Data-driven Li-ion battery health condition assessment

Abstract

Due to the lithium-ion batteries’ suitable energy densities, charge/discharge efficiencies, self discharging rates and cycle durabilities, they are widely used in the transportation power, renewable energy storage systems, mobile communication and other fields. But the battery aging problems may cause serious consequences, so it is necessary to monitor the battery state of health (SOH) in real time. Among three common battery health assessment methods, which are the direct assessment methods, the model based methods and the data-driven methods, the data-driven methods can process a large amount of data, which is a great advantage when the computer nowadays has a strong ability to work with data. Various researches based on the data-driven methods need to extract the health indicators (HIs) from the battery charging and discharging curves, and these researches use the machine learning algorithms or other algorithms to estimate the battery SOH by the HIs. These HIs are obtained by making some changes to the original data of the batteries. There are a few methods which can directly use the charging and discharging data as the input to the algorithms. In order to simplify the complex process of extracting the HIs, in this dissertation, a direct health indicator, Electrochemical Impedance Spectroscopy (EIS), is selected to do the battery health estimation together with the Broad Learning System (BLS). The BLS with hyper-parameter tuning is also studied to see its contribution to the results. The process includes offline modeling and online estimation. And the BLS is compared with other common machine learning algorithms to see which is better in the battery SOH estimation. The results show that the BLS with hyper-parameter tuning has the best testing accuracy, with the minimum RMSE values.