WO2020192166A1 - 一种城市固废焚烧过程二噁英排放浓度软测量方法 - Google Patents

一种城市固废焚烧过程二噁英排放浓度软测量方法 Download PDF

Info

Publication number
WO2020192166A1
WO2020192166A1 PCT/CN2019/122326 CN2019122326W WO2020192166A1 WO 2020192166 A1 WO2020192166 A1 WO 2020192166A1 CN 2019122326 W CN2019122326 W CN 2019122326W WO 2020192166 A1 WO2020192166 A1 WO 2020192166A1
Authority
WO
WIPO (PCT)
Prior art keywords
sub
dxn
model
selection
potential
Prior art date
Application number
PCT/CN2019/122326
Other languages
English (en)
French (fr)
Inventor
汤健
乔俊飞
郭子豪
何海军
Original Assignee
北京工业大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 北京工业大学 filed Critical 北京工业大学
Priority to US16/967,408 priority Critical patent/US12002014B2/en
Publication of WO2020192166A1 publication Critical patent/WO2020192166A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/30—Administration of product recycling or disposal
    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B13/00—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion
    • G05B13/02—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric
    • G05B13/0265—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric the criterion being a learning criterion
    • G05B13/027—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric the criterion being a learning criterion using neural networks only
    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B13/00—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion
    • G05B13/02—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric
    • G05B13/04—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric involving the use of models or simulators
    • G05B13/042—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric involving the use of models or simulators in which a parameter or coefficient is automatically adjusted to optimise the performance
    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B13/00—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion
    • G05B13/02—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric
    • G05B13/04—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric involving the use of models or simulators
    • G05B13/048—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric involving the use of models or simulators using a predictor
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G06N20/10—Machine learning using kernel methods, e.g. support vector machines [SVM]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G06N20/20—Ensemble learning
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02W—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO WASTEWATER TREATMENT OR WASTE MANAGEMENT
    • Y02W90/00—Enabling technologies or technologies with a potential or indirect contribution to greenhouse gas [GHG] emissions mitigation

