A Comparative Analysis of Machine Learning Approaches for State-of-Charge Forecasting for Enhancing Lithium-ion Battery Management in Electric Vehicles

Abstract

The present research is concerned with the analysis of the efficiency of machine learning approaches for state-of-charge forecasting in lithium-ion batteries to help battery management systems become more efficient in electric vehicles. To achieve the goal, four machine learning models, namely Artificial Neural Networks (ANN), Support Vector Machines (SVM), Decision Trees (DT), and Naive Bayes (NB) and the corresponding model training technology were used along with a dataset in which battery parameters, environmental and vehicle operating conditions were implied. Data preprocessing methods, such as cleaning, feature selection, and scaling were also implemented to make the subsequent forecasting procedures more efficient. In terms of the reached state-of-charge level prediction, the efficiency of the employed models was as follows: ANN was measured as the best one with 97.89% of data validity, SVM with 94.5%, DT with 91.22%, and NB with 88.97%. Afterwards, the ANN model was used in real-time along with processing the data collected from the sensors with the purpose of optimizing the vehicle’s work and slowing down the level of battery wear due to broad usage opportunities. As a result, the battery life increased by 2 minutes 6 hours. This fact demonstrates the highest advantages and, therefore, efficiency of the ANN model in terms of electric vehicle operation and battery life optimization. Data preprocessing is mandatory for higher quality and reliability of the machine learning models in terms of state-of-charge level forecasting.