Multi-timescale detection method for dioxin emission concentration in urban solid waste incineration process

Abstract

This invention provides a multi-timescale detection method for dioxin emission concentration in urban solid waste incineration processes. The method includes: constructing a virtual machine mechanistic data generation model under multiple operating conditions; building a virtual machine mechanistic data-driven multi-input multi-output linear regression decision tree combustion state characterization variable model based on the virtual machine mechanistic data; constructing a mechanistic mapping model based on semi-supervised transfer learning based on the characterization variable model; constructing a process data-driven recursive least squares ensemble linear regression decision tree model; and constructing a constrained incremental stochastic weighted neural network fusion model based on the mechanistic mapping model and the recursive least squares ensemble linear regression decision tree model. Based on the above models, dioxin emission concentration is detected at the second, hour, and day levels. The multi-timescale detection method for dioxin emission concentration in urban solid waste incineration processes provided by this invention can achieve DXN emission concentration detection at multiple timescales, providing guidance for existing operating plants.

Description

Technical Field

[0001] This invention relates to the field of urban solid waste treatment technology, and in particular to a multi-timescale detection method for dioxin emission concentration during urban solid waste incineration. Background Technology

[0002] Dioxins (DXN) are persistent organic pollutants with irreversible adverse effects on ecosystems, generated in almost all thermal and combustion processes. DXN is produced as a byproduct during unstable combustion in municipal solid waste incineration (MSWI). With the rapid development of MSWI technology, MSWI processes have become one of the main sources of DXN emissions. Currently, due to limitations in existing detection technologies, real-time monitoring of DXN emission concentrations at industrial sites is difficult. The emergence of industrial big data has spurred the rapid development of modeling methods based on artificial intelligence, machine learning, and deep learning. Therefore, using artificial intelligence technology to detect DXN emission concentrations in MSWI processes has become a feasible solution. Existing research mainly includes models based on statistical learning, models based on classical machine learning, and models based on deep and wide learning. These data-driven (DD) models can establish black-box models of the mapping relationship between independent and dependent variables. However, the performance of DD models is susceptible to insufficient or incomplete coverage of modeling data samples. In contrast, mechanism-driven (MD) models, also known as white-box models, rely on the mechanistic knowledge and principles of momentum, heat, mass, and reaction kinetics of industrial processes. However, due to the complexity and variability of industrial process reaction mechanisms, constructing a suitable MD model for DXN detection remains a challenge.

[0003] To address these issues, existing research has explored ways to achieve higher model performance through mutual compensation between Methodology (MD) and Directed Learning (DD). From a structural perspective, hybrid MD and DD methods can be categorized into cascaded, parallel, and cascade-parallel fusion approaches. Typically, when the underlying industrial process mechanism is well understood, constructing an MD model is the preferred choice; however, in real-world scenarios, due to frequent changes in operating conditions, the accuracy of MD models is often low. Furthermore, fusion strategies combining MD and DD models have not received sufficient attention, and related research is scarce. Moreover, the implementation of MD and DD fusion models should be designed based on the specific characteristics of the target.

[0004] In summary, several unresolved issues remain to be addressed: 1) Existing methods demonstrate the importance of implementing a hybrid modeling strategy suitable for the characteristics of MSWI processes, i.e., MD and DD-driven modeling of DXN is the primary issue to be resolved; 2) Since the generation, emission, and adsorption mechanisms of DXN are still unclear, constructing a mechanistic proxy model for the combustion state characterization variables of the MSWI process is the second issue to be resolved; 3) The typical timescale for offline DXN detection and analysis is hourly, while the timescale for pollution emission control is second-level. Environmental protection departments and plant managers are more concerned with daily or even longer timescales, and the information obtained at different timescales is often inconsistent. Therefore, multi-timescale DXN emission detection is the third issue to be addressed. Thus, designing a multi-timescale detection method for dioxin emission concentration in urban solid waste incineration processes is essential. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-timescale detection method for dioxin emission concentration in urban solid waste incineration processes. This method can achieve DXN emission concentration detection based on a hybrid driving mechanism of virtual mechanisms representing combustion state variables and real process data, enabling DXN emission concentration detection at multiple timescales and providing guidance for existing operating plants.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] A multi-timescale detection method for dioxin emission concentration during urban solid waste incineration includes the following steps:

[0008] Step 1: Build a multi-operational-condition virtual machine physical data generation model based on the coupled FLIC and Aspen plus software to obtain virtual machine physical data under multiple operating conditions;

[0009] Step 2: Construct a virtual machine-driven multi-input multi-output linear regression decision tree model of combustion state representation variables based on virtual machine theoretical data;

[0010] Step 3: Construct a mechanism mapping model based on semi-supervised transfer learning based on the multi-input multi-output linear regression decision tree combustion state characterization variable model;

[0011] Step 4: Construct a process data-driven recursive least squares ensemble linear regression decision tree model;

[0012] Step 5: Construct a constrained incremental stochastic weighted neural network fusion model based on the mechanism mapping model and the recursive least squares ensemble linear regression decision tree model;

[0013] Step 6: Combine the multi-input multi-output linear regression decision tree combustion state characterization variable model, mechanism mapping model, recursive least squares ensemble LRDT model and constrained incremental random weighted neural network fusion model to perform second-level, hour-level and daily-level detection of dioxin emission concentration.

[0014] Optionally, in step 1, a multi-operational-condition virtual machine physical data generation model is built based on the coupled FLIC and Aspen Plus software to obtain virtual machine physical data under multiple operating conditions, specifically as follows:

[0015] The combustion state of the MSWI process was modeled by coupling FLIC and ASPEN Plus software. FLIC simulates solid-phase combustion, which refers to the combustion mechanism of MSW on the furnace bed, while ASPEN Plus simulates gas-phase combustion, which refers to the chemical reaction process of gaseous substances.

[0016] Solid-phase combustion consists of fundamental conservation equations and MSW combustion rates. The fundamental conservation equations include the mass continuity equation, momentum equation, and energy equation. MSW combustion rates encompass moisture evaporation, volatile matter release, volatile matter combustion, and coke combustion. The specific process is as follows:

[0017] The mass continuity equation for MSW combustion is expressed as:

[0018]

[0019]

[0020] In the formula, φ is the bed porosity, and V g and V s V represents the gas velocity and the solid particle velocity. B Let S be the velocity at the moving boundary. s,g S is the rate at which a solid transforms into a gas. s For the mass source term, ρ g and ρ sb The density of the gas and the porous medium;

[0021] Porous media density ρ sb for:

[0022] ρ sb =ρ s (1-φ) (3)

[0023] In the formula, ρ s The solid density of MSW;

[0024] The momentum equation for combustion in porous media is:

[0025]

[0026]

[0027] In the formula, p g Let F(v) be the gas pressure, F(v) be the resistance of the solid to the fluid flow in the porous medium, σ and τ be the normal stress tensor and tangential stress tensor in the bed, respectively, and A be the random disturbance.

