Physics-constrained long-horizon state of health prediction with explicit recovery correction for lithium-ion batteries

Reliable long-horizon state of health (SOH) prognosis of lithium-ion batteries (LIBs) is important for battery management and retire assessment. Reversible capacity recovery can distort recursive SOH trajectories, causing systematic drift errors and biased end-of-life (EOL) timing. To address this issue, we propose a physics-constrained SOH prediction framework that jointly models battery aging evolution and recovery-induced deviations. Firstly, multi-time-scale probe integration (MTPI) is employed to extract three physical features of LIBs directly from charging-pulse relaxation data without signal decomposition. Then, a BiLSTM with an irreversibility-guided loss suppresses recursive predictions through a soft irreversibility prior. Finally, a KWW–Arrhenius-constrained recovery network (KARN) localizes recovery events, estimates their amplitudes, and calibrates the LIB trajectory. The Tsinghua Berkeley Shenzhen Institute (TBSI) dataset contains 32 NCM811 cells tested at 25–55 °C and is divided into cell-disjoint training, validation, and test sets. External validation uses the SNU NMC811/Gr–SiOₓ, UConn LFP/graphite, and CALCE LiCoO₂/graphite datasets. The proposed framework achieves a chain RMSE of 0.97% SOH, an end-point error of 0.89%, and a MAPE of 0.91%. These values correspond to the reductions of 84.1%, 91.4%, and 85.1%, respectively, relative to the degradation-only baseline. Ablation and sensitivity analyses show that recovery detection and drift compensation make the main contributions to the accuracy improvements, whereas KWW–Arrhenius regularization mainly improves the physical plausibility of the learned recovery kinetics. All three representative held-out cells crossed the 80% SOH EOL threshold within the observed test horizon. The proposed framework also reduces the average capacity error from 80.4 mAh to 6.9 mAh. MTPI achieved R 2 = 0.992 on SNU and R 2 = 0.911 on UConn, while a protocol-adapted KARN variant achieved a recovery RMSE of 9.34 mAh on CALCE. These findings support long-horizon health assessment near the retirement stage.