Multi-step surface temperature prediction of lithium-ion batteries based on a CNN-LSTM hybrid architecture

Abstract

Accurate surface temperature prediction of lithium-ion batteries serves as the core prerequisite for feed-forward control in battery thermal management systems. Given that a single prediction model fails to simultaneously capture local temperature fluctuations and long-term thermal accumulation trends within temperature dynamic variations, this study proposes a multi-step temperature prediction method based on the hybrid architecture combining a convolutional neural network and long short-term memory network (CNN-LSTM). In this method, the CNN module is adopted to extract local instantaneous temperature rise features induced by sudden current changes, while the LSTM module is applied to acquire long-range dependency characteristics of thermal accumulation throughout the complete discharge cycle. Meanwhile, the direct multi-step prediction strategy is utilized to output temperature values of the subsequent five-time steps in a single operation. Verified on the discharge simulation data set of lithium-ion batteries, experimental results demonstrate that the mean absolute error (MAE) on the test set reaches 0.364°C, and the MAE of each prediction step remains steadily within the range of 0.35°C to 0.38°C. The prediction errors are mainly distributed between −0.5°C and 0.5°C following a normal distribution, and the coefficient of determination R2 of the training set is 0.836. Additionally, prediction errors of each step do not accumulate obviously with the increase of prediction steps, which confirms the feasibility and superiority of the direct multi-step prediction strategy. This research can provide a practical deployable multi-step prediction scheme for feed-forward control implementation in battery thermal management systems.