State of Charge Estimation of Li-ion Batteries through Efficient Gated Recurrent Neural Networks using Engineered features

Accurate State of Charge(SOC) estimation of Li-ion batteries has been a critical issue in Battery Management Systems(BMS) for the safety and reliability of Battery. There are different methods for estimating SOC, out of which Machine Learning based techniques are becoming more popular because they don't depend on complex Battery modelling aspects. In this paper, Gated Recurrent Units based Recurrent Neural Networks(GRU-RNNs) are used, which can capture the dependency between present output and past inputs, on which the SOC of the battery depends. But GRUs require relatively higher computational power. So the proposed neural network is built with minimum GRU units making it computationally efficient, making it suitable for low cost microcontrollers. The process of feature engineering, where additional input features are obtained from available data, is used to boost the accuracy of the model. The proposed model is able to estimate SOC with a Mean Absolute Error(MAE) of 0.85% on the Panasonic dataset at 25 degrees C. Time taken for one forward pass on Teensy 3.6 is 0.215 seconds.