Study Of Prediction Method Of NOx Concentration On A Rotary Burner Based On Flame Image And Deep Learning

Combustion is an important means of fossil fuel energy utilization and is widely used in various industrial production fields such as energy and power,aeronautics and astronautics,chemical industry,and metallurgy.However,in the process of fossil fuel combustion,a variety of air pollutants are produced,and nitrogen oxide(NOx)is one of them.NOx is easy to cause serious environmental problems such as photochemical smog,acid rain,ozone layer hole and greenhouse effect,which has great harm to human health and ecological environment.With the rapid development of my country’s economy and the rapid growth of industrial and living energy demand,environmental pollution caused by the burning of fossil fuels has become more and more serious.For this reason,the emission of pollutants such as NOx has been strictly restricted.Combustion optimization and flue gas denitrification treatment are the main ways to reduce NOx emissions during the combustion process.Both of these emission reduction methods require accurate monitoring of NOx production.Therefore,it is of great significance to achieve accurate and reliable NOx concentration prediction.This paper proposes a NOx concentration prediction method based on flame images and deep learning,and experiments are carried out to verify the proposed method on a heavy oil combustion device.First,a set of heavy oil combustion test equipment is built,including a rotating cup burner system,an image acquisition system,and a flue gas analyzer.The combustion flame image and NOx concentration are continuously acquired during the operation of the combustion device.Under a variety of combustion conditions,flame images and the corresponding NOx concentration are collected,and an image data set for the NOx concentration prediction method by labeling the flame image data is then established.Then,the structure and principle of two typical deep convolution neural networks,Le Net-5 and VGGNet,are studied in details.On this basis,the NOx concentration prediction models based on flame image and two kinds of deep convolution neural networks are designed,and the influences of activation function,adaptive learning rate algorithm,pooling method and convolution kernel parameters on the prediction accuracy are explored.The experimental results show that the prediction accuracy based on the VGGNet model is better than that of the Le Net-5 model,indicating that the small convolution kernel and deep network structure are beneficial to improve the flame image recognition ability.When using the sigmoid function and the tanh function,the deep convolutional neural network cannot fit the relationship between the flame image and the NOx concentration.Under the same training times,the prediction accuracy of Adam algorithm is better than that of Ada Grad and Adadelta algorithms.The prediction performance of average pooling and maximum pooling is equivalent,and both can be used in the model;The stacked 3?3 convolution kernel has more advantages than the 5?5 convolution kernel,which can improve the nonlinear learning ability of the network;For small sample data,reducing the number of convolution kernels can reduce the training time,while the model can maintain good prediction performance.Finally,the methods of NOx concentration prediction based on machine learning and recurrent neural network are explored respectively.First,DCNN is used to extract the deep features of the flame image,and then machine learning methods such as SVM,GPR,RFR or recurrent neural networks such as Simple RNN,LSTM,GRU are used to analyze the extracted deep features of DCNN to achieve NOx concentration prediction.The experimental results show that:among the three machine learning models,DCNN-SVM has the best prediction performance,and the root mean square error of the test set is 1.30 mg/m~3,which is lower than the prediction model based on static physical features,indicating that DCNN-SVM has good prediction accuracy and overcomes the deficiency of shallow static physical feature expression ability;the root mean square error of DCNN-LSTM is 1.22mg/m~3,which is better than DCNN-Simple RNN and DCNN-GRU,and improves the prediction accuracy of NOx concentration.