design algorithm calculates the heat transfer area. Actually, all of these design variables have a direct and indirect effect on the heat transfer area, i.e. on the capital cost of equipment. Therefore choice of these design variables should be based on cost and the cost minimization algorithm should be included during the design phase only. Thus new design methodology involving cost optimization techniques during the design phase might be a potential area of optimization. Optimization during the operating phase: Energy and other utility functions change over time. Thus the cost of energy taken during the design phase may no longer be valid for equipment running over 10 years. Hence a reoptimization and locating new optimum point may be necessary. As an example, a high reflux ratio was taken to reduce the capital cost of a distillation tower during the design phase when the energy cost was

8.3 Types of Optimization

very low. After 10 years of running, the energy cost may increase fivefold or more and calculation of a new optimum reflux ratio is necessary in order to reduce the distillation cost of steam consumption during running of the plant.

8.3 Types of Optimization Optimization in CPI can also be divided into two major classes, namely (1) steady state optimization and (2) dynamic optimization. 8.3.1 Steady-State Optimization

This optimization is applied when the operation is in a steady state, which essentially means when process conditions like temperature, pressure, and concentration do not change with time. The purpose of this optimization technique is to find the most favorable operating conditions that will maximize profit generation of that process or particular equipment. For example, for a catalytic reactor, a profit function is developed based on its selectivity and yield. This profit function will have many variables, such as temperature, pressure of the reactor, inlet gas composition, gas hourly space velocity through the reactor, concentration of the promoter or inhibitor at the inlet gas, etc. These operating parameters remain more or less constant at steady-state conditions of the process. The optimization algorithm will vary these process conditions within their lower and upper limits, so that the profit function will maximize. In this way, the optimization technique will evaluate the most favorable conditions of operating parameter that will maximize the profit. 8.3.2 Dynamic Optimization

Chemical processes often attain a new steady state when the plant capacity changes. The purpose of dynamic optimization is to find the best path when a process transition takes place from one steady state to another. This could be, for example, when polymer reactors in a polymer plant of ten encounter grade transitions due to business reasons. In one of the polymer reactors all of the operating parameters are set (temperature, pressure, etc.) to produce a particular grade of polymer, say grade A. Now as per market demand, another grade, grade B (say), has to be produced and reactor conditions need to be changed to filfill this requirement. During this grade transition, the process becomes unsteady and starts to produce an off-grade polymer. The purpose of dynamic optimization is to find the best path for transition and also reduce the transition time so that off-grade polymer production is minimized. The optimality function then becomes a time function and the objective is to maximize or minimize the time-averaged performance criteria. Once the optimum trajectory is calculated, these set values are then put in an automated control system to execute them in a real plant. Therefore, for a running plant, parametric optimization is defined as determination of a set of operating parameters ensuring maximization of profit or minimization of

133

134

8 Optimization of Industrial Processes and Process Equipment

production cost. Dynamic optimization determines a change in the operating parameter over a definite time interval such that a predetermined profit function of these variables is maximized.

8.4 Different Methods of Optimization 8.4.1 Classical Method

When a process or profit function can be described by some simple equations involving a small number or variables, it is possible to find the maxima or minima by differentiating and equating the resulting expression to zero. These expressions are then solved to determine optimum conditions. Though simple and straightforward, the main disadvantage of classical methods in an industrial context are as follows: Actual chemical processes in industry are complex and the profit function is usually a complex equation with many variables involved. With such a large number of variables, one runs into the curse of dimensionality and classical methods of differentiation become difficult to use. Also, operating parameters in industry have some lower and higher operating limits and the objective functions have some associated constraints. A classical method is difficult to use for equations involving constraints. Hence, the applicability of using classical methods is very limited when solving an actual complex industrial optimization process and people have shifted to using gradient-based techniques. 8.4.2 Gradient-Based Methods of Optimization

These are numerical optimization techniques that are widely used due to the following characteristics: • There is no need for an analytical method of differentiation of complex objective functions. • They can be implemented easily in computers where the search is for operating variables of constrained and unconstrained non-linear functions. • They are suitable when a function is not known analytically. The essence of this method (say the steepest descent or steepest ascent method) is summarized as follows: 1. 2. 3. 4. 5.

A starting point is selected. The direction of descent (or ascent) from this starting point is determined. The step size is calculated. A small step is taken to determine the next point of search. The value of the objective function at this new point is evaluated and compared with that of the previous point. 6. Based on these comparisons, new directions of movement are determined. As a rule, a new direction is achieved by walking along a gradient. 7. The procedure is repeated until a lowest (or highest) objective function is reached.

