Abstract:In this letter, we train deep learning (DL) models to estimate the state-of-charge (SOC) of lithium-ion (Li-ion) battery directly from voltage, current, and battery temperature values. The deep fully convolutional network model is proposed for its novel architecture with learning rate optimization strategies. The proposed model is capable of estimating SOC at constant and varying ambient temperature on different drive cycles without having to be retrained. The model also outperformed other commonly used DL models such as the LSTM, GRU, and CNN on an open source Li-ion battery dataset. The model achieves 0.85% root mean squared error (RMSE) and 0.7% mean absolute error (MAE) at 25 °C and 2.0% RMSE and 1.55% MAE at varying ambient temperature (–20–25 °C).

SOC Estimation of Li-ion Batteries With Learning Rate-Optimized Deep Fully Convolutional Network