Research On Soft Sensor Modelling Method Using Support Vector Regression(SVR)in Naphtha Cracking Processes

Ethylene is an important basic product in the chemical industry.China’s current annual ethylene capacity is more than 18.218 million tons,which plays an important role in the national economy.The use of naphtha as raw material for cracking to product ethylene production is one of the important ways.The process technology is mainly Steam-Cracking method which 95%of the total ethylene production in the world comes from Steam-Cracking,which is the most developed and widely used cracking technology.The ethylene cracking furnace is the main equipment to realize this technology,but the method exsit the shortcomings of high energy consumption and are easy be influenced by many factors,thus,need to determine the optimal reaction temperature to improve the product yield and product quality.Therefore,real-time measurement of the operating parameters of the ethylene cracking furnace will help to provide a guide for finding the optimal reaction operating parametersi Timely and accurate information of relative contents of key components of naphtha cracking products is an important basis for parameter regulationand yield optimization.There are certain delay from manual analysis in lab,while high equipment and maintenance costs limits the use of online analytical method.Soft sensoring,i.e.regression,based on the historical data is capable for timely perdiction.In recent years,support vector machine(SVM)based on the principle of structural risk minimization has been widely used,and the support vector regression method(SVR)for prediction is developed.It is compareing with the traditional linear learning machine method-Least Square Regression method(PLS)show better learning ability upon the highly nonlinear system such as chemical processes,futher more,because of the application of Support Vector(SV)and Structural Risk Minimization(SRM)concepts makes SVM has better generalization performance and computational efficiency than another widely used nonlinear machine learning method-Artificia Neural Network(ANN),which has a good application prospect.The SVR method was applied to the ethylene cracking production process example to realize real-time sensing of the production process variables of the ethylene cracking furnace.The application results show that the proposed SVR soft measurement method has excellent prediction results.The main research contents are summarized as follows:(1)Research on soft sensor methods,from soft sensor technology classification,several major problem challenges which data-driven soft sensor technology facing,soft sensor technology application fields and several types of typical soft sensor methods.(2)Based on statistical learning theory and support vector machine theory,researched the SVR method,including the latest improvement and application of S VR method,and establish S VR soft measurement model based on parameter optimization;further establish a Adaptive Dynamic Online Support Vector Regression(ADO-SVR)regression algorithm based on moving window technology.(3)In this study,applying the parameter-optimized SVR soft-sensor model established to the ethylene plant production process,establishing a soft-sensor model using the historical monitoring data of the ethylene cracking furnace.Specifically,it includes predicting the product composition of the ethylene cracking furnace and predicting the outlet temperatures of the ethylene cracking furnace tubes.Under the same conditions,in order to further verify the validity of the soft-sensor model established in this paper,the prediction results are compared with the PLS and MLP-ANN methods,especially to verify the prediction of SVR soft-sensor methods and other soft-measurement methods in the case of smaller learning samples.The results of the experiment indicate that the proposed method has better learning ability and generalization ability than the PLS and MLP-ANN methods in this case study,and the advantage is more obvious in small sample learning.The research results effectively realize the demand for soft sensor of process variables in the actual production of the factory,which is of great significance for optimizing the process and improving the economic benefits of the enterprise.