Sparse Self-Attentive Transformer With Multiscale Feature Fusion on Long-Term SOH Forecasting

The estimation of the state of health (SOH) of lithium-ion batteries (LIBs) plays an important role in ensuring the safe and stable operation of LIB management systems. In order to more accurately predict SOH, a model based on a sparse self-attentive transformer (SSAT) with multitimescale feature fusion is proposed. The SSAT follows an encoder-decoder structure construction, and the model inputs are the extracted health indicators and SOH sequences. The encoder stacks three cross-stage partial (CSP)-ProbSparse attention self-attention blocks, between every two CSP-ProbSparse attention blocks, connections are made by dilated causal convolution and max-pooling layers to obtain exponential growth of the sensory field. All the feature maps output from the self-attention blocks are integrated by multiscale feature fusion, and finally, the appropriate feature dimensions are fed to the decoder through a transition layer to obtain the estimation of SOH. Numerous comparative and ablation experiments have demonstrated that the SSAT model achieves superior performance in a wide range of situations.