Study On Soft Sensor Modeling And Parameter Optimization Of Bisphenol A Production Process

As a major research field in process control theory,soft-sensor modeling technique has been widely used in industrial processes.Since the 21st century,industrial production data has received more and more attention,become an important resource.Mastery and utilization ability of data information resources has also become a significant symbol of competitiveness for a company.With extensive application of intelligent instruments in large-scale complex processes of new industries or in the processes of transforming and upgrading traditional industries,huge amount of historical data have been generated and stored in industrial processes.These data can be reasonably employed to establish soft-sensing model for estimating the important but difficult to be online measured process variables and further to realize online monitoring and optimization of the production process,thus achieving green production,energy conservation,consumption reduction and product quality improvement.As an indispensable chemical raw material for many kinds of polymer materials and fine chemical products,Bisphenol A（BPA）is closely related to people’s lives,research on soft-sensor modeling and optimization algorithm for BPA production process can provide much larger competitive advantage and development space for domestic BPA manufacturing.Therefore,this thesis focuses on studies of BPA production process modeling,parameter optimization and model fusion,the detailed research contents are summarized as follows:（1）In order to improve the estimation accuracy of the soft sensor model using relevance vector machine（RVM）,a method based on quantum particle swarm optimization（QPSO）algorithm to optimize the parameters of RVM kernel function is proposed.The proposed method has been applied to establish soft sensor models to estimate the electricity output of a power plant and the BPA concentration of the cracking unite in a BPA production process,respectively.The practical application results demonstrate that the estimation accuracy has been significantly improved by using the optimized RVM model,which verifies the effectiveness of the proposed method.（2）Aiming at problems that the clustering results of K-means are unsatisfactory due to selection uncertainty of initial clustering centers and the density peak clustering algorithm cannot effectively separate the category with small density,a multi-model soft sensor modeling method based on the combined density peak and K-means clustering is proposed.First,the fixed initial clustering centers of the training data are determined by using the density peak clustering algorithm,and a random initial clustering center is combined to serve as the initial clustering centers of the K-means algorithm for separation of the category with small density;second,regression sub-models are established based on the samples of each category by using the random forest regression（RFR）method;finally,the switching method is utilized for multi-model fusion.The proposed method is applied to establish soft sensor models of a sulfur recovery process and a reaction dehydration unit in a bisphenol A（BPA）plant,for estimating the contents of H₂S and BPA,respectively.The application results illustrate that the multi-model soft sensor modeling method based on the combined density peak and K-means clustering can effectively improve the estimation accuracy of the model.（3）To resolve the problems of multi-model fusion modeling method involved with soft sensor development in industrial processes,such as poor fusion performance,low sub-model utilization rate and model prediction accuracy,this thesis puts forward a soft-sensor modeling method based on multi-model fusion and deviation compensation using the least squares support vector machine（LSSVM）algorithm.Firstly,multiple sub-models are established by using DP&K-means clustering method and random forest regression method.Then,the LSSVM method is applied for fusion of the sub-model outputs to obtain the multi-model output.Finally,a compensation model based on the LSSVM algorithm is established to further calibrate the multi-model output.The proposed multi-model modeling method is applied to the soft-sensor modeling of a butane distillation process,as well as the reaction dehydration unit and crystallization unit in a BPA production process,respectively.The simulation results show that the fusion performance and estimation accuracy are substantially improved,demonstrating the effectiveness of the proposed method.