Research On Prediction Method Of NO_x From Utility Boilers Based On Neural Network

In China's energy system,which is still dominated by coal-fired power generation,the release of a large amount of nitrogen oxides is inevitable.The prediction of NO_x emissions from power station boilers is predicted.Inaccurate NO_x emissions are likely to lead to excessive spraying.Ammonia or ammonia is not enough.Excessive ammonia injection not only causes waste of ammonia,but also causes secondary pollution to the environment and causes clogging of air preheaters.Insufficient ammonia injection leads to insufficient catalytic reduction reaction,resulting in excessive NO_x emissions.It will have a serious impact on the economic benefits of the power plant and will also have a negative social impact on the power plant.Modeling of power station boilers using traditional machine learning algorithms will have feature selection and steady state extraction requirements.In order to solve the problem of NO_x prediction under variable operating conditions,a multi-model fusion algorithm SLXM combining linear regression,XGBoost and Long Short-Term memory recurrent neural network is proposed to model the NO_x emission prediction of power plant boilers.The working principle of the fusion model was analyzed and studied in detail,and compared with the existing prediction model under steady state.According to the characteristics of boiler dataset,pointing out the importance of data processing for model prediction,a data preprocessing scheme suitable for power plant boiler dataset is proposed:numerical processing,data cleaning and data for boiler data and various features.Standardization and other pre-processing processes.Finally,the 7-day historical operation data of a unit in guangzhou shajiao c power plant is taken as the training set and the test set,and after preprocessing,it is used for modeling by multi-model fusion algorithm.Research shows that compared with traditional machine learning algorithms,multi-model fusion algorithm SLXM under the condition of variable condition has good prediction ability,the predictions of a multilayer perceptron in test set MAPE is 9.16%,the average error of the SVM prediction results of the overall small compared to the real value,at the same time on the test set MAPE is 7.37%,the average error cannot be put into use in practical engineering,the results of PCA and SVM on the test set is larger than the error of SVM,but SLXM achieved good prediction accuracy on the test set,the average error of 4.28%.