Machine learning modeling for optimization of sulfur compounds separation from fuels: Process optimization for reduction of environmental pollution

Abstract

Development of theoretical models for reduction of sulfur emission and also the material consumption is of great importance for petroleum refinery to obtain high-quality fuels. The latter can be done by employing advanced optimization techniques. In this study, we have developed a modeling methodology for optimization of petroleum refinery in fuel production. Some measured data have been collected for the modeling and computational optimization. Each data point is comprised of four input characteristics: reactor pressure (bar), reactor temperature (°C), initial sulfur concentration (ppm), and dose (g). Outputs for the modeling include sulfur concentration (ppm), emission (%), and HDS cost ($). For modeling, the Adaboost ensemble model is applied on top of the three fundamental models Linear Regression, Gaussian Process Regression, and Bayesian Ridge. On the available dataset, the models are tweaked using the grasshopper optimization algorithm (GOA) method, and then the optimal combination of models and parameters is selected for each output. For sulfur content and emission characteristics, the ADA-GPR model is the most accurate; however, the ADA-BRR algorithm performs the best for calculating the HDS cost. Using these models, the R2-score for outputs is 0.970, 0.950, and 0.999, respectively for sulfur concentration, emission percentage of SO2, and HDS cost.

Keywords

• Petroleum refinery
• SO2 removal
• Clean fuel
• Separation
• Machine learning