Abstract
SUMMARY & CONCLUSIONSThis paper explores artificial intelligence (AI)-based approaches to predict failure cascades initiated by the malfunction of two branches (transmission lines) in power systems. The cascading failure simulator Oracle is used to generate training data under various operating conditions from the Matpower toolbox for power systems including IEEE 39-bus and 118-bus systems. Several machine learning methods are investigated, including decision tree, logistic regression (LogR), k-nearest neighbor (KNN), and support vector machine (SVM), along with a comparative study with the influence model and graph neural networks (GNN). Bayesian optimization is further explored for tuning hyperparameters to enhance the performance of the AI methods. The accuracy of each method is evaluated at both graph and branch levels using metrics of failure size error rate, final state error rate, and failure time error. Additionally, testing time is analyzed to assess the computational efficiency of the AI models. In the 39-bus system, KNN achieves the highest accuracy but exhibits the lowest efficiency. In the 118-bus system, SVM and LogR deliver high accuracy with only a minor trade-off in efficiency. GNN predicts the final state of branches within a 14% error rate and the final step with an error below 0.1 under random conditions.