[0028] The energy equation for combustion in porous media is:

[0029]

[0030]

[0031] In the formula, H g and H s These are the enthalpy of a gas and the enthalpy of a solid, respectively, T g and T s These represent the gas temperature and the solid temperature, respectively, S a h is the particle surface area. s ' is the convective heat transfer coefficient, Q h Q represents the heat loss or gain of a gas. sh For the solid-phase heat source term, q r Let λ be the radiative heat flux density. s The effective thermal conductivity is determined by the material's thermal conductivity λ. s0 Heat transfer λ caused by random particle motion sm Adding them together, we get λ g The thermal diffusivity is calculated using the following formula:

[0032] λ g =λ0+0.5d p V g ρ g C pg (8)

[0033] In the formula, λ0 is the effective thermal diffusivity, and d p C is the particle diameter. pg The heat capacity of the gas mixture;

[0034] Regarding moisture evaporation, on a dry grate, water in the MSW first evaporates through radiation from the flame and hot walls. The evaporation rate of water is expressed as:

[0035]

[0036] In the formula, R evp h represents the rate of water evaporation. s C is the convective mass transfer coefficient. w,s and C w,g H represents the water concentration in the solid and gas phases. evpThe heat of evaporation of solid water, Q cr The heat absorbed by the solid for convective and radiative heat transfer is further expressed as:

[0037]

[0038] In the formula, S a ε is the particle surface area. s σ is the solid-state emissivity / system emissivity. b Boltzmann radiation constant, Ambient temperature;

[0039] Regarding the release, combustion, and combustion of volatiles, assuming that the gaseous products released by the MSW consist of hydrocarbons, CO, CO2, H2O, and char, the process is as follows:

[0040] MSW→Volatile(C m H n ,CO,CO2,H2O)+Char (11)

[0041] The volatiles in the MSW mix with the surrounding air and burn rapidly. The combustion reaction is as follows:

[0042]

[0043] MSW particles volatilize and release to form coke. The main products of coke gasification are CO and CO2. The reaction equation is as follows:

[0044] C(s)+αO2→2(1-α)CO+(2α-1)CO2 (13)

[0045] In the formula, α is a positive constant in the range of 0.5-1;

[0046] Through a solid-state combustion process, the flue gas composition above the MSW bed was determined to be hydrocarbon C. m H n CO, CO2, O2, H2, H2O and N2;

[0047] The process of gas-phase combustion is as follows:

[0048] The gas generated by solid-phase combustion and preheated secondary air are introduced into the furnace to carry out a gas-phase combustion reaction. The reaction process is as follows:

[0049]

[0050] At the furnace outlet, a selective non-catalytic reduction system is used to treat the flue gas, with a corresponding temperature range of 850-1150℃. The reaction process is as follows:

[0051]

[0052] During the heat exchange stage, the flue gas passes through the superheater, evaporator, and economizer, where its temperature drops to 200℃. The flue gas then first enters the desulfurization unit to remove acidic compounds. The specific reaction process is as follows:

[0053]

[0054] Activated carbon is sprayed into the flue gas duct to adsorb heavy metals and DXN. The flue gas then enters the bag filter to complete gas-phase combustion.

[0055] Based on the solid-gas coupling mechanism model, orthogonal experimental design was used to study the operating parameters of the MSWI process, and experiments were conducted in conjunction with the above numerical simulation model to further obtain virtual mechanism data D under multiple operating conditions. MD .

[0056] Optionally, in step 2, a virtual machine theoretical data-driven multi-input multi-output linear regression decision tree model of combustion state representation variables is constructed based on the virtual machine theoretical data, specifically as follows:

[0057] Using LHV, PA, and FC as model inputs, and CO2, CO, and O2 as model outputs, based on virtual machine theoretical data D MD The multi-condition mechanism data used to construct the combustion state characterization model are represented as follows:

[0058] First, the moving window QAF method is used to remove D. MD The abnormal data points in D MD Rewritten as Its length is denoted as Set the fixed sampling time window to t Wind For time window t Wind The data within the dataset was processed using the QAF (Quick AF) method to remove outliers. The removal rules were as follows:

[0059]

[0060] In the formula, x AB Indicates outlier data points, x UB and x DB Q represents the upper and lower boundaries. 1 / 4 (·) and Q 3 / 4 (·) are the first and third quartiles of x;

[0061] If the data point within the time window is greater than x UB or less than x DB If a data point is not found, it is considered an outlier and deleted. Then, after processing using the QAF method, the input and output data for constructing a virtual machine data-driven multi-input multi-output LRDT combustion state characterization variable model are obtained.

[0062] against A model is constructed based on the Multiple-Input Multiple-Output (LRDT) algorithm, which can be represented as:

[0063]

[0064] Furthermore, a multi-input multi-output LRDT model with outputs of CO2, CO, and O2 that can characterize the combustion state of the MSWI process is obtained.

[0065] Optionally, in step 3, the mechanism mapping model module based on semi-supervised transfer learning is constructed based on the multi-input multi-output linear regression decision tree combustion state characterization variable model, specifically including the following steps:

[0066] Step 301: Using combustion state characterization variables as input and dioxins as output, construct a process mapping model driven by real process data;

[0067] Step 302: Label the mechanistic data of the combustion state characterization variables of unlabeled dioxins using a process mapping model;

[0068] Step 303: Structure of the migration process mapping model, input dioxin pseudo-label data, and establish mechanism mapping model 1;

[0069] Step 304: Update the structure of mechanism mapping model 1 to obtain the final semi-supervised transfer learning model MMM2.

[0070] Optionally, in step 4, a process data-driven recursive least squares ensemble linear regression decision tree model is constructed, specifically as follows:

[0071] A process data-driven recursive least squares ensemble linear regression decision tree model is constructed, represented as follows:

[0072]

[0073] Optionally, in step 5, a constrained incremental stochastic weighted neural network fusion model based on the mechanism mapping model and the recursive least squares ensemble linear regression decision tree model is constructed, specifically as follows:

[0074] A constrained incremental stochastic weighted neural network fusion model based on a mechanism mapping model and a recursive least squares ensemble linear regression decision tree model is constructed, which is expressed as follows:

[0075]

[0076] According to specific embodiments provided by the present invention, the following technical effects are disclosed: The present invention provides a multi-timescale detection method for dioxin emission concentration in urban solid waste incineration processes. This method includes: building a multi-operating-condition virtual machine mechanistic data generation model based on coupled FLIC and Aspenplus software; acquiring virtual machine mechanistic data under multiple operating conditions; constructing a virtual machine mechanistic data-driven multi-input multi-output linear regression decision tree combustion state characterization variable model based on the virtual machine mechanistic data; constructing a mechanism mapping model based on semi-supervised transfer learning based on the multi-input multi-output linear regression decision tree combustion state characterization variable model; and constructing a process data-driven recursive least squares set. A linear regression decision tree model is constructed, and a constrained incremental stochastic weighted neural network fusion model based on a mechanism mapping model and a recursive least squares integrated linear regression decision tree model is built. This model combines the multi-input multi-output linear regression decision tree combustion state characterization variable model, the mechanism mapping model, the recursive least squares integrated LRDT model, and the constrained incremental stochastic weighted neural network fusion model to perform second-level, hour-level, and daily-level detection of dioxin emission concentration. It can achieve DXN emission concentration detection based on a hybrid driving force of virtual mechanism and real process data of combustion state characterization variables, and can realize DXN emission concentration detection at multiple time scales, providing guidance for existing operating plants. Attached Figure Description

[0077] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0078] Figure 1 This is a process flow diagram of the MSWI power plant.

