Abstract
•An artificial intelligence-based framework for predicting failure propagation paths.•Perform codified data generation and meticulous hyperparameter optimization.•Performance evaluation using graph and branch-level accuracy and efficiency metrics.•Compare artificial intelligence models using proposed composite reliability index.
As evidenced by large-scale blackouts with severe economic impacts, cascading failures remain a critical challenge that demands intelligent methods to detect hidden propagation paths before they escalate. This work proposes an artificial intelligence (AI)-based framework that encompasses codified data generation, methodical model comparison, meticulous hyperparameter optimization, and holistic performance evaluation for propagation paths prediction. Using training data generated under diverse operating conditions for power systems, including IEEE 39, 89, 118 and 300-bus models, this study investigates the performance of decision tree, logistic regression (LogR), k-nearest neighbor, support vector machine (SVM), random forest, and multilayer perceptron (MLP) in predicting cascading failure paths, alongside a comparative analysis with the influence model and graph neural networks using graph-level and branch-level accuracy metrics and efficiency metrics, as well as a proposed composite reliability index. To further improve model performance, Bayesian optimization is applied for hyperparameters tuning of the AI methods. Comprehensive numerical experiments are performed, revealing that MLP, SVM, and LogR consistently achieve robust performance while the other models exhibit vulnerabilities under moderate load scaling conditions. The proposed framework can support industry stakeholders in selecting and deploying suitable AI methods for cascading failure prediction in real-world power grid applications.