A comparative study of battery state-of-charge estimation using electrochemical impedance spectroscopy by different machine learning methods

Battery state-of-charge (SOC) estimation is an essential function of the battery management system. Most of the current data-driven SOC estimation methods are characterized by voltage or current. Estimating the battery SOC based on battery impedance is an important and emerging technique. This article makes a detailed and specific comparison and discusses impedance-driven SOC estimation methods. First, this article collected four electrochemical impedance spectroscopy datasets that covered different battery types and capacities. Secondly, four commonly used machine learning models and three types of feature extraction strategies are employed to estimate SOC for each dataset. Finally, the SOC estimation results of a single dataset are evaluated in terms of accuracy metrics and computational complexity. At the same time, the performance rank is used to evaluate the models among datasets. The result indicates that the dense neural network model has the highest accuracy metrics but has the largest computational complexity. The generalized linear regression model has the smallest computational complexity but the lowest accuracy metrics. The random forest regression and Gaussian process regression models are more suitable for online scenarios that require faster speed and less memory due to their relatively high accuracy metrics and low computational complexity.