Physics-Informed Hybrid Neural Architecture for Coupled Degradation Modeling and Remaining Useful Life Prediction of LiFePO4 Batteries

Lithium-ion batteries (LIBs) are critical for modern transportation and renewable energy systems. Accurate prediction of their degradation trajectory and remaining useful life (RUL) is essential for reliability and safety. This study proposes a physics-informed neural network (PINN) framework integrating RUL prediction with degradation modeling, featuring three components: (1) A multi-factor aging model incorporating knee-point dynamics, capturing two-phase degradation influenced by depth of discharge, temperature, and C-rate; (2) An end-to-end convolutional neural network (CNN) processing multi-channel charge-discharge profiles (current, voltage, capacity) to jointly predict knee points and RUL, enabling feedback between degradation and RUL outputs; (3) An LSTM-based PINN framework embedding electrochemical constraints into a hybrid loss function, enhancing interpretability and generalizability under limited data. Validated on 132 commercial LiFePO4/graphite batteries under diverse fast-charging protocols, the CNN achieves 75.62-cycle RUL prediction RMSE, while the PINN reaches 0.013 Ah capacity prediction precision. Ablation studies show the model reduces degradation trajectory RMSE by 38.10%–84.71% compared to baselines without physical integration or RUL feedback. This approach bridges data-driven learning and electrochemical principles for robust LIB lifespan estimation.