Data-driven state-of-charge prediction of a storage cell using ABC/GBRT, ABC/MLP and LASSO machine learning techniques

Abstract

Electric vehicles (EVs) will be the dominant technology for the automobile industry due to efficiency and environmental reasons. Lithium-ion batteries lead the energy supply business for the most recent group of EVs and many other electronic consumer devices. One of the most important pieces of information for EV users is the state-of-charge of the battery, also known as SOC. The SOC works like the fuel gauge for the battery. Information about remaining battery capacity is essential to avoid running out of battery power. Battery remaining charge is not easy to estimate, due to non-linear phenomenon inside the battery. This work is concerned with SOC prediction using machine learning techniques. Three machine learning tools, called Artificial Bee Colony-Multilayer Perceptron (ABC/MLP), Artificial Bee Colony gradient boosting regression tree (ABC/GBRT) and Least Absolute Selection and Shrinkage Operator (LASSO) have been used to build models that enable the prediction of the SOC of a storage cell. The predictive results confirm the enhanced performance of the ABC/GBRT-based model over the other methods for SOC prediction. SOC errors remain below 1%, 10% and 17% for ABC/GBRT, ABC/MLP and LASSO, respectively. The goodness of fit, calculated using R2, was 0.99, 0.95 and 0.81 for the three methods, respectively. A comparison of the results obtained using all the methods has also been carried out.