Online joint-prediction of multi-forward-step battery SOC using LSTM neural networks and multiple linear regression for real-world electric vehicles

Prediction of state of charge (SOC) is critical to the reliability and durability of battery systems in electric vehicles. The existing techniques are mostly model-based SOC estimation using experimental data, which are inefficient for learning the unpredictable battery state under complex real-world operating conditions of electric vehicles. This paper presents a novel machine-learning-enabled method to perform real-time multi-forward-step SOC prediction for battery systems using a recurrent neural network with long short-term memories (LSTM). The training results from a yearlong dataset show that the offline LSTM-based model can perform fast and accurate multi-forward-step prediction for battery SOC. Furthermore, the model has excellent practical application effects by taking into account weather and drivers' driving behaviors during real vehicular operating, and the stability, flexibility, and robustness of this method are verified by 10-fold cross-validation. In order to achieve dual control of prediction accuracy and prediction horizon, we proposed an LRLSTM-based joint-prediction strategy while using LSTM and multiple linear regression algorithms, through which an accuracy benchmark can be obtained, and the prediction steps of LSTM within acceptable accuracy can be flexibly controlled. The smooth implementation of multi-forward-step SOC prediction on real-word vehicles can eliminate drivers' mileage anxiety and safeguard vehicle operation in a big way.