Predicting Battery State of Charge: A Machine Learning-based Approach

This paper aims to predict Battery SOC using a machine learning-based approach, employing three distinct methods: Random Forest Regression, Gradient Boosting Regression, and Artificial Neural Networks. The state of charge (SOC) is a critical battery parameter for electric vehicles (EVs), as it indicates the remaining driving range and battery health. Traditional methods frequently face challenges in delivering precise and real-time State of Charge (SOC) estimates. This is due to the intricate and dynamic battery behavior, compounded by variations in driving conditions and ambient temperature. A publicly available dataset of a BMW i3 (60Ah) EV, which contains data from 70 real driving trips in summer and winter season have been utilized for model training & testing. We evaluated the accuracy and robustness of these methods using various statistical metrics and graphical plots. The machine learning models developed in this paper are able to achieve prediction accuracy of more than ${9 8 \%}$ in case of RFR & GBR and $\mathbf{9 0 \%}$ in case of ANN. The proposed approach can provide a fast and reliable way to predict battery SOC for EVs, and can be extended to other types of batteries and applications.