Diversity-Maximizing Strategy for Data Set Curation in Li-Ion Battery Modeling Based on Vehicle Acceleration Data

Accurate modeling of battery dynamics based on operating conditions is increasingly important for advanced battery management. To achieve this, it is necessary to construct and maintain data sets, based on on-board cell measurements, which cover the full State-of-Charge (SOC) range of the cell. However, the performance of the resulting battery model is highly dependent on the specific data included in the data set. In this paper, a data-selective strategy is proposed to construct and curate a data set that produces battery models with high predictive performance on a wide variety of driving profiles. By iteratively replacing certain segments of the data set over multiple discharge cycles, the proposed method aims at maximizing the diversity between the data segments within the data set. Rather than comparing cell data directly, the selection is moved away from the cell level to the vehicle level and is based on comparing specific features of the acceleration profile of the vehicle.Both features extracted from the time domain and frequency domain have been investigated. The proposed selection strategy is used to produce empirical battery models in the Linear Parameter-Varying (LPV) framework based on experimental battery data constructed from synthetic dynamic driving profiles containing city and motorway driving. Experimental results show a reduction in model prediction error of up to $44\%$ compared to constant replacement, and a very high robustness towards the specific order that the driving profiles used for estimation appear in.