Prediction of the Electrical Strength and Boiling Temperature of the Substitutes for Greenhouse Gas SF₆ Using Neural Network and Random Forest

Finding substitutes for sulfur hexafluoride (SF6), a gas with extremely high global warming potential, has been a persistent effort for years in the field of high voltage power equipment, which focuses on the evaluation of the electrical strength and boiling temperature for the practical purpose. Following up the previous proposed linear regression models, this work introduces machine learning algorithms including artificial neural network (ANN) and random forest (RF) as the potential approaches to predict the electrical strength and boiling temperature. Based on a series of descriptors derived from the molecular structure of 74 molecules, the performance of three different methods: multiple linear regression, artificial neural network and random forest are compared and assessed in terms of the sensitivity to the sample size, prediction accuracy and stability, and the interpretability of predictors. Considering the available data are limited, random forest shows superior performance with higher robustness and efficiency. The same approaches were applied to the boiling temperature and random forest produced better results as well. Besides, the variable importance ranked by RF improves understanding of the correlation between the molecular properties and electrical strength. It provides important insights to analyze the properties of the SF6 substitutes during the design and synthesis of the new eco-friendly gases in power equipment.