Function value

8.4 Different Methods of Optimization

9 8 7 6 5 4 3 2 1 0

A4 A2 A1

A3

B1 1

2

3

B2 4

5

6

7

B4

B3 8

9

10 11 12 13 14 15 16 17 18 19 20

Figure 8.1 Different minimum values of a function depending on different starting points

Due to their many comparative advantages over traditional techniques, gradient search algorithms are very popular and widely used to solve an industrial optimization problem. The main drawbacks of these types of algorithms are as follows: A global optimum is difficult to find using this method. The method is applicable only to find local optima. Suppose an objective function is plotted by taking different values of its independent variable, as shown in Figure 8.1. Now if the steepest descent method is applied by taking the starting point at A1, the algorithm will calculate and arrive at point B1 and report it as the function minimum value. Depending on the initial guess of a starting point (say A2, A3 A4, etc.) it will calculate the optimum point at B2, B3, and B4 respectively. Therefore, the final solution depends on the initial guess and is often stuck in the local minima. This is the major disadvantage of gradient search methods. The common way to find a global minimum is to try a heuristic approach where several local minima are found by repeated trial with different starting values. The smallest of all known local minima is then assumed to be the global minima. In this way, the probability of getting a global minimum may be increased but is not guaranteed. Conventionally, various deterministic gradient-based methods are used for optimizing a process model. Most of these methods, however, require that the objective function should be smooth, continuous, and differentiable. The ANN models cannot be guaranteed to be smooth, especially in regions where the input–output data (training set) used in model building is located sparsely. In such situations, an efficient optimization formalism known as a genetic algorithm (GA), which is lenient towards the form of the objective function, can be used. The GA was originally developed as the genetic engineering model mimicks population evolution in natural systems and has been extensively used in chemical engineering. 8.4.3 Non-traditional Optimization Techniques

In spite of their ease of implementation, one major drawback of gradient-based traditional optimization techniques is that the algorithm becomes trapped at local optima depending upon the degree of non-linearity and the initial guess. Since these traditional optimization techniques do not ensure the global optimum, their applicability in solving complex industrial optimization problems has been greatly limited.

135

136

8 Optimization of Industrial Processes and Process Equipment

Hence, in the recent past, non-traditional search and optimization methods based on various metaheuristics and biological natural evolution (evolutionary computation), such as simulated annealing (SA) (Kirkpatrick, Gelatt, & Cecchi, 1983), genetic algorithms (GA) (Goldberg, 1989), and differential evolution (DE) (Price & Storn, 1997), to name a few, have been developed to overcome this problem. The main characteristics that differentiate them from traditional methods is that these algorithms are stochastic in nature, with probabilistic transition rules. These are comparatively new and are gaining popularity due to certain properties that the deterministic algorithms do not have. These are found to have a better global perspective than the traditional methods (Deb, 1995, 1996) and Floudas et al. (1999) listed many test problems on local and global optimization in their handbook. Moreover, when an optimal design problem contains multiple global solutions, designers are not interested in finding just one global optimum solution but as many as possible, for various reasons. First, a design with minimum cost may not be the most desired design as the final design will be judged not only by its cost but also other criteria, like its maintainability, reliability, etc. Second, in an industrial context, designers may not be interested in finding the absolute global solution, i.e. the lowest cost design. A design that is superior with respect to other criteria, like size, ease of maintenance, ease of getting spare parts, size, weight, ease of fabrication, etc., is considered to be a better design even it has a marginally inferior cost than the lowest cost design. Thus, it is always prudent to know about other equally good solutions for later use. Non-traditional optimization techniques generate a lot of optimal solutions that can be equally good. However, if the traditional methods are used to find multiple optimal solutions, they need to be applied a number of times, each time starting from a different initial guess and hoping to achieve a different optimal solution. During the last two decades there has been a growing interest in algorithms based on the principle of evolution (survival of the fittest). A common term, coined recently, refers to such techniques as evolutionary algorithms (EA) or evolutionary computation (EC) methods. The best-known algorithms in this class include genetic algorithms, evolutionary programming, evolution strategies, and genetic programming. There are many hybrid systems that incorporate various features of the above paradigms and consequently are hard to classify, which can be referred to just as EC methods (Dasgupta & Michalewicz, 1997). A brief review of the evolutionary computation techniques was also presented by Babu (2001). Some of these most widely used non-traditional optimization techniques (simulated annealing, genetic algorithms, and differential evolution) are discussed in this chapter.

