Machine-Learning-Based Soft Sensors for LPG Distillation Columns Quality Monitoring: Comparing Strategies for Zero-Inflated Data

Refineries execute a series of interlinked processes. Given the output of a certain unit serves as the input of another one, the failures within these processes affect the quality of the end products and the revenue of the entire refinery. The implementation of realtime quality monitoring helps with accurate decision-making regarding process operations. In this paper, multiple machine-learning models aimed at forecasting the amount of ethylene (C2) and pentane (C5) content in the final product mixture are compared. In particular, strategies to deal with zero-inflated data are explored. The presence of zeroes in the dependent variable usually affects the models' learning and may result in poor performance. Furthermore, the presence of zeroes also affects the metrics used to compute the regressor's prediction quality, making it hard to assess their quality. The research results show a two-fold machine learning model leads to the best outcomes. The approach was validated with a real-world use case from a refinery.