Advanced Machine Learning Approaches for State-of-Charge Prediction of Li-ion Batteries under Multisine Excitation

Abstract

The technological progress of e-Mobility has led to an increased use of Electric Vehicles (EVs), which today satisfies the majority of the customers' demands. Lithium-ion batteries are widely employed in EVs due to their attractive properties. However, the lithium cells, especially in urban traffic are subjected to intense dynamic loads, with a small share of static operation. The Battery Management System (BMS) has a key role in the EV's energy management system and in ensuring reliable operation. Since, a major task of a BMS is to determine the State-of-Charge (SOC) of the cell pack. Its accuracy depends on the model used in the BMS. Typically, simple empirical models are applied for this purpose. Furthermore, these models are parameterized by using standard test measurement data. The SOC prediction issue is still a concern of much research. In the field of energy storage systems, Machine Learning (ML) has recently emerged as a well-established modeling approach. This paper presents a comparative analysis of the performance of State-of-theArt Machine Learning approaches in SOC forecasting with regression, under dynamic loads. Data is generated by performing unique dynamic charge/discharge test by applying multisine signals. The results confirm that a great advantage of Advanced ML models is that they are able to capture important relationships between the variables of interest. Investigations has supported that the State-of-the-Art ML techniques outperform classical ML approaches and are powerful in SOC prediction problems due to their capability of storing past information and catching the cell dynamics, which is critical in predicting the future charge levels.