Definitions

  • the invention belongs to the technical field of solid waste incineration, and particularly relates to a soft measurement method for the concentration of dioxin emission in the process of urban solid waste incineration.
  • Data-driven soft measurement technology can be used for online estimation of difficult-to-detect parameters (such as dioxins in the present invention) that require offline testing [16,17].
  • difficult-to-detect parameters such as dioxins in the present invention
  • offline testing [16,17].
  • literature [18,19] used small sample data collected by European and American research institutions for different types of incinerators many years ago to construct a DXN emission concentration soft-sensing model based on algorithms such as linear regression and artificial neural network (ANN).
  • ANN artificial neural network
  • Least Squares-Support Vector Machine solves the problem by solving linear equations; model hyperparameters can be obtained by single-objective or multi-objective optimization algorithm[22,23,24 ], but these methods are time-consuming and can only get sub-optimal solutions [25]. Therefore, the current research lacks an adaptive mechanism for effective selection of hyperparameters.
  • 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 reach hundreds of dimensions, and are related to the generation, absorption and resynthesis of DXN. Relevant to varying degrees.
  • Literature [26] pointed out that the increase of the model input dimension and the increase of low-value training samples make it difficult to obtain a complete number of training samples. In the field of pattern recognition, it is generally believed that the ratio of the number of training samples to the features should be 2, 5 or 10.
  • Literature [27] defines the ratio of training samples to reduced features after dimensionality reduction, and believes that this value should meet the requirements of building a robust predictive learning model.
  • PCA Principal component analysis
  • the features extracted for different subsystems of the MSWI process can be regarded as multi-source information from multiple views.
  • Theoretical and empirical analysis shows that the soft sensor model constructed with selective integration (SEN) learning mechanism for multi-source information has better predictive stability and robustness, and the difference between sub-models is particularly important [29].
  • Literature [30] summarized the construction strategy of the diversity of integrated sub-models, and pointed out that training sample resampling includes dividing training samples (sample space), dividing or transforming feature variables (feature space), etc., based on the integration of feature space in the model prediction performance The construction strategy is better than the construction strategy based on multiple classifiers. For small samples of multi-source high-dimensional spectrum data, Tang et al.
  • the SEN latent structure mapping model proposed a SEN latent structure mapping model based on selective fusion of multi-source features and multi-condition samples [31,32]; literature [33] and [32] proposed random Sampling the SEN neural network model and latent structure mapping model of the sample space, the literature [34] proposes a general framework for ensemble learning based on subspace, and the literature [35] proposes a multi-scale mechanical signal-oriented random sampling of the sample space in the feature subspace
  • the SEN neural network model proposed in [36] separately constructs the candidate sub-models of the optimized perspective and the optimized selection integrated sub-models and their weights, but the above method does not carry out the study of the model parameter adaptive mechanism.
  • the SEN-LSSVM modeling strategy based on the adaptive hyperparameter selection mechanism and its research in the soft measurement of DXN emission concentration have not been reported with the unsupervised potential features after measurement and reselection as input.
  • the present invention provides a soft measurement method for MSWI process DXN emission concentration based on multi-source potential feature SEN.
  • the potential feature extraction and primary selection module is used to divide MSWI process data into subsystems of different sources according to industrial processes.
  • Principal component analysis (PCA) extracts its potential features separately and conducts preliminary selection of multi-source potential features based on the empirically preset thresholds of principal component contribution rate.
  • the latent feature measurement and reselection module uses mutual information (MI) to measure the correlation between the initially selected latent features and DXN, and adaptively determines the upper and lower limits and thresholds of the latent feature reselection.
  • MI mutual information
  • the adaptive selective integrated modeling module uses the least squares-support vector machine (LS-SVM) algorithm with hyperparameter adaptive selection mechanism based on the potential features of reselection to establish DXN emission concentration sub-models for different subsystems , Adopt the strategy based on branch and bound (BB) and prediction error information entropy weighting algorithm to optimize the selection sub-model and calculate the weight coefficient to construct the SEN soft-sensing model of DXN emission concentration.
  • LS-SVM least squares-support vector machine
  • Figure 5(c) The first and second curves of the hyperparameter adaptive optimization of the flue gas treatment sub-model
  • Figure 5(d) The first and second curves of the superparameter adaptive optimization of the steam electron generation model
  • Figure 5(e) The first and second curves of the hyperparameter adaptive optimization of the smoke emission sub-model
  • Figure 5(f) The first and second curves of the hyperparameter adaptive optimization of the utility sub-model
  • Figure 5(g) The first and second curves of the hyperparameter adaptive optimization of the MSWI full-process sub-model.
  • the main equipment of MSWI includes incinerator, mobile grate, waste pot and exhaust gas treatment equipment.
  • the incinerator converts MSW into residue, dust, flue gas and heat.
  • the mobile grate at the bottom of the incinerator makes MSW effective and complete. Combustion, the steam produced by the waste pot is used to drive the steam turbine to produce electricity, and the dust and pollutants in the flue gas are purified by the exhaust gas treatment equipment and discharged into the atmosphere.
  • the process is shown in Figure 1.
  • the MSWI process includes three stages of DXN production, absorption and emission, which are respectively contained in the flue gas marked G1, G2 and G3. Obviously, the concentrations of DXN contained in the flue gas at these different stages are different.
  • the flue gas temperature should reach at least 850°C and be maintained for more than 2 seconds.
  • the primary air used for incineration of MSW is injected from the bottom of the grate while cooling the grate, and by introducing turbulence and ensuring the supply of excess oxygen, the secondary air can assist in the complete combustion of the flue gas.
  • the incineration slag and waste bottom ash are disposed and collected, and flue gas G1 is discharged at the same time.
  • Activated carbon and lime are injected into the reactor to remove acid gas and absorb DXN and some heavy metals.
  • the flue gas then enters the bag filter.
  • the fly ash generated in the reactor and the bag filter is injected into the mixing equipment, and flue gas G2 is generated at the same time.
  • the induced draft fan sucks the flue gas G2 into the chimney, and then discharges it into the air as flue gas G3.
  • It contains HCL, SO 2 , NOx and HF and other pollutant concentrations that can be detected online in real time, and has the characteristics of long cycle and high cost. Need to test the concentration of DXN offline. From the above description, we can see that the DXN emission concentration is correlated with the easily detectable process variables at different stages of the MSWI process.
  • the MSWI process can be divided into six subsystems, including incineration, boiler, flue gas treatment, steam power generation, flue gas emission, and public engineering assistance.
  • the present invention treats multiple subsystems as multi-source information.
  • the 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 That is, the following relationship exists,
  • I represents the number of subsystems
  • M i represents the number of variables included in the ith sub-system.
  • output data Including N samples (rows), which are derived from the DXN emission concentration detection data of offline testing.
  • the input/output data has a big difference on the time scale: process variables are collected and stored in the DCS system in seconds, and the DXN emission concentration is obtained by offline testing on a monthly/season cycle, so there is N ⁇ M.
  • the present invention proposes a DXN emission concentration soft measurement method based on potential feature SEN modeling, including potential feature extraction and primary selection modules, potential feature measurement and reselection modules, and adaptive selective integrated modeling modules, such as As shown in Figure 2.
  • MI mutual information
  • PCA Potential feature extraction and primary selection module
  • MI Mutual information
  • Adaptive selective integrated modeling module adopting hyperparameter adaptive selection strategy to construct sub-models with the best prediction performance for different subsystems, combining branch and bound (BB) and prediction error information entropy weighting algorithm to adapt
  • BB branch and bound
  • prediction error information entropy weighting algorithm to adapt The purpose of selecting sub-models and calculating their weighting coefficients is to select potential features with good redundancy and complementary relationship to construct sub-models for fusion, so as to improve the prediction performance of the SEN soft sensor model.
  • PCA is used to extract the latent features of high-dimensional input process variables. After normalizing the input data X i with zero mean and 1 variance, it is decomposed into,
  • T means transpose, Indicates the number of potential features extracted for the ith-th subsystem.
  • the calculation formula is as follows:
  • the threshold selected based on experience is denoted as ⁇ Contri , and its default value is 1.
  • the value of 1 indicates that the latent feature is selected for the first time.
  • the potential features obtained in the previous step are extracted in an unsupervised manner, and the features contained in the same subsystem are independent of each other, but the correlation between these features and the DXN emission concentration is not considered, that is, the high contribution rate Latent features are not necessarily strongly correlated with DXN. Still taking the ith subsystem as an example, the potential features of each primary selection The Mutual Information (MI) value between DXN emission concentration and Use the following formula to calculate,
  • the threshold is adaptively determined according to the prediction performance of the soft sensor model.
  • Upper limit of the threshold lower limit And fixed step Use the following formula to calculate,
  • the functions max( ⁇ ) and min( ⁇ ) represent the maximum and minimum values respectively; Represents the number of candidate thresholds determined based on experience, and its default value is 10.
  • the value of 1 indicates that the latent feature is selected again.
  • the potential features will be selected By mapping Transform to a high-dimensional feature space, and then solve the following optimization problem,
  • w i represents the weight coefficient
  • b i represents the offset
  • DXN emission concentration sub-model based on LS-SVM can be expressed as,
  • the hyperparameter adaptive selection mechanism of the above emission concentration sub-model is realized by the following two-step method:
  • the grid search strategy is adopted to take the prediction performance of the sub-model as the objective function, and the initial hyperparameter pair is adaptively selected in the candidate hyperparameter matrix M para
  • the hyperparameter matrix M para is shown below,
  • Step 2 based on the above method Use the following formula to obtain a new candidate hyperparameter set,
  • N ker and N reg indicate the number of new hyperparameters set based on experience; with In order to set the hyperparameter shrinkage and expansion factors based on experience, the default values are both 10.
  • the hyperparameter pair of the ith sub-model is adaptively obtained by adopting the grid search strategy again
  • the set of prediction outputs of the sub-model can be expressed as,
  • f i ( ⁇ ) represents the ith submodel.
  • the above sub-models are adaptively optimized and selected and weighted coefficients are calculated.
  • the best sub-model selection is similar to the weighted optimal feature selection [29].
  • SEN models with an integrated size of 2 to (I-1) can be obtained, and these optimized SEN models are finally sorted and the one with the best prediction performance is used as the final DXN Soft measurement model.
  • the modeling data in this invention comes from the No. 1 furnace of a grate furnace-based MSWI incineration company in Beijing, covering the available DXN emission concentration test samples recorded from 2012 to 2018, the number of which is 39; the corresponding input variables
  • the dimension of is 286 dimensions (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 dimensionality reduction is very necessary.
  • the invention divides the modeling data into two parts, which are used for training and testing respectively.
  • the six subsystems of incineration, boiler, flue gas treatment, steam power generation, flue gas emission, and utility assistance are respectively marked as Incinerator, Boiler, Flue gas, Steam, Stack, and Common.
  • the present invention also uses the MSWI system containing all the variables as a special subsystem for analysis and modeling. Therefore, the present invention contains a total of 7 subsystems.
  • Figure 3 shows that the contribution rate of the first 6 PCs reaches 80%, and the contribution rate extracted by the latent variables of different subsystems is different.
  • the MI method is used to measure the mapping relationship between the primary potential features extracted for different subsystems and the DXN, as shown in Figure 4.
  • Figure 4 shows: (1) The MI value of the first potential feature that can characterize the largest change in the process variable of the subsystem selected by all subsystems is the smallest, indicating that the correlation between these potential features and the DXN emission concentration is relatively high. Weak; (2) Except for the first potential feature, although the contribution rate of other potential features in characterizing process variables is gradually decreasing, there is no obvious rule to follow in characterizing the MI value; (3) In terms of mechanism Analysis shows that the Incinerator, Flue gas, and Stack subsystems are most related to the generation, absorption and emission of DXN, but the MI values of the extracted potential features of these subsystems are less different from other subsystems. It can be seen that there are limitations to making decisions based on MI values alone. Table 2 shows the maximum and minimum MI values of potential features of different subsystems.
  • Table 2 shows: (1) For the maximum value set: the maximum value is derived from the Common Subsystem which is theoretically not directly related to DXN emissions, with a value of 0.8613. Whether it is reasonable or not depends on further combining the model prediction results. Verification; ranked second is the Incinerator subsystem, with a value of 0.8559. This latent variable is theoretically related to the generation of DXN and is relatively reasonable; (2) For the minimum set: the minimum value comes from incineration (MSWI) subsystem, only 0.4429, indicating that separate analysis for different subsystems is still necessary; the maximum value comes from the stack subsystem, which is 0.7182, because other exhaust gases are different from DXN If there is a correlation, this value is more reasonable.
  • MSWI incineration
  • the upper limit of the MI threshold is 0.7882
  • the lower limit is 0.7182
  • the step size is 0.006999.
  • the final threshold is 0.7882
  • the number of potential features and the MI value are shown in Table 3.
  • the sets of candidate regularization parameters and kernel parameters are respectively preselected 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 ⁇ .
  • the number of input features for the sub-models of incineration, boiler, flue gas treatment, steam power generation, flue gas emission, utilities, and MSWI are 5, 2, 1, 3, 2, 6, and 1, respectively.
  • the curves for the first and second time of hyperparameter adaptive optimization using the grid search method are shown in Figure 5.
  • the 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 ⁇ and ⁇ 1362400,158.5 ⁇
  • the root mean square error (RMSE) of the corresponding test data is 0.01676, 0.02302, 0.01348, 0.01943, 0.01475, 0.02261 and 0.02375 respectively.
  • test errors of the SEN model constructed when the integrated size is 2-6 are 0.01345, 0.01332, 0.01401, 0.01460 and 0.01560, respectively.
  • the final integration size of the DXN soft-sensing model is 3.
  • the sub-models corresponding to the selected sub-models are flue gas treatment, flue gas emission and incineration. Theoretically, these three subsystems are related to the absorption, emission and generation of DXN. Judging from the results of the present invention, the effectiveness of all algorithms is verified, and the availability of data is also indicated.
  • Table 4 shows the comparison with the usual PLS single model, PCA-LSSVM single model and different weighting methods.
  • Table 1 shows that the prediction performance of the DXN single model based on PLS and PCA-LSSVM constructed with all process variables is weaker than the SEN modeling method proposed in the present invention, indicating that the strategy of constructing a SEN model based on source features is effective;
  • the method of integrating all sub-models uses the EN method that uses PLS weighting which is stronger than other integrated sub-models, indicating that the PLS algorithm is better in eliminating the collinearity of the sub-models; in addition, the sub-models selected by the SEN model correspond to the They are all related to the generation, absorption and emission mechanism of DXN, indicating the availability of modeling data and the effectiveness of the algorithm.
  • the invention is based on the industrial process data of a certain incineration company in Beijing, adopts PCA and prior knowledge-based potential feature extraction and primary selection, MI and prior knowledge-based primary selection potential feature measurement and selection, and self-adaptation for reselection potential features SEN modeling mechanism, a soft measurement of DXN emission concentration based on multi-source potential feature SEN modeling is proposed, and the simulation verifies the effectiveness of the proposed method.
  • the adaptive adjustment of contribution rate thresholds, MI thresholds, hyperparameters and SEN model structure in combination with the prediction of business trips of the soft-sensing model remains to be studied in depth.
  • the analysis of the DXN emission mechanism needs to be carried out in depth.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Software Systems (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Medical Informatics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Business, Economics & Management (AREA)
  • Human Resources & Organizations (AREA)
  • General Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Automation & Control Theory (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Marketing (AREA)
  • Operations Research (AREA)
  • Quality & Reliability (AREA)
  • Strategic Management (AREA)
  • Tourism & Hospitality (AREA)
  • General Business, Economics & Management (AREA)
  • Economics (AREA)
  • Sustainable Development (AREA)
  • Computer Hardware Design (AREA)
  • Geometry (AREA)
  • Feedback Control In General (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

