Empirical model, capacity recovery-identification correction and machine learning co-driven Li-ion battery remaining useful life prediction

Li-ion battery is the most important energy storage and conversion device. RUL prediction, as an important part of the battery health management system, provides important information for specifying energy control strategies and preventing Li-ion battery failure. In this paper, Empirical Model, Capacity Recovery-Identification Correction and Machine Learning co-driven method was proposed to address the inaccurate and unreliable RUL predictions of Li-ion batteries caused by difference data and non-stationary trends. Firstly, on the basis of using only the historical data of the target Li-ion battery, the acknowledged bi-exponential degradation model was used to generate the guidance sequence, which guides the output of the machine learning model and avoids the unsatisfactory prediction effect caused by difference data. Secondly, the recoverable capacity present during the capacity degradation process was analyzed and identified, and overly aggressive or conservative predictions caused by non-stationary trends were avoided by correcting the historical recoverable capacity. Finally, experiment results on publicly available datasets show that the method proposed in this paper can effectively improve the accuracy of the prediction using historical data, with MAPE of the degradation trend prediction being only 3.79 %, and the average RA of RUL prediction for different failure thresholds remaining at the level of 90 %.