This application is a national stage application of international application number PCT/CN2019/122326, filed Dec. 2, 2019, titled “Soft Measurement Method for Dioxin Emission Concentration In Municipal Solid Waste Incineration Process”, which claims the priority benefit of Chinese Patent Application No. 201910224790.4, filed on Mar. 24, 2019, which is hereby incorporated by reference in its entirety.
The invention relates to the field of solid waste incineration in general, in particular to a soft measurement method of dioxin emission concentration in municipal solid waste incineration (MSWI) process.
Reducing energy consumption and pollution emissions of complex industrial processes based on operational optimization control strategies is a challenge that many industrial process enterprises are facing. Incineration is the main technical solution for municipal solid waste (MSW) treatment. For MSW incineration (MSWI) companies in developing countries, the most pressing issue is how to control the pollution emissions caused by incineration. Among them, the emission of dioxin (DXN), which has “not-in-my-back-yard effect” and “bio-accumulative effects,” commands the highest urgency for emission control from MSWI plants. The main concern of MSWI companies is how to minimize DXN emissions based on optimized operating parameters. At present, setting aside the use of advanced flue gas treatment devices, industrial processes generally adopt the following indirect strategies to control DXN emissions, that is, the “3T1E” criterion with the temperature (T) above 850° C. in the incinerator, more than 2 seconds (T) flue gas residence time, greater degree of turbulence (T) and appropriate excess air coefficient (E). At present, MSWI companies cannot perform operation optimization and feedback control with the direct goal of reducing DXN emissions. The main reasons are: firstly, the first principal model of DXN emission concentration is difficult to build; secondly, offline direct detection based on monthly or seasonal cycles cannot provide real-time feedback of DXN emission concentration. In recent years, research hotspots are turned to indicators/associated substances for online indirect detection of DXN emissions, but the detection equipment is complex with high cost, and the detection lag of these methods make it difficult to use for the operation of MSWI process's optimization and feedback control.
Data-driven soft measurement techniques can be used for online estimation of difficult-to-measure process parameters (such as dioxins in the present invention) that require offline detection. For the MSWI process, Chang et al. used small sample data collected by European and American research institutions for different types of incinerators many years ago, and built a DXN emission concentration soft measurement model based on linear regression, artificial neural network (ANN) and other algorithms. (Chang N B, Huang S H. Statistical modelling for the prediction and control of PCDDs and PCDFs emissions from municipal solid waste incinerators, Waste Management & Research, 1995, 13: 379-400. Chang N B, Chen W C. Prediction of PCDDs/PCDFs emissions from municipal incinerators by genetic programming and neural network modeling, Waste Management & Research, 2000, 18(4):41-351.) In recent years, Bunsan S, combined correlation analysis, principal component analysis (PCA) and ANN and other algorithms to build a DXN emission prediction model based on MSWI process data; (Bunsan S, Chen W Y, Chen H W, Chuang Y H, Grisdanurak N. Modeling the dioxin emission of a municipal solid waste incinerator using neural networks, Chemosphere, 2013, 92: 258-264.) However, ANN is not suitable for building a DXN concentration emission model, mainly because of its inherent shortcomings of easy to fall into local minimum, easy to overfit and poor generalization performance for modeling small sample data. Support vector machine (SVM) algorithm with appropriate hyperparameters can be effectively used for small sample data modeling. To solve the problem of quadratic programming for SVM requirements, the least squares-support vector machine (LS-SVM) overcame it by solving linear equations; the hyperparameters of the model can be obtained by single-objective or multi-objective optimization of algorithms solving, but these methods are time-consuming and can only obtain suboptimal solutions. Therefore, the current research lacks an adaptive mechanism for effective selection of hyperparameters.
