Hybrid Physics and Data-Driven Electrochemical States Estimation for Lithium-ion Batteries

Accurate estimation of electrochemical states plays a fundamental role in guaranteeing safe, reliable, and efficient operations of lithium battery systems. However, current estimation algorithms rely on semi-empirical models that lack physical insights or expensive physics-based models that are hard to implement. Thus, this paper proposes an efficient hybrid physics-based and data-driven electrochemical state estimation framework of lithium-ion batteries by leveraging the advantages and circumventing the disadvantages of physics-based and data-driven models. First, a hybrid physics and data-driven battery model is established through systematic integration of the full-order pseudo-two-dimensional model and long short-term memory recurrent neural network. Second, a data argumentation technique is employed to establish physics-informed training datasets and a battery model with an output error of 10.45mV is trained using the data. Third, the unscented Kalman filter is pertinently designed to detect local concentrations, i.e., state-of-charge-related physical variables. Finally, the estimation performance is comprehensively examined under both simulation and experimental scenarios. The results demonstrate that the proposed framework can accurately provide physically meaningful state variables under a wide operation range, with an estimation error of less than 3% for SOC under standard conditions.