关键词: 锂离子电池, 寿命预测, 自适应白噪声完整集成经验模态分解, 随机森林, 神经网络

Abstract:

Lithium-ion battery's remaining useful life (RUL) is an important indicator of battery health management. Therefore, in this paper, using battery capacity as an indicator of health status, including modal decomposition and machine learning algorithms, a CEEMDAN-RF-SED-LSTM method was proposed to predict lithium battery RUL. First, adaptive white-noise full-ensemble empirical-mode decomposition (CEEMDAN) was used to decompose the battery capacity data. Then, to avoid the influence of the noise in the fluctuation component on the prediction ability of the model and not completely discard the characteristic information in the fluctuation component, this paper used the Random Forest algorithm to obtain important values for each fluctuation component, after which sexual ranking and numerical values were used as weights for each component's ability to explain the original data. Subsequently, the weight value and prediction result obtained by the neural network model constructed by different fluctuation components were weighted and reconstructed, resulting in the RUL prediction of the lithium-ion battery. Next, this research compared the prediction accuracy of the single model and the combined model, followed by the addition of the combined model prediction accuracy of RF, to improve the performance of the five neural networks further, after which the Simple Encoder-Decoder (SED) mechanism was introduced for the two networks with better performance, LSTM and GRU, to better learn the global temporal features and long-range dependencies of sequence data. We finally tested the method's performance using the NASA dataset as the research object. The experimental results showed that although the CEEMDAN-RF-SED-LSTM model performed well in battery RUL prediction, the prediction results had lower errors than the single model.

Key words: Li-ion battery, life prediction, adaptive white noise full ensemble empirical mode decomposition, random forest, neural network