[0079] Figure 2 This is a schematic diagram of the multi-timescale detection method for dioxin emission concentration during urban solid waste incineration, as described in an embodiment of the present invention.

[0080] Figure 3 A schematic diagram of the Aspen Plus model;

[0081] Figure 4 This is a schematic diagram of the MIMO LRDT structure;

[0082] Figure 5 This is a tree structure transformation diagram; Detailed Implementation

[0083] The purpose of this invention is to provide a multi-timescale detection method for dioxin emission concentration in urban solid waste incineration processes. This method can achieve DXN emission concentration detection based on a hybrid driving mechanism of virtual mechanisms representing combustion state variables and real process data, enabling DXN emission concentration detection at multiple timescales and providing guidance for existing operating plants.

[0084] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0085] The process flow of the MSWI power plant is as follows: Figure 1 As shown, it includes five stages: feeding, combustion, heat exchange, flue gas purification, and flue gas emission.

[0086] Currently, the national standard for DXN emissions is 0.1 ng I-TEQ / m³. 3 (Under the conditions of 273K and 11% O2). The generation process of DXN in the MSWI process includes the following stages: (1) MSW containing trace amounts of DXN enters the incinerator; (2) DXN decomposes at high temperature (above 850℃); (3) DXN is synthesized through multiple pathways during the heat exchange process; (4) Fourth, DXN is adsorbed by activated carbon; (5) A concentration of less than 0.1 ng I-TEQ / m³ is discharged from the chimney. 3 DXN.

[0087] Studies have shown that DXN emission concentrations are closely related to the combustion state of the incinerator. From the perspective of DXN formation mechanisms, homogeneous synthesis, precursor synthesis, and de novo synthesis are the main formation pathways. Research indicates that products of incomplete combustion (PICs) are closely related to DXN emissions. Therefore, in the MSWI process, higher PIC levels correlate with a higher probability of DXN formation. Typically, PICs are difficult to detect due to the high-temperature environment within the incinerator. Furthermore, studies have shown that CO concentration in flue gas can be used as an indicator of combustion state, replacing PICs. In industrial settings, CO concentration is typically measured... Figure 1 The CO concentration in the chimney shown is used to assess combustion conditions. CO concentration can serve as a useful indicator of DXN emission concentration, and the relationship between the two can be expressed as:

[0088]

[0089] Among them, y DXN It is the DXN emission concentration, β CO It is a positive constant, x CO The concentration represents the CO emission concentration. The results show that when the CO emission concentration is kept below 10 ppm, the DXN emission concentration can be kept at an acceptable level.

[0090] Further research indicates that excess air (mainly O2 emission concentration) also affects DXN emission concentration, and the approximate linear correlation between the two is shown below:

[0091]

[0092] In the formula, x O2 O2 emission concentration;

[0093] Studies have shown that when the oxygen content in the flue gas is 7%, the DXN concentration exhibits a linear relationship with the combustion temperature and CO concentration, which can be expressed as:

[0094]

[0095] in, and The masses of DXN in water-wall incinerators, modular incinerators, and waste-derived fuel incinerators are respectively x Flow x represents the flue gas flow rate within the flue. HCl x represents the HCl emission concentration. ESP and x Flow This represents a boolean variable. When an electrostatic precipitator or dry scrubber is running during the MSWI process, the boolean variable is set to 1, x. Temps It is the temperature of the chimney;

[0096] The above results indicate that the concentrations of CO and O2 in the flue gas of the MSWI process have a strong characterizing effect on the combustion state and are directly related to the emissions of DXN.

[0097] Furthermore, according to the combustion efficiency calculation method provided in the Hazardous Waste Incineration Emission Standard (GB18484-2020), CO2 also plays an important role in the characterization of combustion status. Therefore, characterizing combustion status requires comprehensive consideration of the concentrations of CO, CO2, and O2.

[0098] like Figure 2 As shown in the embodiment of the present invention, the method for detecting dioxin emission concentration in urban solid waste incineration processes at multiple time scales includes the following steps:

[0099] Step 1: Build a multi-operational-condition virtual machine physical data generation model based on the coupled FLIC and Aspen plus software to obtain virtual machine physical data under multiple operating conditions;

[0100] Step 2: Construct a virtual machine-driven multi-input multi-output linear regression decision tree model of combustion state representation variables based on virtual machine theoretical data;

[0101] Step 3: Construct a mechanism mapping model based on semi-supervised transfer learning based on the multi-input multi-output linear regression decision tree combustion state characterization variable model;

[0102] Step 4: Construct a process data-driven recursive least squares ensemble linear regression decision tree model;

[0103] Step 5: Construct a constrained incremental stochastic weighted neural network fusion model based on the mechanism mapping model and the recursive least squares ensemble linear regression decision tree model;

[0104] Step 6: Combine the multi-input multi-output linear regression decision tree combustion state characterization variable model, mechanism mapping model, recursive least squares ensemble LRDT model and constrained incremental random weighted neural network fusion model to perform second-level, hour-level and daily-level detection of dioxin emission concentration.

[0105] Figure 2 middle, This represents the actual process data used to build the PMM model. Represents the original DXN truth value, y PD This represents the true value of DXN after removing outliers; X MD Represents virtual machine physical data, X PSD X represents actual statistical data. PD This represents the actual process data used to construct a multi-input multi-output LRDT combustion state characterization variable model; This represents the final output value of the DXN detection model.

[0106] The functions of the five models will be introduced separately, including:

[0107] 1. Multi-operation condition virtual machine physical data generation model module, which includes two parts: orthogonal experimental design and numerical simulation model. The former is used to obtain the input data required for the numerical simulation model such as the lower heat value (LHV), primary airflow (PA), and feeding capacity (FC) of MSW under multiple operating conditions. The latter constructs the numerical simulation model by coupling the FLIC and Aspen Plus software used to simulate the solid-phase and gas-phase combustion of MSW.

[0108] 2. The virtual and real data driven multi-input multi-output LRDT combustion state characterization variable model module first removes abnormal data points in the virtual physical data through a fixed time window using a quartile abnormal filter (QAF), and then constructs a multi-input multi-output LRDT model with LHV, PA and FC as inputs and CO2, CO and O2 as outputs.