一种基于多源潜在特征选择性集成(SEN)建模的DXN排放浓度软测量方法。首先，将MSWI过程数据依据工业流程划分为不同来源的子系统，采用主元分析(PCA)分别提取其潜在特征，并依据经验预设的主元贡献率阈值进行多源潜在特征初选；接着，采用互信息(MI)度量初选的潜在特征与DXN间的相关性，自适应确定潜在特征再选的上下限及阈值；最后，基于再选潜在特征，采用具有超参数自适应选择机制的最小二乘-支撑向量机(LS-SVM)算法，建立针对不同子系统的DXN排放浓度子模型，采用基于分支定界(BB)和预测误差信息熵加权算法的策略优化选择子模型和计算权系数，构建DXN排放浓度SEN软测量模型。

Description

一种城市固废焚烧过程二噁英排放浓度软测量方法
本发明由科学技术部国家重点研发计划(No:2018YFC1900801)和国家自然科学基金(No:61573364，61873009)资助。
技术领域
本发明属于固废焚烧技术领域，尤其涉及一种城市固废焚烧过程二噁英排放浓度软测量方法。
背景技术
基于运行优化控制策略降低复杂工业过程的能源消耗和污染排放，是国内外众多流程工业企业所面临的急需解决的难题[1,2,3]。焚烧是进行城市固废(MSW)处理的主要技术手段[4]。对于发展中国家的MSW焚烧(MSWI)企业，最为紧迫的问题是如何控制焚烧造成的污染排放[5,6]，其中急需控制得是：造成焚烧建厂“邻避效应”、具有生物累积效应等特点的剧毒物质二噁英(DXN)的排放浓度[7,8,9]。MSWI企业最主要关注如何基于优化的运行参数实现最小化的DXN排放[10]。目前，除采用先进的尾气处理装置外，工业过程普遍采用如下间接策略控制DXN排放，即“3T1E”准则[11]，即：焚烧炉内高于850℃的温度(T)、超过2秒(T)的烟气停留时间、较大的湍流程度(T)和合适的过量空气系数(E)。当前MSWI企业不能进行以降低DXN排放为直接目标的运行优化和反馈控制，其主要原因：一是DXN排放浓度的机理模型难以构建[12]，二是以月或季为周期的离线直接检测不能提供实时反馈的DXN排放浓度[13]。近年来的研究热点是基于指示物/关联物对DXN排放进行在线间接检测[14,15]，但这些方法固有的设备复杂、造价昂贵、检测滞后等原因导致其仍难以用于MSWI过程的运行优化和反馈控制。
数据驱动软测量技术可用于需要离线化验的难以检测参数(如本发明中的二噁英)的在线估计[16,17]。针对MSWI过程，文献[18,19]通过采用多年前欧美研究机构针对不同类型的焚烧炉采集的小样本数据，基于线性回归、人工神经网络(ANN)等算法构建DXN排放浓度软测量模型。近年来，文献[20]结合相关性分析、主元分析(PCA)和ANN等算法，基于MSWI过程数据构建DXN排放预测模型；但是，ANN并不适用于构建DXN浓度排放模型，主要原因在于其固有的易陷入局部最小、易过拟合和对小样本数据建模泛化性能差等缺点。采用具有合适超参数的支撑向量机(SVM)算法可有效地用于小样本数据建模[21]。针对SVM需求解二次规划的问题，最小二乘-支撑向量机(LS-SVM)通过求解线性等式予以克服；模型超参数可以通过单目标或多目标优化求解算法获得[22,23,24]，但这些方法耗时且只能得到次优解[25]。因此，目前研究缺少对超参数进行有效选择的自适应机制。
通常，MSWI过程包括固废储运、固废焚烧、蒸汽发电和烟气处理等阶段组成的多个子系统，所涉及的过程变量达到数百维，与DXN的产生、吸收和再合成等机理过程在不同程度上具有相关性。文献[26]指出，模型输入维数的增加和低价值训练样本的增加使得完备数量的训练样本难以获得。在模式识别领域，通常认为训练样本的数量与特征之比应该为2、5或10。文献[27]定义了维数约简后的训练样本与约简特征之比，认为该值应满足构建鲁棒预测学习模型的要求。因此，针对具有小样本高维特性的DXN排放浓度建模数据，进行维数约简是非常必要的。基于非监督特征提取的方法获得的特征，虽蕴含原始高维变量中的主要变化却可能导致所提取特征与被预测参数无关。能够提取高维数据蕴含变化的主元分析(PCA)，是目前工业过程在难以检测参数软测量中最为常用的潜在特征提取方法[28]，但贡献率低的主元会导致预测稳定性差[ 29]。
从另外一个视角，针对MSWI过程的不同子系统所提取的特征可视为源于多个视图的多源信息。理论和经验分析表明，面向多源信息采用选择性集成(SEN)学习机制构建的软测量模型具有更佳的预测稳定性和鲁棒性，其中子模型间的差异性尤为重要[29]。文献[30]综述了集成子模型多样性的构造策略，指出训练样本重采样包括划分训练样本(样本空间)、划分或变换特征变量(特征空间)等，在模型预测性能上基于特征空间的集成构造策略优于基 于多分类器的构造策略。针对小样本多源高维谱数据，汤等人提出基于选择性融合多源特征和多工况样本的SEN潜结构映射模型[31,32]；文献[33]和[32]提出了基于随机采样样本空间的SEN神经网络模型和潜结构映射模型，文献[34]提出基于子空间的集成学习通用框架，文献[35]提出了在特征子空间内随机采样样本空间的面向多尺度机械信号的双层SEN潜结构映射模型，文献[36]提出的SEN神经网络模型分别构建优化视角的候选子模型和优化选择集成子模型及其权重，但上述方法未进行模型参数自适应机制的研究。综上可知，以度量和再选后的非监督潜在特征为输入，基于自适应超参数选择机制的SEN-LSSVM建模策略及其在DXN排放浓度软测量中的研究未见报道。
发明内容
综上，本发明提供一种基于多源潜在特征SEN的MSWI过程DXN排放浓度软测量方法，首先，采用潜在特征提取与初选模块将MSWI过程数据依据工业流程划分为不同来源的子系统，采用主元分析(PCA)分别提取其潜在特征并依据经验预设的主元贡献率阈值进行多源潜在特征初选。接着，潜在特征度量与再选模块采用互信息(MI)度量初选的潜在特征与DXN间的相关性，自适应确定潜在特征再选的上下限及阈值。最后，自适应选择性集成建模模块基于再选潜在特征，采用具有超参数自适应选择机制的最小二乘-支撑向量机(LS-SVM)算法，建立针对不同子系统的DXN排放浓度子模型，采用基于分支定界(BB)和预测误差信息熵加权算法的策略优化选择子模型和计算权系数，构建DXN排放浓度SEN软测量模型。
附图说明
图1基于DXN视角的MSWI过程描述；
图2基于潜在特征SEN建模的DXN排放浓度软测量策略；
图3不同子系统的前6个PC的累积贡献率；
图4(a)Incinerator子系统初选潜在特征与DXN间的MI值；
图4(b)Boiler子系统初选潜在特征与DXN间的MI值；
图4(c)Flue gas子系统初选潜在特征与DXN间的MI值；
图4(d)Steam子系统初选潜在特征与DXN间的MI值；
图4(e)Stack子系统初选潜在特征与DXN间的MI值；
图4(f)Common子系统初选潜在特征与DXN间的MI值；
图4(g)MSWI系统初选潜在特征与DXN间的MI值；
图5(a)焚烧子模型超参数自适应寻优的第1次和第2次的曲线；
图5(b)锅炉子模型超参数自适应寻优的第1次和第2次的曲线；
图5(c)烟气处理子模型超参数自适应寻优的第1次和第2次的曲线；
图5(d)蒸汽发电子模型超参数自适应寻优的第1次和第2次的曲线；
图5(e)烟气排放子模型超参数自适应寻优的第1次和第2次的曲线；
图5(f)公用工程子模型超参数自适应寻优的第1次和第2次的曲线；
图5(g)MSWI全流程子模型超参数自适应寻优的第1次和第2次的曲线。
具体实施方试
面向DXN排放过程的MSWI描述
MSWI的主要设备包括焚烧炉、移动炉排、废锅和尾气处理等设备，其中：焚烧炉将MSW转化为残渣、灰尘、烟气与热量，位于焚烧炉底部的移动炉排促使MSW有效和完全燃烧，废锅产生的蒸汽用于推动汽轮机产生电力，烟气中的灰尘和污染物通过尾气处理设备净化后排入大气。其过程如图1所示。
由图1可知，从污染排放视角，MSWI过程包含了DXN产生、吸收和排放共3个阶段，其分别包含在标记为G1、G2和G3的烟气中。显然这些不同阶段的烟气中所包含的DXN浓度具有差异性。通常，为保证焚烧炉内的有害物质能够有效和完全分解，烟气温度至少应该达到850℃并保持2秒以上。用于焚烧MSW的一次风从炉排底部喷入的同时又对炉排进行冷却，并通过引入湍流和保证过量氧的供应使得二次风能够辅助进行烟气的完全燃烧。在烟气冷却过程中，进行焚烧矿渣和废锅底灰的处置和收集，同时排出烟气G1。活性炭和石灰被注入反应器，用于移除酸性气体和吸收DXN及一些重金属，烟气然后再进入袋式过 滤器。在反应器和袋式过滤器内产生的飞灰被注入混涅设备，同时产生烟气G2。引风机将烟气G2吸入到烟囱，进而作为烟气G3排放至空气中，包含HCL、SO 2、NOx和HF等多种能够实时在线检测的污染物浓度，以及具有长周期、高成本等特点需离线化验浓度的DXN。由上述描述可知，DXN排放浓度与MSWI过程不同阶段的易检测过程变量均具有相关性。
由图1可知，MSWI过程可以分为焚烧、锅炉、烟气处理、蒸汽发电、烟气排放、公共工程辅助共6个子系统。对于DXN排放浓度建模，本发明将多个子系统视为多源信息。
建模策略
本发明中，模型输入数据X∈R N×M包括N个样本(行)和M个变量(列)，其源于MSWI过程的不同子系统。将来自第ith个子系统的建模数据表示为
Figure PCTCN2019122326-appb-000001
即存在如下关系，
Figure PCTCN2019122326-appb-000002
Figure PCTCN2019122326-appb-000003
其中，I表示子系统个数，M i表示第ith个子系统包含的变量个数。相应的，输出数据
Figure PCTCN2019122326-appb-000004
包括N个样本(行)，其来源于离线化验的DXN排放浓度检测数据。显然，输入/输出数据在时间尺度上具有较大的差异性：过程变量以秒为单位在DCS系统采集与存储，DXN排放浓度以月/季为周期离线化验获得，故存在N＜＜M。
依据上述情况，本发明提一种基于潜在特征SEN建模的DXN排放浓度软测量方法，包括潜在特征提取与初选模块、潜在特征度量与再选模块、自适应选择性集成建模模块，如图2所示。
在图2中，
Figure PCTCN2019122326-appb-000005
表示从第ith个子系统所采集的全部过程变量；
Figure PCTCN2019122326-appb-000006
表示针对第ith个子系统的全部过程变量采用PCA提取的数量为
Figure PCTCN2019122326-appb-000007
的全部潜在特征；
Figure PCTCN2019122326-appb-000008
表示针对第ith个子系统的全部潜在特征依据设定阈值θ Contri选择的数量为
Figure PCTCN2019122326-appb-000009
的初选潜在特征；
Figure PCTCN2019122326-appb-000010
表示针对第ith个子系统的初选潜在特征与DXN进行互信息(MI)度量后，基于阈值θ MI选择的数量为
Figure PCTCN2019122326-appb-000011
的再选潜在特征；
Figure PCTCN2019122326-appb-000012
和
Figure PCTCN2019122326-appb-000013
