An Informed-Physic Fused Data-driven Approach Based on GRU for Accurate SOC Esitimation

Accurate state of charge (SOC) estimation of batteries helps order reasonable and efficient instruction on the battery management system. With the sharp development of artificial intelligence, data-driven based method has become one of the mainstream trends of SOC estimation. However, its demand for massive training and lack of interpretability are serious obstacles to perfect performance. In this paper, battery inside physic law is incorporated with artificial neural network to enhance the performance of SOC estimation method. Firstly, battery circuit model is built and Adaptive Extended Kalman Filter (AEKF) is used to extract battery parameter containing physic meaning. Then the physic parameters are used as the input of SOC estimation model. Gate recurrent unit (GRU) network is used to establish the model for its simple structure and good properties. Training and testing profiles come from experiments in the laboratory on eight different batteries. At last, the traditional method, which only takes the current and voltage as the inputs, is compared with the proposed method on their accuracy index. The results prove the high precision of the informed-physic fused method. Take one of the test batteries as the example, RMSE and ME are 0.77% and 1.73% for the proposed method, while the RMSE and ME of the traditional method are 2.11 % and 5.42%. The estimation error has been reduced by 55.49% and 61.07% respectively. It displays that the specialized knowledge fusion provides a huge positive effect on the simple data-driven method.