A Model Cluster Adapting to Different Charge Voltage Segments for Battery Capacity Estimation
Qianqian Zhang, Jiangong Zhu, Yixiu Wang, Xuezhe Wei, Haifeng Dai
Abstract
Abstract Accurate capacity estimation is vital for the management of lithium-ion batteries in Electric Vehicles (EVs). Data-driven methods using the battery charging process provide new insights for battery capacity estimation. However, extracting features from the complete or specific charge curves is difficult as the battery charging is related to the behavior of the drivers, e.g., the battery start state of charge (SOC) and end SOC are usually random. Therefore, this study proposes a framework using a model cluster for the capacity estimation of lithium-ion batteries, which uses multi-submodels adapting to different lengths of input charge voltage segments, where input features are extracted. Three datasets (NCA, NCM, and Oxford datasets) are employed to establish the model cluster, and three types of input features and four algorithms are compared. The Random Forest (RF) algorithm combined with the time vector (input feature) achieves the best estimation results on the NCA dataset, in which the Root Mean Square Errors (RMSEs) of most submodels are lower than 1%. Thus, submodels with RMSE lower than 1% are retained to form the model cluster. The NCM dataset is used for the model cluster verification, and all RMSEs are below 0.74%. Three probability distributions of the charging process are constructed based on the three datasets to fit the model cluster to the actual EV operation situation, and the maximum RMSE is 0.403%, which provides a new perspective on the battery capacity estimation for EVs.