A hybrid data-driven method for voltage state prediction and fault warning of Li-ion batteries

Abstract

As the extensive application of electrochemical energy storage (EES), Li-ion battery fault is a key factor reference to the reliable operation and system security, influencing by the environment temperature and battery voltage. To address distinct challenges in lithium-ion battery fault prediction, such as nonlinearity and complex electrochemical reactions within battery state sequence data, a novel 1DCNN-Bi-LSTM hybrid network has been proposed to predict the Li-ion battery fault. Firstly, an 1DCNN module is introduced to extract voltage-related multi-dimension features. Secondly, a Bi-LSTM module is used to learn long-term dependence relationships among fused features while integrating a self-attention mechanism. To further verify the algorithm's effectiveness, a new 18650 battery dataset has been set up under various conditions between day and night. The experimental results show that our model has high accuracy and exemplary performance in various environmental temperatures. The prediction errors for comparative experiments are approximately MAPE of 0.03 %, RMSE of 0.0003 %, MAE of 0.12 %, and R2 of 0.99. Compared with mainstream methods, our prediction result is close to true values, performs better at peaks and valleys, and has higher computational efficiency. Considering the temperature factor and voltage variation, our developed method can be effectively applied to battery management system (BMS).