8.5 Brief Historical Perspective of Heuristic-based Non-traditional Optimization Techniques Valadi and Siarry (2014) give a good historical perspective of various non-traditional optimization techniques. Many scientific applications have been using the heuristic method since the 1940s. However, Rechenberg (1973) established the evolutionary strategy as a separate field to

8.5 Brief Historical Perspective of Heuristic-based Non-traditional Optimization Techniques

solve optimization problems using computers. Fogel and co-workers (1966) introduced evolutionary programming in order to use simulated evolution as a learning process. Genetic algorithm: In the 1970s, Holland (1975) invented the genetic algorithm, which is considered as a milestone in this field, after he published his path-breaking book Adaption in Natural and Artificial Systems. In the same decade, Grover (1977) proposed the new concept of the scatter search method, which introduces new methodology for creating novel solutions and gains tremendous benefit beyond those derived from recourse to randomization. These developments are nowadays collectively called evolutionary algorithms (Bäck, 1996) or evolutionary computation (Bäck, Fogel, & Michalewicz (1997). Simulated annealing and artificial immune system: During 1980 and 1990, the research of the metahuristics algorithm gained its highest momentum. Development of simulated annealing in 1983, pioneered by Kirkpatrick et al. is considered as the first major breakthrough. This technique was inspired by the annealing process of metals. Another important algorithm, i.e. the artificial immune system, was developed in 1986 by Farmer, Packard, & Perelson. Glover, in 1980, initiated the use of memory in a met heuristic algorithm called tabu search, where the search moves are recorded in a tabu list so that future moves would try to avoid revisiting previous solutions. He later published a book on tabu search (Glover & Laguna,1997). Ant colony optimization and genetic programming: In 1992, Dorigo made a significant development in this area by introducing the ant colony optimization (ACO) algorithm. This technique was inspired by the ant’s intelligence using pheromone as a chemical messenger (Dorigo, Maniezzo, & Colomi, 1996). During this time, Koza (1992) also published a book on genetic programming that laid the foundation of a whole new area of machine learning. Particle swarm optimization (PSO) and differential evolution: In 1995, Kennedy and Eberhart developed the PSO algorithm. Finally, in 1996 Storn proposed application differential evolution for optimization and later Storn and Price (1997) developed a vector-based evolutionary algorithm called differential evolution, which has proved to be highly successful for continuous function optimization when compared to genetic algorithms in many applications. Bacteria-foraging algorithm, harmony search and honey-bee algorithm: During the twenty-first century, developments become more interesting. In 2000, Passino (2000, 2002) proposed a bacteria-foraging algorithm for distributed optimization and control applications. In 2001, Geem, Kim, & Loganathan developed harmony search, a music-inspired algorithm. In 2004, Nakrani and Tovey developed a honey-bee algorithm and its application for optimizing Internet hosting centers, while Irizarry (2004) described the LARES algorithm based on an artificial chemical process. Biogeography-based optimization algorithm and firefly algorithm: In 2008, Simon proposed a biogeography-based optimization algorithm inspired by biogeography, which is the study of the distribution of biological species through time and space. Also in 2008, Wu and Saunders (2009) described a group search optimizer, an optimization technique that imitates animal searching behavior. Meanwhile, Yang (2008, 2009) proposed a firefly algorithm, and later Yang and Deb (2009, 2010) developed an efficient cuckoo search algorithm, which demonstrated that its search process is quite effective amongst all other metaheuristic algorithms and had many applications. In addition,

137

138

8 Optimization of Industrial Processes and Process Equipment

in 2010, Yang also proposed a bat algorithm based on the echo location behavior of bats. Many interesting things are still happening in metaheuristic-based algorithmic developments.

8.6 Genetic Algorithm 8.6.1 What is Genetic Algorithm?

Genetic algorithms (GAs) are used primarily for optimization purposes and were first introduced by Holand (1975). They belong to the group of optimization methods known as non-traditional optimization methods. A genetic algorithm is a heuristic-based search optimization algorithm that operates on “survival of the fittest” and was inspired by Charles Darwin’s theory of natural evolution. This algorithm reflects the process of natural selection where the fittest individuals are selected for reproduction in order to produce offspring of the next generation. These are the intelligent exploitation of random search provided with historical data to direct the search into the region of better performance in solution space. They are commonly used to generate high-quality solutions for optimization problems and search problems. GAs do not suffer from the basic setbacks of traditional optimization methods, such as getting stuck in local minima. This is because GAs work on the principle of natural genetics, which incorporates large amounts of randomness and does not allow stagnation. Genetic algorithms combine the “survival of the fittest” principle of natural evolution with the genetic propagation of characteristics to arrive at a robust search and optimization technique. Principal features possessed by the GAs are shown in Figure 8.2. 8.6.2 Foundation of Genetic Algorithms

Genetic algorithms are computerized search and optimization algorithms based on the mechanics of natural genetics and natural selection. Genetic algorithms are based on an analogy with genetic structure and behavior of chromosomes of the population. Genetic algorithms simulate the process of natural selection, which means that those species who can adapt to changes in their environment are able to survive and reproduce and go to the next generation. In simple words, they simulate “survival of the fittest” among individuals of consecutive generations for solving a problem. Each generation consists of a population of individuals and each individual represents a point in search space and a possible solution. Each individual is represented as a string of character/integer/float/bits. This string is analogous to the chromosome. The process of natural selection starts with the selection of the fittest individuals from a population. They produce offspring who inherit the characteristics of the parents and will be added to the next generation. If parents have better fitness, their offspring will be better than their parents and have a better chance of surviving. This process keeps on iterating and, at the end, a generation with the fittest individuals will be found. The foundation of GAs based on this analogy is shown in Figure 8.3.

