Reference Selection for Cross-Condition Lithium-Ion Battery State-of-Health Estimation

Reference measurements can affect battery state-of-health (SOH) estimates across operating conditions. We compare reference protocols using fixed local discharge features, regression settings, training records, and evaluation rules. XJTU Batch-3 provides 4487 training records and Batch-4 provides 4639 test records from eight batteries per batch at five discharge rates. Across 40 battery–rate groups, the reference-free macro-group mean absolute error (MAE) was 0.015887. Original single-1C (B3:r3/B4:r7), early single-1C (B3:r3/B4:r3), and same-rate normalization yielded MAEs of 0.014058, 0.014584, and 0.014643, respectively. Battery-cluster bootstrap intervals for improvement over reference-free had positive lower bounds for original single-1C and fixed 0.5C; intervals for early single-1C and same-rate included zero. The lowest average error did not coincide with the lowest worst-group error. At 3C, same-rate normalization improved on the original single-1C for all eight batteries, with a mean MAE reduction of 0.006663. Separate NASA leave-one-cell-out and XJTU Batch-5 holdout analyses examined local-feature applicability under random load. Reference protocols should specify both operating condition and measurement selection, with conclusions limited by the eight target batteries.