A flexible machine learning based framework for state of charge evaluation

Batteries State-of-Charge (SoC) must be accurately monitored for safe battery operations, and to extend battery life. Machine Learning (ML) algorithms allow to perform the SoC estimation on a data-based approach, avoiding the need for a physical model for each different battery. In this work, a new ML-based framework for the SoC evaluation is proposed, exploiting constant current discharges for model training, rather than the commonly exploited standard drive cycle profiles. This allows avoiding the conversion processes from the drive cycles vehicle acceleration set-point into a current profile, which lead to vehicle-dependent data and the need for a conversion tool. Currents, voltages and temperatures related to different current discharge rates were measured for a Panasonic 18650 Lithium-Ion battery cell. These data were used to train and optimize a Support Vector Regression (SVR) model in the MATLAB environment. Subsequently, different data were combined together to emulate a real vehicle discharge process and were used for evaluating the model. A Root Mean Square Error (RMSE) of 0.564% was obtained, proving that the SVR model trained with constant current discharges data has been capable to estimate the SoC of the tested drive cycles operations.