Lithium-Ion Battery Remaining Useful Life Prediction Based on Data-Driven and Particle Filter Fusion Model

To improve the accuracy and stability of battery remaining useful life (RUL) prediction for lithium-ion batteries, this paper proposes a new convolutional neural network-gated recurrent unit-particle filter (CNN-GRU-PF) fusion prediction model. First, the battery capacity series is decomposed and reconstructed by complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) algorithm and Pearson correlation coefficient method, which reduces the influence of noise on RUL prediction. Then, the capacity is predicted by CNN-GRU, and the CNN-GRU prediction value is used as the observation value of PF, and the prediction error of CNN-GRU is corrected by the state prediction ability of PF. A moving window is used to iteratively update the training set, and the PF optimization value is added to the CNN-GRU training set, forming an iterative training and dynamic updating between them, which improves the long-term prediction performance of CNN-GRU. To verify the effectiveness of proposed method, CNN-GRU-PF model is applied to predict the battery’s RUL. The experiments show that CNN-GRU-PF improves the prediction accuracy of battery B5 by 87.27%, 82.88%, and 55.43% respectively compared with GRU, PF and GRU-PF, and also achieves significant improvement for other batteries. The new model is an effective RUL prediction method with good accuracy and robustness.