SoC estimation of lithium-ion batteries based on machine learning techniques: A filtered approach

Accurate state-of-charge (SoC) estimation is an essential requirement for many situations where Li-Ion batteries (LiBs) are used. This ensures an efficient battery management system (BMS), so the battery can be protected from excessive discharge, and its life span can be maximized. But when it comes to electrified vehicles (xEVs), the SoC estimation accuracy becomes a more critical and indispensable prerequisite. Because the safety of xEVs during driving and the remaining range, which is an indicator of how far the vehicle can go, are directly related to the accurate SoC. However, the complex electrochemical reactions in the battery and the dependence on environmental variables make SoC estimation a challenging task. Traditionally, this is tackled by establishing either electrochemical or electrical battery equivalent models. Both methods suffer from some limitations, such as parameter identification, complex calculations, and model mismatching due to the aging factor. On the other hand, data-driven methods have recently become a popular choice for SoC estimation since they enable building data-based models rather than chemical reactions or equivalent circuit calculations. The model is built based on battery parameters such as current, voltage, battery type, and then used for SoC estimation. However, many studies in the literature examine only a few methods for SoC estimation. Also, these data-driven black box models can lead to outlier data as they are not observers. Thus, the aims of this study are twofold: First, to make a comprehensive comparison based on most of the ML methods. Second, to utilize several filters for outlier removal and measure their effectiveness. For these purposes,18 ML algorithms were handled in three main groups, and SoC estimation results were analyzed. Additionally, five different filters were used to improve the SoC estimation of these methods, and their comparisons were realized. From the results, it is clear that Bagging and ExtraTree algorithms are substantially better than other ML methods for SoC estimation since their Interquartile Range (IQR) is smaller than 3%, performance indices are the lowest ones, and curve matches are the best. Also, Rloess is the best filter among the others, although they all achieved high performance in outlier removal.