Efficiency and Optimality in Electrochemical Battery Model Parameter Identification: A Comparative Study of Estimation Techniques

TL;DR

The paper tackles the challenge of parameter identification in electrochemical battery models where many parameters are not directly measurable. It compares Least Squares (LS), Particle Swarm Optimization (PSO), and Genetic Algorithm (GA) using a discretized Single Particle Model (SPM) with parameter grouping to reduce estimation dimensionality, and datasets generated from real-battery parameters for fitting and validation. The results show that PSO delivers the highest accuracy and stability across validation data, while LS is fastest but highly sensitive to initial guesses and better for fine-tuning with prior knowledge; GA is comparatively slower and less accurate. These findings provide practical guidance for selecting parameter-identification methods in battery modeling and suggest future work on parameter grouping and sensitivity analysis to further reduce estimation effort.

Abstract

Parameter identification for electrochemical battery models has always been challenging due to the multitude of parameters involved, most of which cannot be directly measured. This paper evaluates the efficiency and optimality of three widely-used parameter identification methods for electrochemical battery models: Least Squares Method (LS), Particle Swarm Optimization (PSO), and Genetic Algorithm (GA). Therefore, a Single Particle Model (SPM) of a battery was developed and discretized. Battery parameter grouping was then performed to reduce the number of parameters required. Using a set of parameters previously identified from a real battery as a benchmark, we generated fitting and validation datasets to assess the methods' runtime and accuracy. The comparative analysis reveals that PSO outperforms the other methods in terms of accuracy and stability, making it highly effective for parameter identification when there is no prior knowledge of the battery's internal parameters. In contrast, LS is better suited for minor adjustments in parameters, particularly for aging batteries, whereas GA lags behind in both computational efficiency and optimality with respect to PSO.

Figures (6)

• Figure 1: Distribution of RMSE for the TRR algorithm on the fitting dataset
• Figure 2: Distribution of RMSE for the TRR algorithm on the validation dataset
• Figure 3: Distribution of RMSE for the PSO algorithm on the fitting dataset
• Figure 4: Distribution of RMSE for the PSO algorithm on the validation dataset
• Figure 5: Distribution of RMSE for the GA algorithm on the fitting dataset