Dynamic collaborative optimization method for desulfurization and denitrification system based on physical constraint heterogeneous proxy
By using a physical constraint-based heterogeneous proxy method, the problem of high-precision modeling and optimization of desulfurization and denitrification systems under deep peak shaving conditions was solved, achieving efficient, stable and economical operation within a wide load range, reducing consumption and ammonia slip, and improving the stability and applicability of the system.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING UNIV OF TECH
- Filing Date
- 2026-01-23
- Publication Date
- 2026-05-01
AI Technical Summary
Existing desulfurization and denitrification systems are difficult to model and optimize with high precision under deep peak shaving conditions. They lack physical constraints and dynamic adaptation mechanisms, which leads to decreased desulfurization efficiency, increased absorbent consumption, increased ammonia escape, and increased equipment corrosion risk, making it difficult to meet ultra-low emission requirements.
A physical constraint-based heterogeneous proxy approach is adopted. By constructing a heterogeneous proxy modeling framework that integrates gas-liquid mass transfer mechanism and data-driven model, and combining regeneration nucleus Hilbert space embedding and nucleation distance metric, dynamic environment detection and multi-objective optimization are carried out to achieve high-precision proxy modeling and dynamic operating condition adaptation of desulfurization and denitrification system.
It achieves efficient, stable and economical operation within a wide load range, reduces the consumption of absorbent and reducing agent, suppresses ammonia escape and system fluctuations, and improves the stability and engineering applicability of the desulfurization and denitrification system.
Smart Images
Figure CN121960286A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of flue gas treatment and operation optimization technology for thermal power units, and in particular to a dynamic collaborative optimization method for desulfurization and denitrification systems based on physical constraint heterogeneous proxy. Background Technology
[0002] Against the backdrop of the construction of new power systems, the penetration rate of renewable energy in power systems continues to increase. The operation mode of coal-fired units is gradually shifting from traditional steady-state base load operation to flexible operation modes such as deep peak shaving, wide load fluctuation, and rapid start-up and shutdown. Extensive engineering practice shows that the actual operating load range of a single coal-fired unit has decreased from 80%-100% of its rated capacity to 30% or even lower, with units operating for extended periods in many non-design conditions. Under these operating conditions, the combustion characteristics on the boiler side change significantly, and the flue gas temperature, flow rate, and pollutant formation mechanisms exhibit strong nonlinear and time-varying characteristics, further placing higher demands on the stable operation and efficient control of the downstream desulfurization and denitrification systems.
[0003] During deep peak-shaving operation, rapid fluctuations in flue gas parameters cause the Selective Catalytic Reduction (SCR) and Flue Gas Desulfurization (FGD) processes in the desulfurization and denitrification systems to frequently deviate from design conditions. Specifically, the slurry level, circulation volume, and alkalinity level in the absorber fluctuate significantly with load changes, making it difficult to maintain a stable dissolution and reaction rate of calcium carbonate (CaCO3), resulting in decreased desulfurization efficiency and increased absorbent consumption. Simultaneously, the ammonia-nitrogen ratio, reaction temperature window, and catalyst activity of the SCR system are all strongly affected by changes in flue gas conditions, easily leading to increased ammonia escape, pressure drop fluctuations, and increased equipment corrosion risks. With the continuous advancement of ultra-low emission requirements, the auxiliary energy consumption of the environmental protection island system has increased to 15%-20% of the total unit energy consumption, and its operational performance directly affects the power plant's emission compliance capabilities, operational economy, and equipment safety level.
[0004] At the modeling and optimization level, existing research on desulfurization and denitrification systems mainly relies on computational fluid dynamics (CFD) models or empirical adjustment strategies. While CFD models can characterize physical processes such as gas-liquid mass transfer, chemical reactions, and flow field distribution to a certain extent, their modeling process depends on a large number of geometric and physical property parameters, resulting in high computational costs and long simulation cycles, making it difficult to meet the online optimization requirements under deep peak-shaving conditions. Furthermore, when flue gas conditions or coal quality change, the model needs to be recalibrated, making engineering deployment difficult. In field operation, adjustments are still mainly made manually based on experience, determining operating parameters through repeated experiments. This is not only time-consuming and labor-intensive, but also makes it difficult to guarantee the effectiveness and sustainability of optimization results under conditions of rapid load changes.
[0005] With the development of data-driven methods and intelligent optimization technologies, some studies have attempted to model desulfurization and denitrification processes using surrogate models such as support vector machines, random forests, and neural networks, combined with multi-objective optimization algorithms for optimization. However, these methods are mostly based on historical steady-state data and lack explicit characterization of physical constraints such as gas-liquid mass transfer mechanisms and reaction kinetics. When the load fluctuates drastically or the operating conditions deviate from the training distribution, the model's prediction accuracy and stability decrease significantly. Furthermore, most existing multi-objective optimization methods are designed for static or low-cost problems, typically assuming a fixed optimal solution set, and failing to fully consider the dynamic characteristics of the objective function and optimal operating solution continuously migrating over time under deep peak-shaving conditions in desulfurization and denitrification systems.
