Temporal Convolutional Recombinant Network: A Novel Method for SOC Estimation and Prediction in Electric Vehicles

Abstract

Mileage anxiety is one of the factors affecting the development of electric vehicles (EVs). Accurately estimating and predicting the state of charge (SOC) of power batteries can alleviate this problem. However, due to the complex and variable operating conditions of EVs, SOC estimation is challenging in real-world driving scenarios. To address this issue, we propose a new neural network method called temporal convolutional recombinant network (TCRN) for estimating and predicting the SOC of power batteries. This method adopts a non-normalized temporal convolutional network (TCN) model, which can extract temporal information in parallel computing. It has the advantages of fewer parameters and higher accuracy. To tackle the oscillations in TCN outputs, we design a temporal recombination module (TRM). It optimizes temporal information more effectively by generating time recombination weights, further improving prediction accuracy. The framework's superiority and effectiveness are verified by comparing different models using a dataset of EVs with lithium-ion batteries (LIBs). The proposed method reduces the mean absolute error (MAE) by 23.2% compared to the original TCN, while only increasing the parameters by 2.8%, providing a more accurate SOC estimation. Moreover, it achieves good result in predicting the future SOC, which to some extent alleviates the driver's mileage anxiety.