Data generation and model usage for machine learning-based dynamic security assessment and control

Abstract

The global effort to decarbonise, decentralise and digitise electricity grids in response to climate change and evolving electricity markets with active consumers (prosumers) is gaining traction in countries around the world. This effort introduces new challenges to electricity grid operation. For instance, the introduction of variable renewable energy generation like wind and solar energy to replace conventional power generation like oil, gas, and coal increases the uncertainty in power systems operation. Additionally, the dynamics introduced by these renewable energy sources that are interfaced through converters are much faster than those in conventional system with thermal power plants. This thesis investigates new operating tools for the system operator that are data-driven to help manage the increased operational uncertainty in this transition. The presented work aims to an- swer some open questions regarding the implementation of these machine learning approaches in real-time operation, primarily related to the quality of training data to train accurate machine- learned models for predicting dynamic behaviour, and the use of these machine-learned models in the control room for real-time operation. To answer the first question, this thesis presents a novel sampling approach for generating ’rare’ operating conditions that are physically feasible but have not been experienced by power systems before. In so doing, the aim is to move away from historical observations that are often limited in describing the full range of operating conditions. Then, the thesis presents a novel approach based on Wasserstein distance and entropy to efficiently combine both historical and ’rare’ operating conditions to create an enriched database capable of training a high- performance classifier. To answer the second question, this thesis presents a scalable and rigorous workflow to trade-off multiple objective criteria when choosing decision tree models for real-time operation by system operators. Then, showcases a practical implementation for using a machine-learned model to optimise power system operation cost using topological control actions. Future research directions are underscored by the crucial role of machine learning in securing low inertia systems, and this thesis identifies research gaps covering physics-informed learning, machine learning-based network planning for secure operation, and robust training datasets are outlined.