Study On Modeling Method Of Distillation Column Based On Transfer Learning

The chemical industry penetrates all aspects of national life.In 2018,the total profit of the chemical industry exceeded 600 billion yuan,which is an important part of the national economy.Distillation column is one of the most common and important separation equipment in the petrochemical production process,accounting for 30%to 40%of the total investment in equipment for chemical and petrochemical projects.Its performance is directly related to the investment,capacity,quality,and energy of production equipment.Consumption and cost.However,most sub-sectors in the chemical industry have overcapacity,product competitiveness,and slow technological upgrading.There is still a long way to go to achieve economic transformation and intelligent production.The core problem is that the characteristics of chemical equipment are complex and changeable,and it is difficult to accurately describe them,let alone customize accurate and effective optimization control methods.Due to the complicated technology and numerous products in the chemical industry,and the characteristics of the equipment will change with changes in many factors during operation,the mathematical models of chemical equipment currently established are usually incomplete,cannot be predicted in advance,the results are poor,and the diagnosis is delayed And other defects.How to model complex real-world problems has been a hot topic in the field of controlAiming at the shortcomings of traditional distillation column modeling,such as long modeling period,low prediction accuracy,poor tracking performance,etc.,combined with big data,deep learning and transfer learning technology,this paper proposes efficient processing of massive industrial data.A rectification column modeling solution capable of accurate modeling and dynamic updating.The specific research work is as follows:(1)Aiming at the problem that the traditional feature selection method is difficult to take into account the computational complexity and performance when processing high-dimensional time series data of the distillation column,and is seriously affected by noise,a method for selecting time series data features based on multi-dimensional attention is proposed mAFS).mAFS is composed of two parallel attention generating modules and a learning module.Each time-series feature is scored to obtain a score ranking.For different problems and different environments,the feature subset can be flexibly selected,which can effectively solve the problem of dimensional disasters while ensuring the accuracy of modeling,reducing the complexity of the model and thus the deployment cost.In addition,by observing the weight distribution of the features in the time dimension,the system delay can be obtained,which helps engineers understand the system mechanism.(2)Due to the large equipment of the distillation column,the work of manually modeling each measurement point individually is huge and requires a lot of background knowledge and expert support.At the same time,the production processes of different products are similar but have their own characteristics.There is a lot of repetitive work in modeling alone,which is inefficient.Aiming at the above problems,this paper proposes a method of automatic equipment modeling based on feature migration.Save the existing model parameters.When new equipment or products change,load the saved model parameters without changing the underlying model structure.Use the new data and features to achieve rapid and resource-friendly model construction through fine-tuning.(3)Aiming at the problem that raw materials and equipment evolve with time in the production process and the prediction accuracy of fixed models decreases with time,a dynamic update model construction technology based on sample migration is proposed.When the system detection model accuracy drops to a set threshold,Use the latest data to update the model and realize the smooth switching between the new model and the old model being served to ensure continuous service and effectively track the evolution of the distillation column equipment.The work described in this article has been deployed in two domestic chemical companies to model and predict multiple key parameters of distillation columns.And he entire workflow has been drafted as an industry standard.The results show that the system realized by this method can run stably for a long time and maintain high prediction accuracy under all conditions of the distillation column.It has provided effective guidance for the actual operation scheduling of the plant and verified the feasibility of this method It provides a new way for modeling of distillation column.