Physics-Informed LSTM with Adaptive Parameter Updating for Non-Stationary Time Series: A Case Study on Disconnector Health Monitoring

Accurate prediction of contact temperature in disconnectors is critical for early fault detection. However, purely physics-based models face difficulties in parameter identification, while purely data-driven models often suffer from error accumulation in long-term forecasting. To address these challenges, this paper proposes a novel framework named Hybrid Physics-Informed Long Short-Term Memory (Hybrid-PI-LSTM). Firstly, this paper mathematically formulates the transient heat transfer process as a constrained optimization problem governed by a nonlinear ordinary differential equation (ODE), embedding physical laws into the loss function as a regularization term to promote dynamic consistency. Secondly, to address the inverse problem of parameter drift caused by environmental changes, an Adaptive Parameter Updating (APU) mechanism is introduced. This algorithm utilizes a gradient-based iterative approach to dynamically estimate equivalent physical coefficients (e.g., heat capacity) from observational residuals during inference. Finally, numerical experiments on a real-world dataset demonstrate that the proposed framework significantly outperforms baseline models. Specifically, it achieves a Root Mean Squared Error (RMSE) of 0.283 at a 720-step forecasting horizon, reducing the prediction error by over 35% compared to static-parameter physical models. The results indicate that the proposed adaptive constraint mechanism contributes to enhanced long-term numerical stability and physics-guided parameter tracking.