WD-PINN-based remaining life prediction method for dry-type transformers

Power transformers are pivotal assets in power systems, and accurate prediction of their remaining useful life (RUL) is essential for ensuring safe and stable grid operation. Conventional approaches, including physics based Arrhenius models and data driven machine learning methods, each face inherent limitations. To overcome these chal-lenges, this paper proposes an enhanced physics-informed neural network for dry-type transformer RUL prediction, referred to as WD-PINN (weak-form and dynamically weighted PINN).An engineering-corrected lifetime model is first constructed as the supervisory signal. By applying a multiplicative correction framework, the method generates RUL labels that alleviate the scarcity of authentic failure data. WD-PINN improves upon the conventional PINN architecture through two major advancements. The incorporation of a weak-form conservation consistency constraint transforms point-wise differential requirements into integral expressions, thereby reducing sensi-tivity to higher-order derivatives and measurement noise while enhancing numerical stability under complex operating conditions, including temperature spikes and sudden load variations. In addition, a dynamic weighting mechanism is introduced to auto-matically adjust the balance between data-fitting and physics-based losses. This mechanism prioritizes data fidelity during steady states and strengthens physical con-sistency during regime transitions, enabling effective optimization across diverse op-erating scenarios. Experiments using nearly 10 years of operational data from dry-type transformers demonstrate the superiority of WD-PINN. The model achieves signifi-cantly reduced prediction error, with a mean absolute error (MAE) of 2.5 years, a root-mean-square error (RMSE) of 3.2 years, a mean absolute percentage error (MAPE) of 3.7%, and a coefficient of determination (R 2 ) of 0.98. It remains robust under extreme conditions and consistently outperforms tra-ditional models, while ablation studies confirm the contribution of each component. Overall, the proposed approach delivers high-accuracy, highly adaptive, and robust RUL prediction without compromising physical interpretability, providing a reliable and practical tool for transformer condition assessment. Beyond standalone prediction, the RUL output can serve as an informative input to higher-level grid management and control systems. By integrating the predicted degradation trajectory with coordinated voltage-power control strategies or load redistribution mechanisms, grid operators can proactively adjust operating setpoints before critical ageing thresholds are reached. Such a prediction–control closed-loop framework enables smoother response to sudden load variations and enhances overall system reliability in intelligent power grids.