[0006] In summary, existing modeling and optimization methods for desulfurization and denitrification systems still have significant shortcomings: on the one hand, the systems themselves have strong nonlinearity, strong coupling, and time-varying characteristics, and the online detection of key state variables is subject to noise interference and time delay, making it difficult to support high-precision modeling; on the other hand, existing surrogate models and optimization algorithms generally lack physical constraints and dynamic adaptation mechanisms, making it difficult to achieve stable, accurate, and economical collaborative optimization control under high-cost, multi-objective, and dynamic operating conditions, and thus failing to meet the intelligent and adaptive optimization requirements of desulfurization and denitrification systems for deep peak-shaving operation of coal-fired units. Summary of the Invention
[0007] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a dynamic collaborative optimization method for desulfurization and denitrification systems based on physical constraint heterogeneous proxy. This method can perform physically consistent high-precision proxy modeling, dynamic operating condition perception and traceability, and achieve efficient, stable and economical operation of the desulfurization and denitrification system within a wide load range, while meeting emission constraints and equipment safety requirements.
[0008] To achieve the above objectives, the present invention provides the following solution: a dynamic collaborative optimization method for desulfurization and denitrification systems based on physically constrained heterogeneous proxies, comprising: Collect raw operational data, preprocess and filter the raw operational data to obtain a set of core operational variables with consistent time sequence; The original operating data is optimized and normalized to obtain a normalized dataset, and then the normalized dataset is used to construct a desulfurization and denitrification collaborative prediction proxy model. The running sample is embedded into the regenerating nuclear Hilbert space and dynamic environmental detection is performed to obtain the environmental change judgment result. Then, based on the environmental change judgment result, nucleation selection and kinship pairing are performed in the regenerating nuclear Hilbert space to obtain a diversity-enhanced population. Using the desulfurization and denitrification synergistic prediction proxy model and the environmental change determination results, the diversity-enhancing population is dynamically and collaboratively optimized in multiple objectives, and the Pareto solution set is continuously updated to output the optimal operating strategy.
[0009] Optionally, raw operational data is collected, and the raw operational data is preprocessed and variables are filtered to obtain a set of core operational variables with consistent time series, including: Based on the multi-source online sensor network deployed in the SCR denitrification system and FGD desulfurization system, operating condition variables used to characterize the operating environment and disturbances, as well as controllable operational variables used to optimize decision-making and control execution, are collected to obtain raw operating data; Based on the set process response time window, the original operating data is time-series aligned, outlier detected, and missing value filled to obtain operating samples. The operating samples are initially screened using Spearman rank correlation coefficient to obtain operating variables. Kernel principal component analysis is then used to extract low-dimensional features from the operating variables to obtain a set of core operating variables for modeling and kernelized multi-objective optimization.
[0010] Optionally, the operating variables include flue gas flow rate, flue gas temperature, SO2 concentration at the desulfurization inlet, SO2 concentration at the desulfurization outlet, and NO at the desulfurization inlet. X Concentration, NO at desulfurization outlet X Concentration, absorber level, and system pressure drop; the controllable operating variables include total ammonia injection or total ammonia injection valve opening, ammonia injection grid zone valve opening vector, absorber slurry circulation pump combination and operating frequency, absorber slurry pH setpoint, oxidation blower air volume or outlet pressure, process water makeup water volume, and wastewater discharge flow rate.
[0011] Optionally, the original operating data is subjected to optimization constraints and normalization operations to obtain a normalized dataset. Then, a desulfurization and denitrification synergistic prediction proxy model is constructed using the normalized dataset, including: Based on the original operating data, the decision variable vector, objective function, and operating constraints of the multi-objective optimization are set to obtain the optimization model. The core operation variable set is then normalized to obtain the normalized dataset. Based on the optimized model and the normalized dataset, construct a model for predicting ammonia escape rate and NO, respectively. X A denitrification model consisting of a removal rate prediction model and a desulfurization system consisting of an SO2 removal rate prediction model and a slurry consumption prediction model; A heterogeneous proxy model is constructed based on the denitrification system and the desulfurization system. Physical constraints are introduced into the heterogeneous proxy model to obtain a desulfurization and denitrification collaborative prediction proxy model.
[0012] Optionally, a heterogeneous surrogate model is constructed based on the denitrification system and the desulfurization system. Physical constraints are introduced into the heterogeneous surrogate model to obtain a desulfurization and denitrification synergistic prediction surrogate model, including: In the denitrification system, constraints on the ammonia-nitrogen molar ratio, reaction temperature window, and ammonia escape safety are introduced. In the desulfurization system, constraints on CaCO3 dissolution kinetics, gas-liquid mass transfer efficiency, and slurry pH stability range are introduced to refine the physical constraints. Based on the denitrification system and the desulfurization system with refined physical constraints, an SVR kernel function surrogate model and an FNN nonlinear surrogate model are constructed to obtain a heterogeneous surrogate model. A penalty function or embedded constraint term is introduced into the heterogeneous surrogate model for physical constraints to unify the training objective. Then, the surrogate models are fused through a weighted fusion mechanism to obtain a desulfurization and denitrification collaborative prediction surrogate model. The SVR kernel surrogate model is used to characterize the global nonlinear mapping relationship under low sample conditions, and the FNN nonlinear surrogate model is used to characterize the local dynamic characteristics under strongly nonlinear and strongly coupled conditions.
