Lithium-Ion Battery SOH Prediction with Multi-Feature Extraction Using a BiLSTM-RPTransformer Model

Accurate state of health (SOH) estimation is crucial for battery management systems (BMS). This study proposes a novel hybrid model integrating relative positional encoding into the Transformer architecture and coupling it with a bidirectional LSTM (BiLSTM) for SOH prediction. Key features—current, voltage, and internal resistance—are extracted to characterize battery degradation. The BiLSTM captures both forward and backward temporal dependencies, enhancing sensitivity to short-term dynamics. Historical data are fused with relative positional encodings, dynamically incorporated into each multi-head attention layer to model relative positions and feature correlations. This enables a robust mapping from input features to SOH. Experiments on the CALCE dataset show the model achieves root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) of less than 1\%, 0.7\%, and 1.9\%, demonstrating superior accuracy and robustness, with strong potential for practical SOH estimation applications.