Generally, the MSWI process includes multiple subsystems composed of solid waste storage and transportation, solid waste incineration, steam power generation, and flue gas treatment. The process variables involved hundreds of dimensions, and the DXN generation, absorption, and re-synthesis mechanism processes is relevant in varying degrees. Li D C et al. pointed out that the increase in model input dimension and the increase in low-value training samples make it difficult to obtain a sufficient number of training samples. (Li D C, Liu C W. Extending attribute information for small data set classification. IEEE Transactions on Knowledge and Data Engineering, 2010, 24(3): 452-464.) In the field of pattern recognition, it is generally believed that the ratio of the number of training samples to features should be 2, 5, or 10. Tang et al. defines the ratio of training samples and reduction features after dimensionality reduction, and believes that this value should meet the requirements of constructing a robust predictive learning model. (Tang J, Qiao J F, Chai T Y, et al. Multi-component mechanical signal modeling based on virtual sample generation technology, Acta Automatica Sinica, 2018, 44(9): 1569-1589.) Therefore, for the DXN emission concentration modeling data with high dimensional characteristics of small samples, it is necessary to carry out dimensionality reduction. The features obtained based on the unsupervised feature extraction method, although containing the main changes in the original high-dimensional input features, extracted features may be independent of parameters to be predicted. Principal component analysis (PCA), which can extract changes in high-dimensional data, is currently the most commonly used method for latent feature extraction in soft measurement of difficult-to-measure parameters in industrial processes, but a low contribution rate of the principal component (PC) will lead to poor prediction stability.
From another perspective, features extracted for different subsystems of the MSWI process can be viewed as multi-source information from multiple views. Theoretical and empirical analysis shows that the soft measurement model constructed by the selective ensemble (SEN) learning mechanism for multi-source information has better prediction stability and robustness, and the diversities between the sub-models are particularly important. Gavin Brown et al. summarizes the ensemble construction strategy for increase diversity of ensemble sub-models, and points out that training sample resampling strategy includes dividing training samples (sample space), dividing or transforming input features (feature space), etc., the construction strategy based on feature space is superior to the construction strategy based on multiple classifiers in terms of model prediction performance. (Brown G, Wyatt J, Harris R, Yao X. Diversity creation methods: a survey and categorisation. Information Fusion, 2005, 6: 5-20.) For small sample multi-source high-dimensional spectral data, Tang et al. proposed a SEN projection to latent structure model based on selective fusion of multi-source features and multi-conditions samples; Zhou et al and Tang et al proposed a random-based sampling sample space SEN neural network model and projection to latent structure model. (Zhou Z H, Wu J and Tang W, Ensembling neural networks: many could be better than all,” Artificial Intelligence, 2002, 137(1-2): 239-263; Tang J, Chai T Y, Yu W, Liu Z, Zhou X J. A Comparative study that measures ball mill load parameters through different single-scale and multi-scale frequency spectra-based approaches, IEEE Transactions on Industrial Informatics, 2016, 12(6): 2008-2019.) Ma et al proposes a general framework for ensemble learning based on subspace. (Ma G, Wu L, Wang Y. A general subspace ensemble learning framework via totally-corrective boosting and tensor-based and local patch-based extensions for gait recognition. Pattern Recognition, 2017, 66: 280-294.) Tang et al proposes a two-layer SEN projection to latent structure model in multi-scale mechanical signal oriented random sampling sample space in feature subspace. (Tang J, Qiao J F, Wu Z W, Chai T Y, ZhangJ, Yu W. Vibration and acoustic frequency spectra for industrial process modeling using selective fusion multi-condition samples and multi-source features. Mechanical Systems and Signal Processing, 2018, 99:142-168.) The SEN neural network model proposed by S. Soares et al constructs candidate sub-models for optimization perspective and ensemble sub-models for optimization selection respectively and their weights. (Soares C, Antunes C H, AraUjo R, Comparison of a genetic algorithm and simulated annealing for automatic neural network ensemble development, Neurocomputing, 2013, 21(9):.498-511.) However; none of the above methods involves or adopts the modeling parameter adaptation mechanism. In summary, with the measurement and reselection of unsupervised latent features as input, the SEN-LS-SVM modeling strategy based on the adaptive hyperparameter selection mechanism and its research in soft measurement of DXN emission concentration have not been reported.
