Physics‐Informed LSTM for Degradation Prediction of Metallised Film Capacitors Under Dynamic Stress

Metallised film capacitors (MFCs) are critical components in power electronic systems and are widely used in high‐reliability applications such as aerospace. Their degradation directly affects system reliability. However, effective and generalizable prediction models for MFCs under diverse dynamic stress conditions remain insufficient. This paper proposes a dynamic‐temperature degradation prediction method using constant‐temperature accelerated degradation data. First, a baseline model based on long short‐term memory (LSTM) is established to learn a mapping from constant‐stress trajectories to degradation evolution under dynamic stress profiles. Building upon this, physical information is incorporated by integrating MFC degradation mechanisms to enhance physical consistency and extrapolation stability of prediction results under stress‐switching conditions. Experimental results indicate that, compared to the LSTM model, the physics‐informed LSTM (PI‐LSTM) network model improves R 2 by 15.76% and reduces MAPE by 25.75%, whilst effectively mitigating non‐physical discontinuities at stress transition points. Compared to the temporal convolutional network (TCN), gated recurrent unit (GRU) and gradient boosting regression tree (GBRT) models, it achieves higher goodness of fit and lower prediction error in dynamic stress scenarios and provides a viable modelling approach for degradation prediction and reliability analysis of MFCs under complex operating conditions.