A novel zone-based machine learning approach for the prediction of the performance of industrial flares

Abstract

Industrial flares are used to burn off unwanted gas during operation. If not combusted completely, intermediate products or incomplete combustion products are formed, and they will cause significant environmental and health issues. The EPA Refinery Sector Rule emphasizes smokeless flaring with combustion efficiency (CE) ≥ 96.5% and destruction and removal efficiency (DRE) ≥ 98% for all types of flares in the refineries. In this research, a novel zone-based modeling approach was developed for predicting CE and Opacity of steam assist flares. Flare CE data were partitioned into two zones based on the partition of the carbon and hydrogen atomic ratio (CHR), then random forest (RF) and Catboost algorithms were used to develop CE predictive models, respectively. This CHR-based zone partition has a clear implication in engineering. It was also found out that no zone division for flare Opacity prediction is needed, and both RF and Catboost algorithms generated good prediction results. All the models match extremely well with all the original experimental data. These predictive models under the same zone-partition use either RF or Catboost algorithm can both give superior prediction accuracy. This demonstrates the simplicity, general applicability, and high reliability of the zone-based ML approach.