[0109] 3. The mechanism mapping model module based on semi-supervised transfer learning includes a PMM model driven by real data with combustion state characterization variables as input, a semi-supervised learning process based on PMM, an MMM1 model based on tree structure transfer learning, and an MMM2 model based on incremental learning of model structure. First, a PMM model based on multi-input single-output LRDT is constructed based on real process data with CO2, CO, and O2 as inputs and DXN as output. DXN pseudo-label data is obtained using a semi-supervised learning process based on PMM with unlabeled DXN combustion state characterization variables as input. Then, the MMM1 model based on PMM tree structure transfer learning is obtained using the multi-input single-output LRDT model. Finally, the final MMM2 model based on semi-supervised transfer learning is obtained through structural growth learning of the MMM1 model.

[0110] 4. Integrated LRDT model: Process data is collected from the distributed control system (DCS), and DXN emission concentrations are obtained through online sampling and offline laboratory analysis (HRGC / HRMS). LRDT is used as the basic learner, ensemble is used to improve stability, and iterative Tikhonov regularization is used to improve performance.

[0111] 5. CIRWNN Fusion Model: This model integrates the outputs of LRDT and SSLRDT using a random-weighted neural network (RWNN), and employs a constrained incremental learning strategy to control the incremental learning performance of the model.

[0112] MSWI process data originates from the distributed control system (DCS) in the industrial field. It possesses high-dimensional characteristics and can be represented as follows: The true values ​​of DXN emission concentrations were obtained through online sampling and offline laboratory analysis. The QAF method was applied to remove outlier data points, resulting in accurate high-dimensional small-sample training data.

[0113] In step 1, a multi-running-condition virtual machine physical data generation model is built based on the coupled FLIC and Aspen Plus software to obtain virtual machine physical data under multiple running conditions. Specifically:

[0114] Due to the complexity and variability of MSW components and the frequent fluctuations in operating conditions, it is usually difficult to establish a mechanism model. For the task of DXN concentration modeling, numerical simulation software can simplify the combustion mechanism modeling of the MSWI process. Therefore, the combustion state of the MSWI process is modeled by coupling FLIC and ASPEN Plus software. FLIC simulates solid-phase combustion, which refers to the combustion mechanism of MSW on the furnace bed, while ASPEN Plus simulates gas-phase combustion, which refers to the chemical reaction process of gaseous substances.

[0115] The mechanistic model of MSW combustion mainly consists of two parts: the basic conservation equations and the MSW combustion rate. The basic conservation equations include the mass continuity equation, the momentum equation, and the energy equation. Based on the staged combustion mechanism, the MSW combustion rate is divided into four parts: moisture evaporation, volatile release, volatile combustion, and coke combustion. The modeling process is described below:

[0116] The mass continuity equation for MSW combustion is expressed as:

[0117]

[0118]

[0119] In the formula, φ is the bed porosity, and V g and V s V represents the gas velocity and the solid particle velocity. B Let S be the velocity at the moving boundary. s,g S is the rate at which a solid transforms into a gas. s For the mass source term, ρ g and ρ sb The density of the gas and the porous medium;

[0120] Porous media density ρ sb for:

[0121] ρ sb =ρ s (1-φ) (6)

[0122] In the formula, ρ s The solid density of MSW;

[0123] The momentum equation for combustion in porous media is:

[0124]

[0125]

[0126] In the formula, p gLet F(v) be the gas pressure, F(v) be the resistance of the solid to the fluid flow in the porous medium, σ and τ be the normal stress tensor and tangential stress tensor in the bed, respectively, and A be the random disturbance.

[0127] The energy equation for combustion in porous media is:

[0128]

[0129]

[0130] In the formula, H g and H s These are the enthalpy of a gas and the enthalpy of a solid, respectively, T g and T s These represent the gas temperature and the solid temperature, respectively, S a h is the particle surface area. s ' is the convective heat transfer coefficient, Q h Q represents the heat loss or gain of a gas. sh For the solid-phase heat source term, q r Let λ be the radiative heat flux density. s The effective thermal conductivity is determined by the material's thermal conductivity λ. s0 Heat transfer λ caused by random particle motion sm Adding them together, we get λ g The thermal diffusivity is calculated using the following formula:

[0131] λ g =λ0+0.5d p V g ρ g C pg (11)

[0132] In the formula, λ0 is the effective thermal diffusivity, and d p C is the particle diameter. pg The heat capacity of the gas mixture;

[0133] In addition, the basic conservation model also contains component transport equations and radiation heat transfer equations;

[0134] Regarding moisture evaporation, the combustion rate of the MSW characterizes its state at four different positions on the grate: drying, combustion 1, combustion 2, and burnout. On the dry grate, the water in the MSW first evaporates through radiation from the flame and hot walls. The evaporation rate of water is expressed as:

[0135]

[0136] In the formula, R evp h represents the rate of water evaporation. s C is the convective mass transfer coefficient. w,s and C w,gH represents the water concentration in the solid and gas phases. evp The heat of evaporation of solid water, Q cr The heat absorbed by the solid for convective and radiative heat transfer is further expressed as:

[0137]

[0138] In the formula, S a ε is the particle surface area. s σ is the solid-state emissivity / system emissivity. b Boltzmann radiation constant, Ambient temperature;

[0139] Regarding the release, combustion, and coke combustion of volatiles, on the grate, the release and combustion of volatiles are two separate processes. First, assuming that the gaseous products released by the MSW consist of hydrocarbons, CO, CO2, H2O, and char, the process is as follows:

[0140] MSW→Volatile(C m H n ,CO,CO2,H2O)+Char (14)

[0141] The volatiles in the MSW mix with the surrounding air and burn rapidly. The combustion reaction is as follows:

[0142]

[0143] MSW particles volatilize and release to form coke. The main products of coke gasification are CO and CO2. The reaction equation is as follows:

[0144] C(s)+αO2→2(1-α)CO+(2α-1)CO2 (16)

[0145] In the formula, α is a positive constant in the range of 0.5-1;

[0146] Through a solid-state combustion process, the flue gas composition above the MSW bed was determined to be hydrocarbon C. m H n CO, CO2, O2, H2, H2O and N2;

[0147] For gas-phase reactions, the numerical simulation process using ASPEN Plus is as follows: Figure 3 As shown, the gas-phase combustion process is as follows:

[0148] The gas generated by solid-phase combustion (flow 1) and preheated secondary air (flow 3) are introduced into the furnace to carry out a gas-phase combustion reaction. The reaction process is as follows:

[0149]

[0150] At the furnace outlet, a selective non-catalytic reduction (SNCR) system is used to treat the flue gas, with a corresponding temperature range of 850-1150℃. The reaction process is as follows:

[0151]

[0152] During the heat exchange stage, the flue gas (steam 10) passes through the superheater, evaporator, and economizer, and its temperature drops to 200℃. The flue gas then first enters the desulfurization unit to remove acidic compounds (SO₄). x The specific reaction process is as follows: (with HCl)

[0153]

[0154] Activated carbon (steam 15) is injected into the flue gas duct to adsorb heavy metals and DXN. The flue gas then enters the bag filter to complete gas-phase combustion.

