Study On Soft Sensor Methods Of Quality Parameters For Distillation Process Based On Deep Learning

The petrochemical industry is an important pillar industry in China,and the distillation process is an important operating unit in the petrochemical production process.The optimal operation and advanced control of the distillation process will reduce the production energy consumption and improve the product quality.However,some important quality parameters in the distillation process can’t be directly measured.The study of the soft sensor technology provides a way to solve this problem.In this paper,from the angle of the soft sensor modeling method for quality parameters in the distillation process,we study the problems related to improving the precision of the soft sensor model for quality parameters in the distillation process.Firstly,we establish the dynamic mechanism model of the distillation process and simulate the continuous distillation process and the batch distillation process.According to the dynamic concentration curve of each tray component in the continuous distillation process and the concentration changing curve in steady state,the ratio of feed flow and reflux have a great influence on the industrial production of continuous distillation,which indicates the necessity of variable selection for the subsequent simulation modeling of the actual distillation process.According to the dynamic concentration curve of each tray component in the batch distillation process,it has better applicability to the situation that the variety group changes in some distillation processes because it controls the distillation process according to the reflux ratio.Secondly,we respectively study the basic principle of the three methods of machine learning,the BP neural network,the RBF neural network and the SVM algorithm.Then,we propose a multiple kernel least squares support vector machine algorithm(MKLS-SVM)in this paper,which combines the advantages of the multi-core learning algorithm.After using MSE,RMSE and goodness of fit as evaluation indexes to evaluate the simulation experiment performances of the four machine learning algorithms on the nonlinear function data set and the continuous distillation data set under the noise disturbance.The simulation results show that the MKLS-SVM algorithm is the best on the two data sets,and its performance is better than the other three algorithms.Finally,a soft sensing method based on LASSO and the PSO-DBN neural network algorithm is proposed to solve the problem of dry point soft sensing of aviation kerosene quality parameters in the atmospheric and vacuum distillation process with large noise disturbances.Firstly,we use the LASSO variable selection method to select auxiliary variables to eliminate the auxiliary variables that have little influence on the output variables from the original variables.Then,we use the DBN deep confidence network algorithm for simulation and the PSO algorithm is used to optimize the structural parameters of the DBN,and simultaneously use the BP neural network,the RBF neural network,the SVM algorithm,the MKLS-SVM algorithm and the DBN neural network five methods to carry out simulation experiments on the distillation nominal data set and the industrial distillation data set respectively.The results show that the soft sensor model based on LASSO and the PSO-DBN algorithm has higher accuracy,which provides a research foundation for the design of the optimal operation and advanced control of the distillation process.