Research On NO_x Emission Prediction Of Power Plant Boiler Based On Ensemble Learning Method

In the field of my country’s power generation industry,the current main form of power generation is still thermal power generation,and the emission of power station boiler flue gas is currently one of the main factors causing air pollution.With my country’s strict control of pollutant emissions,precise control of nitrogen oxide(NO_x)emissions from coal-fired units is very important.Selective Catalytic Reduction(SCR)denitrification technology is currently the main flue gas denitrification technology for coal-fired power plants.Accurate and rapid measurement of NO_x content at the entrance of the SCR reactor is the basis for efficient and environmentally friendly emission control of coal-fired units.At present,most thermal power plants use the Continuous Emission Monitoring System(CEMS)to measure the NO_x content of the SCR inlet,but there is a certain time delay when measuring the flue gas concentration,which cannot reflect the NO_x concentration in time.Furthermore,the ammonia injection action cannot be quickly guided for denitration.Therefore,it is particularly important to accurately predict the NO_x content of the SCR inlet in time.This paper takes the NO_x content at the entrance of the SCR as the research object,establishes a boiler NO_x emission prediction model based on the ensemble learning method,and predicts the NO_x content in advance.First,analyze the influencing factors of boiler NO_x emissions and the principle of SCR denitrification,select initial relevant variables for screening,and select 27 variables as model input variables through the mutual information feature selection method,which achieves the purpose of reducing the data dimension and reducing redundant data.After that,the input variables are normalized by Min-max,and the data set is divided,which lays a data foundation for NO_x emission prediction modeling.Secondly,three different models of Gated Recurrent Unit(GRU),Convolutional Neural Networks(CNN),and Multiple Linear Regression(MLR)were established respectively as the base model of subsequent ensemble learning methods.For these three models,the influence of the input step size and the number of network layers in each model on the accuracy of prediction was compared and analyzed.The optimal parameters of each model were determined through experimental comparison,so that each model reached the optimal status.Then,a model based on the Stacking ensemble method was established,which demonstrated the good predictive performance of the Stacking model from different perspectives,and established Long Short-Time Memory(LSTM)and Deep Neural Networks(DNN),Support Vector Regression(SVR),Bagging,Adaboost and other models are compared.The results show that the stacking prediction result is not only better than the base model in the ensemble model,but also better than the above comparison model,which proves the Stacking ensemble model Strong predictive performance.Finally,a model based on the Blending ensemble method is established.The results show that Blending is also better than other comparative models,and the prediction effect of Blending is slightly better than that of the Stacking model.In order to verify the generalization ability of the ensemble model,we selected data from different power plants for experiments,re-tuned the model parameters,and compared them under the same operation steps.The results showed that the ensemble model can still predict NO_x more accurately for different data.The two ensemble methods,Stacking and Blending,are more accurate and effective than other algorithm predictions,and have better prediction results than a single base model,reflecting the generalization capabilities of the two integrated models on different data sets.