[0013] Optionally, the running samples are embedded into the regenerating nuclear Hilbert space and dynamic environmental detection is performed to obtain environmental change determination results. Then, based on the environmental change determination results, nucleation selection and kinship pairing are performed in the regenerating nuclear Hilbert space to obtain a diversity-enhanced population, including: The running samples are mapped to the regenerating kernel Hilbert space, and a radial basis function is defined. Then, the mean embedding vector of the time window of the sample mapping is calculated. The kernel distance metric window distribution change is calculated using the mean embedding vector of the time window. The environmental change judgment threshold is used to determine whether the environment has changed, and the environmental change judgment result is obtained. Determine whether the environment has changed. If so, map the sample candidate solutions to the regenerated and Hilbert spaces to construct a high-dimensional representation of the individual using the radial basis kernel function, and calculate the kernelized distance between any two sample candidate solutions. Based on the calculated nucleation distance, a nucleation diversity measure is introduced on the basis of non-dominated distance and crowding degree to retain candidate solutions of samples with large distribution differences in the regeneration and Hilbert spaces. Then, based on the calculated nucleation distance, the kinship pairing probability is constructed to carry out effective information interaction in different nonlinear regions, resulting in a diversity-enhanced population.
[0014] Optionally, using the desulfurization and denitrification synergistic prediction surrogate model and the environmental change determination results, dynamic synergistic multi-objective optimization is performed on the diversity-enhancing population, and the Pareto solution set is continuously updated to output the optimal operating strategy, including: Based on the aforementioned diversity-enhancing population, a set of representative solutions is selected from the current non-dominated solution set as the initial solution for local search. The initial solution for local search is mapped to the regenerated kernel Hilbert space. The kernel function is used to characterize the local geometric structure. Based on the nucleation gradient direction or nucleation neighborhood structure, local perturbation molecules are constructed to perform small-scale, directional nucleation search and generate initial candidate solutions. The initial candidate solutions are evaluated using the desulfurization and denitrification synergistic prediction proxy model to retain improved solutions that satisfy physical constraints and contain at least one objective. The improved solutions are then incorporated into the current population. Based on the improved solutions, the kernel function parameters, environmental change judgment threshold, and nucleated local search radius are adaptively updated to obtain an adaptive parameter set. Combining the adaptive parameter set, the desulfurization and denitrification collaborative prediction proxy model and the environmental change judgment result are used to execute a multi-objective optimization algorithm to dynamically update the non-dominated solution set, form a Pareto optimal solution set that adapts to the current working conditions, and then select the optimal operating strategy based on the Pareto optimal solution set.
[0015] This invention discloses the following technical effects by providing a dynamic collaborative optimization method for desulfurization and denitrification systems based on physically constrained heterogeneous proxies: 1. By constructing a heterogeneous proxy modeling framework that integrates gas-liquid mass transfer mechanism, reaction kinetic constraints and data-driven model, the problem that traditional pure data models are prone to violating physical laws and have insufficient generalization ability under deep peak shaving conditions is overcome, and stable characterization of the strong nonlinear and strongly coupled dynamic behavior of desulfurization and denitrification systems is achieved.
[0016] 2. By introducing the regenerating kernel Hilbert space embedding and kernelized distance metric mechanism, we have achieved effective perception and discrimination of the time-varying characteristics of the system's operating conditions and optimization objectives, providing a unified high-dimensional feature metric basis for optimization decision-making in dynamic environments.
[0017] 3. By further combining nucleated local search with dynamic collaborative optimization strategies, the optimization process can adaptively adjust the search direction and solution set distribution according to changes in operating conditions, effectively improving the optimization efficiency and solution set tracking capability under high evaluation cost conditions, thereby satisfying SO2 / NOX While achieving ultra-low emission constraints, it reduces the consumption of absorbents and reducing agents, suppresses ammonia escape and system fluctuations, and improves the economy, stability and engineering applicability of desulfurization and denitrification systems in deep peak-shaving operation.
[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0019] 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.
[0020] Figure 1 This is a schematic diagram of the method flow provided in an embodiment of the present invention; Figure 2 A simplified process diagram of the FGD desulfurization system for thermal power units provided in this embodiment of the invention; Figure 3 A simplified process diagram of the SCR flue gas denitrification system for thermal power units provided in an embodiment of the present invention; Figure 4 The overall flowchart of the flue gas treatment and operation optimization method for thermal power units provided in the embodiments of the present invention is shown. Detailed Implementation
[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0022] 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.