表示为第ith个基于LS-SVM的子模型所选择的核参数和正则化参数，即超参数对，本发明将其记为
Figure PCTCN2019122326-appb-000014
表示第ith个子模型的预测输出；y和
Figure PCTCN2019122326-appb-000015
表示DXN排放浓度软测量模型的真值和预测输出。
上述模块的功能是：
(1)潜在特征提取与初选模块：采用PCA提取从不同子系统所采集的过程变量的全部潜在特征，基于依据经验设定的潜在特征贡献率阈值获得多源初选潜在特征，其目的是防止较小贡献率的潜在特征造成模型预测性能不稳定性。
(2)潜在特征度量与再选模块：采用互信息(MI)度量不同子系统的初选潜在特征与DXN排放浓度间的关系，并结合基于软测量模型预测性能自适应确定的阈值获得再选潜在特征，其目的是使得所选的多源潜在特征与DXN排放浓度间具有较好的映射关系。
(3)自适应选择性集成建模模块：采用超参数自适应选择策略构建面向不同子系统的具有最佳预测性能的子模型，结合分支定界(BB)和预测误差信息熵加权算法自适应地选择子模型和计算其加权系数，其目的是选择具有较好冗余与互补关系的潜在特征构建子模型进行融合，以提高SEN软测量模型的预测性能。
潜在特征提取与初选模块
以第ith个子系统为例，首先采用PCA提取高维输入过程变量的潜在特征。将输入数据X i进行零均值1方差的标准化后，将其分解为，
Figure PCTCN2019122326-appb-000016
其中，
Figure PCTCN2019122326-appb-000017
和
Figure PCTCN2019122326-appb-000018
表示第
Figure PCTCN2019122326-appb-000019
个主元(PC)的得分和负载向量，T表示转置，
Figure PCTCN2019122326-appb-000020
表示对第ith个子系统所提取的潜在特征数量，其计算公式如下，
Figure PCTCN2019122326-appb-000021
基于上述表达式，从数据X i所提取的全部潜在特征可表示为，
Figure PCTCN2019122326-appb-000022
其中，
Figure PCTCN2019122326-appb-000023
表示得分矩阵，是数据X i在负载矩阵P i方向上的正交映射；P i采用如下公式表示，
Figure PCTCN2019122326-appb-000024
其中，
Figure PCTCN2019122326-appb-000025
因此，从数据X i所提取的潜在特征可表示为，
Figure PCTCN2019122326-appb-000026
其中，
Figure PCTCN2019122326-appb-000027
进一步，全部潜在特征可表示为，
Figure PCTCN2019122326-appb-000028
研究表明，采用贡献率较小的潜在变量建模会导致模型预测性能的不稳定。此处，将与第
Figure PCTCN2019122326-appb-000029
个负载向量
Figure PCTCN2019122326-appb-000030
对应的特征向量记为
Figure PCTCN2019122326-appb-000031
相应的第
Figure PCTCN2019122326-appb-000032
个潜在特征
Figure PCTCN2019122326-appb-000033
的贡献率
Figure PCTCN2019122326-appb-000034
采用如下公式计算，
Figure PCTCN2019122326-appb-000035
将依据经验选择的阈值记为θ Contri，其默认取值为1。采用如下规则对全部潜在特征进行初次选择，
Figure PCTCN2019122326-appb-000036
其中，
Figure PCTCN2019122326-appb-000037
表示第
Figure PCTCN2019122326-appb-000038
个潜在特征是否被选中的标记值，其值为1表示该潜在特征被初次选中。
因此，将针对第ith个子系统的初选潜在特征表示为，
Figure PCTCN2019122326-appb-000039
进一步，全部初选潜在特征Z FeSe1st可表示为，
Figure PCTCN2019122326-appb-000040
潜在特征度量与再选模块
上一步骤中所获得的初选潜在特征是采用非监督方式提取，并且同一子系统所包含的特征是相互独立的，但未考虑这些特征与DXN排放浓度间的相关性，即贡献率高的潜在特征并不一定与DXN间的相关性强。仍以第ith个子系统为例，将每个初选潜在特征
Figure PCTCN2019122326-appb-000041
与DXN排放浓度间的互信息(MI)值标记为
Figure PCTCN2019122326-appb-000042
采用如下公式计算，
Figure PCTCN2019122326-appb-000043
其中，
Figure PCTCN2019122326-appb-000044
和p prob(y)表示
Figure PCTCN2019122326-appb-000045
和y的边际概率密度；
Figure PCTCN2019122326-appb-000046
表示联合概率密度；
Figure PCTCN2019122326-appb-000047
表示条件熵，H(y)表示信息熵。
依据软测量模型的预测性能自适应确定阈值。阈值的上限值
Figure PCTCN2019122326-appb-000048
下限值
Figure PCTCN2019122326-appb-000049
及固定步长
Figure PCTCN2019122326-appb-000050
采用如下公式计算，
Figure PCTCN2019122326-appb-000051
Figure PCTCN2019122326-appb-000052
Figure PCTCN2019122326-appb-000053
其中，函数max(·)和min(·)分别表示取最大值和最小值；
Figure PCTCN2019122326-appb-000054
表示依据经验确定的候选阈值的数量，其默认值为10。
将选定的阈值记为θ Contri，其值在
Figure PCTCN2019122326-appb-000055
和
Figure PCTCN2019122326-appb-000056
之间以DXN软测量模型的预测性能为准则进行自适应选择。
采用如下规则对初选的潜在特征进行再次选择，
Figure PCTCN2019122326-appb-000057
其中，
Figure PCTCN2019122326-appb-000058
表示第
Figure PCTCN2019122326-appb-000059
个潜在特征是否被选中的标记值，其值为1表示该潜在特征被再次选中。
进一步，将针对第ith个子系统的再选潜在特征表示为，
Figure PCTCN2019122326-appb-000060
因此，全部再选潜在特征Z FeSe2nd可表示为，
Figure PCTCN2019122326-appb-000061
自适应选择性集成建模模块
以第ith个子系统为例，描述基于再选潜在特征
Figure PCTCN2019122326-appb-000062
和模型超参数对
Figure PCTCN2019122326-appb-000063
构建DXN排放浓度子模型的过程。
首先，将再选潜在特征
Figure PCTCN2019122326-appb-000064
通过映射
Figure PCTCN2019122326-appb-000065
变换到高维特征空间，然后求解如下优化问题，
Figure PCTCN2019122326-appb-000066
其中，w i表示权重系数，b i表示偏置，
Figure PCTCN2019122326-appb-000067
是第nth个样本的预测误差。
采用拉格朗日方法，可得如下公式，
Figure PCTCN2019122326-appb-000068
其中，
Figure PCTCN2019122326-appb-000069
表示拉格朗日算子向量，
Figure PCTCN2019122326-appb-000070
表示预测误差向 量。
对上述公式进行求解，
Figure PCTCN2019122326-appb-000071
将所采用的核函数表示如下，
Figure PCTCN2019122326-appb-000072
进一步，将LS-SVM问题转换为求解以下线性等式系统，
Figure PCTCN2019122326-appb-000073
通过求解上述公式，得到β i和b i。
进而，基于LS-SVM构建的DXN排放浓度子模型可表示为，
Figure PCTCN2019122326-appb-000074
上述排放浓度子模型的超参数自适应选择机制采用下述的两步法实现：
第1步，采用网格搜索策略以子模型的预测性能为目标函数，在候选超参数矩阵M para中自适应选择初始超参数对
Figure PCTCN2019122326-appb-000075
超参数矩阵M para如下所示，
Figure PCTCN2019122326-appb-000076
其中，k＝1,…,K,K表示候选核参数的数量；r＝1,…,R,R表示候选惩罚参数的数量；
Figure PCTCN2019122326-appb-000077
表示由第kth个核参数和第rth个惩罚参数组成的超参数对，也超参数矩阵M para中的第jth个参数对，即存在
Figure PCTCN2019122326-appb-000078
j＝1,…,J,J＝K×R表示超参数矩阵M para中的全部超参数对的个数。因此，对于初次采用网格搜索策略所选择的超参数对
Figure PCTCN2019122326-appb-000079
是矩阵M para中一个元素，即存在
Figure PCTCN2019122326-appb-000080
第2步，基于上述方法选择的
Figure PCTCN2019122326-appb-000081
采用如下公式获得新的候选超参数集合，
Figure PCTCN2019122326-appb-000082
Figure PCTCN2019122326-appb-000083
其中，
Figure PCTCN2019122326-appb-000084
和
Figure PCTCN2019122326-appb-000085
表示新的候选超参数集合，分别对应核参数向量和惩罚参数向量；N ker和N reg表示依据经验设定的新的超参数的数量；
Figure PCTCN2019122326-appb-000086
和
Figure PCTCN2019122326-appb-000087
为依据经验设定的超参数收缩和扩放因子，其默认值均为10。
通过再次采用网格搜索策略自适应获得第ith个子模型的超参数对
Figure PCTCN2019122326-appb-000088
对全部子系统执行上述过程，子模型预测输出的集合可表示为，
Figure PCTCN2019122326-appb-000089
其中，f i(·)表示第ith个子模型。
