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 method for state-of-charge (SOC) estimation under domain shift is introduced. The approach employs a lightweight sequence model—two stacked LSTMs with temporal attention—augmented by regime-aware cues and physics-guided regularization. Inputs comprise terminal voltage, current, and an auxiliary polarization feature, together with a binary rest flag that highlights near-zero-current samples. Windows are loss-weighted to prioritize rest and transition segments, where relaxation and load switching yield high information content. The objective supplements a Huber data term with two soft physical priors: (i) sign consistency between the window-average current and the predicted SOC increment, and (ii) magnitude consistency with Coulomb counting over the window. Adaptation proceeds in two stages—head-only, then full fine-tuning—using only a single labeled cycle from the target cell. Evaluation on an unseen pulse–rest protocol demonstrates low errors, with MAE = 0.0030 and RMSE = 0.0043 in 0–1 SOC units (also reported as 0.7538% and 0.8247% of full scale, relative to the SOC span considered). Attention maps consistently peak at pulse onsets and early rest, aligning with electrochemical intuition and providing transparent diagnostics. Ablation results indicate that regime cues and physics priors tighten error tails, while staged adaptation closes most of the transfer gap with minimal target labels. The pipeline requires no OCV map or protocol labels during adaptation, is compact enough for embedded deployment, and offers auditable behavior through regime-wise metrics and attention visualizations. This combination yields a practical recipe for low-label SOC transfer across batteries and chemistries.