The present disclosure provides a multi-source latent feature SEN-based soft measurement method for DXN emission concentration in MSWI process. First, the latent feature extraction and primary selection module is used to divide the MSWI process data into subsystems of different sources according to industrial processes. Principal component analysis (PCA) is used to separately extract the subsystems' latent features and conduct multi-source latent feature primary selection according to the threshold value of the principal component contribution rate preset by experience. Subsequently, the latent feature evaluation and reselection module uses mutual information (MI) to measure the correlation between the latent features of the primary selection and the DXN, and adaptively determine the upper and lower limits and thresholds of latent feature reselection. Finally, based on the reselection latent features, the adaptive selective ensemble modeling module uses a least squares-support vector machine (LS-SVM) algorithm with a hyperparameter adaptive selection mechanism and establishes DXN emission concentration sub-models for different subsystems; adopts the strategy based on branch and bound (BB) and prediction error information entropy weighting algorithm to optimize the selection of sub-models and calculate the weight coefficients, to build a DXN emission concentration SEN soft measurement model.
MSWI's main equipment includes incinerator, mobile grate, waste boliler and flue gas treatment equipment, among which: incinerator converts MSW into residue, dust, flue gas and heat, and mobile grate located at the bottom of incinerator makes MSW effective and complete combustion, the steam generated by the waste boiler is used to drive the steam turbine to generate electricity, and the dust and pollutants in the flue gas are purified by the flue gas treatment equipment and discharged into the atmosphere. The process is shown in
As can be seen from
As can be seen from
In the present invention, model input data X∈R<N×M> includes N samples (rows) and M variables (columns), which are derived from different subsystems of the MSWI process. Represent the modeling data from the “ith” subsystem as Xi∈RN×M
Wherein, I represents the number of subsystems, Mi represents the number of variables contained in the ith subsystem. Correspondingly, the output data y={yn}n=1N includes N samples (rows), which are derived from the DXN emission concentration detection data of offline testing. Obviously, the input/output data has great differences on the time scale: the process variables are collected and stored in DCS system in seconds, and DXN emission concentration is obtained by offline testing on a monthly/quarterly cycle, so that N=M.
Based on the above situation, the present invention proposes a soft measurement method of DXN emission concentration based on latent feature SEN modeling, including latent feature extraction and primary selection module, latent feature evaluation and reselection module, adaptive selective ensemble modeling module, such as
In
The functions of the above modules are:
Taking the ith subsystem as an example, PCA is first used to extract the latent features of high-dimensional input process variables. input data Xi is normalized to zero mean 1 variance, it is decomposed into:
Xi=t1
Wherein, tm
MFeAlli=rank(Xi) (4)
Based on the above expression, all the latent features extracted from the data Xi can be expressed as:
Ti=[t1
wherein, Ti∈RN×M
Pi=[p1
wherein Pi∈RM×M
Therefore, the latent features extracted from the data Xi can be expressed as,
Wherein ZFeAlli∈RN×M
Further, all latent features can be expressed as:
ZFeAll=[ZFeAll1,L, ZFeAlli,L, ZFeAllI]={ZFeAlli}i=1I (8)
Studies have shown that modeling with latent variables with a small contribution rate can lead to instability in model prediction performance.
