An ensemble CNN-BiLSTM-attention and BiGRU adaptive-weighted model for state of charge prediction in real electric vehicles

ABSTRACT To guarantee safe electric vehicle system functioning and extend battery life, it is essential to estimate the state-of-charge (SOC) of electric vehicle battery systems. Existing SOC prediction models struggle to balance long-term trends and short-term dynamic changes in time series predictions. To address this issue, we propose a hybrid CNN-BiLSTM-Attention and BiGRU model (CABLG) that improves SOC prediction accuracy by integrating long- and short-term features. Validation using real-world and laboratory data shows superior performance compared to benchmark models. The experimental results indicate that CABLG model exhibits excellent performance. In real driving data, the model achieves the best prediction results in winter conditions with MAE of 0.0945, MSE of 0.0187, and RMSE of 0.1369. In laboratory conditions, the model performs well with MAE of 0.0018, MSE of 4.7962E-06, and RMSE of 0.0022. The results clearly indicate CABLG model’s excellent performance in SOC prediction and also demonstrate its potential in real-world applications.