结合基于分支定界(BB)的最优化选择算法和基于误差的信息熵加权算法对上述子模型进行自适应的优化选择和计算加权系数。在给定候选子模型和加权算法后，最佳子模型选择与加权类似最优特征选择[29]。面向有限数量的候选子模型，通过多次运行优化和加权算法，可获得集成尺寸为2到(I-1)的SEN模型，最后排序这些优化的SEN模型并将预测性能最佳的作为最终DXN软测量模型。
假定最终DXN软测量模型的集成尺寸为I sel，其预测输出值
Figure PCTCN2019122326-appb-000090
可由下式计算：
Figure PCTCN2019122326-appb-000091
其中，
Figure PCTCN2019122326-appb-000092
表示第i selth个优化选择的子模型，
Figure PCTCN2019122326-appb-000093
和
Figure PCTCN2019122326-appb-000094
表示其对应的加权系数及预测值，
Figure PCTCN2019122326-appb-000095
和
Figure PCTCN2019122326-appb-000096
表示子模型
Figure PCTCN2019122326-appb-000097
的超参数和输入特征。
对比式(29)可知，存在如下关系，
Figure PCTCN2019122326-appb-000098
利用子模型的预测值和真值，
Figure PCTCN2019122326-appb-000099
采用基于预测误差信息墒的加权算法得到，如下所示，
Figure PCTCN2019122326-appb-000100
其中，
Figure PCTCN2019122326-appb-000101
Figure PCTCN2019122326-appb-000102
其中，
Figure PCTCN2019122326-appb-000103
和
Figure PCTCN2019122326-appb-000104
表示第n th个样本基于第i selth个优化选择的子模型
Figure PCTCN2019122326-appb-000105
的预测值和相对预测误差，
Figure PCTCN2019122326-appb-000106
表示第i selth个优化选择的子模型的预测误差的信息熵。
应用研究
本发明中的建模数据源于北京某基于炉排炉的MSWI焚烧企业的1#炉，涵盖了2012～2018年所记录可用的DXN排放浓度检测样本，其数量为39个；相应的输入变量的维数为286维(包含了MSWI过程的全部过程变量)。可见，输入特征数量远远超过建模样本数量，进行维数约简非常有必要。本发明将建模数据等分为两部分，分别用于训练和测试。
本发明中，将6个子系统焚烧、锅炉、烟气处理、蒸汽发电、烟气排放和公用工程辅助分别的标记为Incinerator、Boiler、Flue gas、Steam、Stack和Common。为表示焚烧过程变量的整体变化特性，本发明将包含全部变量的MSWI系统也作为一个特殊的子系统进行分析和建模。因此，本发明共包含7个子系统。
潜在特征提取与初选结果
采用PCA提取的7个子系统的前6个潜在特征的累积贡献率如图3所示。
图3表明，前6个PC的贡献率达到了80％，并且不同子系统的潜在变量所提取的贡献率具有差异性。
基于单个PC的贡献率不小于1％的准则，初选潜在特征的主元个数及其贡献率如表1所示。
表1 初选潜在特征的主元个数及其贡献率
Figure PCTCN2019122326-appb-000107
由表1可知，不同子系统所初选的潜在特征的数量为13、6、9、8、6、12和15。由于PCA属于非监督特征提取方法，这些所提取特征仅是描述了输入数据的变化，其与DXN间的映射关系需要进行进一步的度量。
潜在特征度量与再选结果
采用MI方法度量针对不同子系统所提取的初选潜在特征与DXN间的映射关系，如图4所示。
图4表明：(1)全部子系统所初选的能够表征所在子系统过程变量最大变化的第1个潜在特征的MI值是最小的，表明了这些潜在特征与DXN排放浓度间的相关性较弱；(2)除第1个潜在特征外，其他潜在特征在表征过程变量的贡献率上虽然是逐渐下降的，但在表征MI值上却没有明显的规律可循；(3)从机理上分析，与DXN的产生、吸收和排放最为相关的是Incinerator、Flue gas和Stack子系统，但这些子系统在所提取的潜在特征的MI值与其他子系统的差别较小。可见，仅仅是基于MI值进行决策存在局限性。不同子系统的初选潜在特征的MI值的最大值和最小值如表2所示。
表2 不同子系统的初选潜在特征的MI值的最大值和最小值集合
Figure PCTCN2019122326-appb-000108
表2表明：(1)针对最大值集合：最大值源于理论上与DXN排放并无直接关系的公共辅助(Common)子系统，值为0.8613，其是否合理有待于进一步的结合模型预测结果进行验证；排在第2位的是焚烧(Incinerator)子系统，其值为0.8559，该潜在变量理论上与DXN的产生相关，是比较合理的；(2)针对最小值集合：最小值源于焚烧(MSWI)子系统，仅为0.4429，表明针对不同的子系统进行单独分析还是比较必要的；最大值源于烟气排放(Stack)子系统，其为0.7182，由于其他的排放气体与DXN间是存在相关性的，此值也是较为合理的。
由表2可知，MI阈值的上限为0.7882，下限为0.7182，步长为0.006999。结合阈值的上下限和步长，最终确定的阈值为0.7882，再选的潜在特征的数量和MI值如表3所示。
表3 再选潜在特征数量和MI值
Figure PCTCN2019122326-appb-000109
自适应选择集成建模
本发明中，将候选正则化参数与核参数的集合分别预先选择为{0.0001,0.001,0.01,0.1,1,10,100,1000,2000,4000,6000,8000,10000,20000,40000,60000,80000,160000}和{0.0001,0.001,0.01,0.1,1,10,100,1000,1600,3200,6400,12800,25600,51200,102400}。
结合上文可知，焚烧、锅炉、烟气处理、蒸汽发电、烟气排放、公用工程、MSWI全流程子模型的输入特征数量分别为5、2、1、3、2、6和1。采用网格搜索方法进行超参数自适应寻优的第1次和第2次的曲线如图5所示。
基于上述结果，上述子模型自适应选择的超参数对分别为{109,109}、{10000,25.75}、{5.950,0.0595}、{30.70,2.080}、{5.950,0.5950}、{1520800,22816}和{1362400,158.5}，对应的测试数据的均方根误差(RMSE)分别为0.01676、0.02302、0.01348、0.01943、0.01475、0.02261和0.02375。
采用基于BB和预测误差信息熵加权算法的寻优和加权策略，在集成尺寸为2～6时所构建的SEN模型的测试误差分别为0.01345、0.01332、0.01401、0.01460和0.01560。最终DXN软测量模型的集成尺寸为3，其选择的子模型所对应的子系统为烟气处理、烟气排放和焚烧，理论上这3个子系统与DXN的吸收、排放和生成相关。从本发明结果看，验证了所以算法的有效性，也表明了数据的可用性。
比较结果
与通常采用PLS单模型、PCA-LSSVM单模型以及不同加权方法的比较如表4所示。
表4 不同建模方法的结果统计
Figure PCTCN2019122326-appb-000110
表1表明，采用全部过程变量构建的基于PLS和PCA-LSSVM的DXN单模型的预测性能均弱于本发明所提的SEN建模方法，说明采用基于源特征构建SEN模型的策略是有效；同时，集成全部子模型的方法出采用PLS加权具有强于其他集成全部子模型的EN方法，表明PLS算法在消除子模型的共线性方面较好；此外，SEN模型选择的子模型所对应的子系统均与DXN的产生、吸收和排放机理相关，表明了建模数据的可用性和算法的有效性。
本发明基于北京某焚烧企业的工业过程数据，采用基于PCA和先验知识的潜在特征提取和初选、基于MI和先验知识的初选潜在特征度量和选择和面向再选潜在特征的自适应SEN建模机制，提出了基于多源潜在特征SEN建模的DXN排放浓度软测量，仿真验证了所提方法的有效性。在结合软测量模型的预测出差自适应调整贡献率阈值、MI阈值、超参数和SEN模型结构等方面还有待于深入研究。此外，结合DXN排放机理的分析也有待于深入进行。
参考文献
[1]柴天佑。复杂工业过程运行优化与反馈控制.自动化学报,39(11)(2013)1744-1757.
[2]T.Y.Chai,J.L.Ding,G.Yu,H.Wang.Integrated optimization for the automation systems of mineral processing[J].IEEE Transactions on Automation Science&Engineering,11(4)(2014)965-982.
[3]T.Y.Chai,S.J.Qin,H.Wang.Optimal operational control for complex industrial processes[J].Annu.Rev.Control,38(1)(2014)81-92.
[4]Arafat H A,Jijakli K,Ahsan A,2015.Environmental performance and energy recovery potential of five processes for municipal solid waste treatment[J].Journal of Cleaner Production,2015,105:233-240
[5]Yuanan H,Hefa C,Shu T.The growing importance of waste-to-energy(WTE)incineration in China's anthropogenic mercury emissions:Emission inventories and reduction strategies[J].Renewable and Sustainable Energy Reviews,2018,97:119-137.
[6]Huang T,Zhou L,Liu L,Xia M.Ultrasound-enhanced electrokinetic remediation for removal of Zn,Pb,Cu and Cd in municipal solid waste incineration fly ashes[J].Waste Management,2018，75:226-235.
[7]P.H.Jones,J.Degerlache,E.Marti,G.Mischer,M.C.Scherrer,W.J.Bontinck,H.J.Niessen,The global exposure of man to dioxins-a perspective on industrial-waste incineration[J],Chemosphere 26(1993)1491-1497.
