Soft Sensing Of Dioxin Emission Concentration For Municipal Solid Waste Incineration

Municipal solid waste(MSW)is a kind of solid waste that can not be avoided due to the needs of daily life.Municipal solid waste incineration(MSWI)has significant advantages in the harmless,reduction and resource-based treatment.However,the MSW process produces a very toxic persistent organic pollutant with strong chemical and thermal stability,namely,dioxin(DXN).It can damage the chromosomes of organisms and further lead to cell carcinogenesis.Moreover,it has significant accumulation and amplification effect in organisms.Thus,DXN has a huge potential harm to human health and ecological environment.In addition,DXN is also the main reason for the "not in my backyard" effect of MSWI.Researches show that the realtime detection of DXN emission concentration is one of the key factors to realize the operation optimization of MSWI process in terms of reducing pollution emission.In actual industrial process,DXN can only be directly detected by off-line assay in monthly/quarterly period or longer uncertain period.The accuracy of this method is high,however it has many disadvantages,such as large lag time scale,high detection cost and so on.Another method,i.e.,the online indirect detection method based on indicator/correlator of DXN,it can shorten the time scale of detection lag to the level of hours.However,there are some shortcomings such as complex composition of detection equipment,difficult to maintain and insufficient stability of mapping model.In addition,the formation and emission mechanism of DXN is complex,so it is difficult to build an accurate mathematical model.In actual industry,the process variables related to the generation,adsorption and emission of DXN are up to hundreds of dimensions.Moreover,they have strong redundancy and non-linear relationship among each other.Although the number of process variabels collected and stored by the control system is large,the samples that can be labeled for the construction of DXN emission concentration model are extremely rare.Thus,this leads to the poor generalization performance of the existing data-driven soft measurement method.Therefore,this DXN modeling can be positioned as a kind of intelligent modeling problem based on small sample high-dimensional sparse labeled data.Moreover,there are few researches based on actual industrial process data for modeing DXN.Therefore,this paper studies the soft sensing method of DXN emission concentration by using a large number of unlabeled and a small number of labeled samples.These research results can provide support for the operation optimization of MSWI process.The main research work and innovations are as follows:1.Aiming at the problem that mechanism knowledge of DXN emission is difficult to obtain,a numerical simulation strategy of DXN emission concentration of MSWI process based on commercial process simulation software is proposed.Numerical simulation of the DXN emission process in the incinerator is made based on heat and material balance process design standards.The preliminary comparison and analysis with the original thermal mass balance design data is given out.It shows that more in-depth simulation and analysis will provide mechanism support for DXN emission concentration soft measurement。2.Aiming at the problem of high dimension and strong collinearity among the process variables,a multi-layer feature selection method for DXN emission concentration soft measurement is proposed.Firstly,from the perspective of correlation between single feature and DXN,combined with correlation coefficient and mutual information,a comprehensive evaluation value index is constructed to realize the first layer feature selection.Secondly,from the perspective of multi feature redundancy and feature selection robustness,Genetic algorithm partial least square method(GA-PLS)feature selection algorithm with multiple run-times are used to make the second layer feature selection.Finally,combined with the statistical frequency of the second layer feature selection,model prediction performance and mechanism knowledge,the third layer feature selection is carry out.The validity of the proposed method is verified by using the DXN data of an actual MSWI process.3.Aiming at the scarcity of labeled samples in modeling data,a soft sensing method of DXN emission concentration based on the virtual sample generation technology of general trend diffusion(MTD)is proposed.Firstly,the input/output sample space is expanded by MTD technology based on Euclidean distance of real sample sub-region.Secondly,the virtual sample input is generated by equal interval interpolation,and the virtual sample outpu is generated by combining mapping model and pruning mechanism.Further,a new type virtual sample output is obtained by improved random weight neural network implicit layer interpolation method.And its responding virtual sample input is obtained by inverse deduction.Finally,these complementary input/output virtual samples and the original real samples are mixed to expand the sample capacity.The validity of the proposed VSG method is verified by using the DXN data of an actual MSWI process.4.Aiming at the problem that a large number of unlabeled samples in the actual industrial process have not been used effectively,a soft sensing method of DXN emission concentration based on unsupervised unlabeled sample learning is proposed.Firstly,model pre-training is carried out based on a large number of unlabeled samples to extract the hidden knowledge contained in these samples.Secondly,based on the labeled modeling samples,an adaptive learning rate error back propagation algorithm is proposed to iteratively fine tune the pre training model.Moreover,neuron random deactivation(dropout)mechanism is proposed to improve the robustness of the soft measurement model.The validity of the proposed method is verified by using the DXN data of an actual MSWI process.