Prognostics and Health Management in Energy and Power Systems
Description
Key insights and practical guidance on transitioning to clean energy while meeting increasing energy demands, covering AI developments and more
Prognostics and Health Management in Energy and Power Systems explores two highly topical subjects, energy transition and the latest advances in Artificial Intelligence, and provides insights and practical guidance for a smooth transition to clean, low-carbon energy while simultaneously continuing to meet the ever-increasing demand for energy.
The first part of this book is completely devoted to the challenges, trends, and Asset Management requirements for the energy transition and explains why the energy system of the future must be resilient, autonomous, anticipatory, and situation-aware. The second part of the book presents key developments in recent years and shows the gradual shift from a collection of monolithic architectures for narrow, singular tasks to a set of modular, reconfigurable architectures capable of handling different types of tasks. An industrial case study is illustrated in the third part of the book, showing that Large-Scale Foundation models represent a promising technique to support the Prognostics and Health Management of the energy system.
This book includes information on:
- Key differences between reliability and resilience, covering Low-Impact, High-Probability events and High-Impact, Low-Frequency events
- Important factors in the operation of current and future power plants and substations, including software, complexity, human error, data, and maintenance
- Modularity, reliability, and explainability of Large-Scale Foundation models
- Transformer-based Deep Neural Networks, covering Attention Mechanisms, Positional Encoding, and input-output data embedding
- Graph-based approaches to prognostics of complex machinery with sparse Run-to-Failure data, covering diagnostics feature extraction and graph dataset generation
Prognostics and Health Management in Energy and Power Systems is an essential forward-thinking reference for engineers and researchers working in the energy sector with an interest in AI techniques and Machine Learning.
More details
Other editions
Additional editions
Persons
Ryad M. Zemouri, Ph.D, is a Data Scientist at Hydro-Québec's Research Institute (IREQ), Canada. Previously, he was an Associate Professor at the University of Cnam, Paris. His research interests include machine learning and artificial neural networks, with a particular interest in industrial applications of machine learning to prognosis and health management (PHM). He has published nearly 100 papers in various international conferences and journals.
Jean Raymond, ing., Ph.D., M.Sc.A., is a RAMS Engineer in Hydro-Québec's Expertise, Engineering and Standardization, Canada. He has over 34 years of experience as a telecom network and systems engineer. He was responsible for the long-term development of its transport and power systems. He actively contributes to international standards groups (IEC, IEEE), and leads several committees. He has authored over twenty publications. Jean is involved in modernizing university programs in RAMS and Asset Management.
Dragan Komljenovic, ing., Ph.D, is a Senior Research Scientist at Hydro-Québec's Research Institute (IREQ), specializing in reliability, resilience, asset management, and risk analysis. He previously served as a reliability and nuclear safety engineer at the Gentilly-2 nuclear power plant, also part of Hydro-Québec. Dragan actively collaborates with several universities and has authored over 120 peer-reviewed journal and conference papers. He is a Fellow of the International Society of Engineering Asset Management (ISEAM).
Content
List of Figures xi
List of Tables xvii
Abstract xix
About the Authors xxi
Preface xxiii
Acknowledgments xxv
Notations xxvii
About the Companion Website xxix
1 Introduction 1
1.1 The Energy Transition: Toward a Highly Interconnected System of Systems 1
1.2 The Power Plant and Substation of the Future: Toward Situational Awareness 2
1.3 The New Paradigm in AI: The Emergence of the Large-scale Foundation Models 3
1.4 Topics and Organization of the Book 4
Part I Challenges, Trends, and Asset Management Requirements for the Energy Transition 7
2 Energy Transition and Digital Transformation 9
2.1 Introduction 9
2.2 Digital Transformation 11
2.3 Energy Transition 12
2.4 Arrival of DERs 13
2.5 Lifecycle Requirements, Expectations, and Speed of New Technologies, Introduction in the Electric System 14
3 Asset Management and Resilience 15
3.1 Introduction 15
3.2 Asset Management 15
3.3 Resilience 17
3.4 Combining AM and Resilience: Resilience-based AM 18
3.5 Key Differences Between Reliability and Resilience 20
3.6 The Link Between DTs, Reliability, LCM, and AM 21
