关键词: 机器学习, SOH, 混合模型, LSTM

Abstract:

The traditional method of predicting the state of health (SOH) of a battery is generally based on historical data. Predicting the real-time state of a lithium battery or estimating its remaining service life is difficult. Aiming at the real-time prediction of battery SOH, we introduce a large amount of real-vehicle battery data (combined with machine learning and ampere-hour integration method to model and predict SOH) to process features and train data. On the basis of the model test results, this article proposes a real-time SOH hybrid prediction model combining the LightGBM and CatBoost algorithms. By verifying the hybrid model with two real vehicles as the carrier, the measured absolute average error of the real-time SOH prediction is 0.009. Our research intends to obtain the SOH attenuation curve to predict the remaining battery life. Therefore, we establish a long short-term memory (LSTM) neural network model to predict the future decay curve of battery SOH, characterized by the difference in SOH within a fixed time interval. This reduces the fluctuation of the difference and ensures that the data have similar distribution laws. By verifying the real-time monitoring data set provided by an original equipment manufacturer, the absolute average error of the future attenuation curve prediction is 0.021. The overall results show that the real-time SOH prediction model and the remaining life prediction model of the lithium battery studied in the article have high prediction accuracy. The battery user can better grasp the real-time status of the lithium battery and provide a basis for relevant decision making.

Key words: machine learning, SOH, combined model, LSTM