A comparative study of data-driven thermal fault prediction using machine learning algorithms in air-cooled cylindrical Li-ion battery modules

Failure to mitigate thermal runaway (TR) early can result in TR propagation due to cell-to-cell heat interactions, resulting in module/pack fires. In this study, a comparative investigation is conducted on the TR cell position prediction performance by different machine learning (ML) algorithms for 32-cell cylindrical air-cooled LiB modules in aligned, staggered, and cross arrangement cells. The ML models are trained on temperature data from the sensors optimized using the Pearson Correlation Coefficient approach with a threshold of 0.85. The air temperature data in the battery module under different operating and faulty conditions were generated from experimentally validated numerical models and recorded by the sensors initially dispersed in single mid-plane and multiple arbitrary planes of the battery domain. The base ML algorithms chosen for the study comprise the k-Nearest Neighbors, Random forest, Gradient boosting, and Long short-term memory classification algorithms. The developed models are further subjected to 5-fold cross-validation and external testing using random test cases and subsequently compared for prediction accuracy with the error metrics, training, and prediction times. A Stacked Ensemble learning model is built and tested for accuracy based on the base models to improve the overall predictive accuracy. The study concludes that the RF model outperformed other base models owing to its 100% accuracy across all cell arrangements, high consistency across validation folds, and the lowest training and prediction time of 22 s and 0.59 s, respectively. The study identifies the best-fit ML model for early fault detection and preventing catastrophic accidents.