A hybrid system for Lithium-ion battery State-of-Charge univariate forecasting

In a world where Lithium-ion batteries are in increasing demand, appropriate management techniques have become more relevant than ever to be able to control possible hazards, performance reliability and early degradation. Several techniques have been studied in order to achieve a proper estimation of battery State-of-Charge (SoC, battery's performance metric). In this study, a weighted average ensemble model is proposed to consider linear and non-linear components of the time series using ARIMA+SVR residual forecasting while adding the capacity of LSTM to process large datasets properly. One hundred output voltage and current values measured at 1C-rate where forecasted comparing the results of 7 models found in literature to the proposed one. Later on, SoC was computed using the Coulomb's counting method based on the mentioned current forecast. Forecasting using the proposed weighted average ensemble of ARIMA+SVR and LSTM shows lower MSE and MAPE error value than the ones achieved in comparison methods for both current and voltage time series. SoC resulting values also showed better MSE, MAPE and R2 than the values obtained by ARIMA+SVR (0.14401%, 3.31933%, and 13.40696% improvement) and LSTM (0.01450%, 0.45319%, and 0.71939% improvement) on their own.