Recurrent Neural Network Modeling And Swarm Intelligence Scheduling Of Refinery Hydrogen System

Intelligent Factory and Industry 4.0 have become a strategic direction for the future development.The optimization of production scheduling problem is one of the important link to establish intelligent factory,which is the bond between the production management and process control.In petrochemical industry,on the one hand,because the ratio of processing heavy crude and sour crude is increasing,environmental protection laws continuously strengthen,production quality requirements are more and more high,on the other hand,the heavy oil market gradually atrophic,light oil market continues to grow,which led to wide application of deep processing technology such as the hydrocracking,hydrotreating crude.Oil processing depth and hydrogenation rate increased.Hydrogen consumption of refineries is increasing,hydrogen cost has become the second to the cost of crude oil in refineries cost.Therefore,how to efficiently use of hydrogen resources,meet the demand of hydrogen at the same time reduce the cost of hydrogen system has the vital significance to improve the comprehensive benefit of refineries.The optimization of refinery hydrogen system including structure design optimization and scheduling optimization.This paper study scheduling optimization of refinery hydrogen system,and establish a mixed integer nonlinear programming(MINLP)model of refinery hydrogen system by using a hybrid modeling approach.The hybrid modeling approach merge Data-driven Modeling based on the Long Short-Term Memory with Mechanism Modeling based on material conservation.Then this paper puts forward a solution of swarm intelligence with linear programming algorithm to solve MINLP model.The effectiveness of the proposed algorithm is verified by standard test functions.Verifying the feasibility and effectiveness of the proposed model and strategy in this paper by simulation study of refinery.The main contents of this paper are summarized below:1)Establish a prediction model for hydrogen consumption in hydrogen system based on long short-term memory data-driven modeling.Because hydrogen trap device and reaction process is complex,the establishment of the mechanism of the model need much chemical reaction mechanism and the accumulation of experience.And huge numbers of parameters,uncertainty and disturbance,we use RNN to establish data-driven predictive model based on LSTM.According to the data of hydrogen consumption,we make a one-step prediction and time series multi-step prediction,the results showed that the average prediction error of one-step forecast around 1.5%,and within 2%for multi-step prediction.Compared with the mechanism model,it simplify a lot of work and the predicted results are more accurate.2)Establish a MINLP model of hydrogen scheduling optimization for the entire hydrogen system,using a hybrid modeling approach merging data-driven modeling based on the long short-term memory with mechanism modeling based on material conservation.The total hydrogen system cost within a period of time sequence in the future as the objective function,including source of hydrogen production cost,compressor power costs,tail hydrogen combustion efficiency,hydrogen gas pipe network abnormal capacity penalty term,hydrogen source replacement penalty term and compressor start and stop penalty term.Constraints are considered the hydrogen source,hydrogen trap,purification device,gas pipe network,compressor,which include hydrogen material conservation,hydrogen flow conservation,plant process or production capacity constraints,device state,etc.3)Propose a new swarm intelligence with linear programming algorithm to solve the complex MINLP problems.The algorithm includes the outer layer using an improved algorithm of particle swarm algorithm and inner layer using the simplex algorithm module,outer module will be converted to LP problems to the inner module,the inner module will result feedback to outer module to update problem.In outer algorithm module,dual fitness function filter particle not meet the constraints,select the optimal particle and enhance boundary searching ability.Introduce particle mutation probability and dual mutation strategy to prevent the swarm from falling into local optimum and enhancing the ability of local optimization.The fitness function introduces robust optimization to eliminate sharp optimal value.The algorithm is applied to the standard MINLP function,and the results show that the algorithm is superior to other algorithms and proves the effectiveness of the algorithm.