A physics-informed CNN-LSTM for state-of-health reconstruction and remaining-useful-life prediction of lithium-ion batteries

Abstract

Accurate remaining-useful-life (RUL) prediction is important for battery safety, maintenance planning, and second-life assessment, but models trained on a small set of cells can overfit degradation patterns that do not transfer to unseen cells or operating conditions. This paper revises and evaluates a physics-informed CNN-LSTM framework for lithium-ion battery prognostics. Four incremental-capacity descriptors are combined with the state-of-health (SOH) signal to form a compact sequence representation. A convolutional-recurrent backbone predicts normalized RUL while an auxiliary SOH head is regularized by a learnable empirical capacity-fade law. The two physics parameters are optimized jointly with the neural network, allowing the constraint to adapt to the observed cell population rather than being fixed a priori. Experiments on the CALCE CS2 cells show that the proposed model obtains the best average leave-one-cell-out performance among the compared methods, with RMSE/MAE of 0.2488/0.2212. The improvement over Gaussian process regression (GPR) is small, indicating that GPR remains a strong baseline on this compact dataset, while the improvement over neural baselines is more pronounced. Cross-C-rate and CALCE-to-MIT transfer experiments reveal that generalization across operating conditions and battery chemistries remains challenging. The learned physics parameters remain finite and evolve during training, suggesting that the physics term is active rather than degenerate. The results support the usefulness of physics-informed regularization, while also identifying limitations under severe degradation-rate distribution shift.