Research On Dynamic Modeling Of Wax Oil Hydrogenation Unit Based On LSTM And Construction Of WEB Platform

The rough way of economic growth in our country has caused many problems of energy and environmental protection,and currently,the state attaches great importance to these problems.This puts forward new requirements for the production and development of various industries,especially the oil refining and petrochemical industry.Hydrogen as a very valuable clean energy,it is not only an important raw material in petroleum refining and petrochemical hydrogenation,but also one of the by-products.With the aggravating trend of heavier and deteriorated of the crude oil,and the improvement of the national standard for the quality of refined oil,refinery and petrochemical enterprises need to consume more hydrogen for heavy oil cracking and desulfurization and denitrification,which has led to a sharp increase in the cost of hydrogen.And the cost of hydrogen has now become the second cost factor of the enterprise.How to reduce the cost of hydrogen has become a major research topic.Hydrogenation units as major hydrogen consumption member,modeling of them can effectively monitor the use of hydrogen,and has guiding significance for the regulation of the hydrogen production system and hydrogen supply system.The wax oil hydrogenation unit,one of the hydrogenation units,is analyzed and modeled in this article.New hydrogen flow prediction model based on LSTM(Long-Short Term Memory)for wax oil hydrogenation unit is established,and to improve this model a new model is proposed--CC-LASSO-LSTM.And the validity of the model is verified on the actual data set of the refinery.Aiming at the applicability of the model under variable working conditions,a new modeling method-adaptive CC-LASSO-LSTM is presented.As the practical requirements for the model,a new hydrogen flow prediction platform based on WEB is established.There are several aspects of the main research content of this study:1.For prediction of new hydrogen flow rate for wax oil hydrogenation unit,the wax oil hydrogenation unit,models which based on data are established to predict the new hydrogen flow rate in next time.Based on the field data set of oil refining enterprises,new hydrogen flow prediction models of wax oil hydrogenation unit based on SVR(Support Vector Regression),BPNN(Back Propagation Natural Network),RNN(Recurrent Neutral Network)and LSTM are established respectively.The experimental results show that the LSTM model has the best prediction effect because of its advantage in processing time series data,the RMSEP(Root Mean Square Error of Prediction)is 427.855 Nm3/h and the MAPE(Mean Absolute Percentage Error)is 1.281%,and CNN model comes next.BPNN and SVR are not suitable for this problem because of poor stability and low prediction accuracy.2.Considering the characteristics selected by operators may be redundancy and collinearity,a new modeling method-CC-LASSO-LSTM is put forward,which uses two step feature selection strategy,and finally 6 characteristics are selected.Then using LSTM to model them,the RMSEP and MAPE of prediction is 401.182 Nm3/h and 1.233%,which comparing LSTM model reduces by 6.000%and 3.747%respectively.3.Aiming at the problem of inadequate applicability of CC-LASSO-LSTM model under variable conditions,a new modeling method with adaptive mechanism is proposed-adaptive CC-LASSO-LSTM,and to verifying the effectiveness of the new method on an actual data set.Under the condition that the working condition is constant,the prediction effect of the adaptive CC-LASSO-LSTM model is slightly higher than that of CC-LASSO-LSTM model.In the case of variable conditions,the model prediction effect is greatly improved,the RMSEP is 412.62 Nm3/h and the MAPE is 1.535%,comparing CC-LASSO-LSTM model reduces by 15.165%and 14.405%.3.As the platformization requirement of refineries,the new hydrogen flow prediction platform based on WEB is established.Four major modules-login verification module,data display module,new hydrogen flow prediction module and message notification module are integrated in this WEB platform.For prediction,a variety of models can be used,such as LSTM,RNN,SVR,etc.