Abstract

A novel data-driven, soft sensor based on support vector regression (SVR) integrated with a data compression technique was developed to predict the product quality for the hydrodesulfurization (HDS) process. A wide range of experimental data was taken from a HDS setup to train and test the SVR model. Hyper-parameter tuning is one of the main challenges to improve predictive accuracy of the SVR model. Therefore, a hybrid approach using a combination of genetic algorithm (GA) and sequential quadratic programming (SQP) methods (GA–SQP) was developed. Performance of different optimization algorithms including GA–SQP, GA, pattern search (PS), and grid search (GS) indicated that the best average absolute relative error (AARE), squared correlation coefficient (R 2), and computation time (CT) (AARE = 0.0745, R 2 = 0.997 and CT = 56 s) was accomplished by the hybrid algorithm. Moreover, to reduce the CT and improve the accuracy of the SVR model, the vector quantization (VQ) technique was used. The results also showed that the VQ technique can decrease the training time and improve prediction performance of the SVR model. The proposed method can provide a robust, soft sensor in a wide range of sulfur contents with good accuracy.

Highlights

  • Sulfur compounds are one of the most important impurities in crude oil and various petroleum fractions

  • The results showed that the vector quantization (VQ) technique can decrease the training time and improve prediction performance of the support vector regression (SVR) model

  • Note that when catalyst deactivation occurs during the time, the outlet temperature would change for the same input conditions

Read more

Summary

Introduction

Sulfur compounds are one of the most important impurities in crude oil and various petroleum fractions. Online determination of sulfur concentration in the end product is difficult or impossible due to the limitations in process technology and measurement techniques. This index as the key indicator of process performance is normally determined by offline sample analysis in laboratories or online hardware analyzers that are mostly expensive with high maintenance costs. Data-driven models are based on the data taken from the processing plants, and describe the real process conditions (Kadlec et al 2009, 2011). These data-driven models can be developed more quickly with less expense.

Objectives
Results
Conclusion
Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call