Real-Time Predicting the Low-Temperature Performance of WLTC-Based Lithium-Ion Battery Using an LSTM-PF Sequential Ensemble Model

Predicting an abnormally rapid decline in battery capacity in low-temperature environments is important for maintaining battery stability and performance. This study introduces a method that integrates cycling tests under various current conditions with deep neural network algorithms to identify and predict in real-time the trend of battery capacity reduction in low-temperature conditions (
$- 10~^{\circ }$
C). For this method, 18 feature data points were included, consisting of the test environment and conditions, as well as geometric and statistical features. The importance of these features was analyzed using the Random Forest (RF) algorithm, and the top 12 feature data points were selected to improve the efficiency and accuracy of the Long Short-Term Memory (LSTM) model. Furthermore, we applied a sequential ensemble technique that uses the output of the LSTM model as the input for the particle filter, significantly improving the performance of the prediction model. The approach was used to predict the capacity of the tested battery using C-rate transformation based on the WLTC. The results showed an error rate of 0.9% and an RMSE of 0.0048, representing a 25% decrease in the error rate and a 48% reduction in the RMSE compared with those predicted by the LSTM model.