[0155] Based on the aforementioned solid-gas coupling mechanism model, through orthogonal experimental design of the MSWI process operating parameters and experimental implementation combined with the aforementioned numerical simulation model, virtual mechanism data D under multiple operating conditions can be further obtained. MD Here, six operating parameters (FC, grate speed, air intake in zones 1, 2, 3, and 4), two component parameters (moisture content, CHO element ratio), and two micro-parameters (particle size, particle mixing coefficient), totaling ten factors, were selected for orthogonal experiments. Five levels were set for each factor, resulting in an orthogonal experimental design with 10 factors and 5 levels, ultimately yielding 49 operating conditions. To obtain mechanistic data under more operating conditions, an additional 6-factor, 5-level orthogonal experimental design was conducted based on the above 49 operating conditions (for a total of 42 operating conditions). Therefore, mechanistic data for 49*42=2058 operating conditions can be obtained here, and these are denoted as...

[0156] In step 2, a virtual machine-driven multi-input multi-output linear regression decision tree model for combustion state representation variables is constructed based on virtual machine theoretical data. Specifically:

[0157] To construct a model capable of characterizing combustion states, LHV, PA, and FC were used as model inputs, and CO2, CO, and O2 were used as model outputs, based on virtual physical data D. MD The multi-condition mechanism data used to construct the combustion state characterization model are represented as follows:

[0158] First, the moving window QAF method is used to remove D. MD The abnormal data points in D MD Rewritten as Its length is denoted as Set the fixed sampling time window to t Wind For time window t Wind The data within the dataset was processed using the QAF (Quick AF) method to remove outliers. The removal rules were as follows:

[0159]

[0160] In the formula, x AB Indicates outlier data points, x UB and x DB Q represents the upper and lower boundaries. 1 / 4 (·) and Q 3 / 4 (·) are the first and third quartiles of x;

[0161] If the data point within the time window is greater than x UB or less than x DB If a data point is not found, it is considered an outlier and deleted. Then, after processing using the QAF method, the input and output data for constructing a virtual machine data-driven multi-input multi-output LRDT combustion state characterization variable model are obtained.

[0162] Regarding mechanistic data A Multiple-Input Multiple-Output (MIMO) LRDT algorithm is proposed for model construction. MIMO LRDT is an improvement on Multiple-Input Single-Output (MISO) LRDT, aiming to enable its non-leaf nodes and leaf nodes to process multiple output information simultaneously. Its structure is the same as CART, consisting of feature selection and linear regression. Figure 4 As shown;

[0163] In such Figure 4 middle, and This represents the 1st, 2nd, and 3rd non-leaf nodes. and These are the 1st and (T / 2nd)th multi-output leaf nodes. The feature set of the (T / 2) path nodes;

[0164] For ease of explanation later, the MIMO mechanism data has been rewritten as follows: in, and The learning process of the MIMO LRDT method is shown below;

[0165] Based on characteristics Dataset in each non-leaf node It is divided into two child nodes, namely the left node. and right node It can be represented as:

[0166]

[0167] In the formula, and They represent and The number of samples in the sample, Let i be the i-th eigenvector. For vectors The value;

[0168] In order to find the optimal Need to traverse the dataset The mean squared error (MSE) of all points and the calculated target value can be expressed as:

[0169]

[0170] In the formula, Let (n,i) represent the loss function value of MSE, and (n,i) represent the i-th feature value of the n-th sample in the dataset during the k-th iteration.

[0171] The first non-leaf node Determined by the minimum MSE of all points, it is:

[0172]

[0173] In the formula, The optimal coordinates of the dataset are represented by K, and the maximum value of the iteration is represented by K.

[0174] According to formulas (21)-(23), the intermediate nodes (T / 2-1) can be obtained. There are T paths from the root node to all leaf nodes. The node information contained in each path is significantly different. Therefore, the input features for selecting leaf nodes are based on the nodes in the path, as shown below:

[0175]

[0176] In the formula, It is the training dataset for the t-th leaf node. This represents the t-th path of the t-th leaf node. express The number of samples, express The dimension;

[0177] Typically, CART leaf nodes only use target information; that is, this type of tree model does not consider feature space information when predicting leaf nodes. To comprehensively utilize information from both the feature space and the target space, a linear regression method is used to calculate the multi-target prediction output of the leaf nodes. The linear regression function is given below:

[0178]

[0179] In the formula, It is the multiple output of the t-th leaf node. Representing feature space data, This represents the weight matrix that needs to be solved;

[0180] Then, the regularized least squares cost function is used. Calculate the weight matrix

[0181]

[0182] In the formula, λ is the regularization parameter and satisfies λ≥0;

[0183] The weight matrix is ​​solved using the gradient descent method. The calculation is as follows:

[0184]

[0185] Taking Targth as an example, the solution process is as follows:

[0186]

[0187] Furthermore, set The Targ We can obtain:

[0188]

[0189] Therefore, the weight matrix is ​​obtained. The predicted value for the t-th leaf node can be calculated using linear regression; finally, the MIMO LRDT model can be expressed as...

[0190]

[0191] Furthermore, a multi-input multi-output LRDT model with outputs of CO2, CO, and O2 that can characterize the combustion state of the MSWI process is obtained.

[0192] In step 3, a mechanism mapping model module based on semi-supervised transfer learning is constructed based on the multi-input multi-output linear regression decision tree combustion state characterization variable model. This specifically includes the following steps:

[0193] Step 301: Using combustion state characterization variables as input and DXN as output, construct a process mapping model (PMM) driven by real process data;

[0194] From process data X PD The inputs are CO2, CO, and O2, and the emission concentration y of DXN is used. MD As output, learning data for a real-data-driven PMM model is obtained based on the MISO LRDT algorithm. Detailed modeling details are shown in Table 1. Here, the final real-data-driven model is denoted as Θ. PMM (·);

[0195] Table 1 Modeling steps of the PMM model

[0196]

[0197]

[0198] Step 302: Label the mechanistic data of the combustion state characterization variables of the unlabeled DXN using a process mapping model;

[0199] The mechanistic data for the combustion state characterization variables of unlabeled DXN are from the multi-input multi-output LTDT model f. MIMOLRDT The output of (·) can be expressed as by As Θ PMM The process of obtaining DXN pseudo-labeled data from the input of the model is as follows:

[0200]

[0201] Furthermore, and By combining the data, we obtain the pseudo-labeled DXN dataset.

[0202] Step 303: Structure of the migration process mapping model. Input DXN pseudo-labeled data and establish Mechanism mapping model 1 (MMM1).

[0203] Transfer learning is used to convert the trained PMM model into an MMM1 model. The method involves directly using the non-leaf node information from the PMM model as the node information for the MMM1 model without modifying the information of each non-leaf node. A diagram illustrating this process is shown below. Figure 5 As shown;

[0204] like Figure 5As shown, the pseudo-label data is input into the PMM structure, and each non-leaf element is used as prior knowledge. The pseudo-label dataset D... MMM The network is divided into two child nodes, resulting in T routes. The leaf nodes of the MMM1 model are retrained based on the input pseudo-labeled data to better learn unlabeled targets, thus obtaining the MMM1 model Θ. MMM1 (·), see Table 2 for details.

