State of charge aligned dual branch learning for reliable cross vehicle state of health prediction from fragmented charging data

Real-world electric vehicle fleet data consist of fragmented charging records with irregular sampling, heterogeneous operating conditions, and pack-level inconsistency. For battery state of health (SOH) prediction, these characteristics introduce time-coordinate mismatch and non-degradation-related variations, thereby reducing model transferability to unseen vehicles. To address this problem, this study proposes an end-to-end deep learning framework for SOH point and interval prediction. The framework reparameterizes charging data from the time domain to the state of charge (SOC) domain and constructs mask-aware SOC-aligned token sequences, improving comparability among charging segments with different durations, sampling frequencies, and SOC ranges. A dual-branch architecture learns complementary degradation representations. The temporal branch captures multi-scale SOC-dependent response variations and trajectory-level degradation patterns, whereas the inconsistency-aware branch incorporates pack-level voltage and temperature dispersion as complementary degradation cues. For uncertainty-aware prediction, quantile regression with a structural monotonicity constraint generates ordered prediction intervals, and conformalized quantile regression is applied as a post-hoc calibration step when empirical coverage correction is required. Experiments on two electric vehicle fleet datasets show that the proposed method achieves the best performance across four point-prediction metrics. Root mean square errors are 0.960 ± 0.025% on Dataset A and 1.981 ± 0.095% on Dataset B. On Dataset B, independent conformal calibration yields a prediction interval coverage probability of 0.8679 ± 0.0487 and a Winkler score of 8.9384 ± 0.5322. These results indicate improved interval reliability under stronger distribution shift and demonstrate the potential of the framework for uncertainty-aware fleet battery health assessment.