8.6 Genetic Algorithm

They are zeroth order optimization methods requiring only the scalar values of the objective function,

Capability to handle nonlinear, complex and noisy objective functions,

They perform global search and thus are more likely to arrive at or near the global optimum, and

Their search procedure being stochastic, GAs do not impose pre-conditions, such as smoothness, differentiability and continuity on the objective function form.

Figure 8.2 Principle features possessed by a genetic algorithm

Individuals in population compete for resources and mate

A set of solutions for a problem is considered and the set of best ones out of them is selected.

Those individuals who are successful (fittest) then mate to create more offspring than others

This notion of natural selection is applied for a GA based search problem.

Genes from “fittest” parent propagate throughout the generation, that is sometimes parents create offspring who are better than either parent.

Thus each successive generation is more suited for their environment.

Figure 8.3 Foundation of the genetic algorithm

139

140

8 Optimization of Industrial Processes and Process Equipment

8.6.3 Five Phases of Genetic Algorithms

The five phases considered in a genetic algorithm are shown in Figure 8.4. 8.6.3.1 Initial Population

The process begins with a set of individuals which is called a population. Each individual is a solution to the problem we want to solve. An individual is characterized by a set of parameters (variables) known as genes. Genes are joined into a string to form a chromosome (solution). In a genetic algorithm, the set of genes of an individual is represented using a string, in terms of an alphabet. Usually, binary values are used (string of 1s and 0s). We say that we encode the genes in a chromosome. The evolution process in a GA is termed generation. 8.6.3.2 Fitness Function

The fitness function determines how fit an individual is (the ability of an individual to compete with other individuals). It gives a fitness score to each individual. The probability that an individual will be selected for reproduction is based on its fitness score. In each generation, the fitness of an individual, typically the value of an objective function, is computed for individuals in the current population. 8.6.3.3 Selection

The idea of a selection phase is to select the fittest individuals and let them pass their genes to the next generation. Two pairs of individuals (parents) are selected based on their fitness scores. Individuals with high fitness have a better chance to be selected for reproduction. A fitness-based stochastic selection is applied in the current population in order to construct the parent population of individuals of equal size. This selection process assumes that more fit individuals will have more copies in the parent population. Most popular selection mechanisms are roulette-wheel selection, rank selection, tournament selection, and stochastic universal sampling. 8.6.3.4 Crossover

Crossover is the most significant phase in a genetic algorithm. For each pair of parents to be mated, a crossover point is chosen at random from within the genes. Offspring are created by exchanging the genes of parents among themselves until the crossover point is reached. The new offspring are added to the population.

Mutation Crossover Selection Initial population

Fitness function

Figure 8.4 Five main phases of a genetic algorithm

8.6 Genetic Algorithm

Mutation occurs to maintain diversity within the population and prevent premature convergence. 8.6.3.5 Termination

The algorithm terminates if the population has converged (does not produce offspring that are significantly different from the previous generation). Then it is said that the genetic algorithm has provided a set of solutions to our problem. GA has been applied to a variety of optimization problems in engineering and science and more recently it has extended to data mining and machine learning applications and the rapidly growing bioinformatics area. 8.6.4 The Problem Definition

Consider a maximization problem: Maximize f (x) subject to constraints xi,lb ≤ xi ≤ xi,ub where i = 1, 2, … , n where xi,lb and xi,ub are the lower bounds and upper bounds of xi The aim is to find the maximum value of f (x) by using a GA. This can be considered as an unconstrained optimization problem. 8.6.5 Calculation Steps of GA (Babu, 2004)

All metaheuristic algorithms including GA use a certain tradeoff between randomization and local search. In the context of solving optimization problems, they find decent solutions to difficult problems in a reasonable amount of time, but there is no guarantee that optimal solutions can always be reached. Almost all metaheuristic algorithms tend to be suitable for global optimization. GA algorithms consist of two important components, namely, exploitation and exploration. In exploration, the algorithm tries to generate diverse solutions to explore the search space globally, whereas in exploitation the algorithm focuses the search in a local region knowing that a current good solution is found in this region. A good balance between exploitation and exploration should be found in selecting the best solutions to improve the rate of algorithm convergence. A good combination of these two important components usually ensures that global optimum can be achieved. The following well-defined steps need to be performed to implement a GA algorithm. 8.6.5.1 Step 1: Generating Initial Population by Creating Binary Coding

In this step, a certain number of initial guesses have to be made to generate first generation populations and then these guess values have to transform into a binary format. Coding is the method by which the variables xi are coded into string structures. Select the string length for binary coding based on the accuracy required (say a string length is selected 10). Accuracy depends on string length and can be calculated using the following formula: xi,ub − xi,lb Accuracy = 2li where li is the length of the binary code.