[0205] Table 2 Modeling steps of MMM1 model

[0206]

[0207]

[0208] Step 304: Update the structure of mechanism mapping model 1 to obtain the final semi-supervised transfer learning model MMM2.

[0209] Due to mechanism data N MD Much larger than process data N Model3 Therefore, the number of leaf nodes in the transfer model is much larger than the threshold in normal training. Here, the tree structure update algorithm is applied to the MMM1 model;

[0210] First, the mechanistic data D MMM The data is input into the MMM1 model, and the prediction performance of each leaf node is calculated, as shown below:

[0211]

[0212] In the formula, e MMM The MSE of the T-leaf node in the MMM1 model;

[0213] Secondly, calculate the average value of the MSE vector and identify leaf nodes that are greater than the average value as needing further updates;

[0214]

[0215] Finally, the selected leaf nodes are updated by continuous growth until the number of leaf node samples is less than or equal to the initially set threshold, thus obtaining the MMM2 model. The update process here is consistent with the training process of MISO LRDT, as detailed in Table 3.

[0216] Table 3 Modeling steps of the MMM2 model

[0217]

[0218]

[0219] In step 4, a recursive least squares ensemble linear regression decision tree model driven by process data is constructed, specifically as follows:

[0220] MSIW process data is acquired from DCS, and its high-dimensional representation can be expressed as follows: The true values ​​of DXN emission concentrations were obtained through online sampling and offline laboratory analysis, and were used for... The QAF method is applied to remove outlier data points, resulting in high-dimensional small sample training data.

[0221] First, bootstrapping sampling is used to sample from the training data D. PD Multiple subsets are generated from it, which can be represented as:

[0222]

[0223] In the formula, For the p-th training subset, the number of samples and features is equal to the original D. PD ;

[0224] Secondly, based on the aforementioned P subset Construct P MISO LRDT models. The training process of MISO LRDT is similar to that of MIMO LRDT, except that the intermediate nodes are determined and the leaf nodes are calculated.

[0225] In order to search for the optimal A single output of MSE can be represented as:

[0226]

[0227] According to the MIMO LRDT, the p-th MISO LRDT of the t-th path of the t-th leaf node is... Furthermore, the p-th input data of the t-th leaf node of this MISO LRDT is Therefore, the linear regression of the t-th leaf node is expressed as:

[0228]

[0229] in,

[0230] Weight To improve the performance of the ensemble LRDT model, values ​​are assigned to each MISO LRDT model. For a P MISO LRDT model, it can be represented as:

[0231]

[0232] in, For the p-th MISO LRDT model, express The p-th weight;

[0233] The iterative Tikhonov regularization is applied to the weights of each MISO LRDT model, as shown below:

[0234]

[0235] in, Let B be the weight vector and B be the residual vector. The following relationship exists between them:

[0236]

[0237] A process data-driven recursive least squares ensemble linear regression decision tree model is constructed, represented as follows:

[0238]

[0239] In step 5, a constrained incremental stochastic weighted neural network fusion model based on the mechanism mapping model and the recursive least squares ensemble linear regression decision tree model is constructed, specifically as follows:

[0240] Using the MD model (f MIMOLRDT (·) and Θ MMM (·)) and the output of the DD model (f BaggingLRDT Using (·) as input, a CIRWNN fusion modeling strategy is proposed;

[0241] Given the training dataset is in, This represents the output of the MMM model. This represents the output of the ensemble LRDT model, y DXN This represents the actual value of DXN emission concentration;

[0242] CIRWNN is a single-hidden-layer feedforward neural network algorithm, whose input weights ω In The bias value σ is randomly generated from a uniform distribution (usually set to [-μ,μ] and μ>0), and the system input is... Assume the number of hidden layer neurons is L - 1. The predicted output can be expressed as:

[0243]

[0244] in, This is a basic RWNN model with L-1 hidden layer neurons, where ψ(·) represents the activation function, and ω Out The weights of the neurons in the L-1 hidden layer;

[0245] The output weights ω are obtained using the least squares algorithm. Out As shown below:

[0246]

[0247] In the formula, Φ L-1 (·) represents the matrix output by the hidden layer;

[0248] During the incremental process of hidden layer neurons, the constraint inequality is:

[0249]

[0250] Where <·,·> represent scalar products, r and γ L For positive constants, when 0 < r < 1 and and Φ L The calculation is as follows:

[0251]

[0252]

[0253] The new output of the hidden layer is [Φ L-1 (·),Φ L (·)], output weight ω Out Using formula (42), based on the above constraints, CIRWNN performs incremental learning until the learning accuracy reaches a satisfactory level.

[0254] Finally, the constrained incremental stochastic weighted neural network fusion model based on the mechanism mapping model and the recursive least squares ensemble linear regression decision tree model is obtained, which is expressed as:

[0255]

[0256] Step 6: Combine the multi-input multi-output linear regression decision tree combustion state characterization variable model, mechanism mapping model, recursive least squares ensemble LRDT model, and constrained incremental stochastic weighted neural network fusion model to perform second-level, hourly-level, and daily-level detection of dioxin emission concentrations. Specifically:

[0257] 1. Second-level scale (process data only): DXN concentration prediction based solely on process data, with real-time process data MIMOLRDTf input in integrated LRDT mode. MIMOLRDT (·) can be used to obtain the current DXN emission concentration;

[0258] 2. Hourly Scale (Process Data Only): Using the MMM2 model, the integrated LRDT model, and the CIRWNN fusion model, the strategy is as follows: real-time process data is input into the integrated LRDT model to obtain the first DXN emission concentration; then, CO2, CO, and O2 from the process data are input into the MMM2 model to obtain the second DXN emission concentration; finally, the first and second DXN emission concentrations are input into the CIRWNN fusion model to obtain the final DXN concentration.

[0259] 3. Daily Scale (with process and production data as input): All existing models are used here. The strategy is as follows: First, the production data is input into the MIMO proxy model to obtain CO2, CO, and O2 outputs; next, CO2, CO, and O2 are input into the MMM model to obtain the second DXN emission concentration; then, the process data is input into the integrated LRDT model to obtain the first DXN emission concentration; finally, the first and second DXN emission concentrations are input into the CIRWNN fusion model to obtain the final DXN concentration.

[0260] This invention also provides an embodiment to verify the method proposed in this invention, where the MSW samples are taken from an actual MSWI plant. Industrial composition and elemental analysis, as compositional parameters, are inputs to FLIC and are also factors in the orthogonal experimental design.

[0261] The real DXN dataset obtained from the MSWI plant contains 116 process variables and 1 DXN ground truth value, with a total sample size of 82, covering the period from 2016 to 2020.

[0262] The root mean square error (RMSE) and the coefficient of determination (R²) are used. 2 Two evaluation metrics are used to evaluate the performance of the proposed method, and the calculation method is as follows:

[0263]

[0264]

[0265] The model results generated from the multi-operational-condition virtual machine theoretical data are as follows:

[0266] The multi-condition orthogonal experimental design consisted of two parts: the first part involved 10 factors at 5 levels (49 cases), and the second part focused solely on the operating parameters, further designing 5 factors at 5 levels (42 cases). After manual processing, 1960 multi-condition mechanistic data were obtained.