The feature vector corresponding to the mFeAllith load vector Pm
The threshold selected according to experience is recorded as θContri, and its default value is 1. Use the following rules to select all latent features for the first time:
wherein, ξm
Therefore, the latent features of the primary selection for the ith subsystem are expressed as:
Further, all latent features of the primary selection ZFeSe1st can be expressed as:
ZFeSe1st=[ZFeSe1st1,L, ZFeSe1sti,L, ZFeSe1stI]={ZFeSe1sti}i=1I (12)
The preliminary selected latent features obtained in the previous step are extracted using unsupervised methods, and the features contained in the same subsystem are independent of each other, but the correlation between these features and DXN emission concentration is not considered, that is, high contribution latent features are not necessarily strongly correlated with DXN. Take the ith subsystem as an example, the mutual information (MI) value between each primary feature zm
Wherein, pprob(zm
The threshold is adaptively determined according to prediction performance of the soft measurement model. The upper limit of the threshold θContriUplimit the lower limit of θContriDownlink and the fixed step size θContriStep are calculated using the following formula:
Wherein, the functions max(⋅)and min(⋅) represent the maximum and minimum values respectively; NContriStep represents the number of candidate thresholds determined based on experience, and the default value is 10.
The selected threshold is recorded as θContri, and its value is adaptively selected between θContriUplimit and θContriDownlimit based on the prediction performance of the DXN soft measurement model.
Use the following rules to reselect the latent features of the primary selection:
Wherein, ξm
Further, express the reselected latent features for the ith subsystem as:
Therefore, all the re-selected latent features zFeSe2nd can be expressed as:
ZFeSe2nd=[ZFeSe2nd1,L, ZFeSe2ndi, L, ZFeSe2ndI]={ZFeSe2ndi}i=1I (19)
Taking the ith subsystem as an example, describe the process of constructing a DXN emission concentration sub-model based on the reselection of latent features zFeSe2nd and model hyperparameter pair {Keri, Regi}.
First, transform the reselected latent feature {(zFeSe2ndi)n}n=1N into a high-dimensional feature space by mapping φ(⋅), and then solve the following optimization problem:
Wherein, wi represents the weight coefficient, bi represents the offset, and ζni is the prediction error of the nth sample.
Using Lagrangian method, the following formula can be obtained:
Wherein, βi=[β1i,L, βni,L, βNi] represents the Lagrange operator vector, and ζi=[ζ1i,L, ζni,L, ζNI] represents the prediction error vector.
Calculate the above formula:
The adopted kernel function can be expressed as follows:
Ωkeri(zFeSe2ndi,(zFeSe2ndi)n)=<φ(zFeSe2ndi)·φ((zFeSe2ndi)n)< (23)
Further, LS-SVM problem is converted into solving the following linear equation system:
By solving the above formula, βi and bi are obtained.
Furthermore, DXN emission concentration sub-model based on LS-SVM can be expressed as:
Hyperparameter adaptive selection mechanism of the above DXN emission concentration sub-model is implemented using the following two-step method:
The first step is to adopt the grid search strategy with the prediction performance of the sub-model as objective function, and adaptively select the initial hyperparameter pair {(Kerinitial)i, (Reginitial)i} in the candidate hyperparameter matrix Mpara The hyperparameter matrix Mpara can be expressed as follows:
Where k=1,L,K; K represent the number of candidate kernel parameters; r=1,L,R; R represent the number of candidate penalty parameters; [Kerk, Regr] represents the kth kernel parameter and rth penalty parameter. The formed hyperparameter pair is also the jth parameter pair in the hyperparameter matrix Mpara, that is, there are Mparaj=[Kerk, Regr], j=1,L,J, J=K×R means that all hyperparameter pairs' number in the hyperparameter matrix Mpara. Therefore, the hyperparameter pair {(Kerinitial)i, (Reginitial)i} selected for the first time using the grid search strategy is an element in matrix Mpara, i.e., there is {Kerinitial, Reginitial}∈Mpara. In the second step, based on the {(Kerinitial)i, (Reginitial)i} selected by the above method, a new set of candidate hyperparameters is obtained using the following formula:
Wherein, (Kervector)i and (Regvector)i represent new candidate hyperparameter sets, corresponding to kernel parameter vector and penalty parameter vector respectively; Nker and Nreg represent the number of new hyperparameters set based on experience; ksuparadown and ksuparaup represents the hyperparameter shrinkage and expansion factor based on experience, the default value is 10.
