A multi-stage dynamic heterogeneous agent model combustion optimization method for a thermal power unit boiler system

Abstract

The present application belongs to the technical field of boiler combustion system optimization of thermal power generating units, and provides a multi-stage dynamic heterogeneous agent model combustion optimization method for a boiler system of a thermal power generating unit, comprising: data acquisition and core operation variable screening, multi-objective optimization problem construction, normalization processing, preliminary fitting, local refinement, weighted fusion, knowledge base construction, knowledge transfer and retraining, global optimization and combined updating; the present application realizes high-precision characterization of the complex nonlinear characteristics of the combustion system by constructing a heterogeneous modeling framework comprising a Kriging model, a least squares support vector machine and an Elman recurrent neural network, and using a multi-stage dynamic training and self-adaptive weight fusion mechanism; by introducing an adaptive transfer learning strategy, knowledge transfer and model retraining are realized using historical operating condition data, thereby significantly reducing the modeling and simulation cost.

Description

Technical Field

[0001] This invention relates to the field of combustion system optimization technology for thermal power unit boilers, and in particular to a multi-stage dynamic heterogeneous proxy model combustion optimization method for thermal power unit boiler systems. Background Technology

[0002] Fluid dynamics, abbreviated as CFD, is used to describe the dynamics of combustion. Thermal power units still face significant challenges during deep peak-shaving operation, including poor combustion stability, decreased thermal efficiency, and increased pollutant emissions. A report from a power research institute indicates that when the unit load is below 40% of its rated capacity, the combustion process in the furnace exhibits strong nonlinear characteristics, making conventional optimization strategies based on design conditions inadequate to adapt to dynamic load changes. Due to the combined effects of fluctuating coal quality, equipment aging, and constraints on peak-shaving rates, boiler thermal efficiency can decrease by 5% to 8%, and NOx emissions can rise significantly. X Emission concentrations have risen to 2 to 3 times the normal load, severely restricting the economic efficiency and environmental compliance rate of generating units. At the modeling level, existing research mostly uses computational CFD models to simulate boiler combustion mechanisms. For example, a work published by a university shows that this model can simultaneously consider turbulent combustion, radiative heat transfer, and multiphase flow effects, and has high prediction accuracy for temperature fields and component distribution. However, it has significant limitations in engineering applications: first, model construction depends on furnace geometry and coal quality parameters, which are difficult for power plants to obtain long-term; second, a single simulation calculation usually takes more than 4 hours, lacking real-time capability; and third, parameters need to be recalibrated when coal quality or load changes, causing difficulties in practical deployment. Experiments at a university show that when the load change rate exceeds 3% / min, the prediction error of the CFD model exceeds 12%. On-site testing and adjustment remains the mainstream combustion optimization method. An operation report from a large power generation group shows that typical combustion optimization tests require maintaining stable unit load for 4 to 6 hours, finding the optimal operating point by adjusting variables such as the opening of dampers at each level and the proportion of burnout air. Such methods are not only time-consuming and reliant on human experience, but also prone to failure under conditions of frequent coal quality fluctuations. During the deep peak-shaving phase, unit loads need to change rapidly according to grid commands, leaving insufficient time for systematic optimization tests. Statistical data shows that over 70% of thermal power plants still adopt conservative operating strategies during deep peak-shaving, sacrificing economic efficiency for safety, resulting in an increase in coal consumption for power generation of 8g / kWh to 12g / kWh. Traditional data-driven methods also suffer from accuracy and adaptability issues. A thermal power research institute used support vector machines to establish a combustion optimization model, which performed well at fixed load points, but the prediction error increased significantly after load fluctuations exceeded 10% of rated capacity. Researchers at a university attempted to improve nonlinear representation capabilities through deep neural network modeling, but this required relying on more than 10... 5Annotated data sets are difficult to obtain on-site at power plants. With the widespread application of artificial intelligence algorithms in power systems, scholars and enterprises both domestically and internationally are exploring new paths for intelligent combustion optimization. However, in general, these methods mostly rely on closed models or rule bases under specific operating conditions, lacking the ability for cross-objective collaborative optimization and cross-unit knowledge transfer under extreme low load and rapid ramp / reduction load conditions, making it difficult to meet the adaptive optimization requirements for deep peak shaving of next-generation coal-fired power units.