141

142

8 Optimization of Industrial Processes and Process Equipment

(a) Generate a random binary number of selected length (say 1111 0101 10 is generated). (b) Decode the value of the binary number by using the following formula: (c) li −1 ∑ 2i bi Decoded value = i=0

where bi is the binary digit in the coding and bi ∈ (0, 1). For example, for a binary code (0101), the decoded value will be (0101) = (1) 20 + (0) 21 + (1) 22 + (0) 23 = 5 (d) Calculate the value of xi from decoded value by the following formula: xi,ub − xi,lb xi = xi,lb + (decoded value) 2li − 1 Therefore, by following the above guidelines, it is possible to generate a number of guesses or, in other words, an initial population of coded points that lie in the given range of the function. Then we move on to the next step, that is, calculation of the fitness. 8.6.5.2 Step 2: Evaluation of Fitness

As GA works on Darwin’s principle of “survival of fittest,” it is essential to evaluate the fitness of every population. Only the fittest population will carry forward on to the next generations and those having a lower fitness will be discarded. In this way, every generation will enrich the objective function and slowly move towards a desired target. As has already been mentioned, GAs work on the principle of “survival of the fittest.” This in effect means that the “good points” or the points that yield maximum values for the function are allowed to continue in the next generation, while the less profitable points are discarded from our calculations. GAs maximize a given function, so it is necessary to transform a minimization problem to a maximization problem before we can proceed with our computations. Depending upon whether the initial objective function needs to be maximized or minimized, the fitness function is changed in the following ways: F(x) = f (x) for maximization problem 1 for minimization problem 1 + f (x) It should be noted that this transformation does not alter the location of the minimum value. The fitness function value for a particular coded string is known as the string’s fitness. This fitness value is used to decide whether a particular string carries on to the next generation or not. F(x) =

8.6.5.3 Step 3: Selecting the Next Generation’s Population

The GA operation begins with a population of random strings. In the next generations, populations that are more fit need to be chosen. The GA carried out the following three operations to select possible populations for the next generation: 1. Reproduction 2. Crossover 3. Mutation

8.6 Genetic Algorithm

Step 3a: Reproduction The reproduction operator selects the strings for the next gener-

ation. The main objective of this operation is the formation of a “mating pool,” where the fitter strings are copied in a probabilistic manner. The rule mimics the survival of the fittest theory: the probability of selection into the mating pool is proportional to the fitness of the strings. The probability of selection of the ith string into the mating pool is given by F Probabilityi = ∑n i j=1

Fj

where F i is the fitness of the ith string. F j is the fitness of the jth string, and n is the population size. The average fitness of all the strings is calculated by summing the fitness of individual strings and dividing by the population size, and is represented by the symbol F avg : ∑n Fi Favg = i=1 n It is obvious that the string with the maximum fitness will have the most copies in the mating pool. This is implemented using the “roulette wheel selection.” The algorithm of this procedure is as follows: Roulette wheel selection: Step 1: Step 2: Step 3: Step 4:

Using F i calculate probabilityi . Calculate the cumulative probability, Cum_probabilityi . Generate n random numbers (between 0 and 1). Copy the string that represents the chosen random number in the cumulative probability range into the mating pool. A string with higher fitness will have a larger range in the cumulative probability and so has more probability of getting copied into the mating pool.

At the end of this implementation, all the strings that are fit enough would have been copied into the mating pool and this marks the end of the reproduction operation. Step 3b: Crossover The main objective of a crossover operation is to introduce some

amount of randomness into the population in order to avoid getting trapped in local minima. In the crossover operation, new strings are formed by exchange of information among strings of the mating pool. For example, strings are chosen at random and a random crossover point is decided; the crossover is performed using the method shown in Figure 8.5. It is evident that, using this method, better or worse children strings can be formed. If worse children are formed, then they will not survive for long, since they will 00 001

