A multi-time-resolution attention-based interaction network for co-estimation of multiple battery states

Effective and reliable battery management systems rely on accurate co-estimation of multiple battery states, including state of charge (SOC), state of health (SOH), and remaining useful life (RUL). However, this task presents significant challenges due to the varying time resolutions and complex interactions between these states across different timescales, especially when historical battery data is unavailable. To address these challenges, we propose a novel end-to-end multi-time-resolution attention-based interaction network (MuRAIN) for the co-estimation of multiple battery states, directly utilizing current cycling data without the need for historical data. The MuRAIN approach incorporates a multi-resolution patching module that intelligently extracts features with varying timescales from the cycling data. An interactive learning module is then designed to model the intricate interactions between features at different timescales and update the feature sequence accordingly in parallel. Finally, an encoder module, leveraging the multi-head self-attention mechanism, coupled with a state estimator, is used to estimate the battery’s multiple states based on the updated feature sequence from the interactive learning module. To validate the effectiveness of the proposed approach, we employ three benchmark datasets comprising both full-cycling cells with diverse cycling protocols and shallow-cycling cells with varying cycling depths. The results show that (i) for full-cycling cells, MuRAIN achieves mean absolute errors (MAEs) of 0.45%, 0.45%, and 63 cycles for estimated SOC, SOH, and RUL respectively, which are reduced by approximately 22%, 33%, and 30%, compared to a state-of-the-art method based on CNN-BiLSTM; (ii) For shallow-cycling cells, MuRAIN achieves average MAEs of 0.69% and 0.37% for estimated SOC and SOH, respectively, with reductions of about 61% and 45% compared to the CNN-BiLSTM method.