Machine Learning Approaches for Close to Real Time Assessment of Transient Stability and Dynamic Behaviour of Uncertain Power System

Abstract

Modern power systems tend to operate closer to stability boundaries due to efficiency improvements and economic reasons, constrained by numerous uncertainties and driven by the integration of Renewable Energy Sources (RES) for power generation. The widespread adoption of Wide Area Measurement Systems (WAMS) and Phasor Measurement Units (PMUs) provides on-line measurements, resulting in the acquisition of a vast amount of data at every time instant. On-line Transient Stability Assessment (TSA) has become increasingly desirable for rapidly mapping dynamic behaviour and delivering precise support information for corrective control actions. This thesis contributes to the development and application of data analytics tools for on-line TSA of uncertain power systems with a high penetration level of RES. Utilising a dataset generated through probabilistic Monte Carlo (MC) simulations that consider various sources of uncertainty, it first proposes a set of new indices for fast assessment of system stability. The indices have been shown not only to be suitable to adequately indicate the transient stability of the system but also to define the stability boundary. A Composite Index (COMI) is then developed combining the features of previously defined indices to enhance the accuracy of the assessment of system stability. The development of composite index using Principal Component Analysis (PCA), is additionally offering a general approach for any composite index development with automatic weight allocation. In the second part the thesis presents a feasibility analysis of the use of multiple Machine Learning Techniques (MLTs) for predicting specific index values is performed. The results demonstrate that, with appropriate tuning of MLT hyper-parameters, stability index values can be accurately predicted, thereby expediting on-line TSA compared to solely relying on newly developed stability indices. Moreover, a Recurrent Neural Network (RNN) with Long Short-Term Memory (LSTM) units is developed for real-time prediction of post-fault rotor angle trajectories. Using a sliding time window that prespecifies input parameters and prediction targets, the LSTM-RNN can predict rotor angle responses at any time instant following a fault inception. Practical issues related to realistic PMU data are considered, showcasing the feasibility of the proposed LSTM-RNN when applied in actual systems. Results indicate that this method can accurately predict future rotor angle responses and exhibits robustness against imperfections in realistic PMU data. The primary outcomes of this research encompass the development of various data analytics tools for on-line TSA, including stability indices and MLTs. The focus then shifts from transient stability status to the stability boundary and actual post-fault rotor angle responses, compared to conventional methods. This research provides more accurate information regarding the post-fault power system, enabling further analysis and facilitating more informed control actions.

Date of Award	14 Dec 2023
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Jovica Milanovic (Main Supervisor) & Haiyu Li (Co Supervisor)