[8]Li X,Zhang C,Li Y,Zhi Q.The Status of Municipal Solid Waste Incineration(MSWI)in China and its Clean Development[J].Energy Procedia,2016,104:498-503.
[9]Phillips K,Longhurst P J,Wagland S T.Assessing the perception and reality of arguments against thermal waste treatment plants in terms of property prices[J].Waste Management.2014,34(1):219-225.
[10]Zhang H J,Ni Y W,Chen J P,Zhang Q.Influence of variation in the operating conditions on PCDD/F distribution in a full-scale MSW incinerator[J].Chemosphere,2008,70(4):721-730.
[11]Mukherjee A,Debnath B,Ghosh S K.A Review on Technologies of Removal of Dioxins and Furans from Incinerator Flue Gas[J].Procedia Environmental Sciences,2016,35:528-540.
[12]B.R.Stanmore,Modeling the formation of PCDD/F in solid waste incinerators,Chemosphere 47(2002)565-773.
[13]乔俊飞，郭子豪，汤健.面向城市固废焚烧过程的二噁英排放浓度检测方法综述[J].自动化学报,2019(在审).
[14]M.Pandelova,D.Lenoir,and K.W.Schramm,Correlation between PCDD/F,PCB and PCBz in coal/waste combustion Influence of various inhibitors,Chemosphere 62(2006)1196-1205.
[15]B.K.Gullett,L.Oudejans,D.Tabor,A.Touati,and S.Ryan,Near-real-time combustion monitoring for PCDD/PCDF indicators by GC–REMPI–TOFMS[J][J],Environmental Engineering Science 46(2012)923-928.
[16]W.Wang,T.Y.Chai,W.Yu,Modeling component concentrations of sodium aluminate solution via hammerstein recurrent neural networks[J],IEEE Transactions on Control Systems Technology,20(2012)971-982.
[17]J.Tang,T.Y.Chai,W.Yu,L.J.Zhao,Modeling load parameters of ball mill in grinding process based on selective ensemble multisensor information[J],IEEE Transactions on Automation Science&Engineering,10(2013)726-740.
[18]Chang N B,Huang S H.Statistical modelling for the prediction and control of PCDDs and PCDFs emissions from municipal solid waste incinerators[J].Waste Management&Research,1995,13,379-400.
[19]Chang N B,Chen W C.Prediction of PCDDs/PCDFs emissions from municipal incinerators by genetic programming and neural network modeling[J].Waste Management&Research,2000,18(4)41-351.
[20]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[J].Chemosphere,2013,92:258-264.
[21]T.A.F.Gomes,R.B.C.Prudêncio,C.Soares,A.L.D.Rossi,A.Carvalho,Combining meta-learning and search techniques to select parameters for support vector machines[J].Neurocomputing 75(1)(2012)3-13.
[22]J.Tang,T.Y.Chai,W.Yu and L.J.Zhao,Feature extraction and selection based on vibration spectrum with application to estimate the load parameters of ball mill in grinding process[J],Control Engineering Practice,20(10)(2012)991-1004.
[23]C.Soares,.A hybrid meta-learning architecture for multi-objective optimization of svm parameters[J].Neurocomputing,143(143)(2014)27-43.
[24]G.Yu,T.Y.Chai,X.C.Luo.Multiobjective production planning optimization using hybrid evolutionary algorithms for mineral processing[J].IEEE Transact.Evolut.Comput.15(4)(2011)487-514.
[25]S.Yin,J.Yin,Tuning kernel parameters for svm based on expected square distance ratio[J].Information Sciences,370-371(2016)92-102.
[26]Li D C,Liu C W.Extending attribute information for small data set classication.IEEE Transactions on Knowledge and Data Engineering,2010,24(3):452-464
[27]汤健,乔俊飞,柴天佑,等.基于虚拟样本生成技术的多组分机械信号建模[J].自动化学报,2018,44(9):1569-1589.
[28]Jian Tang,Zhuo liu,Jian Zhang,Zhiwei Wu,Tianyou Chai,Wen Yu.Kernel latent feature adaptive extraction and selection method for multi-component non-stationary signal of industrial mechanical device[J],Neurocomputing,2016,216(C):296-309.
[29]汤健，田福庆，贾美英，李东。基于频谱数据驱动的旋转机械设备负荷软测量[M],国防工业出版社，2015年6月，北京
[30]Gavin Brown,Jeremy Wyatt,Rachel Harris,Xin Yao.Diversity creation methods:a survey and categorisation[J].Information Fusion 6(2005)5–20
[31]Tang Jian,Chai Tian-you,Yu Wen,Zhao Li-jie.Modeling Load Parameters of Ball Mill in Grinding Process Based on Selective Ensemble Multisensor Information[J].IEEE Tranzaction on Automation Science and Engineering,2013,10(3):726-740.
[32]Jian Tang,Tianyou Chai,Wen Yu,Zhuo Liu,Xiaojie Zhou.A Comparative study that measures ball mill load parameters through different single-scale and multi-scale frequency spectra-based approaches[J],IEEE Transactions on Industrial Informatics,2016,12(6):2008-2019.
[33]Zhou ZH,Wu J and Tang W,Ensembling neural networks:many could be better than all,”Artificial Intelligence[J],2002，137(1-2):239-263.
[34]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[J].Pattern Recognition,2017.
[35]Jian Tang,Junfei Qiao,ZhiWei Wub,Tianyou Chai,Jian Zhang,Wen Yu.Vibration and acoustic frequency spectra for industrial process modeling using selective fusion multi-condition  samples and multi-source features.Mechanical Systems and Signal Processing,99,1-27,2018.
[36]S.Soares,C.H.Antunes,R.Araújo,“Comparison of a genetic algorithm and simulated annealing for automatic neural network ensemble development,”Neurocomputing,Vol.21,No.9,pp.498-511,May.2013.

