A Physics-Informed and Interpretable Learning Framework for Smart O&M of Power Equipment

We propose a physics-informed and interpretable learning framework for smart operation and maintenance (O&M) of substation assets. The method (i) constructs a mechanism-aligned Health Index (HI) from thermo-electro-mechanical indicators under soft physical constraints; (ii) fuses SCADA/online, infrared (IR), dissolved-gas analysis (DGA), and partial-discharge (PD) data with substation topology through temporal encoders and a graph neural network; (iii) pairs TFT/GBDT learners with monotonicity and SHAP for auditable decisions; (iv) enables cross-site collaboration via federated transfer and lightweight domain adapters; and (v) attaches conformal prediction intervals to remaining useful life (RUL) for calibrated uncertainty. On an 8-site cohort (2020–2025; 58 transformers, 91 breakers), our approach improves early-warning F1@80% recall by +0.058 (7-day horizon) and + 0.061 (14-day), increases lead-time compliance to 0.812/0.768, and maintains alarm rates within budget. For RUL, it reduces MAE/RMSE to 8.11/13.98 days and achieves 90.70% coverage with ~18% narrower intervals than quantile regression. Ablations identify HI and graph modeling as primary contributors, and a field case demonstrates operational impact (6.50 outage-hours avoided, 28.00% fewer emergency work orders, $18,700.00 cost avoidance). The framework offers a practical path to physics-consistent, privacy-preserving, and portable O&M analytics across substations.