By adopting the grid search strategy to adaptively obtain the hyperparameter pair of the ith sub-model {Keri, Regi}.
Performing the above process on all subsystems, the set of sub-model prediction outputs can be expressed as:
Wherein, fi(⋅) represents the ith sub-model.
Combining the optimization selection algorithm based on branch and bound (BB) and the prediction error information entropy weighting algorithm, the above sub-models are adaptively optimized and weighted coefficients are calculated. Given the candidate sub-model and weighting algorithm, the optimal sub-model selection is similar to the weighted optimal feature selection. For a limited number of candidate sub-models, by running optimization and weighting algorithms multiple times, SEN models with integrated sizes of 2 to (I−1) can be obtained, Finally, these optimized SEN models are ranked and the best predicted performance is taken as the final DXN soft measurement model.
Assuming that the ensemble size of final DXN soft measurement model is Isel, its predicted output value ŷ can be calculated by the following formula:
Wherein, fi
Compare with equation (29), there has the following relationship:
Using predicted value and true value of the sub-model, wi
Wherein:
Wherein, (ŷi
The modeling data in the present invention is derived from the 1 #furnace of a grate based MSWI incineration enterprise in Beijing, covering the available DXN emission concentration detection samples recorded from 2012 to 2018, the number of which is 39; the corresponding input variables dimensions is 286 (including all process variables of the MSWI process). It can be seen that the number of input features far exceeds the number of modeling samples, and it is necessary to carry out dimensionality reduction. The invention divides the modeling data into two parts, which are used for training and testing respectively.
In the present invention, six subsystems of incineration, boilers, flue gas treatment, steam power generation, stack emissions, and common engineering are respectively marked as Incineration, Boiler, Flue gas, Steam, Stack, and Common. In order to represent the overall change characteristics of the incineration process variables, the present invention uses the MSWI system containing all variables as a special subsystem for analysis and modeling. Therefore, the present invention contains a total of 7 subsystems.
The cumulative contribution rate of the first 6 latent features of 7 subsystems extracted by PCA is shown in
Based on the criterion that the contribution rate of a single PC is not less than 1%, the number of primary features and their contribution rates are shown in Table 1.
It can be seen from Table 1 that the number of latent features selected by different subsystems is 13, 6, 9, 8, 6, 12, and 15. Since PCA belongs to an unsupervised feature extraction method, these extracted features only describe the change of input data, and the mapping relationship between it and DXN needs to be further measured.
The MI method is used to measure the mapping relationship between the primary features extracted for different subsystems and the DXN, which is shown in
Table 2 shows that: (1) For the maximum value set: the maximum value is derived from a Common subsystem that is theoretically not directly related to DXN emissions, and the value is 0.8613. Whether it is reasonable needs to be further combined with the model prediction results verification; the second place is the Incinerator subsystem, which has a value of 0.8559. This latent variable is theoretically related to the generation of DXN, which is more reasonable; (2) For the minimum set: the minimum is derived from incineration (MSWI) subsystem, which is 0.4429, indicating that separate analysis for different subsystems is still necessary; the maximum value comes from the stack subsystem of the flue gas emission (Stack), which is 0.7182, because other emissions are between DXN and If there is a correlation, this value is more reasonable.
As can be seen from Table 2, the upper limit of the MI threshold is 0.7882, the lower limit is 0.7182, and the step size is 0.006999. Combining the upper and lower limits of the threshold and the step size to determine the final threshold is 0.7882. Twice selection of the number and MI value of latent features are shown in Table 3.
In the present invention, the set of candidate regularization parameters and kernel parameters are pre-selected as {0.0001, 0.001, 0.01, 0.1, 1, 10, 100, 1000, 2000, 4000, 6000, 8000, 10000, 20000, 40000, 60000, 80000, 160000} and {0.0001, 0.001, 0.01, 0.1, 1, 10, 100, 1000, 1600, 3200, 6400, 12800, 25600, 51200, 102400}.
