Safety-oriented CNN-attention-GRU hybrid network for remaining useful life prediction of lithium-ion batteries

Abstract

Average error metrics do not show whether a battery RUL model becomes overly optimistic near the end of life, where delayed intervention is most harmful. Prediction intervals are also difficult to interpret when calibration uses test-battery information or when high coverage is obtained only through wide intervals. This study evaluates CMDH, a compact CNN-Attention-GRU model, under a strict battery-level protocol on the NASA lithium-ion battery aging dataset. The main setting uses B0005, B0006, and B0007 for training and B0018 for final testing, with a sequence length of 50, three input channels, validation-only calibration, and MC dropout with 20 passes. CMDH shows competitive point prediction accuracy and lower late-stage overestimation under the reported protocol. A robustness test on B0025, B0026, and B0027 to B0028 supports the same late-stage pattern. The prediction intervals remain conservative rather than sharp.