4 Challenges and Issues Surrounding the Operation of Current and Future Power Plants and Substations 25
4.1 Introduction 25
4.2 Reliability and Asset Management 27
4.3 Different Designs 29
4.4 Sensor Proliferation 29
4.5 Dynamic Systems 29
4.6 Cohabitation of Current and New-generation Technologies 29
4.7 Software 30
4.8 Complexity 32
4.9 Behavioral Nonlinearity of Components and Systems 35
4.10 System of Systems 35
4.11 Human Factors 36
4.12 Data 37
4.13 Different Operational Time Ranges of the Electric Network 37
4.14 Possible Multistates of a Component 38
4.15 Maintenance 38
4.16 Hidden Failures 39
4.17 Degradation Process and Obsolescence of Electric and Mechanical Components or Systems 40
4.18 Climate Change, Extreme Weather Events, and Others 40
4.19 Complete Life Cycle of Component/System 43
4.20 Prescriptive Maintenance or Knowledge-based Maintenance 43
4.21 Regulation Evolution 44
4.22 Prosumers 45
4.23 Potential Consequences of Energy Transition 46
4.24 Remaining Technical Gaps for Electric Power Utilities 47
Part II Large-scale Foundation Models 51
5 From Shallow Machine Learning to the Requirements of Large-scale Foundation Models 53
5.1 Introduction 53
5.2 ANNs: Theoretical Foundations 54
5.3 A Brief History of AI: The Main Developments 57
5.4 Trustworthiness of AI Systems 61
6 Main Elements of Large-scale Foundation Models: Theoretical Backgrounds 77
6.1 Introduction 77
6.2 Modular Learning 78
6.3 Transformer-based DNNs 82
6.4 Self-supervised Learning 87
6.5 Multimodal Fusion 90
6.6 Multitask Learning 93
6.7 Graph-oriented Approaches 93
6.8 Conclusion 99
7 Main Elements of Large-scale Foundation Models: A Practical and Literature Review 101
7.1 Introduction 101
7.2 Transformer Architecture-based Deep Neural Network 101
7.3 Self-supervised Learning 104
7.4 Multimodal Fusion 107
7.5 Multitask Learning 109
7.6 Graph-oriented Approaches 110
7.6.1 Anomaly Detection 114
7.6.2 Diagnostics 114
7.6.3 Prognostics 114
7.7 Conclusion and Synthesis 116
8 Combining Situational Awareness and LSF Models to Support the Energy Transition 119
8.1 Introduction 119
8.2 The Target of Future Power Plants and Substations 120
8.3 What Is the Situational Awareness? 121
8.4 Incorporating the SA to the Power Plant/Substation of the Future 122
8.5 Conclusion 124
9 Toward a New PHM Process 125
9.1 The Concept of PHM Process 125
9.2 Integrating ML into the PHM Process 126
9.3 The Situational Awareness Integrated to the PHM Process 128
9.4 Conclusion 130
Part III Industrial Case Study 131
10 Hydro-generators Prognostics and Health Management 133
10.1 Introduction 133
10.2 Description of the Case Study 133
10.3 Overview of the Global Methodology 142
11 Set of Deep Learning Models for Feature Extraction 145
11.1 Introduction 145
11.2 Feature Extraction from Visual Inspection Data 145
11.3 Feature Extraction from Text Data 149
11.4 Feature Extraction from PD 152
11.5 Conclusion 155
12 Set of AI-Experts with Deep Modular Learning 157
12.1 Introduction 157
12.2 Description of the AI-Experts 158
12.3 Managing the Mixture-of-AI-Experts 161
12.4 Experimental Results 164
12.5 Conclusion 169
12.6 Appendix 169
13 Graph-based Approach for Prognostics of Complex Machinery with Sparse Run-to-failure Data 175
13.1 Introduction 175
13.2 Preliminaries and Assumptions 176
13.3 Diagnostics Feature Extraction 177
13.4 Graph Structure Definition 178
13.5 Graph Dataset Generation for the Prognostics Considering the Sparse RTF Data 179
13.6 Assigning a Likelihood for Each Edge 180
13.7 Graph-based Forecasting Model 181
13.8 Experimental Results 184
13.9 Conclusion 190
Part IV Conclusion 191
14 Conclusion 193
14.1 What to Keep in Mind 193
14.2 Future Directions 195
Acronyms 199
Glossary 203
References 205
List of Figures
1.1 The high-level organization of the book. 2.1 Major players in the smart grid shift. 2.2 Impacts of digital transformation in utility activities. 2.3 Duck curve from 2023 California Independent System Operator. 3.1 Scope of AM. 3.2 The reliability program and its relationship to AM and LCM. 3.3 AM activities supporting UN sustainability goals. 3.4 Resilience concept. 3.5 Resilience-based AM. 3.6 AIM's 10-box capabilities model. 3.7 Mapping the Global Forum on Maintenance and Asset Management (GFMAM) AM landscape subjects against the ten capabilities. 4.1 Historical and emerging electric grid features. 4.2 Drivers of energy transition and digital transformation. 4.3 Industrial revolutions with the evolution of the principal standards and technical standards of the AM process. 4.4 Comparison of bathtub hardware and software curves. 4.5 Characteristics of CASs. 4.6 Examples of critical infrastructure interdependencies. 4.7 Representation of the entire electrical smart grid SoS. 4.8 Different operational time ranges of the electric network. 