Comparative data-driven prediction of thermal performance of cylindrical lithium-ion battery packs

Abstract

Effective thermal management is essential for ensuring the safety, longevity, and performance of lithium-ion batteries used in electric vehicles. While experimental studies and computational fluid dynamics simulations provide highly accurate results, their application in real-time prediction is limited by high computational costs. In this study, data-driven machine learning methods, which are computationally efficient forecasting tools, are used with a dataset obtained from an experimental study of a cylindrical lithium-ion battery thermal management system with an aligned 7 & times; 2 configuration. This dataset consists of 230 data points covering a wide range of working conditions, including Reynolds numbers from 988 to 4700, inlet air temperatures from 20 degrees C to 40 degrees C, and applied heat fluxes from 0.2 to 11.0 kW/m2. To predict the average wall temperature and Nusselt number, different machine learning algorithms such as support vector regression, random forest, extreme gradient boost, and multilayer perceptron are employed. The results show that support vector regression exhibits the most consistent performance for average wall temperature, which has a relatively linear profile, with a coefficient of determination value of 0.9979 and a mean absolute percentage error of 1.2434%, and a root mean square error of 0.6408 degrees C. On the other hand, the random forest model presents the most reliable predictions (with a mean absolute percentage error of 3.6457%, a coefficient of determination of 0.9855, and a root mean square error of 5.3598) for Nusselt number, which has a regime-dependent nonlinear heat transfer behavior. For the random forest and support vector regression models, cross-validation standard deviations of 0.0083 and 0.0005, respectively, confirm their generalization capability and consistency. The choice of a specific algorithm ensures that these models effectively capture fundamental heat transfer behavior without empirical correlations. The results show that the target variable's physical characteristics specify the selection of an optimal algorithm. Unlike previous studies that typically compare algorithms on a single output variable, this study establishes output-specific model selection principles. Also, it provides a computationally efficient alternative to conventional numerical and experimental approaches within the examined operating range.