Soft Sensing And Real-time Control Of Crude Products Properties Based On NARX Neural Network

As the first process of refinery process,crude distillation affects the quality,yield,and energy consumption of the entire refinery process.In order to enhance the efficiency and competitiveness of petrochemical enterprises,the companies are vigorously promoting the advanced control technologies for crude ditillation unit,which can efficiently increase the production yields,improve the stability of the crude distillation process and reduce the resources waste with the products quality guaranteed.Usually,the real-time data of key properties of products are needed in the implementation of advanced control,however not only the online meters but also the offline assay analysis cannot meet the basic requirements of real-time measurement.Model identification technology,known for neural network,can establish the soft sensor to timely forecast the required products properties data,which lays the foundation for the implementation of advanced control strategy.In this paper,NARX neural network is used to establish the soft sensor of products properties,and the method for correction of the bad points in prediction is proposed.On this basis,the NARX neural network linearization control is used to realize the real-time control of products properties.The main works in this paper are below:(1)The research background and significance of crude properties soft sensing and real-time control are introduced.From the perspective of nonlinear model and neural network,it introduces the development and application of NARX neural network;(2)The steady state and dynamic models of the pre-distillation tower and atmospheric tower in the crude distillation are established as the replacement of actual plant in the follow-up works.The modeling process on Aspen plus and Aspen Dynamics is introduced in detail and the simulation results are consistent with the plant data by adjusting the parameters of the models.In the dynamic model,basic control loops are implemented for subsequent experiments;(3)A new kind of soft sensor based on NARX neural network is proposed,which has a better performance than traditional closed-loop NARX network in prediction.The multi-step prediction correction algorithm can effectively remove the bad points with big errors in prediction,and significantly reduce the mean square error of prediction results.Validation experiments were carried out on the combined simulation platform of Simulink and Aspen Dynamics;(4)NARX neural network linearization control strategy is implemented for real-time control of the crude products quality.The saturation modified correction method is demonstrated to reduce the system residual in controlling process.The simulation results of three different situations show that the correction method can effectively decrease the system residual,and the varying parameter method can reduce the settling time of the system.