Data-Driven Aging Modeling of a Light Electric Vehicle Battery Based on Synthetic Load Cycles and Physics-Based Aging Simulations

Aging of lithium-ion batteries is one of the key features addressed through battery design and management, since it corresponds to continuous decrease in battery energy and power densities. Predicting battery aging during its exploitation in a vehicle by means of detailed, physics-based modeling is challenging due to mathematical and computational complexity, which makes data-driven methods attractive. However, the outputs of data-driven models, such as the capacity change, are hard to measure after every load cycle. This paper, therefore, proposes a data-driven modeling methodology that combines the physics-based and data-driven approaches. A limited amount of recorded field data of a light electric vehicle (a 3.5 kW electric scooter) is used for generating a large amount of synthetic load cycles, which are then inputted to a physical aging simulation model implemented within the GT-AutoLion environment (21 aging simulations with 41,347 load cycles in total). With this approach, the capacity change after every load cycle can be calculated and used as the output of linear regression and alternative neural network models, whose inputs are relevant load cycle features systematically selected from a rich set of feature candidates. The model covers two key aging processes, growth of the solid electrolyte interphase and lithium plating on a graphite anode, the latter employing a neural network-based classifier of plating activation. The data-driven models are surrogates of the physics-based reference model, meaning that the reported errors express their agreement with the reference simulation, and not with the measured capacity fade. On four held-out test simulations, which recombine the recorded micro-cycles of the training simulations but were not used for model identification or selection, the linear regression model of the capacity loss due to interphase growth reaches a per-cycle coefficient of determination of 0.85 and a cumulative error of at most 0.20% of the nominal capacity. The cumulative capacity loss of the complete model deviates from the reference by at most 0.36% of the nominal capacity on the test simulations, partly owing to compensating errors of the lithium-plating sub-models, 0.33% on a simulation driven by the recorded load data of the same scooter, and 1.34% on the validation simulations.