Battery energy storage systems are vital for a variety of applications, with a particularly important role in facilitating the widespread use of renewable energy resources and electric vehicles. To ensure the safety and optimal performance of these devices, analyzing their operation through physical and data-driven models is essential. While physical models can effectively model the underlying physicochemical processes, their complexity often renders them impractical for real-time onboard diagnostics. Conversely, data-driven models are usually more flexible and easier to implement, but they lack a physical description of the battery. In response to these challenges, this work models the battery state using a single particle model as a baseline for subsequent predictions made with neural networks. To achieve this, two neural networks were leveraged: one to apply a correction to the voltage obtained from the physical model, and the second to evaluate the battery state of health and aging states. The novel hybrid model, integrating a single particle model with two neural networks, consistently outperformed both the individual single particle model and the two neural networks in isolations. Results were benchmarked using two real-world battery cycling datasets, including one collected in-house. The hybrid model consistently outperformed the individual neural networks in terms of voltage prediction accuracy, as evidenced by lower root mean square error (RMSE) values. Notably, in four out of the five cases where the analysis was stratified by battery manufacturer, the RMSE was reduced by at least 50 % and up to tenfold with the hybrid approach. Given the promise of this new hybrid model, it is expected that the present work will pave the way for advanced modeling of batteries.
Pepe et al. (Wed,) studied this question.