[0023] like Figure 1-Figure 4 As shown, this invention provides a dynamic collaborative optimization method for desulfurization and denitrification systems based on physically constrained heterogeneous proxies, comprising: Step 1: Collect raw operational data, preprocess and filter the raw operational data to obtain a set of core operational variables with consistent timing; specifically including: 1.1 Based on the multi-source online sensor network deployed in the SCR denitrification system and FGD desulfurization system, operating condition variables used to characterize the operating environment and disturbances, as well as controllable operational variables used to optimize decision-making and control execution, are collected to obtain raw operating data.
[0024] The operating variables include flue gas flow rate, flue gas temperature, SO2 concentration at the desulfurization inlet, SO2 concentration at the desulfurization outlet, and NO at the desulfurization inlet. X Concentration, NO at desulfurization outlet X Concentration, absorber level, and system pressure drop; the controllable operating variables include total ammonia injection or total ammonia injection valve opening, ammonia injection grid zone valve opening vector, absorber slurry circulation pump combination and operating frequency, absorber slurry pH setpoint, oxidation blower air volume or outlet pressure, process water makeup water volume, and wastewater discharge flow rate.
[0025] Operating parameters of the denitrification system include: total ammonia injection or opening degree of the total ammonia injection control valve, opening vector of the ammonia injection grid zone valve, NOx concentration at the inlet and outlet of the denitrification reactor, oxygen content in the flue gas, and flue gas temperature.
[0026] The operating parameters of the desulfurization system include: the combination and frequency of circulating pumps, the pH setting value of the absorber slurry, the air volume or pressure of the oxidation blower, the process water makeup water volume, the wastewater discharge flow rate, and the SO2 concentration at the inlet and outlet of the desulfurization tower.
[0027] 1.2 Based on the set process response time window, the original operating data is time-series aligned, outlier detected, and missing value filled to obtain operating samples. The operating samples are initially screened using Spearman rank correlation coefficient to obtain operational variables. Kernel principal component analysis is then used to extract low-dimensional features from the operational variables to obtain a set of core operational variables for modeling and kernelized multi-objective optimization.
[0028] Multi-objective optimization problems include at least: environmental performance objectives aimed at meeting pollutant emission constraints, comprehensive operating cost objectives aimed at reducing system operation and reagent consumption, and equipment safety objectives aimed at ensuring the safe and stable operation of key equipment.
[0029] Step 2: Perform optimization constraints and normalization operations on the original operating data to obtain a normalized dataset, and then use the normalized dataset to construct a desulfurization and denitrification collaborative prediction surrogate model; specifically including: 2.1 Based on the original operational data, a decision variable vector for multi-objective optimization, an objective function for multi-objective optimization, and operational constraints are defined to obtain an optimization model. The core operational variable set is then normalized to obtain a normalized dataset. Specifically: Decision variables for a multi-objective optimization problem are defined, including: total ammonia injection or opening of the ammonia injection control valve, opening of the ammonia injection grid zone valve, combination and operating frequency of the absorber slurry circulation pump, pH setting of the absorber slurry, air volume or outlet pressure of the oxidation blower, process water makeup water volume, and wastewater discharge flow rate.
[0030] The operational constraints for the multi-objective optimization problem are defined, including: pollutant emission limit constraints, absorber level constraints, slurry pH safety range constraints, ammonia escape concentration constraints, equipment operating load constraints, and process operation boundary constraints.
[0031] Define an optimization decision variable vector corresponding to the overall optimization objective. ; in, A t This refers to the total ammonia injection volume or the total valve opening. A z The opening degree of the ammonia injection grid zone valve. P c It is a combination of slurry circulation pumps. f c The operating frequency of the slurry circulation pump, pH set Set the absorber slurry setpoint. Q ox This refers to the air volume or air pressure of the oxidation fan. W a This refers to the amount of process water makeup water. W d This refers to the wastewater discharge flow rate.
[0032] Based on the optimization objective, decision variables, and operational constraints, a multi-objective optimization model for the dynamic collaborative optimization of the desulfurization and denitrification system is constructed, providing a mathematical foundation for subsequent surrogate model training and optimization solutions. Specifically: ; in, x To optimize the decision variable vector, x L and x U These are the lower and upper bounds of the decision variable, respectively.
[0033] The optimization objective function is: f 1( x () indicates the outlet of the denitrification system NO x The goal is to maximize the removal rate; f 2( x () indicates the outlet of the desulfurization system SO 2. The goal is to maximize the removal rate; f 3( x This represents the objective of minimizing the overall operating cost of the desulfurization and denitrification system, which includes the cost of reducing agent consumption, absorbent consumption, electricity consumption, and wastewater treatment. f 4( x The target is to minimize the operational risks of key equipment, which includes the ammonia injection system, circulating slurry pump, oxidation blower, and absorption tower body.
