A degradation-aware SegRNN method optimized by ASHA for lithium-ion battery SOH prediction

Accurate multi-step state-of-health (SOH) prediction is important for assessing the future degradation of lithium-ion batteries, but noisy and small-sample sequences make it difficult to maintain both point-wise accuracy and trajectory continuity. Most existing multi-step SOH prediction methods directly output the SOH values at several future cycles. In this paper, a degradation-aware segment recurrent neural network (SegRNN) divides the historical SOH sequence into consecutive segments, and a gated recurrent unit (GRU) encodes their temporal evolution. A cumulative degradation increment prediction (CDIP) head predicts the SOH change at each future cycle, and the predicted changes are accumulated from the last observed SOH to obtain the future SOH sequence. A degradation-aware training strategy (DATS) combines SOH prediction error with adjacent-step change error to maintain the degradation trend under noise and local fluctuations. Reversible instance normalization (RevIN) reduces distribution differences between input windows, and the asynchronous successive halving algorithm (ASHA) selects the model configuration for each battery. Experiments on the National Aeronautics and Space Administration (NASA) and Center for Advanced Life Cycle Engineering (CALCE) battery datasets demonstrate the effectiveness of the proposed method for noisy multi-step SOH prediction.