Unveiling the performance impact of module level features on parallel-connected lithium-ion cells via explainable machine learning techniques on a full factorial design of experiments
Gabriele Piombo,Simone Fasolato,Robert Heymer,Marc Hidalgo,Mona Faraji Niri,Simona Onori,James Marco
DOI: https://doi.org/10.1016/j.est.2024.110783
IF: 9.4
2024-02-16
Journal of Energy Storage
Abstract:Parallel string performance imbalances are unavoidable due to manufacturing-related cell-to-cell inhomogeneities (e.g. capacity, internal resistance), suboptimal pack and cooling system design. Understanding the most important features at a single cell and module level influencing the heterogeneity propagation inside the modules/packs is therefore crucial to limiting the phenomenon. In this article, a methodology combining well-established non-invasive single-cell characterisation tests with data-driven modelling tools is proposed. Two batches of twenty new, commercial NMC and NCA cells are first characterised to identify out-of-manufacture internal resistance and capacity distributions. Then, a 54 test condition full-factorial Design of Experiment campaign on four cells ladder-parallel connected modules is performed. The experiments inform about how the cells' current, State of Charge, temperature distributions and time to self-balance under 0.75C constant current discharge loads are affected by interconnection resistance, operating temperature, different chemistry combination and ageing. The multivariate linear model analysis confirms that combining NMC and NCA cells in parallel is possible both for first and second life applications. Nevertheless, mixing different chemistries and including an aged cell show a detrimental effect on the balanced performance of the module. The application of Explainable Machine Learning techniques such as SHAP, Partial Dependence Plots and Individual Conditional Expectation closes the gap between data-driven models' interpretability against traditional black box models while maintaining the advantage of capturing highly non-linear control-response relationships. According to the results, the interconnection resistance is the most relevant contributor to heterogeneous performance within the string. In the first and middle phases of the discharge, the distributions of internal resistance and capacity impact the load imbalance across the cells, respectively. Increasing the operating temperature contributes to exacerbate the thermal gradient in the string.
energy & fuels