00 110

11 110

11 001

Parent

Children

Figure 8.5 Mechanism of crossover

Crossover point

143

144

8 Optimization of Industrial Processes and Process Equipment

be eliminated by survival of the fittest principles implemented by the reproduction operation. However, what if the majority of the new strings formed are worse? Then the next generation population will be inferior to the current generation. To avoid this situation, we ensure that many strings in a population are selected for crossover. A crossover probability (pc) is used to select strings eligible for a crossover. Therefore, only (100 pc)% of the strings are used in the crossover. The remaining (1 – pc)% of the strings are not used in the crossover. This ensures that some of the good strings from the mating pool remain unchanged. The procedure can be summarized as follows: Step 1: Select (100 pc)% of the strings out of the mating pool at random. These strings will be eligible for crossover operation. Step 2: Select pairs of strings at random (generate random numbers that map the string numbers and select accordingly). Step 3: Decide a crossover point in each pair of strings (again, this is done by a random number generation over the length of the string and the appropriate position is decided according to the value of the random number). Step 4: Perform the crossover on the pairs of strings by exchanging the appropriate bits. Step 3c: Mutation Mutation involves making changes to some of the population mem-

bers directly, that is, by flipping randomly selected bits in certain strings. The aim of mutation is to promote the exploitation of nearby search spaces so that better solutions can be found in the nearby area. Mutation is performed by deciding a mutation probability (pm) and selecting strings on which mutation is to be performed, at random. The procedure can be summarized as follows: Step 1: The approximate number of mutations to be operated is given by n/pm. Step 2: Generate random numbers to decide whether mutation is to be performed on a particular population member or not. This is decided by a “coin toss.” That is, select a random number range to represent true and one to represent false. If the outcome is true, perform mutation; if false, do not. Step 3: If the outcome found in step 2 is true for a particular population member, then generate another random number to decide the mutation point over the length of the string. Once decided, flip the bit corresponding to the mutation point. With the end of mutation, the strings obtained represent the next generation (refer to Figure 8.5). The same operations are carried out on this generation until the optimum value is encountered. 8.6.6 Advantages of GA Against Classical Optimization Techniques

1. In a GA, the coding discretizes the search space even if the function is continuous. One major advantage of the GA is that even discontinuous functions can be optimized using GA. However, in most traditional methods, differentiation is to be carried out and this is not possible for discontinuous functions (Babu, 2004). 2. GAs work with a population of points instead of a single point, as in most traditional methods. 3. Using GAs, multiple optimal solutions can be reached, whereas using traditional methods, only a single optimum can be obtained. 4. For implementing GAs, no auxiliary information about the function is required. 5. GAs are probabilistic, while most traditional methods are deterministic.

8.7 Differential Evolution

8.7 Differential Evolution 8.7.1 What is Differential Evolution (DE)?

Differential evolution (DE) is another evolutionary algorithm that has some comparative advantage over GA and has recently been applied to many process engineering problems. The main advantage of DE over GA is that in DE the implementation direct floating points number is used without any coding as compared to GA, where only integers are used, which require coding to convert it into a binary format (Babu, 2004). Robustness, efficiency, simplicity, and its ability to find the global optimum of a function with ease and accuracy have made the DE popular in recent times. DE algorithms are faster than GA ones as they are computationally less intensive than a GA. The coding required in GA for mutation and crossover is quite complicated due to the use of a binary number. This problem is not encountered in DE as this algorithm does not use binary coding. 8.7.2 Working Principle of DE (Babu, 2004)

The main working principle of DE is similar to GA and follows survival of the fittest principles. Therefore, as in GA, a population-based search is performed to arrive at an optimum. The population size is denoted by NP. The dimension of each vector is denoted by D. The main operation is the NP number of competitions that are to be carried out to decide the next generation. To start with, we have a population of NP vectors within the range of the objective function. We select one of these NP vectors as our target vector. We then randomly select two vectors from the population and find the difference between them (vector subtraction). This difference is multiplied by a factor F (specified at the start) and added to a third randomly selected vector. The result is called the noisy random vector. Subsequently, crossover is performed between the target vector and the noisy random vector to produce the trial vector. Then, a competition between the trial vector and the target vector is performed and the winner is replaced in the population. The same procedure is carried out NP times to decide the next generation of vectors. This sequence is continued until some convergence criterion is met. This summarizes the basic procedure carried out in a differential evolution. The details of this procedure are described below. 8.7.3 Calculation Steps Performed in DE

The main task is to minimize an objective function of dimension D. The weighting constant F and the crossover constant CR are known. Refer to Figure 8.6 for the description of the calculation steps. Figure 8.7 depicts the schematic diagram of the differential evolution. 8.7.4 Choice of DE Key Parameters (NP, F, and CR)

NP should be 5–10 times the value of D, that is, the dimension of the problem. Choose F = 0.5 initially. If this leads to premature convergence, then increase F. The range

145

146

8 Optimization of Industrial Processes and Process Equipment

Step 1 Generate NP random vectors as the initial population: Generate (NP x D) random numbers and linearize the range between 0 and 1 to cover the entire range of the function. From these (NP x D) random numbers, generate NP random vectors, each of dimension D, by mapping the random numbers over the range of the function.

