Uncovering the predominant factors influencing the transient stability margin in power systems with increasing renewable generation using interpretable machine learning

This thesis is centred around the impact of increasing penetrations of Renewable Energy Sources (RES)—which are often connected to power networks via Power Electronic Converters (PEC)—and disconnection of conventional Synchronous Generation (SG) on transient stability. The significantly different dynamic characteristics between these types of generation mean that the dynamic response of power systems to faults is changing. The increasingly large number of variables (and complex interactions between them) make understanding the factors that influence transient stability very difficult with existing tools. Without full knowledge of the factors and mechanisms responsible for impacting transient stability boundary, operational blind spots may exist, and effective transient stability enhancement measures are not easy to develop. Moreover, the intermittent nature of RES means that the transient stability boundary changes over very short timescales. This drives a need for more frequent Transient Stability Assessment (TSA) that can capture the full dynamic response of Converter Interfaced Generation (CIG). However, existing TSA tools (namely Transient Energy Function (TEF) and Time Domain Simulations (TDS)) are limited in that they can either not capture the dynamic response of CIG or are computationally expensive to run and extraction of meaningful insights complicated. This increases the requirement for the design and development of new tools for TSA—the primary focus of this thesis. Machine Learning (ML) can provide explicit mappings of complex functions and accelerate computationally heavy tasks and has been used for TSA in recent years. However, research has mainly focused on error reduction and the computational savings that can be achieved—rather than obtaining detailed insights into the predominant factors influencing the transient stability boundary. To gain such insights, this thesis proposes the use of Interpretable Machine Learning (IML) an emerging area of research—to provide detailed explanations of complex ML models trained to predict the transient stability margin at each location on a power network. Insights can be used to understand better the main power system variables that impact transient stability, the interactions between them, and locational trends. Such tools can bolster existing knowledge of the transient stability boundary and/or infer new information that can be used to enhance situational awareness and develop stability enhancement measures.This thesis is centred around the impact of increasing penetrations of Renewable Energy Sources (RES)—which are often connected to power networks via Power Electronic Converters (PEC)—and disconnection of conventional Synchronous Generation (SG) on transient stability. The significantly different dynamic characteristics between these types of generation mean that the dynamic response of power systems to faults is changing. The increasingly large number of variables (and complex interactions between them) make understanding the factors that influence transient stability very difficult with existing tools. Without full knowledge of the factors and mechanisms responsible for impacting transient stability boundary, operational blind spots may exist, and effective transient stability enhancement measures are not easy to develop. Moreover, the intermittent nature of RES means that the transient stability boundary changes over very short timescales. This drives a need for more frequent Transient Stability Assessment (TSA) that can capture the full dynamic response of Converter Interfaced Generation (CIG). However, existing TSA tools (namely Transient Energy Function (TEF) and Time Domain Simulations (TDS)) are limited in that they can either not capture the dynamic response of CIG or are computationally expensive to run and extraction of meaningful insights complicated. This increases the requirement for the design and development of new tools for TSA—the primary focus of this thesis. Machine Learning (ML) can provide explicit mappings of complex functions and accelerate computationally heavy tasks and has been used for TSA in recent years. However, research has mainly focused on error reduction and the computational savings that can be achieved—rather than obtaining detailed insights into the predominant factors influencing the transient stability boundary. To gain such insights, this thesis proposes the use of Interpretable Machine Learning (IML) an emerging area of research—to provide detailed explanations of complex ML models trained to predict the transient stability margin at each location on a power network. Insights can be used to understand better the main power system variables that impact transient stability, the interactions between them, and locational trends. Such tools can bolster existing knowledge of the transient stability boundary and/or infer new information that can be used to enhance situational awareness and develop stability enhancement measures