Forecasting Methanol-to-olefins Product Yields based on Relevance Vector Machine with Hybrid Kernel and Rolling-windows

Abstract

Listen

Translate article

Light olefins (ethylene and propylene) have become prominent in chemical industries. Forecasting of the yields of light olefins plays a crucial role in monitoring and optimizing the Methanol-to-olefins (MTO) process. In this work, we introduce an approach for forecasting the yields of ethylene and propylene in the MTO process with the Relevance Vector Machine (RVM) model, which is uniquely enhanced with hybrid kernels and a rolling window methodology. Through an in-depth analysis of 32 independent variables and their pairwise differences, our research pinpoints temperature and pressure as the most critical factors influencing the yields of ethylene and propylene, respectively. The model showcases satisfactory predictive accuracy and reasonable interpretability compared with the traditional statistical and popular machine learning models, marking a step forward in the predictive modeling of chemical engineering processes.