Multi-objective particle swarm optimization and training of datasheet-based load dependent lithium-ion voltage models

In this study, a datasheet-based lithium-ion (Li-ion) battery cell model is built, trained and validated by utilizing Multi-Objective Particle Swarm Optimization (MOPSO). Firstly, in order to determine the most suitable solver among various multi-objective solutions, a first-order (1RC) Thevenin model is adopted. At this stage, a thorough comparison is carried out among the Mean Squared Errors (MSE), Mean Signed (MSD) and Average Deviations (AD) of five study cases, where the proposed solver showed both up to 71%,96% and 54% improvements respectively compared to the previously reported optimal result, and capability to converge over the whole available capacity of the cells. Secondly, the optimal solver selection is followed by a comparative study among various equivalent circuit models (1RC, 2RC and 3RC), which are thoroughly evaluated upon computational time, model complexity and accuracy. Thirdly, the proposed optimized Thevenin Equivalent Circuit Model (ECM) is parameterized and compared to an experiment-based model of a commercial high-power Li-ion cell, for both charge and discharge processes. Outcomes indicate an improved performance of the proposed methodology which is ultimately utilized to train a dataset obtained from the manufacturer and successfully apply it on several unspecified current profiles. As a result, a wide range of accurate voltage-model outputs is acquired from a few input profiles, indicating the overall robustness of the method and rendering it as a strong candidate for non-dynamic loading battery applications.