A Multi-Step Predictive Model to Estimate Li-Ion State of Charge for Higher C-Rates

Abstract

Energy Storage or battery management systems for Li-ion batteries require accurate prediction of state of charge (SOC). Existing methods predict SOC for a given charging/discharging rate (C-rate) using experimentally obtained values of cell current and voltage. However, in scenarios where there is a lack of such historical data, these methods perform poorly because of inadequate training data. This paper proposes a combinatorial model involving autoregressive integrated moving average (ARIMA) and a nonlinear autoregressive network with exogenous inputs (NARX-net). ARIMA is used to first predict cell current and cell voltage for the desired higher C-rate (C/10) only using the voltage and current from historical, lower C-rates (C/2 to C/8) of an actual 3.7V, 3.5Ah Li-ion battery. The NARX-net is used to predict SOC using the voltage and current values predicted by ARIMA. To train NARX-net, four algorithms are used, and their performance is evaluated by comparing the predicted SOC values with those obtained experimentally for C/10. Results show that the proposed data-driven model is effective at predicting SOC for Li-ion batteries given some preliminary historical data on current and voltage of previous, lower C-rates.