SOC estimation of Li-ion battery using convolutional neural network with U-Net architecture

Abstract

State-of-charge (SOC) is critical to the safe operation and energy management of electric vehicles. Data-driven SOC estimation algorithms all require a period of data to ensure convergence of the estimation results and cannot accurately estimate the SOC values near the starting point. We propose a SOC estimation method based on the U-Net architecture that can handle variable-length input data and output equal-length SOC estimation results, including accurate SOC of the starting point. Symmetric padding convolutional layer was proposed to address the boundary effect of Convolutional Neural Networks (CNN) and improve the accuracy of SOC estimation at the edges. We also propose a total variation loss function, which improves the stability of the estimation only by optimizing the loss function without increasing the model complexity, and significantly reduces the maximum error. The model was trained using dynamic drive cycle data at five constant temperatures, and the model has high accuracy at both constant and variable temperature conditions. The proposed method can estimate the SOC at constant temperatures with mean absolute error (MAE) within 1.1% and root-mean-square error (RMSE) within 1.4%. This method also can estimate SOC at varying temperatures with MAE within 1.5% and RMSE within 1.8% under different driving conditions.