[0003] In summary, the current research and control of boiler combustion systems in thermal power units still faces the following challenges: On the one hand, the combustion process exhibits significant time delays and strong coupling characteristics, leading to delays and noise interference in the online detection of key parameters such as furnace temperature, oxygen content, and flue gas composition. Furthermore, some detection instruments suffer from high maintenance costs and poor stability, making high-frequency and accurate condition monitoring difficult. On the other hand, existing optimization algorithms have limitations in handling multi-objective conflicts and complex nonlinear relationships under deep peak-shaving conditions. They lack multi-stage modeling and knowledge transfer mechanisms, resulting in insufficient model accuracy and generalization ability. This hinders real-time, accurate, and low-cost intelligent optimization and control of the combustion system, thus restricting the improvement of efficient and clean operation levels in thermal power units. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a multi-stage dynamic heterogeneous proxy model combustion optimization method for boiler systems of thermal power units. This method aims to solve the technical problems of high optimization calculation costs, insufficient model accuracy, and poor system adaptability caused by the strong nonlinearity of the boiler combustion system, multi-objective conflicts, and frequent changes in operating conditions during the deep peak shaving process of thermal power units.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] A combustion optimization method using a multi-stage dynamic heterogeneous surrogate model for boiler systems in thermal power units includes:

[0007] The operation data of the target thermal power unit under different load and coal quality conditions are collected, and the operation data are sequentially subjected to outlier detection, missing value completion, correlation screening, and kernel principal component analysis to obtain the core operation variables.

[0008] Build a system to maximize boiler thermal efficiency and minimize NO X A multi-objective optimization problem with emissions as the objective;

[0009] The core operational variables are normalized.

[0010] Based on the multi-objective optimization problem, the core operational variables are explored globally and initially fitted using the Kriging model and least squares support vector machine to obtain a preliminary solution space.

[0011] The dynamic time delay characteristics of the preliminary solution space are captured and locally refined using an Ullman recurrent neural network to obtain a dynamically refined solution space.

[0012] A high-precision surrogate model is obtained by fusing the preliminary solution space and the dynamic refined solution space using an adaptive weighting mechanism.

[0013] A knowledge base of load, coal quality, and operating conditions was constructed based on historical optimization cases.

[0014] Based on the load-coal quality-operating condition knowledge base, similarity matching and transfer parameter adjustment are used to perform model knowledge transfer and retraining on the high-precision proxy model to obtain a transfer proxy model.

[0015] The migration surrogate model is globally optimized using a multi-objective evolutionary optimization algorithm until the convergence condition is met, and the optimal combination of process settings is obtained. The optimal combination of process settings is then updated in the combustion controller of the target thermal power unit.

[0016] The present invention discloses the following technical effects:

[0017] This invention provides a multi-stage dynamic heterogeneous surrogate model combustion optimization method for boiler systems of thermal power units. By constructing a heterogeneous modeling framework that includes a Kriging model, a least-squares support vector machine, and an Ullman recurrent neural network, it solves the problem that traditional models cannot simultaneously consider global fitting ability and temporal response characteristics, achieving high-precision representation of the complex nonlinear characteristics of the combustion system. By introducing an adaptive transfer learning strategy, it solves the problem of high remodeling and calibration costs in traditional methods, realizing cross-scenario knowledge transfer and autonomous learning. By adopting a surrogate model-driven multi-objective optimization framework, it solves the balance problem between combustion efficiency and pollutant emissions, improving the operational economy and cleanliness of thermal power units during the deep peak shaving phase. Attached Figure Description

[0018] 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.

[0019] Figure 1 This is a schematic diagram of the combustion optimization process of a multi-stage dynamic heterogeneous proxy model for boiler systems of thermal power units provided in an embodiment of the present invention;

[0020] Figure 2 A process flow diagram of boiler combustion side operation in a thermal power unit provided in an embodiment of the present invention;

[0021] Figure 3 The algorithm operation flowchart provided in the embodiments of the present invention;

[0022] Figure 4 The graph shows the algorithm test results provided in the embodiments of the present invention. Detailed Implementation

[0023] 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.

[0024] The purpose of this invention is to provide a multi-stage dynamic heterogeneous proxy model combustion optimization method for boiler systems of thermal power units, aiming to solve the technical problems of high optimization calculation cost, insufficient model accuracy, and poor system adaptability caused by the strong nonlinearity of the boiler combustion system, multi-objective conflicts, and frequent changes in operating conditions during the deep peak shaving process of thermal power units.

[0025] 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.

