Research On Forecast Method Of FCCU Pollution Emissions Based On Attention Mechanism

Atmospheric environment is one of the environments that human beings depend on for survival,and the problem of air pollution is not only related to the sustainable development of social economy,but also seriously affects people’s health.Nitrogen oxide(NO_x)is one of the main air pollutants.NO_x produced by fluid catalytic cracking unit(FCCU)is also one of the main sources of pollutants in petrochemical enterprises.Therefore,it is important to establish accurate pollutant prediction models to prevent the occurrence of FCCU pollution events and environmental protection.Taking the NO_x emission data of FCCU in domestic petrochemical refineries A and B as the research object,this thesis mainly completed the following research work:In this thesis,the prediction of NO_x emission concentration is divided into short-term prediction and medium and long-term prediction according to the different time steps of the prediction output.The combination of attention mechanism and long short term memory neural network(LSTM)is used for short-term prediction.Compared with LSTM,the root mean square error(RMSE)of the short-term prediction model integrated with attention is reduced by 5.99%,the average absolute error(MAE)is reduced by7.66%,and the coefficient of determination(R2)increased by 1.45%;The combination of attention and sequence to sequence(seq2seq)is used for medium and long-term prediction.Compared with seq2seq,the RMSE,MAE and R~2 of the medium and long-term prediction model integrated with attention are decreased or increased by 2.58%,8.32%and 0.83%respectively.In order to more truly simulate the NO_x emission in FCCU,according to the NO_x generation mechanism,the production factors related to NO_x emission are added to construct a multivariable time series.A hybrid model is proposed based on CNN,LSTM and attention,it reached 18.4972,13.0641 and 0.8915 in RMSE,MAE and R~2,respectively,which were the best among the contrast models.In addition,in order to solve the problem that it is difficult to model accurately in some petrochemical enterprises with the lack of data,using transfer learning methods,the rich production data of plant A is used to provide a pre-trained model,which is transferred to plant B to fine-tune the model to complete the prediction.The experiments show that compared with the retraining model,the RMSE is reduced by 38.97%,the MAE is reduced by 44.19%,the R~2 is increased by 81.48%,and the training time is shortened by14.2%,which not only saves the training time,but also significantly improves the accuracy of model prediction.Experiments show that the neural network model integrating attention can effectively improve the prediction of NO_x emission concentration in FCCU;In addition,the method of transfer learning can also solve the disadvantage of large data demand of deep learning technology,and provide a certain solution for the modeling and prediction of similar small sample data.