[0034] The optimization decision variables include: Total ammonia injection volume or total ammonia injection valve opening in the denitrification system A t Ammonia injection grid zone valve opening A z ; Combination of slurry circulation pumps in desulfurization system P c and operating frequency f c pH set value of absorber slurry pH set Oxidation fan air volume or air pressure Q ox Process water makeup volume W a Wastewater discharge flow rate W d .
[0035] The constraints include: The constraints include pollutant emission constraints, system operation stability constraints, equipment safe operation constraints, and physical boundary constraints for operational variables; wherein, the pollutant emission constraints are used to limit the concentration of nitrogen oxides at the outlet of the denitrification system and the concentration of sulfur dioxide at the outlet of the desulfurization system to meet the emission standards.
[0036] 2.2 Based on the optimized model and the normalized dataset, construct a prediction model for ammonia slip rate and NO, respectively. X The denitrification model consists of a removal rate prediction model and the desulfurization system consists of an SO2 removal rate prediction model and a slurry consumption prediction model.
[0037] 2.3 Based on the denitrification system and the desulfurization system, a heterogeneous surrogate model is constructed. Physical constraints are introduced into the heterogeneous surrogate model to obtain a desulfurization and denitrification synergistic prediction surrogate model. Specifically, this includes: 2.3.1 In the denitrification system, constraints on the ammonia-nitrogen molar ratio, reaction temperature window, and ammonia escape safety are introduced. In the desulfurization system, constraints on CaCO3 dissolution kinetics, gas-liquid mass transfer efficiency, and slurry pH stability range are introduced to refine the physical constraints.
[0038] 2.3.2 Based on the denitrification system and the desulfurization system with refined physical constraints, an SVR kernel function surrogate model and an FNN nonlinear surrogate model are constructed to obtain a heterogeneous surrogate model. A penalty function or embedded constraint term is introduced into the heterogeneous surrogate model for physical constraints to unify the training objective. Then, the surrogate models are fused through a weighted fusion mechanism to obtain a desulfurization and denitrification collaborative prediction surrogate model.
[0039] The SVR kernel surrogate model is used to characterize the global nonlinear mapping relationship under low sample conditions, and the FNN nonlinear surrogate model is used to characterize the local dynamic characteristics under strongly nonlinear and strongly coupled conditions.
[0040] Step 2 is specific, for example: Based on the aforementioned core operational variables, a prediction model for ammonia slip rate in the denitrification system was constructed respectively. NO x SO2 removal efficiency prediction model, SO2 removal efficiency prediction model and slurry consumption prediction model of desulfurization system.
[0041] The heterogeneous surrogate model includes at least a kernel function surrogate model based on support vector regression and a nonlinear surrogate model based on feedforward neural network. The support vector regression model is used to characterize the global nonlinear mapping relationship under low sample conditions, and the feedforward neural network model is used to characterize the local dynamic characteristics under strong nonlinear and strong coupling conditions.
[0042] During the training of the surrogate model, constraints on the desulfurization and denitrification process mechanism are introduced, including the following physical constraints: Constraints on ammonia-nitrogen molar ratio, reaction temperature window, and ammonia escape safety in denitrification systems; Gas-liquid mass transfer efficiency constraints in desulfurization systems, slurry pH Stability constraints and oxidation reaction completeness constraints; The physical constraints are introduced into the loss function of the surrogate model in the form of a penalty function or an embedded constraint term to form the training objective of the physical constraint surrogate model, the expression of which is: ; in, i For proxy model parameters, E d For the prediction error term based on runtime data, c k For the first kA physical constraint function, l k These are the corresponding constraint weight coefficients. K This represents the total number of physical constraints.
[0043] By comprehensively evaluating the prediction accuracy and constraint satisfaction of different surrogate models on historical operating condition samples, a desulfurization and denitrification collaborative prediction surrogate model is constructed using a weighted fusion or adaptive selection mechanism for subsequent multi-objective optimization and rapid response calculation under dynamic environments.
[0044] Step 3: Embed the running sample into the regenerating nuclear Hilbert space and perform dynamic environmental detection to obtain environmental change determination results. Then, based on the environmental change determination results, perform nucleation selection and kinship pairing in the regenerating nuclear Hilbert space to obtain a diversity-enhanced population; specifically including: 3.1 The running samples are mapped to the regenerating kernel Hilbert space, and the radial basis function is defined. Then, the time window mean embedding vector of the sample mapping is calculated. The kernelized distance metric window distribution change is calculated using the time window mean embedding vector. The environmental change judgment threshold is used to determine whether the environment has changed, and the environmental change judgment result is obtained.
[0045] 3.2 Determine whether the environment has changed. If so, map the sample candidate solutions to the regenerated and Hilbert spaces to construct a high-dimensional representation of the individual using the radial basis kernel function, and calculate the kernelized distance between any two sample candidate solutions.
