Analysis of Optimal Deep Neural Network for Enhanced Battery SOC Estimation

This paper presents a deep neural network (DNN) approach for enhanced state of charge (SOC) estimation of Lithium-Ion (Li-ion) battery for eMobility. A series of DNN models are developed with varying numbers of hidden layers trained with dynamic drive profiles, to test and analyze the performance on untrained drive profiles. Analysis indicates that DNNs with four, five, and six hidden layers are found to be top performers for SOC estimation. To examine the performance and generalization capability of the suggested DNN architecture for battery SOC estimation, the model is tested with three different dynamic drive cycles at different temperatures using error metrics. The root mean squared error (RMSE) and mean squared error (MSE) of 1.57% and 0.024%, respectively, were observed when the model was tested with the Dynamic Stress Test dataset. The maximum RMSE and MSE of 2.64% and 0.069%, respectively, were observed when tested with the Federal Urban Driving Schedule dataset. Further, the results of SOC estimation using suggested DNNs are compared with other advanced techniques to showcase the enhanced performance of the suggested architecture.