Robust prediction and optimization of gasoline quality using data-driven adaptive modeling for a light naphtha isomerization reactor

Abstract

In this research, a data-driven adaptive model is developed to predict the variables indicating gasoline quality in the light naphtha isomerization process and determine the optimal conditions leading to improved gasoline quality. To this end, an integrated method based on double-level similarity criterion and support vector regression (DLS-SVR) is proposed. The variables that indicate gasoline quality are research octane number (RON), benzene volume percentage (BVP), and Reid vapor pressure (RVP). In addition to the influential operating variables of pressure, temperature, feed weight hourly space velocity (WHSV), and hydrogen to naphtha feed molar ratio, the model considers benzene's feed concentration and cycloparaffin content. Experiments are conducted using commercial Pt/Al2O3-CCl4 catalyst in a pilot-scale packed-bed reactor. The developed model's predictive performance and generalization ability are compared with the response surface methodology, support vector regression, and double-level locally weighted extreme learning machine through the fivefold cross-validation technique. The generalized DLS-SVR predicts gasoline's RON, BVP, and RVP with R2 = 0.901, 0.959, and 0.931 and RMSE = 0.055, 0.061, and 0.053, respectively, indicating that its performance is superior to alternative generalized models. The optimal conditions are computed using the DLS-SVR model and co-evolutionary particle swarm optimization algorithm (CPSO). The optimal operation of the reactor yielded a 6.78-unit increase in gasoline RON and a minimum BVP of 0.394 %. The results demonstrate that the proposed DLS-SVR model can accurately predict the variables indicating the quality of isomerate gasoline. © 2022 Elsevier Ltd