Performance Comparison of FCN, LSTM and GRU for State of Charge Estimation

The longevity and safety of lithium-ion (Li-ion) batteries depend on accurate State of Charge (SoC), a key and vital characteristic in battery management systems (BMS) of electric vehicles (EV). This work involves implementing and comparing the performance of Fully Convoluted Neural Network (FCN), Long Short-term Memory (LSTM) and Gated Recurrent Unit (GRU) for SoC estimation. The models for SoC estimation, based on these algorithms, are applied to the US06 Highway driving Schedule, the eVTOL battery dataset and the BMW i3 dataset from IEEE Dataport. The models are trained to predict the SoC when voltage, current and temperature are given as inputs using Jupyter Notebook and libraries like Keras and Tensorflow. The effectiveness of the models is assessed using quantitative performance metrics such as Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), time and memory consumed by the models to train and predict the SoC. The GRU model shows better results with an RMSE score of 0.0013 and MAE score of 0.009 compared to the LSTM and FCN model. The work can be further extended by implementing and analysing more models and more battery profiles facilitating an easier selection of best suited model for estimation of SoC.