Data-efficient parameter identification of electrochemical lithium-ion battery model using deep Bayesian harmony search

Abstract

Lithium-ion batteries have been used in many applications owing to their high energy density and rechargeability. It is very important to monitor the internal physical parameters of the lithium-ion battery for safe and efficient usage, because this can help estimate the state of the battery, develop battery aging models, and schedule optimal operation of batteries. Parameter optimization methods using an accurate electrochemical battery model are much less expensive than direct parameter measurement methods, such as post-mortem methods. Thus, many model-based parameter optimization methods have been developed so far. However, most of these methods are random search methods that are based on heuristic rules, which leads to data-inefficient parameter identification. This means that they require many time-consuming battery model simulation runs to identify optimal parameters. Herein, a novel learning-based method is proposed for data-efficient parameter identification of lithium-ion batteries. A deep Bayesian neural network is used to efficiently identify optimal parameters. The simulations and experimental data validation show that the proposed method requires much fewer battery model simulation runs to identify optimal parameters than existing methods such as genetic algorithms, particle swarm optimization, and the Levenberg-Marquardt algorithm. The parameter estimation error of the proposed method is about 10 times lower than that of the second-best algorithm.