Step 6 Calculate the cost of the trial vector and the target vector: For a minimization problem, calculate the function value directly and this is the cost. For a maximization problem, transform the objective function f(x) using the rule, F(x) = 1/[ 1 + f(x)] and calculate the value of the cost. Steps 1–6 are continued till some stopping criterion is met.

Step 2 Choose a target vector from the population of size NP: First generate a random number between 0 and 1. From the value of the random number decide which population member is to be selected as the target vector (Xi)

Step 5 Perform crossover between Xi and Xʹc to find Xt, the trial vector: Generate D random numbers. For each of the D dimensions, if the random number is greater than CR, copy the value from Xi into the trial vector; if the random number is less than CR, copy the value from Xʹc into the trial vector.

Step 3 Choose two vectors at random from the population and find the weighted difference: Generate two random numbers. Decide which two population members are to be selected (Xa, Xb). Find the weighted difference between the two vectors as F(XaXb)

Step 4 Find the noisy random vector: Generate a random number. Choose a third random vector from the population (Xc). Obtain the noisy random vector (Xʹc) as • Noisy random vector Xʹc = Xc + F(Xa – Xb)

Figure 8.6 Calculations steps performed in DE (Babu, 2004)

of values of F is 0 < F < 1.2, but the optimal range is 0.4 < F Xt ?

Xt

Next generation population

Figure 8.7 Schematic diagram of DE

Step 3: Mutation and crossover: In the first column of Table 8.2, i is populated with a population counter of 1 to 10. Three population points p, q, and r, located in the second, third, and fourth columns of Table 8.2, are filled with random numbers (between 0 and 10) such that i ≠ p ≠ q ≠ r. A parameter j is selected randomly for mutation ( j = 0, 1) and is shown in column 5 in Table 3.2. Columns 6 and 7 are filled up with random numbers [0, 1].

147

148

8 Optimization of Industrial Processes and Process Equipment

Table 8.1 Initial Population of x 1 and x 2 and Their Fitness Sr no.

x1

x2

Fitness

1

7.50

3.80

1178.611

2

8.50

4.20

1682.505

3

3.70

4.60

−224.007

4

8.00

6.90

−501.912

5

4.20

2.80

212.0384

6

2.00

5.70

−984

7

2.60

6.10

−1265.82

8

2.30

0.40

42.4554

9

6.90

7.00

−1215.37

10

0.40

2.90

−65.4961

Table 8.2 Mutation and Crossover 1

2

3

4

5

6

7

8

9

10

11

12

Random number (between 0–1)

×l

×2

×1

×2

Fitness

0.049

0.00

0.00

i

p

q

r

j

Random number (between 0–1)

1

5

10

7

0

0.833

2

10

2

9

1

0.808

0.635

17.00

5.34

3

5

6

1

1

0.001

0.612

0.74

8.45

4

1

10

1

0

0.730

0.286

0.00

0.00

5

8

9

2

1

0.617

0.271

21.00

3.36

0.00 0.80

−791.97 −5066.53 0.00

5.00

319.74

6

5

3

4

1

0.150

0.668

11.20

22.85

5.10

6.10

−857.09

7

1

6

4

1

0.735

0.723

18.20

30.85

3.20

2.00

111.10

7.50

5.60

457.68

8

4

9

10

0

0.010

0.679

0.00

2.17

9

3

4

6

1

0.807

0.901

62.10

56.78

10

2

5

4

0

0.909

0.238

0.00

0.00

−22.23 0.00

If the cell in column 6 is less than CR (0.5) then the crossover constant mutates the parameters by the following equation: trial[j] = x1 [r][j] + 0.8(x1 [p][j] − x1 [q][j]) where 0.8 is the assumed value of F. If the cell in column 6 is not less than CR (0.5) then the trial value is calculated by following equation: trial[j] = x1 [i][j] Columns 8 and 9 of table 8.2 are filled using this method. If any values of columns 8 and 9 go beyond their limits then a new random number between (0 and 10) is assumed in columns 10 and 11.

8.8 Simulated Annealing

Table 8.3 New Generation Populations Sr no.

x1

1

0.00

0.00

0.00

2

0.80

5.34

–791.97

3

0.74

8.45

–5066.53

4

0.00

0.00

0.00

5

4.20

2.80

212.04

6

2.00

5.70

–984.00

7

2.60

6.10

–1265.82

8

0.00

2.17

–22.23

9

6.90

7.00

–1215.37

10

0.40

2.90

–65.50

x2

Fitness

Step 4: Evaluation: Objective function values, that is fitness, are calculated and tabulated in column 12 of Table 8.2 for the vector obtained after mutation and crossover. Step 5: Selection: Compare the fitness values in column 12 of Table 8.2 and column 4 of Table 8.1 and select the least cost vector for the next generation as the present problem target is the minimization of the objective function. In this way the populations of second generations are selected as shown in Table 8.3. Step 6: Repeat: Steps 3 to 5 are repeated until a maximum generation or some termination criteria is met. After some generations f (x) is minimized.

