Abstract
Accurate prediction of partial discharge (PD) activity is important for assessing transformer insulation conditions during IEC 60270 acceptance testing. This study proposes a Hybrid Particle Whale Optimization Algorithm-optimized feedforward neural network (HPWOA-FNN) that combines Particle Swarm Optimization (PSO) for rapid initial exploration with Whale Optimization Algorithm (WOA) refinement. The model uses nine operational, temporal, and autoregressive features extracted from 23,142 PD records collected during 49 test sessions on 18 power transformers. HPWOA-FNN was compared with PSO-FNN, WOA-FNN, and Stochastic Fractal Search Algorithm-optimized FNN (SFSA-FNN) under a common architecture and iteration budget. Robustness was assessed using ten repeated 80/20 splits and five-fold group validation via transformer. Across the repeated splits, HPWOA-FNN achieved an RMSE of 28.11 ± 1.45 pC, an R2 of 0.696 ± 0.031, an AUC of 0.953 ± 0.007 for PD > 100 pC, and an AUC of 0.949 ± 0.009 for PD > 300 pC. Friedman tests indicated overall algorithm differences for RMSE, R2, and AUC at 100 pC. Pairwise Wilcoxon–Holm tests showed statistical equivalence between HPWOA and PSO, while HPWOA significantly outperformed SFSA for RMSE, R2, and AUC at 100 pC. Group validation without the transformer-identity feature achieved an RMSE of 27.71 ± 4.83 pC and R2 of 0.695 ± 0.153, indicating that predictive performance was not dependent on transformer-identity memorization. The results support HPWOA-FNN as a robust continuous PD prediction method while also showing that clipping the training target at 300 pC limits direct magnitude extrapolation for severe discharges above the acceptance threshold.