Evolution of parameters in the Doyle-Fuller-Newman model of cycling lithium ion batteries by multi-objective optimization

An electrochemical model having high-fidelity parameters enables a correct description of the performance of lithium ion batteries. Hence, developing an accurate and high-efficiency parameter identification method is crucial to precisely predicting the battery's state of health. In the past, most studies on estimation of such parameters took the discharge voltage as the only target to be fitted. Here, the first-order derivative of the discharge curve, i.e., the dQ/dV curve, is proposed as another fitting target since this new curve is directly related to battery aging. Four different objective functions, associated with the discharge curve and its derivative curve, are used to perform the multi-objective optimization, where the two algorithms, namely the genetic algorithm and the deep neural network are employed. We show that, by using the genetic algorithms, the mean absolute errors of the discharge curve for each cycle are lower than 0.07 (V), while the errors of the ICA curve are below 0.32 (Ah/V), both of which show a good convergence. The deep neural network also leads to excellent result. We present the evolutions of 13 identified parameters and demonstrate that the initial lithium ion concentration in the negative electrode dominates the cycle age of the tested batteries.

Automated summary

Developing an accurate and high-efficiency parameter identification method is crucial for predicting the state of health of lithium ion batteries. This study proposes using the first-order derivative of the discharge curve as another fitting target and employs genetic algorithm and deep neural network for multi-objective optimization.