Lithium-ion battery state-of-health estimation based on TVFEMD and BiLSTM-Attention

The safe operation of lithium-ion batteries depends on correct State-of-Health (SOH) assessment. Data-driven SOH estimation stands as the main research area in modern battery assessment. This research used Pearson correlation coefficients to evaluate five health features, which were extracted from battery charge-discharge curves, incremental capacity curves, and electrochemical characteristics. The research introduces a dual-threat estimation design that links Time-Variant Filter-based Empirical Mode Decomposition (TVFEMD) with a Bidirectional Long Short-Term Memory with Attention Mechanism (BiLSTM-Attention) framework to process the data. Principal component analysis (PCA) decreased the input feature dimensions, which brought substantial computational performance improvements without affecting prediction accuracy significantly. The proposed model was compared with traditional recurrent neural network models on two public datasets (Oxford and CALCE datasets). The results demonstrate that our model achieved the lowest values across all three error metrics for all twelve tested battery cells, with all errors remaining below 1 %. While PCA-based feature dimension reduction typically leads to increased error, this issue was effectively mitigated in our approach - the maximum error variation before and after dimensionality reduction was only 0.24 %, with most batteries showing error fluctuations around 0.1 %. Furthermore, we validated the model's transfer capability by testing it on an independently measured battery from our laboratory, where all three error metrics remained below 0.63 % (and below 1.1 % after PCA processing), outperforming all comparative models. These experimental results confirm that the proposed TVFEMD-BiLSTM-Attention model can effectively predict battery SOH across different application scenarios.