Interpretable machine learning modeling of temperature rise in a medium voltage switchgear using multiphysics CFD analysis

Abstract

In recent decades, leading companies and research groups have extensively conducted Multiphysics computational fluid dynamics (CFD) analyses to evaluate temperature rise in switchgear systems, aiming to meet type-testing requirements specified in IEC standards. However, the complex interaction of geometrical and operational parameters presents significant challenges in interpreting these methods. Artificial intelligence (AI) algorithms have gained attention in various engineering fields to address similar issues. This paper investigates the influence of four key manufacturing and operational parameters, both categorical and continuous, on temperature rise in a medium voltage (MV) switchgear case study. A CFD-based dataset was created from these parameters to target maximum temperature, facilitating the study's objective. Several models for temperature rise estimation, including extreme gradient boosting (XGBoost), support vector regression, decision tree, and random forest, were compared. An explainable artificial intelligence (XAI) technique, Shapley Additive Explanations (SHAP), was applied to the best-performing model to evaluate the importance of each feature in predicting maximum temperature. The results revealed that XGBoost provided the most accurate predictions, with a scatter band (SB) of ±1.01 and average R2 values of 99.98 % and 96.59 % for the training and testing sets, respectively. SHAP analysis identified the most significant variables affecting temperature prediction as current, air velocity, duct area, and switchgear conditions, in that order.