A temperature- and impedance-aware LSTM–PINN framework for physically consistent battery SOH prediction

This paper introduces a physics-informed neural network (PINN) and long short-term memory (LSTM) prediction system based on an offline and laboratory-supervised prediction of the state of health (SOH) of a lithium-ion battery. The model proposed combines an electrochemical degradation concept and temporal sequence learning, thus allowing the network to predict SOH development during early-life pathways and conserve consistency with the trends of physical ageing. Time dependencies of degradation dynamics are represented by the LSTM component, and structures the required behavioral limitations imposed by the physics-informed head by Arrhenius-type temperature acceleration, impedance-growth effects as well as monotonic degradation kinetics. The framework uses an adaptive degradation-rate scaling mechanism that integrates explicitly the real-time increase of impedance with thermal kinetics, this enables it to adapt to the changing operating conditions. These constraints lessen dependence on the entirely empirical fit and prevent non-physical outputs so often found in the unconstrained data-driven models. Cycling data evaluation shows that with the hybrid LSTM–PINN design, the root-mean-square error and mean absolute error of SOH trajectory prediction are at $$1.01\%$$ and $$0.6\%$$ respectively, and shows physical plausibility exists in conditions of different temperatures and impedances. The paper also emphasizes how the incorporation of the electrochemical ageing laws in the sequence-learning framework can be used to increase the stability of long-horizons and increase the interpretability of battery health prognostics.