Machine Learning-Based Rapid Post-Earthquake Performance Assessment Method of Composite Insulators in Substations

Previous strong earthquakes have severely damaged composite insulators in substations, highlighting the urgent need for rapid, intelligent post-earthquake assessment methods to support power system recovery. This paper proposes a machine learning (ML)-based approach to predict the mechanical response of composite insulators using seismic intensity measures (IMs). A simulation model is developed and checked against the first three natural frequencies from on-site dynamic tests, and then used to generate datasets for training the ML model. To select model hyperparameters, Particle Swarm Optimization (PSO) is employed for hyperparameter tuning. A typical ±800 kV surge arrester with composite insulators serves as a case study. Three ML architectures are evaluated to compare their predictive performance. Based on the best-performing model, Shapley Additive Explanations (SHAP) are applied to identify the most influential IMs. Under the reported evaluation procedure, results show that PSO improves prediction accuracy across all models. The Random Forest model achieves the highest precision, predicting the stress response of composite insulators in under one second. SHAP analysis highlights the dominant IMs affecting seismic responses at both global and local levels. The proposed method provides rapid FE-based stress estimates to support post-earthquake assessment; direct stress validation, independent-event testing, and validated damage thresholds remain necessary for safety or usability decisions.