[0046] 3.3 Based on the calculated nucleated distance, a nucleated diversity measure is introduced on the basis of non-dominated distance and crowding degree to retain the candidate solutions of samples with large distribution differences in the regeneration and Hilbert spaces. Then, based on the calculated nucleated distance, the kinship pairing probability is constructed to carry out effective information interaction in different nonlinear regions and obtain a diversity-enhanced population.
[0047] Step 3 is specific, for example: The system operation samples within a continuous time window are mapped to the regenerative kernel Hilbert space, and a high-dimensional feature representation of the samples is constructed using a kernel function; the kernel function used is the radial basis function, and its expression is: ; in, x i , x j These are system state samples collected at different times. s This is the kernel width coefficient.
[0048] The mean embedding vector of the sample distribution within adjacent time windows in the reproducing kernel Hilbert space is calculated based on the kernel function. The expression for the mean embedding vector is: ; in, m t For the first t The mean point of the generation population in the regenerating kernel Hilbert space. The feature mapping induced by the kernel function, that is, the individual x i ( t Mapping of ) in the regenerating kernel Hilbert space; N t For the first t Number of samples within a time window; x i ( t ) indicates the first t The first generation i One solution.
[0049] The kernelized distance between the embedding vectors of the means of adjacent time windows is calculated to characterize the degree of change in the system's operational distribution. The kernelized distance is defined as follows: ; Where H represents the regenerating kernel Hilbert space; The nucleation distance is compared with a preset change judgment threshold. When the nucleation distance exceeds the threshold, it is determined that the system operating environment has changed, and the subsequent multi-objective optimization process triggers population re-initialization, proxy model update or local search enhancement mechanism to adapt to the dynamic operating condition changes of the desulfurization and denitrification system.
[0050] The candidate solutions in the current optimization population are mapped to the regenerating kernel Hilbert space, and a high-dimensional feature representation between individuals is constructed through the kernel function. The kernelized distance between any two candidate solutions in the reproducing kernel Hilbert space is calculated based on the kernel function, and the kernelized distance is defined as follows: ; Where xi and xj are two candidate solutions for the population. The kernel function is consistent with the kernel function used for environmental change detection in claim 5.
[0051] Based on the nucleated distance, a nucleated diversity metric is constructed. On the basis of non-dominated ranking and crowding evaluation, the nucleated diversity metric is introduced for the individual selection process. As a supplement to the diversity evaluation of the target space, the nucleated diversity metric is used to characterize the distribution differences of candidate solutions in the nonlinear feature space and to preferentially retain candidate solutions with larger distribution differences in the regenerating kernel Hilbert space.
[0052] During the crossover and recombination stage, a kinship pairing probability function is constructed based on the nucleation distance between individuals. This reduces the crossover probability between individuals with smaller nucleation distances and increases the crossover probability between individuals with larger nucleation distances, thereby promoting effective information exchange between different nonlinear feature regions.
[0053] Step 4: Using the desulfurization and denitrification synergistic prediction proxy model and the environmental change determination results, perform dynamic synergistic multi-objective optimization on the diversity-enhancing population, continuously update the Pareto solution set, and output the optimal operating strategy. Specifically, this includes: 4.1 Based on the aforementioned diversity-enhancing population, a set of representative solutions is selected from the current non-dominated solution set as the initial local search solution. The initial local search solution is mapped to the regenerated kernel Hilbert space. The kernel function is used to characterize the local geometric structure. Based on the nucleation gradient direction or nucleation neighborhood structure, local perturbation molecules are constructed to perform small-scale, directional nucleation search and generate initial candidate solutions.
[0054] 4.2 The initial candidate solutions are evaluated using the desulfurization and denitrification synergistic prediction proxy model to retain improved solutions that satisfy physical constraints and contain at least one objective. The improved solutions are then incorporated into the current population. Based on the improved solutions, the kernel function parameters, environmental change judgment threshold, and nucleated local search radius are adaptively updated to obtain an adaptive parameter set.
[0055] 4.3 Combining the adaptive parameter set, the desulfurization and denitrification collaborative prediction proxy model and the environmental change judgment result are used to execute a multi-objective optimization algorithm to dynamically update the non-dominated solution set, form a Pareto optimal solution set adapted to the current working conditions, and then select the optimal operating strategy based on the Pareto optimal solution set.
[0056] Step 4 is specific, for example: After determining that the operating environment of the desulfurization and denitrification system has changed, a representative set of solutions is selected from the current non-dominated solution set as the initial solution for local search.
[0057] The initial solution is mapped to the regenerating kernel Hilbert space, and the local geometry of the solution in the high-dimensional feature space is characterized by the kernel function.
[0058] Based on the kernelized gradient direction or kernelized neighborhood structure of the solution in the regenerated kernel Hilbert space, a local perturbation operator is constructed to perform a small-scale, directional kernelized search on the initial solution to generate candidate solutions.
