Data-Driven Approaches for Cascading Failure Mitigation in Power Networks

Abstract

Mitigating cascading failures is a critical objective for supporting informed decision-making and guiding strategic investments at both the planning and operational stages of power networks, as such failures can trigger large-scale blackouts and incur severe socioeconomic consequences. Traditional mitigation strategies, typically constrained to reinforcing a subset of components due to budgetary limitations, depend heavily on the precise identification of critical elements within the network. While existing approaches can be effective, they often suffer from high computational costs and limited adaptability to the rapidly evolving dynamics of real-time failure propagation. This thesis introduces a series of data-driven frameworks designed to develop and evaluate efficient algorithms for cascading failure mitigation. These frameworks address key computational challenges from two dimensions: reducing the combinatorial search space for critical component selection, and minimizing reliance on repeated, costly cascading failure simulations. Furthermore, the scope of study is extended to the development of real-time prediction and mitigation frameworks that can operate under emergency conditions, enabling proactive responses during the progression of cascading events. The proposed frameworks are applicable across both long-term planning and real-time operational contexts, offering a comprehensive solution to enhance the resilience of modern power systems.

First, the identification of critical branches for cascading failure mitigation is formulated as a multi-objective optimization problem. To address the computational challenges arising from the vast combinatorial search space, a novel graph-based framework, the failure propagation graph (FPG), is proposed to accurately capture the critical failure propagation patterns based on cascading failure simulation. Leveraging this representation, a failure-propagation-graph greedy-search (FPG-GS) algorithm is developed, reducing the search space by iteratively selecting the most critical branches with respect to their impact on failure propagation. Experimental results on multiple benchmark power systems demonstrate that FPG-GS outperforms existing methods in both mitigation effectiveness and computational efficiency. Notably, the FPG-GS algorithm can also provide a computationally tractable algorithm for large-scale networks where conventional algorithms often become computationally infeasible.

Then, the computational challenge due to expensive risk assessments of cascading failure is addressed by developing a graph neural network (GNN)-based framework. The framework uses a novel GNN named bus and branch graph attention network (BB-GAT) for efficient risk prediction, reducing reliance on computationally expensive cascading failure simulations. To enhance the generalization capability of the risk prediction model, the BB-GAT model is first developed by jointly encoding bus-level and branch-level features into a graph representation. Second, a pre-training strategy is adopted to reduce the dependency on large amounts of task-specific data. After training BB-GAT, the framework implements a three-step optimization pipeline including statistical search space pruning, BB-GAT-based candidate screening, and final simulation validation. Experimental results demonstrate that the proposed method achieves risk mitigation comparable to the Non-dominated Sorting Genetic Algorithm II (NSGA-II) while significantly outperforming it in computational efficiency.

Finally, as the occurrence of failures remains inevitable, early prediction of failure evolution is essential to enable system operators to take timely and informed actions that prevent escalation into large-scale blackouts. To address this need, the mitigation framework is extended to a real-time setting by developing a power flow forecasting model that operates during cascading failure propagation. The model leverages Transformer-based architectures to jointly capture the temporal dynamics and spatial dependencies underlying the evolution of cascading failures. It provides accurate forecasts of power flow during the early post-fault stages, enabling proactive mitigation through early warning mechanisms. Experimental results validate the effectiveness of the proposed framework in reducing the power outages caused by cascading failures.

Date of Award	10 Nov 2025
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Chi Kong TSE (Supervisor)