Joint Prediction of Lithium Battery State of Health and Remaining Useful Life Based on Indirect Health Features Optimization and Multi-Model Fusion

Lithium-ion batteries have become essential for new energy vehicles and energy storage power stations due to their high energy density, low self-discharge rate, and reliable cleanliness. However, lithium batteries gradually age with increasing cycle counts, leading to decreased battery Performance. State of health (SOH) and remaining useful life (RUL) are important indicators for evaluating battery aging Status, and accurate predictions of these metrics are crucial for the safe Operation of energy storage Systems. However, choosing suitable indirect health features (IHF) for SOH and RUL predictions is challenging. Data-driven models generate uncertainties, resulting in inaccurate predictions of SOH and RUL. Therefore, this paper proposes a Joint prediction method for lithium battery SOH and RUL based on indirect health feature optimization and multi-model fusion.Firstly, multiple health factors (HF) are extracted from charging voltage curves, and IHF are obtained through feature concatenation and attention mechanism optimization, denoted as feature Co-HF. Then, Bayesian model averaging (BMA) is introduced to address uncertainties in the prediction process. With support vector regression (SVR) and long short-term memory (LSTM), SVR-BMA and LSTM-BMA models are constructed. Additionally, the input features for RUL prediction are extracted from SOH capacity predictions using the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) method, enabling Joint prediction of SOH and RUL.The proposed method is validated using the University of Maryland's CALCE lithium battery dataset. Compared with different HFs and predictions of each sub-model, the Co-HF capacity prediction results are closer to the measured values, and the tracking effect of capacity trends is ideal. The mean absolute error (MAE) and root mean squared error (RMSE) of SOH predictions using the proposed fusion model are 0.008 1 and 0.012 0, respectively, and the MAE and RMSE of RUL predictions are 2.56 and 4.63, respectively. The prediction errors are lower than those of each sub-model. In addition, the error tolerance of SOH and RUL predictions is within 5%at SP=100/200/300. Thus, the proposed method significantly reduces the MAE and RMSE of SOH and RUL predictions.From the experimental analysis, it can be concluded that: (1) the optimized feature Co-HF comprehensively reflects the aging information in lithium battery charging voltage, which is suitable for SOH predictions. (2) The proposed model effectively reduces prediction errors of lithium battery SOH and RUL, improving the accuracy and reliability of predictions. Future work will focus on selecting indirect health features that reflect special battery capacity degradation laws and adopting better algorithms to solve model parameters. © 2024 China Machine Press. All rights reserved.