8.8 Simulated Annealing 8.8.1 What is Simulated Annealing?

This optimization technique resembles the cooling process of molten metals through annealing. The atoms in molten metal can move freely with respect to each other at high temperatures. However, the movement of the atoms becomes restricted as the temperature is reduced. The atoms start to get ordered and finally form crystals having the minimum possible energy. Formation of the crystals mostly depends on the cooling rate. If the temperature is reduced at a very fast rate, the crystalline state may not be achieved at all; instead, the system may end up in a polycrystalline state, which may have a higher energy than the crystalline state. Therefore, in order to achieve the absolute minimum energy state, the temperature needs to be reduced at a slow rate. The process of slow cooling is known as annealing in metallurgical parlance. 8.8.2 Procedure (Babu, 2004)

The simulated annealing procedure simulates this process of slow cooling of molten metal to achieve the minimum value of a function in a minimization problem. The cooling phenomenon is simulated by controlling a temperature-like parameter introduced using the concept of the Boltzmann probability distribution. According to the

149

150

8 Optimization of Industrial Processes and Process Equipment

Boltzmann probability distribution, a system in thermal equilibrium at a temperature E T has its energy distributed probabilistically according to P(E) = e− kT , where k is the Boltzmann constant. This expression suggests that a system at high temperatures has an almost uniform probability of being in any energy state, but at low temperatures it has a small probability of being in a high energy state. Therefore, by controlling the temperature T and assuming that the search process follows the Boltzmann probability distribution, the convergence of an algorithm can be controlled. It was suggested to implement the Boltzmann probability distribution in simulated thermodynamic systems. This can also be used in the function minimization context. Let us assume that at any instant the current point is x(t) and the function value at that point is E(t) = f (x(t) ). Using the Metropolis algorithm, we can say that the probability of the next point being at x(t + 1) depends on the difference in the function values at these two points or on ΔE = E(t + 1) – E(t) and is calculated using the Boltzmann probability distribution: P(E(t + 1)) = min[1, e−E∕kT ] If ΔE ≤ 0, this probability is 1 and the point x(t + 1) is always accepted. In the function minimization context, this makes sense because if the function value at x(t + 1) is better than that at x(t) the point x(t + 1) must be accepted. An interesting situation results when ΔE > 0, which implies that the function value at x(t + 1) is worse than that at x(t) . According to many traditional algorithms, the point x(t + 1) must not be chosen in this situation. However, according to the Metropolis algorithm, there is some finite probability of selecting the point x(t + 1) even though it is worse than the point x(t) . However, this probability is not the same in all situations. This probability depends on the relative magnitude of P(E) and T values. If the parameter T is large, this probability is more or less high for points with largely disparate function values. Thus, any point is almost acceptable for a large value of T. On the other hand, if the parameter T is small, the probability of accepting an arbitrary point is small. Thus, for small values of T, the points with only a small deviation in the function value are accepted. Simulated annealing is a point-by-point method. The algorithm begins with an initial point and a high temperature T. A second point is created at random in the vicinity of the initial point and the difference in the function values at these two points is calculated. If the second point has a smaller function value, the point is accepted; otherwise E the point is accepted with probability e− kT . This completes one iteration of the simulated annealing procedure. In the next generation, another point is created at random in the neighborhood of the current point and the Metropolis algorithm is used to accept or reject the point. In order to simulate the thermal equilibrium at every temperature, a number of points (n) are usually tested at a particular temperature before reducing the temperature. The algorithm is terminated when a sufficiently small temperature is obtained or a small enough change in function values is found. 8.8.3 Algorithm

The basic algorithm and calculation sequence of simulated annealing is shown in Figure 8.8 (Babu, 2004). The initial temperature T and the number of iterations (n)

8.9 Case Study: Application of the Genetic Algorithm Technique

Step 1 Choose an initial point x(0), a termination criterion ε. Set T sufficiently high, number of iterations to be performed at a particular temperature n, and set t = 0.

Step 2 Calculate a neighboring point x(t + 1) = N(x(t)). Usually, a random point in the neighborhood is created.

Step 4 If absolute value of x(t + 1) –x(t) < ε. and T is small, terminate; else if (t mod n) = 0, then lower T according to a cooling schedule. Go to step 2; else go to step 2.

Step 3 If ΔE = E(t + 1) –E(t)