Comparison and integration of machine learning based ethylene cracking process models

Ethylene is an essential petrochemical industry product produced in a complex steam cracking process. Fast, accurate predictions of ethylene cracking depths depend on accurate naphtha cracking models. This paper compares three machine learning models based on a support vector regression (SVR), a k-nearest neighbor regression, and an extreme gradient boosting (XGBoost) to predict the ethylene cracking depth. Several industrial datasets are screened to identify the critical variables controlling the process using the density-based spatial clustering of applications with noise (DBSCAN) and a local abnormal factor detection algorithm. These three models are then trained and combined into an ensemble model to provide better predictions. The ensemble model combines the advantages of the three models and reduces the overfitting, the sensitivity to noise and other shortcomings. The ensemble model then has better prediction stability and generalization ability. The ensemble model predictions have R²=0.955 and an average absolute percentage error of about 0.23%, which is sufficient for process research and industrial applications.