Based on the above, the number of input features of the Incineration, Boiler, Flue gas, Steam, Stack, Common, and MSWI whole process sub-models are 5, 2, 1, 3, 2, 6, and 1, respectively. The first and second curves of the hyper-parameter adaptive optimization using grid search method are shown in
Based on the above results, hyperparameter pairs adaptively selected by the above sub-models are {109,109}, {10000,25.75}, {5.950,0.0595}, {30.70,2.080}, {5.950,0.5950}, {1520800,22816} {1362400,158.5}, the corresponding root mean square error (RMSE) of the test data is 0.01676, 0.02302, 0.01348, 0.01943, 0.01475, 0.02261 and 0.02375.
Using optimization and weighting strategy based on BB and prediction error information entropy weighting algorithm, the test errors of SEN model constructed when integration size is 2-6 are 0.01345, 0.01332, 0.01401, 0.01460 and 0.01560, respectively. The final ensemble size of DXN soft measurement model is 3, and the corresponding subsystems of selected sub-model are flue gas treatment, stack emission and incineration. In theory, these three subsystems are related to absorption, emission and generation of DXN. From the results of present invention, validity of all algorithms is verified, and availability of data is also indicated.
The comparison with usual PLS single model, PCA-LSSVM single model and different weighting methods is shown in Table 4.
Table 3 shows that the prediction performance of DXN single model based on PLS and PCA-LSSVM constructed with all process variables is weaker than SEN modeling method proposed by the present invention, indicating that the strategy of building a SEN model based on multi-source features is effective. The method of ensemble all sub-models adopts the PLS weighted method that is stronger than other ensemble (EN) all sub-models, indicating that the PLS algorithm is better in eliminating the collinearity of sub-models; in addition, the subsystem corresponding to sub-model selected by SEN model are related to the generation, absorption and emission mechanism of DXN, indicating the availability of modeling data and the effectiveness of the algorithm.
The invention based on industrial process data of an incineration enterprise in Beijing, and uses latent feature extraction and primary selection based on PCA and prior knowledge, primary latent feature evaluation and selection based on MI and prior knowledge, and self-adaptive latent feature-oriented adaptation based on SEN modeling mechanism, a soft measurement of DXN emission concentration based on multi-source latent feature SEN modeling is proposed, and the simulation verifies the effectiveness of proposed method. The adjustment of contribution rate threshold, MI threshold, hyperparameters, and SEN model structure combined with the soft measurement model's predicted performance adaptive adjustment still needs further study. In addition, the analysis of combined DXN emission mechanism also needs to be carried out in depth.
Number | Date | Country | Kind |
---|---|---|---|
201910224790.4 | Mar 2019 | CN | national |
Filing Document | Filing Date | Country | Kind |
---|---|---|---|
PCT/CN2019/122326 | 12/2/2019 | WO |
Publishing Document | Publishing Date | Country | Kind |
---|---|---|---|
WO2020/192166 | 10/1/2020 | WO | A |
Number | Name | Date | Kind |
---|---|---|---|
20090193936 | Lu et al. | Aug 2009 | A1 |
Number | Date | Country |
---|---|---|
103455635 | Dec 2013 | CN |
107944173 | Apr 2018 | CN |
107944173 | Apr 2018 | CN |
108549792 | Sep 2018 | CN |
108549792 | Sep 2018 | CN |
109960873 | Jul 2019 | CN |
2001242154 | Sep 2001 | JP |
Entry |
---|
JP2001242154A, translation (Year: 2001). |
CN108549792A, translation (Year: 2018). |
CN107944173A, translation (Year: 2018). |
Number | Date | Country | |
---|---|---|---|
20210233039 A1 | Jul 2021 | US |