A hybrid deep learning model for lithium-ion batteries state of charge estimation based on quantile regression and attention

Abstract

Accurate estimation of the state of charge (SOC) is essential to ensure the safe and efficient utilization of lithium-ion batteries. However, external factors can affect the SOC, making it challenging to achieve precise estimations. To address this issue, we propose a deep learning model that integrates the parallel computing capabilities of temporal convolutional networks with the robust learning abilities of gated recurrent units. This model incorporates an attention mechanism to dynamically assign weights and focus on salient features based on correlations in historical information. Furthermore, we employ the quantile regression loss function during network training, thereby equipping the neural network model with the capability to directly generate interval estimates. The results show that the MAE and RMSE of the lithium iron phosphate battery, lithium cobalt oxide battery and ternary lithium battery dataset are below 1.45 and 1.98, 1.12 and 1.25, 0.61 and 0.75 respectively. The comparison analysis with other SOC estimation methods demonstrates its exceptional accuracy in point estimation and high reliability in confidence interval estimation. Furthermore, it showcases remarkable generalization ability across diverse temperatures, operating conditions, battery lifetimes, and battery types.