[0267] Under baseline conditions, the solid and gas temperatures on the grate indicate that the initial bed height of the MSW at the inlet is 592 mm, and its residence time on the grate is 1 hour and 22 minutes. As the grate moves to the right, the mass and volume of the MSW gradually decrease until it is burned into ash. At the center of the grate, the bed temperature is very high, with the highest possible temperatures for the solid and gas phases reaching 1122 K and 1332 K, respectively. The temperature gradually decreases as the combustion process completes.

[0268] After obtaining 1960 sample data points, based on data from an actual MSWI power plant, the final inputs were selected as MSW lower calorific value (LHV), primary air volume (PA), and feed rate (FC), with outputs of CO2, CO, and O2. Therefore, 1960 mechanistic data points with 3 inputs and 3 outputs were obtained. The influence of LHV on PA, LHV on FC, and PA on FC on the experimental results shows that the influence of the three outputs on the three outputs is not negligible, which is consistent with the influence of LHV, PA, and FC on the combustion state of an actual MSWI plant. However, outliers were found in the 1960 samples from the FC and PA outlet regions. Therefore, QAF with a sampling window of 10 was used to process the outliers. After removing outliers using the QAF method, 1846 samples were finally obtained.

[0269] The results of the virtual machine data-driven multi-input multi-output LRDT combustion state characterization variable model are as follows:

[0270] The 1846 samples were divided into three equal parts: the first two parts were used for the training set (1231), and the last part was used for the test set (615).

[0271] Each method was repeated 30 times, and the best result among the 30 was taken as the statistical result.

[0272] The results of the hybrid-driven model based on semi-supervised transfer learning are as follows:

[0273] To construct a mapping model between CO2, CO, and O2 and DXN concentrations, 1846 unlabeled data points should first be labeled. Therefore, the previous MISO LRDT (PMM) model should be constructed using process data with labeled data, taking CO2, CO, and O2 as inputs and DXN concentration as output. The PMM model was trained with an RMSE of 0.0013 and R... 2 It is 0.9188.

[0274] The PMM model was used to label 1846 unlabeled mechanistic data points, resulting in 1846 pseudo-labeled data points. These data points were divided into three equal parts: the first two parts were used for the training set (1231), and the last part was used for the test set (615). Next, the 1231 training data points were input into the PMM model to achieve tree structure transfer, resulting in the MMM1 model. Then, the MMM1 model was updated based on the leaf node errors to obtain the MMM2 model.

[0275] Integrated LRDT modeling results: In addition to the memory effect of DXN, the process of detecting DXN emission concentration involves many steps, which may lead to outliers in the detection results. QAF with a sampling window of 15 was used to process outliers in the DXN content. By removing outliers, the data distribution of DXN concentration became more intuitive. Therefore, the final total number of DXN samples was 74, and 8 outliers were removed. The 74 samples were equally divided into three parts: the first two parts were the training set (50), and the last two parts were the test set (24). Each method was repeated 30 times, and the best result among the 30 was taken as the statistical result.

[0276] The results of the CIRWNN fusion model are as follows:

[0277] Using the outputs of ensemble LRDT and MMM2 LRDT as input, CIRWNN demonstrates accurate learning capability on the training set. Furthermore, CIRWNN's generalization performance on the test set (RMSE 0.0016, R² 0.7437) outperforms the results of modeling using ensemble LRDT alone (RMSE 0.0017, R² 0.7202). Moreover, compared to existing methods, this fusion model achieves higher prediction accuracy. These results demonstrate the effectiveness of the proposed fusion model.

[0278] This invention provides a multi-timescale detection method for dioxin emission concentrations during urban solid waste incineration processes. The method includes methods based on coupled FLIC and Aspen... The Plus software is used to build a virtual machine physiological data generation model under multiple operating conditions, acquire virtual machine physiological data under multiple operating conditions, and construct a virtual machine physiological data-driven multi-input multi-output linear regression decision tree combustion state characterization variable model. Based on the multi-input multi-output linear regression decision tree combustion state characterization variable model, a mechanism mapping model based on semi-supervised transfer learning is constructed. A process data-driven recursive least squares ensemble linear regression decision tree model is constructed. A constrained incremental stochastic weighted neural network fusion model based on the mechanism mapping model and the recursive least squares ensemble linear regression decision tree model is constructed. The multi-input multi-output linear regression decision tree combustion state characterization variable model, mechanism mapping model, recursive least squares ensemble LRDT model and constrained incremental stochastic weighted neural network fusion model are combined to perform second-level, hour-level and daily detection of dioxin emission concentration. It can realize DXN emission concentration detection based on the virtual machine physiological data of combustion state characterization variables and real process data, and can realize DXN emission concentration detection at multiple time scales, providing guidance for existing operating plants.

