State of Health Prediction of Lithium-Ion Batteries Using Combined Machine Learning Model Based on Nonlinear Constraint Optimization

Accurate State of Health (SOH) estimation of battery systems is critical to vehicle operation safety. However, it's difficult to guarantee the performance of a single model due to the unstable quality of raw data obtained from lithium-ion battery aging and the complexity of operating conditions in actual vehicle operation. Therefore, this paper combines a long short-term memory (LSTM) network with strong temporality, and support vector regression (SVR) with nonlinear mapping and small sample learning. A novel LSTM-SVR combined model with strong input features, less computational burden and multiple advantage combinations is proposed for accurate and robust SOH estimation. The nonlinear constraint optimization is used to assign weights to individual models in terms of minimizing the sum of squared errors of the combined models, which can combine strengths while compensating for weaknesses. Furthermore, voltage, current and temperature change curves during the battery charging were analyzed, and indirect health features (IHFs) with a strong correlation with capacity decline were extracted as model inputs using correlation analysis and principal component analysis (PCA). The NASA dataset was used for validation, and the results show that the LSTM-SVR combined model has good SOH estimation performance, with MAE and RMSE all less than 0.75% and 0.97%.