The Research On NO_x Prediction Algorithms Based On The Flame Radical Imaging

In order to meet the environmental problems caused by thermal power plants,it is important to study the techniques for the combustion optimization,predict and reduce pollutant emissions during the combustion process.Studies on NO_x prediction algorithm in fossil fuel combustion have been undertaken computationally and experimentally for many years.The development of an accurate model for NO_x emission prediction is crucial.With the advances in the visualization techniques for burner flames,it is possible to establish a NO_x prediction model based on the flame radicals which act as intermediates in fuel thermal and chemical reactions.This thesis,focusing on the NO_x prediction algorithm based on flame radical image and machine learning algorithms,presents a thorough research on the following work.1?On the basis of analyzing the characteristics of the Zernike moment,a new NO_x prediction algorithm based on the Zernike moment and LSSVR(Least Square Support Vector Regression)is proposed,i.e.the image feature extraction algorithm is designed based on the Zernike moment,and the LSSVR is deployed to describe the nonlinear relationship between the characteristics of flame radicals and NO_x emissions.In order to minimize the root mean square error(RMSE)of the prediction model,the optimization procedure can adjust the image feature extraction process and the concrete form of the image feature by means of the order of the Zernike moment.Then the proposed meth od is utilized to predict the NO_x emission using the experimental data on a biomass-gas fired test rig.The results show that the variable structure image features can improve the performance of the prediction model.2?Considering that the order of the Zernike moment has the narrow range,a new variable structure image feature is designed based on contourlet transform,the best M-term approximation and the zero-trees structure.The new NO_x prediction algorithm are then proposed based on the contourlet transform and LSSVR.The prediction algorithm can optimize the number of the retained contourlet coefficients to minimize the RMSE.Compared to the order of the Zernike moment,the number of the retained contourlet coefficients can provide a larger search space.The results show that the optimization of the structure image feature can improve the accuracy of the prediction algorithm.3?On the basis of analyzing the characteristics of the multiscale framework of the contourlet transform,the image decomposition structure is applied to the NO_x prediction algorithm.A new NO_x prediction algorithm is established based on the multiscale framework of the contourlet transform,the Zernike moment and LSSVR.The RMSE of the prediction model can be minimized by optimizing the number of the directional sub-bands.The results show that the optimization of the image decomposition structure can improve the performance of the prediction algorithm.4?Considering the prediction deviation caused by the multiscale framework of the contourlet transform,a new image decomposition structure is designed based on the non-negative matrix factorization(NMF).The corresponding prediction algorithm is proposed based on the NMF,Zernike moment and LSSVR.The RMSE can be minimized by optimizing the number of the independent components obtained in the image decomposition process.The results demonstrate the feasibility and advantage of the image decomposition structure.5?On the basis of the in-depth study of deep learning(DL),a new algorithm is designed to predict the NO_x emissions.The developed DL-based prediction model contains successive stages for implementing t he feature extraction,feature fusion and emission prediction.The fine-tuning based on the prediction is introduced to adjust the process of the feature fusion.The results suggested that the DL based prediction algorithm obtain the good performance.However,it is time-consuming due to the complicated learning framework.For this problem,the new DL based prediction algorithm is proposed based on the deep Boltzmann machine model and LSSVR-based ensemble model.The deep Boltzmann machine model is used to learn the image features quickly.The LSSVR-based ensemble model is used to reduce the effect of the simplified learning framework.