Longitudinal proxy pack benchmark generation for remaining useful life assessment of light electric vehicle battery packs in swapping systems

Abstract Battery swapping enables rapid energy replenishment for light electric vehicles (LEVs), but high-turnover pack circulation requires reliable pack-level health and remaining useful life ( RUL ) assessment. Publicly reusable longitudinal ageing datasets for complete LEV packs remain scarce. We develop a proxy-pack data-generation and learning framework that organizes independently tested cell-ageing records into fixed-member trajectories representing a 48 V-class 16-series configuration. Each trajectory retains 16 physical-cell identities over their common-cycle sequence. Cycle-level member states are aggregated into pack features for predicting a conservative minimum-member proxy RUL, with test barcodes isolated before combinatorial augmentation. Across ten barcode-level splits in each of three source cohorts, same-source tests with endpoints within the augmented training lifetime range achieved a median R² of 0.909 and a median Spearman correlation of 0.997. Tests beyond this range retained strong degradation ordering but showed larger absolute errors. A secondary mixed-source audit of 80 held-out trajectories achieved a median R² of 0.714 and a median Spearman correlation of 0.992. A matched composition-control analysis yielded a median R² by change of -0.015 after balancing barcode reuse and limiting composition overlap, indicating limited sensitivity to training-composition sampling. A complementary physical audit using one fixed 12S3P pack confirmed cell-to-branch-to-pack electrical closure and revealed persistent voltage limitation and migration of the maximum-temperature location. Proxy-pack generation therefore provides a reproducible approach for constructing longitudinal member-aware benchmarks from cell-ageing records and evaluating pack-health and RUL models across varied compositions and lifetime conditions.