4.9 Possible organizational structure of maintenance. 4.10 Combined probabilities - failure and obsolescence. 4.11 Billion-dollar weather/climate disasters in the United States in 2021. 4.12 Impact of climate change on the aging of assets. 4.13 Interactions between all aspects of a component/system life cycle. 4.14 Prescriptive or KBM. 4.15 Integrated power system. 5.1 The idea of foundation models. 5.2 The biological inspiration of the formal neuron proposed by McCulloch and Pitts. 5.3 Geometrical view of the loss function as a surface sitting over weight space. The local gradient of the loss surface is given by the vector 5.4 The way to solve a task from data rather than defining how to solve it from prior knowledge. 5.5 The emergence of the DML. 5.6 The LSF models. 5.7 Modular DML. Light gray (green) components illustrate different routing functions. 5.8 IND uncertainty introduced by human labeling, example taken from. 5.9 IND domain shift, example taken from. 5.10 OOD domain shift, example taken from. 5.11 Visualization of aleatoric and epistemic uncertainty for regression model. 5.12 Data uncertainty visualization through the entropy. 5.13 Example of an ensemble of six prediction AI models used to improve the PRPD source recognition. 5.14 The trade-off between accuracy and precision. 5.15 Prediction pipeline conditioned by the OOD test: If , then the tested unknown data is still considered to be in the distribution. This gives a high level of confidence to the prediction made by the AI system. In the contrary, if , the OOD test is positive resulting to a low level of confidence. This means that the AI system needs to be updated by adding the OOD to the train set in order to cover the new domain. 5.16 Different types of the uncertainty sources and the corresponding quantification methodologies. 5.17 Example of adversarial samples generated by crafting the original input. Example (a) and (b) extracted from, (c) from. 5.18 Classification of adversarial attacks. 5.19 Overview of the adversarial GAN. 5.20 Concepts and importance of XAI. 5.21 The trade-off between accuracy and XAI. 5.22 Taxonomy of explainable ML. 5.23 The distribution of the different XAI methods used in the field of energy and power systems. A considerable research interest for model-agnostic compared to model-specific (intrinsic) techniques. 6.1 Paradigm shift in the field of ML: toward multi-model, multitask and multimodal architectures. 6.2 Overall representation of modular learning. 6.3 Different routing techniques. (a) Fixed routing. The modules are selected based on predefined logic or a priori knowledge. (b) Routing by learning with binary selection of modules: the modules are selected or not. (c) Routing by learning with weighted module selection. The modules are selected and aggregated according to a weighting. 6.4 The original architecture of a transformer. 6.5 Illustration of self-attention mechanism. 6.6 The vision transformer - ViT architecture. 6.7 Taxonomy of SSL. 6.8 Principle of generative learning techniques. 6.9 Process of reconstructing a temporal signal. The output representation element of the encoder is the embeddings representations of the masked input time series. 6.10 Process of predicting a temporal signal. 6.11 Developing pairs of similar and dissimilar samples for contrastive learning. 6.12 Illustration of multimodal learning from three modalities: vision, audio, and text. 6.13 Multimodal learning: direct fusion. 6.14 Multimodal learning: fusion by concatenation. The features extracted from each modality are concatenated. 6.15 Multimodal learning: fusion by attention mechanism. 6.16 The classic methods of MTL. (a) Hard sharing of parameters. (b) Soft sharing of parameters. 6.17 Illustration of the convolution operation and the graph convolution operation. (a) The result of the standard convolution operation is the sum of the pointwise multiplications of a subset of the signal with the convolution kernel . (b) The result of the graph convolution operation is the sum of the node features with the product of the adjacent node features. 6.18 Illustration of the encoding function of a graph. Two nodes with similarity in the original graph will have geometrically close embedding in the latent space. 6.19 Overview of the whole encoding and decoding procedures. The encoder allows to obtain the embedding of the node , while the decoder reconstructs the neighborhood of the node from . 7.1 Typical application of the vision transformer architecture for fault diagnostics. Here a multichannel signal transformer (MSiT) is applied for cross-machine fault diagnosis (CMFD) of complex equipment such as the computer numerical control (CNC) spindle motors. The transformer captures the channel correlation features of the unidirectional patch (UDP) obtained from the multichannel...