[0059] During the local search process, the physical constraint proxy model described in claim 4 is introduced to quickly evaluate the feasibility of candidate solutions and the direction of target improvement, and only retains solutions that satisfy physical constraints and obtain improvement on at least one optimization objective.
[0060] The improved solution obtained through nucleated local search is incorporated into the current population to guide the multi-objective optimization process to converge rapidly toward the new Pareto front after environmental changes.
[0061] Based on the degree of change in the nucleation distance of system operating samples in the regenerating kernel Hilbert space within adjacent time windows, the kernel width parameter of the radial basis kernel function is adaptively adjusted. s This allows it to be dynamically updated as the scale of the system's operational distribution changes. The update rule is expressed as follows: ; in, s t For the first t The kernel width parameter corresponding to each time window α As a smoothing factor, x i , x j For the first t System status samples within a time window.
[0062] Based on the statistical distribution characteristics of the kernel distance between adjacent time windows, the environmental change judgment threshold is adaptively updated. e t This is used to distinguish between normal system fluctuations and significant changes in operating conditions, and its update rule is expressed as follows: ; in, m d and s d Let represent the mean and standard deviation of the historical nucleation distance, respectively. β This is the threshold adjustment coefficient.
[0063] After detecting a change in the system operating environment, the local search radius for nucleation is adaptively adjusted based on the nucleation distance amplitude. r t This is used to control the perturbation range of the nucleation local search, and its update rule is expressed as: ; in, r min and r max These are the lower and upper limits of the local search radius, respectively. c The attenuation coefficient is... D t For the first t The nucleation distance within a time window.
[0064] By adaptively updating the kernel width parameter, change judgment threshold, and local search radius, the kernelization representation capability, environmental change sensitivity, and local search intensity of the regenerated kernel Hilbert space are adaptively adjusted according to the dynamic operating conditions of the desulfurization and denitrification system, thereby improving the stability and response efficiency of the dynamic collaborative optimization process.
[0065] Therefore, this invention provides a dynamic collaborative optimization method for desulfurization and denitrification systems based on physical constraint heterogeneous proxy, which can perform physically consistent high-precision proxy modeling, dynamic operating condition perception and traceability, and achieve efficient, stable and economical operation of desulfurization and denitrification systems in a wide load range, under the premise of meeting emission constraints and equipment safety requirements.
[0066] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.
[0067] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present 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 the present invention.
Claims
1. A dynamic collaborative optimization method for desulfurization and denitrification systems based on physical constraint heterogeneous proxy, characterized in that, include: Collect raw operational data, preprocess and filter the raw operational data to obtain a set of core operational variables with consistent time sequence; The original operating data is optimized and normalized to obtain a normalized dataset, and then the normalized dataset is used to construct a desulfurization and denitrification collaborative prediction proxy model. The running sample is embedded into the regenerating nuclear Hilbert space and dynamic environmental detection is performed to obtain the environmental change judgment result. Then, based on the environmental change judgment result, nucleation selection and kinship pairing are performed in the regenerating nuclear Hilbert space to obtain a diversity-enhanced population. Using the desulfurization and denitrification synergistic prediction proxy model and the environmental change determination results, the diversity-enhancing population is dynamically and collaboratively optimized in multiple objectives, and the Pareto solution set is continuously updated to output the optimal operating strategy.
2. The dynamic collaborative optimization method for desulfurization and denitrification systems based on physical constraint heterogeneous proxy as described in claim 1, characterized in that, Raw operational data is collected, preprocessed, and variables are filtered to obtain a set of core operational variables with consistent time series, including: Based on the multi-source online sensor network deployed in the SCR denitrification system and FGD desulfurization system, operating condition variables used to characterize the operating environment and disturbances, as well as controllable operational variables used to optimize decision-making and control execution, are collected to obtain raw operating data; Based on the set process response time window, the original operating data is time-series aligned, outlier detected, and missing value filled to obtain operating samples. The operating samples are initially screened using Spearman rank correlation coefficient to obtain operating variables. Kernel principal component analysis is then used to extract low-dimensional features from the operating variables to obtain a set of core operating variables for modeling and kernelized multi-objective optimization.
3. The dynamic collaborative optimization method for desulfurization and denitrification systems based on physical constraint heterogeneous proxy as described in claim 2, characterized in that, The operating variables include flue gas flow rate, flue gas temperature, SO2 concentration at the desulfurization inlet, SO2 concentration at the desulfurization outlet, and NO at the desulfurization inlet. X Concentration, NO at desulfurization outlet X Concentration, absorber level, and system pressure drop; the controllable operating variables include total ammonia injection or total ammonia injection valve opening, ammonia injection grid zone valve opening vector, absorber slurry circulation pump combination and operating frequency, absorber slurry pH setpoint, oxidation blower air volume or outlet pressure, process water makeup water volume, and wastewater discharge flow rate.
