PINN-LSTM Transformer Oil Temperature Time Series Prediction Model Based on Power Bias

Accurate oil temperature prediction is critical for condition assessment of power transformers. To address the limitations of conventional models in accuracy and robustness, we propose Attention-PINN-LSTM, a hybrid framework synergizing physical thermodynamics with deep learning. The methodology features a dual-path architecture: (1) A physics-based predictor derived from discretized thermal differential equations serves as the baseline, explicitly incorporating non-uniform sampling intervals to handle irregular measurements; (2) A residual-compensation LSTM with power-biased attention dynamically corrects model deviations, where the attention mechanism prioritizes load fluctuation patterns through learnable power-weighted coefficients. Based on on-site data of operating transformers, our model achieved an R2 of 0.95 in multi-step prediction scenarios, significantly better than traditional LSTM and transformer baselines, while demonstrating consistent reliability and robustness under sudden load changes. This physics guided paradigm establishes a robust framework for modeling interpretable time series in power systems.