Real-time Prediction of Cascading Failures in Power Systems
Blackouts in power systems cause major financial and societal losses, which necessitate devising better prediction techniques that are specifically tailored to detecting and preventing them. Since blackouts begin as a cascading failure (CF), an early detection of these CFs gives the operators ample time to stop the cascade from propagating into a large-scale blackout. In this thesis, a real-time load-based prediction model for CFs using phasor measurement units (PMUs) is proposed. The proposed model provides load-based predictions; therefore, it has the advantages of being applicable as a controller input and providing the operators with better information about the affected regions. In addition, it can aid in visualizing the effects of the CF on the grid. To extend the functionality and robustness of the proposed model, prediction intervals are incorporated based on the convergence width criterion (CWC) to allow the model to account for the uncertainties of the network, which was not available in previous works. Although this model addresses many issues in previous works, it has limitations in both scalability and capturing of transient behaviours. Hence, a second model based on recurrent neural network (RNN) long short-term memory (LSTM) ensemble is proposed. The RNN-LSTM is added to better capture the dynamics of the power system while also giving faster responses. To accommodate for the scalability of the model, a novel selection criterion for inputs is introduced to minimize the inputs while maintaining a high information entropy. The criteria include distance between buses as per graph theory, centrality of the buses with respect to fault location, and the information entropy of the bus. These criteria are merged using higher statistical moments to reflect the importance of each bus and generate indices that describe the grid with a smaller set of inputs. The results indicate that this model has the potential to provide more meaningful and accurate results than what is available in the previous literature and can be used as part of the integrated remedial action scheme (RAS) system either as a warning tool or a controller input as the accuracy of detecting affected regions reached 99.9% with a maximum delay of 400 ms. Finally, a validation loop extension is introduced to allow the model to self-update in real-time using importance sampling and case-based reasoning to extend the practicality of the model by allowing it to learn from historical data as time progresses.
cascading outage, artificial intelligence, smart grid, neural network, higher statistical moments
Doctor of Philosophy (Ph.D.)
Electrical and Computer Engineering