Virtual sensing of catalytic naphtha reforming process under uncertain feed conditions

Abstract

Stable operation of a naphtha reforming process is always desired to get high Research Octane Number (RON) of its end product, gasoline. Uncertainty in process variables, i.e. temperature, pressure, feed composition, makes the process unstable and result in lower RON of gasoline. Soft sensors (virtual sensors) have been used to get stable process operation. In this study, a soft sensor is developed through the ensemble learning method, i.e., boosting, for prediction of RON value of the naphtha reforming process. Prediction performance of the boosted model is compared with an Artificial Neural Networks (ANN) model; the boosted model outperformed the ANN model. For analyzing the effect of process uncertainty, sensitivity analysis is performed using Fourier amplitude sensitivity test (FAST) and Sobol technique. In addition, Polynomial Chaos Expansion (PCE) is used to analyze the collective effect of inputs uncertainty on the model output. The proposed methods of soft sensors development, sensitivity analysis and uncertainty analysis are validated through real process data of a petroleum refinery. The results are highly accurate and suitable for industrial applications.