Physics-informed neural networks for electrode-level state estimation in lithium-ion batteries

Abstract

An accurate estimation of the internal states of lithium-ion batteries is critical to improving the reliability and durability of battery systems. Data-driven methods have exhibited enormous potential for precisely capturing electric and thermal cell dynamics with a low computational cost. However, challenges remain regarding accurate and low-cost data acquisition as electrode-level states are unmeasurable with conventional sensors. This paper presents a hybrid state estimation method for lithium-ion batteries integrating physics-based and machine learning models to leverage their respective strengths. An electrochemical-thermal model is developed and experimentally verified, which is employed to generate a large quantity of data, i.e., voltage, current, temperature and internal electrochemical states, under a comprehensive operating condition matrix including various load profiles and temperatures. These data are fed to train a deep neural network to estimate the internal concentrations and potentials in the electrodes and the electrolyte at different spatial positions. The results show that the proposed approach is capable of bridging spatial, temporal and chemical complexity and achieves a maximum error of 2.93% for all the estimated states under new ambient temperatures, indicating high reliability and generalization ability with solid robustness to input noises and outperforming the one-dimensional network under both normal and noisy conditions.