Claims (4)

  1. 一种城市固废焚烧过程二噁英排放浓度软测量方法，其特征在于，首先，采用潜在特征提取与初选模块将MSWI过程数据依据工业流程划分为不同来源的子系统，采用主元分析(PCA)分别提取其潜在特征，并依据经验预设的主元贡献率阈值进行多源潜在特征初选；接着，潜在特征度量与再选模块采用互信息(MI)度量初选的潜在特征与DXN间的相关性，自适应确定潜在特征再选的上下限及阈值；最后，自适应选择性集成建模模块基于再选潜在特征，采用具有超参数自适应选择机制的最小二乘‐支撑向量机(LS‐SVM)算法，建立针对不同子系统的DXN排放浓度子模型，采用基于分支定界(BB)和预测误差信息熵加权算法的策略优化选择子模型和计算权系数，构建DXN排放浓度SEN软测量模型。
  2. 如权利要求1所述的城市固废焚烧过程二噁英排放浓度软测量方法，其特征在于，潜在特征提取与初选模块的工作过程如下：以第ith个子系统为例，首先采用PCA提取高维输入过程变量的潜在特征。将输入数据X i进行零均值1方差的标准化后，将其分解为，
    Figure PCTCN2019122326-appb-100001
    其中，
    Figure PCTCN2019122326-appb-100002
    和
    Figure PCTCN2019122326-appb-100003
    表示第
    Figure PCTCN2019122326-appb-100004
    个主元(PC)的得分和负载向量，T表示转置，
    Figure PCTCN2019122326-appb-100005
    表示对第ith个子系统所提取的潜在特征数量，其计算公式如下，
    Figure PCTCN2019122326-appb-100006
    基于上述表达式，从数据X i所提取的全部潜在特征可表示为，
    Figure PCTCN2019122326-appb-100007
    其中，
    Figure PCTCN2019122326-appb-100008
    表示得分矩阵，是数据X i在负载矩阵P i方向上的正交映射；P i采用如下公式表示，
    Figure PCTCN2019122326-appb-100009
    其中，
    Figure PCTCN2019122326-appb-100010
    因此，从数据X i所提取的潜在特征可表示为，
    Figure PCTCN2019122326-appb-100011
    其中，
    Figure PCTCN2019122326-appb-100012
    进一步，全部潜在特征可表示为，
    Figure PCTCN2019122326-appb-100013
    将与第
    Figure PCTCN2019122326-appb-100014
    个负载向量
    Figure PCTCN2019122326-appb-100015
    对应的特征向量记为
    Figure PCTCN2019122326-appb-100016
    相应的第
    Figure PCTCN2019122326-appb-100017
    个潜在特征
    Figure PCTCN2019122326-appb-100018
    的贡献率
    Figure PCTCN2019122326-appb-100019
    采用如下公式计算，
    Figure PCTCN2019122326-appb-100020
    将依据经验选择的阈值记为θ Contri，其默认取值为1，采用如下规则对全部潜在特征进行初次选择，
    Figure PCTCN2019122326-appb-100021
    其中，
    Figure PCTCN2019122326-appb-100022
    表示第
    Figure PCTCN2019122326-appb-100023
    个潜在特征是否被选中的标记值，其值为1表示该潜在特征被初次选中，
    因此，将针对第ith个子系统的初选潜在特征表示为，
    Figure PCTCN2019122326-appb-100024
    进一步，全部初选潜在特征Z FeSe1st可表示为，
    Figure PCTCN2019122326-appb-100025
  3. 如权利要求2所述的城市固废焚烧过程二噁英排放浓度软测量方法，其特征在于，潜在特征度量与再选模块的工作过程如下：以第ith个子系统为例，将每个初选潜在特征
    Figure PCTCN2019122326-appb-100026
    与DXN排放浓度间的互信息(MI)值标记为
    Figure PCTCN2019122326-appb-100027
    采用如下公式计算，
    Figure PCTCN2019122326-appb-100028
    其中，
    Figure PCTCN2019122326-appb-100029
    和p prob(y)表示
    Figure PCTCN2019122326-appb-100030
    和y的边际概率密度；
    Figure PCTCN2019122326-appb-100031
    表示联合概率密度；
    Figure PCTCN2019122326-appb-100032
    表示条件熵，H(y)表示信息熵，
    依据软测量模型的预测性能自适应确定阈值。阈值的上限值
    Figure PCTCN2019122326-appb-100033
    下限值
    Figure PCTCN2019122326-appb-100034
    及固定步长
    Figure PCTCN2019122326-appb-100035
    采用如下公式计算，
    Figure PCTCN2019122326-appb-100036
    Figure PCTCN2019122326-appb-100037
    Figure PCTCN2019122326-appb-100038
    其中，函数max(·)和min(·)分别表示取最大值和最小值；
    Figure PCTCN2019122326-appb-100039
    表示依据经验确定的候选阈值的数量，其默认值为10，
    将选定的阈值记为θ Contri，其值在
    Figure PCTCN2019122326-appb-100040
    和
    Figure PCTCN2019122326-appb-100041
    之间以DXN软测量模型的预测性能为准则进行自适应选择，
    采用如下规则对初选的潜在特征进行再次选择，
    Figure PCTCN2019122326-appb-100042
    其中，
    Figure PCTCN2019122326-appb-100043
    表示第
    Figure PCTCN2019122326-appb-100044
    个潜在特征是否被选中的标记值，其值为1表示该潜在特征被再次选中，
    进一步，将针对第ith个子系统的再选潜在特征表示为，
    Figure PCTCN2019122326-appb-100045
    因此，全部再选潜在特征Z FeSe2nd可表示为，
    Figure PCTCN2019122326-appb-100046
  4. 如权利要求3所述的城市固废焚烧过程二噁英排放浓度软测量方法，其特征在于，自适应选择性集成建模模块的工作过程如下：
    以第ith个子系统为例，首先，将再选潜在特征
    Figure PCTCN2019122326-appb-100047
    通过映射
    Figure PCTCN2019122326-appb-100048
    变换到高维特征空间，然后求解如下优化问题，
    Figure PCTCN2019122326-appb-100049
    其中，w i表示权重系数，b i表示偏置，
    Figure PCTCN2019122326-appb-100050
    是第nth个样本的预测误差，
    采用拉格朗日方法，可得如下公式，
    Figure PCTCN2019122326-appb-100051
    其中，
    Figure PCTCN2019122326-appb-100052
    表示拉格朗日算子向量，
    Figure PCTCN2019122326-appb-100053
    表示预测误差向量。
    对上述公式进行求解，
    Figure PCTCN2019122326-appb-100054
    将所采用的核函数表示如下，
    Figure PCTCN2019122326-appb-100055
    进一步，将LS-SVM问题转换为求解以下线性等式系统，
    Figure PCTCN2019122326-appb-100056
    通过求解上述公式，得到β i和b i，
    进而，基于LS-SVM构建的DXN排放浓度子模型可表示为，
    Figure PCTCN2019122326-appb-100057
    上述排放浓度子模型的超参数自适应选择机制采用下述的两步法实现：
    第1步，采用网格搜索策略以子模型的预测性能为目标函数，在候选超参数矩阵M para中自适应选择初始超参数对
    Figure PCTCN2019122326-appb-100058
    超参数矩阵M para如下所示，
    Figure PCTCN2019122326-appb-100059
    其中，k＝1,…,K,K表示候选核参数的数量；r＝1,…,R,R表示候选惩罚参数的数量；
    Figure PCTCN2019122326-appb-100060
    表示由第kth个核参数和第rth个惩罚参数组成的超参数对，也超参数矩阵M para中的第jth个参数对，即存在
    Figure PCTCN2019122326-appb-100061
    j＝1,…,J,J＝K×R表示超参数矩阵M para中的全部超参数对的个数，因此，对于初次采用网格搜索策略所选择的超参数对
    Figure PCTCN2019122326-appb-100062
    是矩阵M para中一个元素，即存在
    Figure PCTCN2019122326-appb-100063
    第2步，基于上述方法选择的
    Figure PCTCN2019122326-appb-100064
    采用如下公式获得新的候选超参数集合，
    Figure PCTCN2019122326-appb-100065
    Figure PCTCN2019122326-appb-100066
    其中，
    Figure PCTCN2019122326-appb-100067
    和
    Figure PCTCN2019122326-appb-100068
    表示新的候选超参数集合，分别对应核参数向量和惩罚参数向量；N ker和N reg表示依据经验设定的新的超参数的数量；
    Figure PCTCN2019122326-appb-100069
    和
    Figure PCTCN2019122326-appb-100070
    为依据经验设定的超参数收缩和扩放因子，其默认值均为10，
    通过再次采用网格搜索策略自适应获得第ith个子模型的超参数对
    Figure PCTCN2019122326-appb-100071
    对全部子系统执行上述过程，子模型预测输出的集合可表示为，
    Figure PCTCN2019122326-appb-100072
    其中，f i(·)表示第ith个子模型，
    结合基于分支定界(BB)的最优化选择算法和基于误差的信息熵加权算法对上述子模型进行自适应的优化选择和计算加权系数。在给定候选子模型和加权算法后，最佳子模型选择与加权类似最优特征选择。面向有限数量的候选子模型，通过多次运行优化和加权算法，可获得集成尺寸为2到(I-1)的SEN模型，最后排序这些优化的SEN模型并将预测性能最佳的作为最终DXN软测量模型，
    假定最终DXN软测量模型的集成尺寸为I sel，其预测输出值
    Figure PCTCN2019122326-appb-100073
    可由下式计算：
    Figure PCTCN2019122326-appb-100074
PCT/CN2019/122326 2019-03-24 2019-12-02 一种城市固废焚烧过程二噁英排放浓度软测量方法 WO2020192166A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US16/967,408 US12002014B2 (en) 2019-03-24 2019-12-02 Soft measurement method for dioxin emission concentration in municipal solid waste incineration process

