Research On Emission Prediction Of Combustion Process Based On Flame Digital Image Processing

A lot of coal have been consumed in thermal power plant which is the main form of thermal electricity production in China. Sulfur oxides (SOx), nitrogen oxides (NOx) and other pollutants, produced by coal combustion, endanger the human living environment and health. Combustion flame not only relates to the boiler safety and economy, but also contains important information about combustion energy conversion and combustion pollutants formation. Flame radicals are intermediates of combustion process, they contain a wealth of information on combustion. Therefore, this paper presents some models of NOx emission prediction of combustion by collecting digital images of the combustion flame radicals, image processing and extracting feature extraction. The main content is as follows:Firstly, it is introduced that formation mechanism of pollutants in the combustion process and the factors that affect the combustion pollutants generated. Based on coal-fired power plant combustion data, principal component analysis is used to analyze various factors that affect the combustion pollutants generated and the importance of each factor are determined. The combustion flame temperature is obtained by using two-color method based on the spectral distribution of the flame.Secondly, flame radicals digital image, including OH*, CN*, CH* and C2*, have been collected in laboratory combustion device. Three denoising methods are used to process the radical image, which include neighborhood average, median filtering and multi-image average. Multi-image average method is adopted to process radical image after comparsion and analysis of three methods. Then features are extracted.Finally, three kinds of soft measurement method, which includ neural networks (including BP neural network and RBF neural network), support vector machine and relevance vector machine, are used to establish prediction model of NOx emission based on feature values of flame radical image and the combustion flame temperature The prediction accuracy of the three methods and the advantages and disadvantages are compared and summaried. The results show that support vector machine and relevance vector machine have a better predict result than neural network.