Research on Minimization of Data Set for State of Charge Prediction

Abstract

The quick estimation and prediction of lithium-ion batteries’ (LIBs) state of charge (SoC) are attracting growing attention, since the LIB has become one of the most essential power sources for daily consumer electronics. Most deep learning methods require plenty of data and more than two LIB parameters to train the model for predicting SoC. In this paper, a single-parameter SoC prediction based on deep learning is realized by cleaning the data for lithium-ion battery parameters and constructing the feature matrix based on the cleaned data. Then, by analyzing the feature matrix’s periodicity and principal component to obtain two kinds of the original eigenmatrix’s substitution matrices, the two substitutions are fused to obtain an excellent prediction effect. In the end, the minimization method is verified with newly measured lithium battery data, and the results show that the MAPE of the SoC prediction reaches 0.96%, the input data are reduced by 93.33%, and the training time is reduced by 96.68%. Fast and accurate prediction of the SoC is achieved by using only a minimum amount of voltage data.

Keywords

• long short-term memory (LSTM)
• state of charge
• principal components analysis (PCA)
• data fusion