Modeling Of Hydrocracking Reactor Based On Convolution Neural Network

Under the background that domestic economic and social development is constrained by resource and environment problems,and under the guidance of energy development goals formulated in the 13th Five-Year Plan,petrochemical enterprises are facing an important task of energy conservation and emission reduction.Hydrogen is the second largest cost factor after crude oil.Improving the utilization rate of hydrogen and reducing the consumption of hydrogen have become an important research topic.Based on the national major fund proj ect and in view of the demand of reducing hydrogen consumption in a petrochemical enterprise,this paper combines DCS data of the plant with in-depth learning,and carries out the research of new hydrogen flow forecasting and modeling in the form of modeling and simulation,which provides guidance for the intelligent scheduling of hydrogen gas network in the enterprise.In this paper,a new hydrogen flow prediction model based on Convolutional Neural Networks(CNN)is established by using Lasso method to select correlative variables and time series matching of variables with Cross-correlation Function(CCF).The validity of the model is verified by actual production data.Then a CNN model with self-tuning parameters is proposed to improve the adaptability and applicability of the model.The main contents of this paper are as follows:Firstly,in view of the problems of feature redundancy and timing mismatch in actual DCS data sets of factories,the data are preprocessed.Lasso method was used to screen the original features of the data set,and finally six correlative variables were selected.Then CCF was used to calculate the correlation and delay between the correlative variables and the new hydrogen flow.The correctness of the selection of variables was verified and the foundation for the sequential data rearrangement of subsequent CNN modeling was laid.Secondly,for hydrocracking unit,a black box model is established to predict the new hydrogen flow rate at the next moment.Based on the pre-processed data set,the new hydrogen flow prediction model of hydrocracking unit was established by using four machine learning methods:Multiple Linear Regression(MLR),Back Propagation Neural Network(BPNN),Recurrent Neural Network(RNN)and CNN.The experimental results show that CNN model has the best prediction effect,low prediction accuracy and good stability.RNN,as a commonly used time series data processing method,is second only to CNN in accuracy.BPNN prediction results lack stability and accuracy and MLR.Neither of them is suitable for the modeling of the device.Thirdly,a CNN model with self-tuning parameters is proposed to solve the problem of working condition change in actual production process.The working conditions are divided into stable and variable conditions,and a new performance evaluation index of the model is put forward.The data of stable conditions are divided according to prior knowledge to collect the performance index sequence.After determining the confidence interval,if the evaluation index exceeds the interval,the model parameters are corrected by Fine Tune method,which is verified on the actual data set containing variable conditions.Compared with the general CNN model,the root mean square error and the average absolute percentage error are reduced by 9.11%and 11.70%,the prediction effect of the model is significantly improved,which lays the foundation for the application of deep learning to the actual industrial production process.