Transfer learning for state of charge estimation across batteries and chemistries: a lightweight, physics-guided LSTM with regime-aware temporal attention and staged adaptation

A cross-battery, cross-chemistry framework for state-of-charge (SOC) estimation under domain shift is proposed. A compact sequence encoder, composed of two stacked LSTM layers with a single-head temporal attention block, is employed together with regime-aware cues and physics-guided regularization. Terminal voltage, current, an auxiliary H∞-observer SOC channel, and a binary rest flag are used as inputs, while training windows are reweighted to emphasize rest and transition regimes where IR-drop and early relaxation signatures provide the highest identifiability. The loss function is formulated as a Huber term augmented by two Coulomb-consistency priors enforcing sign alignment between average current and SOC change, and magnitude alignment with charge integration. Adaptation is conducted in two stages—head-only calibration followed by low-rate full-network fine-tuning—using a single labeled cycle from the target cell without requiring OCV maps or protocol annotations. Evaluation on held-out Hybrid Pulse Power Characterization (HPPC) sequences demonstrates sub-percent errors across chemistries: MAE = 0.0030 and RMSE = 0.0043 (0.75%/0.82%) for NCA (Molicel INR-21700-P45B) and MAE = 0.0022, RMSE = 0.0028 (0.72%/0.55%) for LFP (JGPFR26650). The final model remains lightweight (~ 31 k parameters, ~ 1.9 MFLOPs, ~ 8–9 ms/sample latency) while ensuring interpretability through regime-wise metrics and attention visualizations that peak at pulse onsets and early rest. Ablation studies confirm that regime cues, physics priors, and staged adaptation collectively reduce error tails and close the transfer gap under minimal supervision, yielding a practical solution for embedded SOC estimation across batteries and chemistries.