Prediction of Lithium Battery Voltage and State of Charge Using Multi-Head Attention BiLSTM Neural Network
Haiwen Xi, Taolin Lv, Jincheng Qin, Mingsheng Ma, Jingying Xie, Shigang Lu, Zhifu Liu
Abstract
Predicting battery states such as the voltage and state of charge (SOC) can help us monitor lithium batteries more efficiently during usage. This study proposed a predictive model for the lithium battery voltage and SOC by combining a second-order RC equivalent circuit model with a multi-head attention Bidirectional Long Short-Term Memory (MHA-BiLSTM) neural network. The equivalent circuit model simulates long-term charge–discharge cycles in Simulink, providing essential data for model training. The BiLSTM model, enhanced by the multi-head attention mechanism, is used for accurate short-term predictions of the battery voltage and SOC. The experimental results demonstrate that the proposed MHA-BiLSTM model outperforms other models in the prediction accuracy, achieving an R2 of 0.91, with the lowest RMSE of 0.0567 and MAPE of 0.0095. This hybrid approach effectively captures the dynamic behavior of the battery and reduces predictive errors, making it a promising solution for battery health monitoring and management.