Li-ion battery health diagnosis based on electrochemical model

Abstract

Secondary batteries are widely leveraged in the rapidly growing electric vehicle (EV) and energy-storage system (ESS) markets. Among the battery types, Li-ion batteries are extensively used, owing to their advantages of high energy and power densities, low self-discharge rate, and lack of memory effect. However, they tend to be susceptible to thermal runaways caused by undesirable operations, because their thermal stability is inherently lower than other batteries. Moreover, as Li-ion batteries become more widespread, the number of associated accidents also increases significantly. To use such delicate batteries safely and efficiently, it is essential to accurately estimate their health states.

This thesis deals with several methods for diagnosing the health states of Li-ion batteries from various angles based on an electrochemical model, especially a pseudo-2-dimensional (P2D) model, which precisely expresses the dynamics of a Li-ion battery. First, a data-driven state-of-health (SOH) estimation scheme for Li-ion batteries with reference performance test (RPT)-reduced experimental data is proposed. In this work, it is proved that the capacity of a Li-ion battery can be effectively estimated using only scarce data acquired through minimal number of RPT performing, unlike most of existing studies have employed abundant reference data. Second, some parameters of the P2D model for effective and practical diagnosis of Li-ion batteries are selected as aging parameters and identified. In this study, some major parameters of the P2D model were identified through a genetic algorithm (GA). Based on the results, the parameters suitable for diagnosing the battery health state were selected as aging parameters, and their physical meanings were considered. Lastly, a novel P2D model parameter identification method, genetic algorithm and neural network cooperative optimization (GANCO), is proposed. Many existing studies that employ meta-heuristic methods to identify model parameters did not efficiently utilize data generated during optimization. In this work, an 1-dimensional convolutional neural network (1D CNN) periodically trains the sample data obtained from previous iterations to learn thedynamics between the known input current and the corresponding simulated voltage. Then the trained network recommends highly probable parameter candidates to the GA. In such manner, an 1D CNN and a GA cooperate and complement each other in GANCO.