Research On Soft Sensing Of Carbon Deposition In DMTO Process Catalyst Based On Data-Driven

In the Methanol to Olefins(DMTO)process developed by Dalian Institute of Chemical Physics,the parameter of coke deposition on catalysts is extremely important.It not only affects the catalyst activity and life,which in turn affects production stability and product quality,but also affects the pore size of the catalyst,thereby affecting the yield distribution of ethylene and propylene.Therefore,fast and accurate measurement of the coke deposition parameter on catalysts in DMTO process is of great significance.Based on the existing research on the DMTO process,this paper takes the reaction-regeneration system of the process as the research object and carries out soft sensing modeling of the coke deposition parameter of the key process parameter,the catalyst,on this process.For the DMTO process,which is characterized by strong nonlinearity,high dimensionality,redundancy,and complexity,this paper builds three soft sensing models of the coke deposition parameter of the process based on the historical data-driven method from two machine learning modeling perspectives.These models achieve fast and accurate measurement of the coke deposition parameter on catalysts in the process and provide guidance for subsequent production control optimization.Specifically as follows:(1)Starting from a global modeling perspective,the standard learning method was adopted for modeling.To deal with the high-dimensional features of the DMTO process,the normalized mutual information(NMI)method was used to select and reduce the input feature variables.For the strong nonlinear and redundant complex features of the DMTO process,the ensemble learning method,Gradient Boosting Regression Tree(GBR),was used.Multiple weak learners were used to learn the internal hidden information of the input mapping to the output,and a strong learner was formed to achieve accurate measurement of the target parameters.The NMI-GBR model was combined and a NMI-GBR soft measurement global model framework for the catalyst carbon deposition parameters of the DMTO process was established.To verify the effectiveness of the model,the production record data of a DMTO unit with an annual processing capacity of1.8 million tons of methanol was used for modeling,validation,and evaluation over the past four years.The experimental results show that in the prediction of the catalyst carbon deposition parameters on the test set,the key indicators MAPE,Max AE,and R~2 of the NMI-GBR soft measurement model were 1.4752%,0.4465,and 0.9181,respectively,and all indicators were significantly better than those of traditional soft measurement methods such as Bayesian Ridge Regression(BR),Support Vector Regression(SVR),and Linear Regression(LR).In addition,after optimizing the model training,the NMI-GBR method can also provide the importance of the returned feature variables.Combined with process mechanism knowledge,it can provide some direction for subsequent production regulation.(2)Starting from the local modeling perspective to further improve the adaptability and interpretability of the NMI-GBR model in the DMTO process catalyst carbon deposition parameter,this paper proposes an instantaneous learning soft measurement model EDBR-JITL based on Euclidean distance Bayesian Ridge regression.The EDBR-JITL model is based on the idea of generating similar outputs from similar inputs,and uses the Euclidean distance as a similarity measure to search for similar samples in the historical data as a temporary local data set for each query sample.A linear model Bayesian Ridge regression is used to establish the mapping relationship from the input to the output in the temporary local data set.Then,the target parameter of the query sample is measured accurately according to the mapping relationship,and finally,the query sample is included in the historical database for updating.In addition,the simple linear model is used to enhance the interpretability of the local modeling,and each sample is modeled and updated separately to enhance the adaptability of the model.Similarly,using production data for experimental verification,the results show that in the prediction of the catalyst carbon deposition parameter of the test set,the key error indicators MAPE,Max AE,R~2,and CT of the EDBR-JITL soft measurement model are 1.406%,0.6209,0.9219,and 22.4427 ms,respectively,which are better than those of the traditional soft measurement methods Bayesian ridge regression,support vector regression,partial least squares regression,and their respective models combined with instantaneous learning methods.The key error indicator MAPE of EDBR-JITL model is reduced by 4.6909%,and the R~2 is increased by 0.4139%compared to the NMI-GBR model.In addition,the EDBR-JITL model can provide a linear weight formula for calculating the catalyst carbon deposition parameter value from each input parameter,which can explain the specific impact(positive/negative impact and weight size)of the input feature parameters on the soft measurement results and provide guidance for subsequent production optimization.(3)To further improve the feasibility,robustness,and versatility of the EDBR-JITL model for catalyst carbon deposition parameter in DMTO process,improvements were made in the construction of the temporary local dataset and historical database update parts based on the EDBR-JITL model,proposing the dynamic adaptive window strategy and historical database adaptive update strategy,and building the DAW-BJITL catalyst carbon deposition parameter soft measurement framework.In terms of constructing the temporary local database,the DAW-BJITL model dynamically sets the lower limit of similarity and the adaptive local window size for each query sample at each time,avoiding problems such as zero samples and introduction of unrelated samples when modeling edge-distributed sample data in actual production,while solving the problems of large computation and time consumption for concentrated-distributed sample data.In terms of historical database update,the DAW-BJITL model adaptively identifies and classifies new samples,reducing the management and computational burden of the historical database.Similarly,using production data for experimental verification,the results show that in the prediction of the catalyst carbon deposition parameter for the test set,the key indicators of the DAW-BJITL soft measurement model,including MAPE,Max AE,R~2,and CT,are 1.362%,0.6209,0.9249,and 21.0806 ms,respectively,which are better than the similarity strategy and fixed window strategy,with MAPE reduced by 2.2254%and3.1294%,respectively;in particular,for the samples with adaptive windows,the DAW-BJITL soft measurement model has MAPE indicators reduced by 14.4%and 18.2%compared to the similarity strategy and fixed window strategy,and R~2 indicators increased by 2.35%and 2.14%,respectively.In summary,the NMI-GBR soft sensing model is suitable for stable production processes with accumulated historical data,as it requires only one-time modeling and has excellent real-time performance.The EDBR-JITL model is adaptable and highly interpretable,making it suitable for production processes with multiple modes.The DAW-BJITL model not only improves the accuracy of predicting catalyst carbon deposition parameters but also has stronger feasibility,robustness,and generalization,making it suitable for a wider range of production processes.