[0279] Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. Furthermore, those skilled in the art will recognize that, based on the ideas of this invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A multi-timescale detection method for dioxin emission concentration during urban solid waste incineration, characterized in that, Includes the following steps: Step 1: Build a multi-operational-condition virtual machine physical data generation model based on the coupled FLIC and Aspen plus software to obtain virtual machine physical data under multiple operating conditions; Step 2: Construct a virtual machine-driven multi-input multi-output linear regression decision tree model of combustion state representation variables based on virtual machine theoretical data; Step 3: Construct a mechanism mapping model based on semi-supervised transfer learning based on the multi-input multi-output linear regression decision tree combustion state characterization variable model; Step 4: Construct a process data-driven recursive least squares ensemble linear regression decision tree model; Step 5: Construct a constrained incremental stochastic weighted neural network fusion model based on the mechanism mapping model and the recursive least squares ensemble linear regression decision tree model; Step 6: Combine the multi-input multi-output linear regression decision tree combustion state characterization variable model, mechanism mapping model, recursive least squares ensemble LRDT model and constrained incremental random weighted neural network fusion model to perform second-level, hour-level and daily-level detection of dioxin emission concentration; In step 2, a virtual machine-driven multi-input multi-output linear regression decision tree model for combustion state representation variables is constructed based on virtual machine theoretical data. Specifically: Using LHV, PA, and FC as model inputs, and CO2, CO, and O2 as model outputs, based on virtual machine theoretical data... The multi-condition mechanism data used to construct the combustion state characterization model are represented as follows: ; First, the moving window QAF method is used to remove... Abnormal data points in the data will Rewritten as Its length is denoted as ( Set the fixed sampling time window to For time windows The data within the dataset was processed using the QAF (Quick AF) method to remove outliers. The removal rules were as follows: （17） In the formula, Indicates outlier data points. and Indicates the upper and lower boundaries. and yes The first and third quartiles; If the data points within the time window are greater than or less If a data point is not found, it is considered an outlier and deleted. Then, after processing using the QAF method, the input and output data for constructing a virtual machine data-driven multi-input multi-output LRDT combustion state characterization variable model are obtained. ; against A model is constructed based on the Multiple-Input Multiple-Output (LRDT) algorithm, which can be represented as: （18） Furthermore, a multi-input multi-output LRDT model with outputs of CO2, CO, and O2 that can characterize the combustion state of the MSWI process is obtained; In step 3, a mechanism mapping model module based on semi-supervised transfer learning is constructed based on the multi-input multi-output linear regression decision tree combustion state characterization variable model. This specifically includes the following steps: Step 301: Using combustion state characterization variables as input and dioxins as output, construct a process mapping model driven by real process data; Step 302: Label the mechanistic data of the combustion state characterization variables of unlabeled dioxins using a process mapping model; Step 303: Structure of the migration process mapping model, input dioxin pseudo-label data, and establish mechanism mapping model 1; Step 304: Update the structure of mechanism mapping model 1 to obtain the final semi-supervised transfer learning model MMM2; In step 4, a recursive least squares ensemble linear regression decision tree model driven by process data is constructed, specifically as follows: A process data-driven recursive least squares ensemble linear regression decision tree model is constructed, represented as follows: （19）； In step 5, a constrained incremental stochastic weighted neural network fusion model based on the mechanism mapping model and the recursive least squares ensemble linear regression decision tree model is constructed, specifically as follows: A constrained incremental stochastic weighted neural network fusion model based on a mechanism mapping model and a recursive least squares ensemble linear regression decision tree model is constructed, which is expressed as follows: （20）； Process data is collected from the distributed control system, and true DXN emission concentration data are obtained. A recursive least squares ensemble linear regression decision tree model based on multiple MISOLRDT models is constructed. That is, integrating LRDT to predict DXN concentration on a second-scale based on real-time process data; Simultaneously, using the outputs of integrated LRDT and MMM models as inputs and the true values ​​of DXN emission concentrations as outputs, a constrained incremental stochastic weighted neural network fusion model CIRWNN was constructed based on a stochastic weighted neural network. In the second-scale concentration prediction process: DXN concentration prediction is performed based solely on process data, while real-time process data is input into MIMOLRDT in integrated LRDT mode. Obtain the current DXN emission concentration; In the hourly-scale concentration prediction process, real-time process data is input into the integrated LRDT to obtain the first DXN emission concentration. CO2, CO and O2 in the process data are input into the MMM2 model to obtain the second DXN emission concentration. The first and second DXN emission concentrations are input into the CIRWNN fusion model to obtain the final DXN concentration. In daily-scale concentration prediction, production data is input into a multi-input multi-output linear regression decision tree combustion state characterization variable model to obtain CO2, CO, and O2 outputs. CO2, CO, and O2 are then input into an MMM model to obtain the second DXN emission concentration. Process data is input into an integrated LRDT model to obtain the first DXN emission concentration. The first and second DXN emission concentrations are then input into a CIRWNN fusion model to obtain the final DXN concentration.

2. The multi-timescale detection method for dioxin emission concentration during urban solid waste incineration according to claim 1, characterized in that, In step 1, a multi-running-condition virtual machine physical data generation model is built based on the coupled FLIC and Aspen Plus software to obtain virtual machine physical data under multiple running conditions. Specifically: The combustion state of the MSWI process was modeled by coupling FLIC and ASPEN Plus software. FLIC simulates solid-phase combustion, which refers to the combustion mechanism of MSW on the furnace bed, while ASPEN Plus simulates gas-phase combustion, which refers to the chemical reaction process of gaseous substances. Solid-phase combustion consists of fundamental conservation equations and MSW combustion rates. The fundamental conservation equations include the mass continuity equation, momentum equation, and energy equation. MSW combustion rates encompass moisture evaporation, volatile matter release, volatile matter combustion, and coke combustion. The specific process is as follows: The mass continuity equation for MSW combustion is expressed as: （1） （2） In the formula, The porosity of the bed, and For gas velocity and solid particle velocity, The velocity at the moving boundary, The rate at which a solid transforms into a gas. For quality source items, and The density of the gas and the porous medium; Porous media density for: （3） In the formula, The solid density of MSW; The momentum equation for combustion in porous media is: （4） （5） In the formula, For gas pressure, The resistance of solids to fluid flow in porous media. and These are the normal stress tensor and the tangential stress tensor in the bed, respectively. For random perturbations; The energy equation for combustion in porous media is: （6） （7） In the formula, and These are the enthalpy of a gas and the enthalpy of a solid, respectively. and These are the gas temperature and the solid temperature, respectively. For particle surface area, The convective heat transfer coefficient is... Indicates the heat loss or gain of the gas. For solid-phase heat source term, For radiative heat flux density, The effective thermal conductivity is determined by the thermal conductivity of the material. Heat transfer caused by random particle motion Adding them together, The thermal diffusivity is calculated using the following formula: （8） In the formula, For the effective thermal diffusivity, The particle diameter is The heat capacity of the gas mixture; Regarding moisture evaporation, on a dry grate, water in the MSW first evaporates through radiation from the flame and hot walls. The evaporation rate of water is expressed as: （9） In the formula, The rate of water evaporation reaction. The convective mass transfer coefficient, and For the concentrations of water in the solid and gas phases, The heat of evaporation of solid moisture. The heat absorbed by the solid for convective and radiative heat transfer is further expressed as: （10） In the formula, For particle surface area, Solid-state emissivity / system emissivity Boltzmann radiation constant, Ambient temperature; Regarding the release, combustion, and combustion of volatiles, assuming that the gaseous products released by the MSW consist of hydrocarbons, CO, CO2, H2O, and char, the process is as follows: （11） The volatiles in the MSW mix with the surrounding air and burn rapidly. The combustion reaction is as follows: （12） MSW particles volatilize and release to form coke. The main products of coke gasification are CO and CO2. The reaction equation is as follows: （13） In the formula, It is a positive constant in the range of 0.5-1; By the solid phase combustion process, the flue gas components above the MSW bed are hydrocarbons C m H n , CO, CO2, O2, H2, H2O and N2; The process of gas-phase combustion is as follows: The gas generated by solid-phase combustion and preheated secondary air are introduced into the furnace to carry out a gas-phase combustion reaction. The reaction process is as follows: （14） At the furnace outlet, a selective non-catalytic reduction system is used to treat the flue gas, with a corresponding temperature range of 850-1150℃. The reaction process is as follows: （15） During the heat exchange stage, the flue gas passes through the superheater, evaporator, and economizer, where its temperature drops to 200℃. The flue gas then first enters the desulfurization unit to remove acidic compounds. The specific reaction process is as follows: （16） Activated carbon is sprayed into the flue gas duct to adsorb heavy metals and DXN. The flue gas then enters the bag filter to complete gas-phase combustion. Based on the solid-gas coupling mechanism model, orthogonal experimental design was used to study the operating parameters of the MSWI process, and experiments were conducted in conjunction with the above numerical simulation model to further obtain virtual mechanism data under multiple operating conditions. .

Images