Modeling Of Residue Hydrogenation Unit Based On Deep Recurrent Neural Network

As one of the important fossil energy sources,petroleum is the most important pillar of the national economy.However,when it comes to the energy consumption of petroleum,we have to face the severe situation of enormous energy consumption and low energy utilization.Under the strategic plan of National Petroleum "13th Five-Year Plan",oil refineries are in urgent dual needs of economic and environmental benefits This paper takes the Major projects of "Experimental platform and application verification for refinery plant-wide optimization" as an opportunity to explore the industrial application of deep learning technology by using the process data of residue hydrotreating unit,and establish a mobile platform for production monitoring to help grassroots field staff monitor the production process and indicators in time.In this paper,a prediction model for fresh hydrogen flow forecasting is established,and the following research work is carried out:(1)In order to accurately predict the new hydrogen flow rate needed for the operation of the plant,a fresh hydrogen flow prediction model is established based on the Long Short Term Memory(LSTM)and Gated Recurrent Unit(GRU).After data preprocessing,feature selection,modeling,validation and analysis,the superiority and validity of recurrent neural network structures are verified in regression task in industrial process with complex mechanism and huge time-delay.(2)A GRU prediction model with deep structure(Deep-GRU prediction model),is established.Compared with shallow GRU prediction model,Deep GRU prediction model has better performance in mean absolute percentage error(MAPE),root mean square error(RMSE)and accurate prediction rate(Ace Rate)on the same parameter conditions and data sets.The deep structure model has better ability to extract and express the potential characteristics of industrial processes with complex mechanism,various variables and time series,which can be extended to all similar modeling tasks in the field of petroleum refining and chemical industry.(3)To ease the difficulty of training the framework of deep GRU structure,together with the common characteristics of industrial data,a semi-supervised learning strategy based on deep GRU model(SSDGRU)is proposed with the help of auto-encoder.And the SSDGRU modeling process is described in detail.Finally,a prediction model for fresh hydrogen flow forecasting based on CC-LASSO-SSGRU is established with residue hydrotreating data,which verifies the superiority of semi-supervised lea:rning strategy in improving the training speed of deep model and the stability of accurate precision prediction.(4)With the information of previous research and analysis of refinery enterprises'demand for real-time production monitoring,a mobile production monitoring platform implementation scheme is formed based on the three-layer architecture of data layer,application layer and user layer.Mobile production monitoring platform is mainly focus on functions of real-time data presentation and abnormal alarm.It integrates real-time data display module,model prediction module and alarming module.Achieving transparency and information management of production data,the Mobile production monitoring platform supports operators to view the real-time modeling results and abnormal information on their mobile devices.