Machine learning and impedance spectroscopy for battery state of charge evaluation

Abstract

The Lithium-ion batteries market is rapidly growing. Estimating the batteries State of Charge (SOC) and their State of Health (SOH) is a challenging but crucial task, which Artificial Intelligence (AI) techniques can manage when trained with appropriate data. Physical measurements such as current, voltage and temperature during battery discharge are conventionally used as inputs of AI algorithms to provide an estimation of SOC. In this work, the effect of the battery impedance measurement on the training of a Support Vector Machine (SVM) has been studied. Electrochemical Impedance Spectroscopy (EIS) has been employed for in-situ impedance measurements at different frequencies to consider the effects of each perturbation. The obtained complex impedance values along with the measured current, voltage and temperature data, have been evaluated as features of a training set for an SVM in its regression form (SVR). To allow for simultaneous data acquisition, a module composed of 16 battery cells connected in series has undergone a total of 15 discharge cycles. Several SVR models have been trained with a variety of feature combinations, to evaluate the effect of different impedance information on the resulting model. When using the same battery cell for training and testing, the addition of magnitude and phase of the 100 Hz impedance to the input vector decreased the Root Mean Square Error (RMSE) of the estimated SOC from 1.34% to 1.09%. On the other hand, the same SVR model showed an RMSE of 1.23% when using different (but nominally identical) cells for testing.