PLS-Net: A Physics-Embedded LSTM-Siamese Network for Early RUL Prediction in Lithium-Ion Batteries with Explainable AI Analysis

The increasing adoption of electric vehicles intensifies the need for precise prognostics of lithium-ion battery capacity and end-of-life, but the task remains constrained due to aging mechanisms over time. To overcome these issues, this study presents PLS-Net, a Physics embedded LSTM Siamese Network that integrates domain knowledge with deep learning. As part of the feature extraction process, battery deterioration signals are analyzed to derive multivariate higher-order attributes and uses physics-based elements like resistance and capacity fading which serves as an input to PLS-Net. A Siamese architecture with cross-pair learning allows the network to simulate both self and cross battery interactions, increasing tolerance to data deficits. Unlike conventional models PLS-Net leverages both physical cognition and relational learning to make more accurate predictions. Experimental analysis on the NASA battery dataset shows that PLS-Net with limited historical samples starting from the 50th cycle of Battery B5, achieving a MSE of 3.73 × 10 −5 , MAE of 0.0036, MAPE of 0.0025, and an R 2 score of 99.78%, validating PLS-Net as scalable and reliable framework for predictive battery management. Also, Explainable AI using SHAP values and latent embeddings shows that the model’s predictions are substantially impacted by critical degradation features, improves that interpretability and confidence in RUL estimations.