4. The dynamic collaborative optimization method for desulfurization and denitrification systems based on physical constraint heterogeneous proxy as described in claim 3, characterized in that, The original operating data is optimized and normalized to obtain a normalized dataset. Then, a desulfurization and denitrification synergistic prediction proxy model is constructed using the normalized dataset, including: Based on the original operating data, the decision variable vector, objective function, and operating constraints of the multi-objective optimization are set to obtain the optimization model. The core operation variable set is then normalized to obtain the normalized dataset. Based on the optimized model and the normalized dataset, construct a model for predicting ammonia escape rate and NO, respectively. X A denitrification model consisting of a removal rate prediction model and a desulfurization system consisting of an SO2 removal rate prediction model and a slurry consumption prediction model; A heterogeneous proxy model is constructed based on the denitrification system and the desulfurization system. Physical constraints are introduced into the heterogeneous proxy model to obtain a desulfurization and denitrification collaborative prediction proxy model.
5. The dynamic collaborative optimization method for desulfurization and denitrification systems based on physical constraint heterogeneous proxy as described in claim 4, characterized in that, A heterogeneous surrogate model is constructed based on the denitrification system and the desulfurization system. Physical constraints are introduced into the heterogeneous surrogate model to obtain a desulfurization and denitrification synergistic prediction surrogate model, including: In the denitrification system, constraints on the ammonia-nitrogen molar ratio, reaction temperature window, and ammonia escape safety are introduced. In the desulfurization system, constraints on CaCO3 dissolution kinetics, gas-liquid mass transfer efficiency, and slurry pH stability range are introduced to refine the physical constraints. Based on the denitrification system and the desulfurization system with refined physical constraints, an SVR kernel function surrogate model and an FNN nonlinear surrogate model are constructed to obtain a heterogeneous surrogate model. A penalty function or embedded constraint term is introduced into the heterogeneous surrogate model for physical constraints to unify the training objective. Then, the surrogate models are fused through a weighted fusion mechanism to obtain a desulfurization and denitrification collaborative prediction surrogate model. The SVR kernel surrogate model is used to characterize the global nonlinear mapping relationship under low sample conditions, and the FNN nonlinear surrogate model is used to characterize the local dynamic characteristics under strongly nonlinear and strongly coupled conditions.
6. The dynamic collaborative optimization method for desulfurization and denitrification systems based on physical constraint heterogeneous proxy as described in claim 5, characterized in that, The running samples are embedded into the regenerating nuclear Hilbert space and subjected to dynamic environmental monitoring to obtain environmental change determination results. Then, based on these environmental change determination results, nucleation selection and kinship pairing are performed in the regenerating nuclear Hilbert space to obtain a diversity-enhanced population, including: The running samples are mapped to the regenerating kernel Hilbert space, and a radial basis function is defined. Then, the mean embedding vector of the time window of the sample mapping is calculated. The kernel distance metric window distribution change is calculated using the mean embedding vector of the time window. The environmental change judgment threshold is used to determine whether the environment has changed, and the environmental change judgment result is obtained. Determine whether the environment has changed. If so, map the sample candidate solutions to the regenerated and Hilbert spaces to construct a high-dimensional representation of the individual using the radial basis kernel function, and calculate the kernelized distance between any two sample candidate solutions. Based on the calculated nucleation distance, a nucleation diversity measure is introduced on the basis of non-dominated distance and crowding degree to retain candidate solutions of samples with large distribution differences in the regeneration and Hilbert spaces. Then, based on the calculated nucleation distance, the kinship pairing probability is constructed to carry out effective information interaction in different nonlinear regions, resulting in a diversity-enhanced population.
7. The dynamic collaborative optimization method for desulfurization and denitrification systems based on physical constraint heterogeneous proxy as described in claim 6, characterized in that, Using the aforementioned desulfurization and denitrification synergistic prediction surrogate model and the environmental change determination results, dynamic synergistic multi-objective optimization is performed on the diversity-enhancing population, and the Pareto solution set is continuously updated to output the optimal operating strategy, including: Based on the aforementioned diversity-enhancing population, a set of representative solutions is selected from the current non-dominated solution set as the initial solution for local search. The initial solution for local search is mapped to the regenerated kernel Hilbert space. The kernel function is used to characterize the local geometric structure. Based on the nucleation gradient direction or nucleation neighborhood structure, local perturbation molecules are constructed to perform small-scale, directional nucleation search and generate initial candidate solutions. The initial candidate solutions are evaluated using the desulfurization and denitrification synergistic prediction proxy model to retain improved solutions that satisfy physical constraints and contain at least one objective. The improved solutions are then incorporated into the current population. Based on the improved solutions, the kernel function parameters, environmental change judgment threshold, and nucleated local search radius are adaptively updated to obtain an adaptive parameter set. Combining the adaptive parameter set, the desulfurization and denitrification collaborative prediction proxy model and the environmental change judgment result are used to execute a multi-objective optimization algorithm to dynamically update the non-dominated solution set, form a Pareto optimal solution set that adapts to the current working conditions, and then select the optimal operating strategy based on the Pareto optimal solution set.