Performance Analysis on Artificial Neural Network Based State of Charge Estimation for Electric Vehicles

Abstract

In the recent years, Artificial Neural Networks (ANNs) have gained wider interest in estimating the State of charge (SOC) of Li-ion batteries used in electric vehicles. As the ANN configurations proposed in recent literature were trained under different training parameters and datasets, a fair comparison cannot be made by directly referring to the prediction errors reported. Thus, the SOC prediction performance of the ANNs proposed in the recent years were investigated, by training with same training parameters and dataset (US06 vehicle dynamic profile from the Centre of Advanced Life Cycle Engineering). Results show that the testing dataset Mean Squared Error (MSE) for using only Convolutional Neural Network (CNN) is 3.140%, whereas combining CNN with Long Short-Term Memory Networks (LSTM-RNN) is 1.820%, and CNN with Gate Recurrent Unit (GRU-RNN) is 1.819% MSE. Therefore, it is evident that in-cooperation of any form of recurrent architecture in an ANN configuration contributes to better SOC prediction. The results also highlight that inclusion of a bidirectional recurrent architecture such as Bidirectional LSTM-RNN (MSE: 0.927%) and attention mechanism such as the combination of LSTM-RNN with attention (MSE: 0.004%) contribute to better SOC prediction. Overall, the performance analysis conducted shows that there is a need in further research investigation on integrating different types of bidirectional recurrent architecture and attention mechanism with other ANNs and evaluate the SOC prediction performance as compared to previously proposed ANN configurations. Successful testing and implementation would contribute to increased battery life span and reduced maintenance costs, leading to increased usage of EVs.