Research On Online Estimation Methods Of Product Quality Index In Chemical Process

Advanced Control is widely used in chemical process control for its effectiveness. Monitoring the product quality index is required for advanced control, such as product composition, gas content, etc. However, limited to some factors like measuring condition and the cost to purchase and maintenance of instruments, it's not always possible to measure these variables through on-line analyzers. Therefore, online estimation is in need for certain variables.As a problem of modeling, the characteristic of online estimation of product quality index for chemical process is: the process condition is always varying and the sample size for training is often small. In this paper, the solution for'online estimation of product quality index of atmospheric column'is studied. The property of the crude oil varies, which causes model mismatch and the decrease of prediction precision. First, the situation that crude oil varies is simulated through using flowsheeting software'Hysys', and the validity of building predictor after classification is studied. Then the research is carried for the real condition and the solution is proposed for the more complicated industry data, and the main work focused on the following two aspects.1 For the problem of varying crude oil, the classification and cluster methods is proposed. Different model is built respectively for similar crude oil, and the proper predictor is online selected. For classification, several classifiers are compared, including RBF-Classfier, SVM-Classfier and AdaBoost. The best classification accuracy is approached by AdaBoost, with the best prediction precision, which decreased the prediction error by 17.8%. Considering the class labels are not available in some condition, clustering is used and the prediction error is decreased by 12.8%. The experiment indicates that for problems like the varying crude oil, classification or clustering before building the preditor is very useful to improving the prediction precision. 2 For the problem of the small size sample usually occurring in chemical process, especially in the the subclass after classification or clustering, the solution for small size sample modeling is applied. According to the statistical learning theory, the SVR (Supported Vector Regression) which has powerful generality is choesd with proper kernel function, and the prediction precision is improved for 19.1%. As the PLS model is found'unstable'in small size sample through experiment, this is just accord with the requirement of Bagging. As a result, Bagging improves the predition precision for 9.9%. SVR and Bagging provide the possible solution for the small size sample modeling in chemical process.