Online state of charge estimation of lithium-ion battery using surrogate model based on electrochemical model

As electric vehicles (EVs) become popular, the use of lithium-ion batteries is increasing. However, it is difficult to ensure the reliable use of the battery because the state of charge (SoC) of a lithium-ion battery cannot be measured directly. SoC estimation is necessary to ensure the safety of battery systems by preventing over-charge and discharge of the battery and to extend battery life through efficient use. The optimal battery balancing strategy, estimation of the remaining driving range of EV, and Vehicle to Grid strategy can be developed through accurate SoC estimation. Therefore, it is necessary to study how to estimate the SoC of the battery in real-time. As methods for estimating SoC, equivalent circuit model, electrochemical model, and recently artificial neural network-based model are being studied. The electrochemical model is not suitable for real-time use due to its high computational complexity. Artificial neural network-based models require a large amount of data for learning, but most of the data is collected in the lab, which can lead to a lack of data and is difficult to ensure accuracy in such cases. In this study, a surrogate model based on electrochemical models for SoC estimation is developed to solve computational complexity problems of electrochemical models and the accuracy problems of data-driven models due to data dependence. Based on the 1C discharge experimental data, parameter identification of the electrochemical model was performed using a Genetic Algorithm. Output variables such as Li-ion amount in the negative electrode and voltage for various drive cycle loads are derived from this model to create a sufficient amount of data for model training. The surrogate model is trained using these output variables and compared with data-driven models in terms of accuracy and computational complexity. Based on the Long Short Term Memory (LSTM) architecture, the artificial neural network-based model was trained gradually with driving cycle data and then compared with the surrogate model at each step. As a result, the surrogate model based on an electrochemical model using fewer data showed feasible computational complexity and high accuracy. Through the proposed method, it is expected to accurately estimate battery SoC in real-time and make the battery use efficiently.