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910224790.4A CN109960873B (zh) 2019-03-24 2019-03-24 一种城市固废焚烧过程二噁英排放浓度软测量方法
CN201910224790.4 2019-03-24

Publications (1)

Publication Number Publication Date
WO2020192166A1 true WO2020192166A1 (zh) 2020-10-01

Family

ID=67024801

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/122326 WO2020192166A1 (zh) 2019-03-24 2019-12-02 一种城市固废焚烧过程二噁英排放浓度软测量方法

Country Status (2)

Country Link
CN (1) CN109960873B (zh)
WO (1) WO2020192166A1 (zh)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112541242A (zh) * 2020-12-23 2021-03-23 燕山大学 一种弹性提升城市电与气综合能源系统扩展规划方法
CN113189891A (zh) * 2020-11-08 2021-07-30 北京工业大学 基于双向安全隔离的城市固废焚烧过程半物理仿真平台
CN114518737A (zh) * 2022-02-28 2022-05-20 北京工业大学 面向mswi过程多模态数据驱动预测模型的仿实时验证平台
CN114661754A (zh) * 2022-01-20 2022-06-24 福建中兴电子科技有限公司 一种基于分数引导正则化网络的水质污染无监督预警方法
CN115358152A (zh) * 2022-08-26 2022-11-18 绍兴市再生能源发展有限公司 垃圾焚烧气体控制与反馈调节系统和方法

Families Citing this family (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109960873B (zh) * 2019-03-24 2021-09-10 北京工业大学 一种城市固废焚烧过程二噁英排放浓度软测量方法
WO2023083009A1 (zh) * 2021-11-11 2023-05-19 北京工业大学 一种面向mswi过程的多入多出回路控制半实物仿真实验平台
CN114398836A (zh) * 2022-01-19 2022-04-26 北京工业大学 基于宽度混合森林回归的mswi过程二噁英排放软测量方法
CN114564894A (zh) * 2022-03-04 2022-05-31 北京工业大学 基于残差拟合机制简化深度森林回归的炉排炉mswi过程二噁英排放软测量方法

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20090193936A1 (en) * 2008-02-01 2009-08-06 Honeywell International, Inc. Methods and apparatus for an oxygen furnace quality control system
CN103455635A (zh) * 2013-09-24 2013-12-18 华北电力大学 基于最小二乘支持向量机集成的热工过程软测量建模方法
CN107944173A (zh) * 2017-12-05 2018-04-20 北京工业大学 一种基于选择性集成最小二乘支撑向量机的二噁英软测量系统
CN108549792A (zh) * 2018-06-27 2018-09-18 北京工业大学 一种基于潜结构映射算法的固废焚烧过程二噁英排放浓度软测量方法
CN109960873A (zh) * 2019-03-24 2019-07-02 北京工业大学 一种城市固废焚烧过程二噁英排放浓度软测量方法

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105910111B (zh) * 2016-06-15 2018-01-16 重庆盎瑞悦科技有限公司 一种具有废气二次燃烧干燥点火装置的固废焚烧系统及处理方法
US10751723B2 (en) * 2017-04-26 2020-08-25 Adr Technology B.V. Method and apparatus for liberating particles from moist MSWI ash
CN108062566A (zh) * 2017-12-13 2018-05-22 北京工业大学 一种基于多核潜在特征提取的智能集成软测量方法

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20090193936A1 (en) * 2008-02-01 2009-08-06 Honeywell International, Inc. Methods and apparatus for an oxygen furnace quality control system
CN103455635A (zh) * 2013-09-24 2013-12-18 华北电力大学 基于最小二乘支持向量机集成的热工过程软测量建模方法
CN107944173A (zh) * 2017-12-05 2018-04-20 北京工业大学 一种基于选择性集成最小二乘支撑向量机的二噁英软测量系统
CN108549792A (zh) * 2018-06-27 2018-09-18 北京工业大学 一种基于潜结构映射算法的固废焚烧过程二噁英排放浓度软测量方法
CN109960873A (zh) * 2019-03-24 2019-07-02 北京工业大学 一种城市固废焚烧过程二噁英排放浓度软测量方法

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113189891A (zh) * 2020-11-08 2021-07-30 北京工业大学 基于双向安全隔离的城市固废焚烧过程半物理仿真平台
CN113189891B (zh) * 2020-11-08 2023-01-13 北京工业大学 基于双向安全隔离的城市固废焚烧过程半物理仿真平台
CN112541242A (zh) * 2020-12-23 2021-03-23 燕山大学 一种弹性提升城市电与气综合能源系统扩展规划方法
CN112541242B (zh) * 2020-12-23 2024-01-12 燕山大学 一种弹性提升城市电与气综合能源系统扩展规划方法
CN114661754A (zh) * 2022-01-20 2022-06-24 福建中兴电子科技有限公司 一种基于分数引导正则化网络的水质污染无监督预警方法
CN114661754B (zh) * 2022-01-20 2024-05-03 福建中兴电子科技有限公司 一种基于分数引导正则化网络的水质污染无监督预警方法
CN114518737A (zh) * 2022-02-28 2022-05-20 北京工业大学 面向mswi过程多模态数据驱动预测模型的仿实时验证平台
CN114518737B (zh) * 2022-02-28 2024-04-16 北京工业大学 面向mswi过程多模态数据驱动预测模型的仿实时验证平台
CN115358152A (zh) * 2022-08-26 2022-11-18 绍兴市再生能源发展有限公司 垃圾焚烧气体控制与反馈调节系统和方法

Also Published As

Publication number Publication date
CN109960873A (zh) 2019-07-02
CN109960873B (zh) 2021-09-10
US20210233039A1 (en) 2021-07-29

Similar Documents

Publication Publication Date Title
WO2020192166A1 (zh) 一种城市固废焚烧过程二噁英排放浓度软测量方法
Tang et al. The multi-objective optimization of combustion system operations based on deep data-driven models
CN108549792B (zh) 一种基于潜结构映射算法的固废焚烧过程二噁英排放浓度软测量方法
Borunda et al. Bayesian networks in renewable energy systems: A bibliographical survey
Xia et al. Dioxin emission prediction based on improved deep forest regression for municipal solid waste incineration process
Ummel CARMA revisited: an updated database of carbon dioxide emissions from power plants worldwide
WO2020228215A1 (zh) 基于多层特征选择的固废焚烧过程二噁英排放浓度软测量方法
CN111144609A (zh) 一种锅炉废气排放预测模型建立方法、预测方法及装置
WO2021159585A1 (zh) 一种二噁英排放浓度预测方法
Braun et al. A neuro-genetic approach for modeling and optimizing a complex cogeneration process
Dong et al. Evaluation of TPGU using entropy-improved TOPSIS-GRA method in China
Wen et al. Dioxin emission prediction from a full-scale municipal solid waste incinerator: Deep learning model in time-series input
Wang et al. Application of genetic algorithm-back propagation for prediction of mercury speciation in combustion flue gas
Movahed et al. Modeling and optimization of NO emission for a steam power plant by data‐driven methods
CN109978011A (zh) 一种城市固废焚烧过程二噁英排放浓度预测系统
Fu et al. Prediction model of desulfurization efficiency of coal-fired power plants based on long short-term memory neural network
Cui et al. Multi-condition operational optimization with adaptive knowledge transfer for municipal solid waste incineration process
CN115526433A (zh) 一种基于集成混合模型的电厂再热烟气挡板操作预测方法
US12002014B2 (en) Soft measurement method for dioxin emission concentration in municipal solid waste incineration process
Chongwatpol et al. Applying analytics in the energy industry: A case study of heat rate and opacity prediction in a coal-fired power plant
Zhou et al. Scenario Analysis of Carbon Emissions of China's Power Industry Based on the Improved Particle Swarm Optimization-Support Vector Machine Model.
Zheng et al. Prediction of NOx concentration from coal combustion using LS-SVR
Liu et al. A prediction method of NOx in thermal power plants using GC-LSTM neural network
Xu et al. Low-cost sensor outlier detection framework for on-line monitoring of particle pollutants in multiple scenarios
Tang et al. Dioxin Emission Concentration Forecasting Approach Based on Latent Feature Extraction and Selection for Municipal Solid Waste Incineration

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19921466

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19921466

Country of ref document: EP

Kind code of ref document: A1

122 Ep: pct application non-entry in european phase

Ref document number: 19921466

Country of ref document: EP

Kind code of ref document: A1