[0026] Figure 1 This is a schematic diagram of the multi-stage dynamic heterogeneous proxy model combustion optimization process for boiler systems of thermal power units provided in an embodiment of the present invention, such as... Figure 1 As shown, this invention provides a multi-stage dynamic heterogeneous proxy model combustion optimization method for boiler systems of thermal power units, comprising:

[0027] Step 100: Collect the operating data of the target thermal power unit under different load and coal quality conditions, and perform outlier detection, missing value completion, correlation screening, and kernel principal component analysis on the operating data in sequence to obtain the core operating variables;

[0028] Step 200: Construct a system to maximize boiler thermal efficiency and minimize NO X A multi-objective optimization problem with emissions as the objective;

[0029] Step 300: Normalize the core operational variables;

[0030] Step 400: Based on the multi-objective optimization problem, use the Kriging model and least squares support vector machine to perform global exploration and preliminary fitting of the core operational variables to obtain a preliminary solution space;

[0031] Step 500: Use an Ulman recurrent neural network to capture and locally refine the dynamic time delay characteristics of the preliminary solution space to obtain a dynamically refined solution space;

[0032] Step 600: The preliminary solution space and the dynamic refined solution space are fused using an adaptive weighting mechanism to obtain a high-precision surrogate model;

[0033] Step 700: Construct a load-coal quality-operating condition knowledge base based on historical optimization cases;

[0034] Step 800: Based on the load-coal quality-operating condition knowledge base, similarity matching and transfer parameter adjustment are used to perform model knowledge transfer and retraining on the high-precision proxy model to obtain the transfer proxy model;

[0035] Step 900: Use a multi-objective evolutionary optimization algorithm to perform global optimization on the migration surrogate model until the convergence condition is met, obtain the optimal process setpoint combination, and update the optimal process setpoint combination to the combustion controller of the target thermal power unit.

[0036] Specifically, this embodiment provides a multi-stage dynamic heterogeneous proxy model combustion optimization method for boiler systems of thermal power units, comprising the following steps:

[0037] S1: Definition of runtime data acquisition and optimization problem:

[0038] By collecting operating data of thermal power units under different loads and coal quality conditions, key operating variables affecting boiler combustion performance are identified, and a system is established to maximize boiler thermal efficiency and minimize NO. X A multi-objective optimization problem with emissions as the objective.

[0039] S2: Construction of a Multi-Stage Dynamic Heterogeneous Proxy Model

[0040] Global exploration and preliminary fitting were performed using the Kriging model and Least Squares Support Vector Machine (LSSVM). Then, the Elman Recurrent Neural Network was introduced to capture and refine the dynamic time delay characteristics of the combustion process locally. Finally, an adaptive weighting mechanism was used to achieve multi-model fusion and form a high-precision surrogate model.

[0041] S3: Application of Adaptive Transfer Learning Strategy:

[0042] A knowledge base of "load-coal quality-operating condition" is built based on historical optimization cases. The model knowledge transfer and retraining are realized through similarity matching and transfer parameter adjustment, thereby improving the model's generalization ability and computational efficiency.

[0043] S4: Cooperative optimization and control decision output:

[0044] A multi-objective evolutionary optimization algorithm is executed under the surrogate model to perform global optimization on the fusion surrogate model. When the convergence condition is met, the optimal combination of process setpoints for the boiler combustion system of the thermal power unit is output for actual combustion control.

[0045] refer to Figure 2 The system uses a multi-source online sensor network deployed on the boiler body and its auxiliary equipment to collect raw coal characteristic data and unit operating parameters in real time. The raw coal characteristic data includes industrial analysis indicators (moisture (Mad), ash (Aad), volatile matter (Vdaf), with testing standards conforming to GB / T212-2008). The unit operating parameters cover boiler temperature field distribution (covering key temperature measurement points such as furnace outlet, screen, and high-pressure reheat, with a temperature measurement accuracy of ±1.5%), air-coal ratio (range 0.8 to 1.5), oxygen content (range 2% to 6%), damper opening (0 to 100%), burnout air ratio (10% to 40%), and SCR outlet NO. X Concentration, etc. The collected data is transmitted to the plant-level monitoring information system (MIS) via the OPC protocol and aligned in time and space based on the unit's operating characteristics. The alignment time window width ΔT is dynamically adjusted according to the unit's load change rate, and the specific formula is as follows:

[0046]

[0047] Among them, T char This refers to the characteristic combustion reaction time of coal, with a typical range of 5 to 15 minutes.

[0048] Preferably, outlier detection and missing value completion are performed during the data preprocessing stage; initial screening is conducted using the Spearman rank correlation coefficient, retaining values ​​that are relevant to the target variables (boiler thermal efficiency, coal consumption, NO). X For candidate variables that are significantly correlated with emission concentration (|ρ|≥0.35), the Spearman correlation coefficient is calculated using the following formula:

[0049]

[0050] Where, d i =R(X i )-R(Y i Let be the rank difference between variable X and target Y in the i-th sample. n is the selected sample size, taking the operating data of the unit within a continuous time range. Subsequently, feature compression is performed using the L1 regularized random forest algorithm.

[0051] Specifically, to support adaptive transfer learning, a "load-coal quality-operating condition" knowledge base is further constructed. This knowledge base maps coal quality and operating condition parameters to a low-dimensional feature space based on kernel principal component analysis (KPCA). The kernel function used is a Gaussian kernel, as shown in the following equation:

[0052]

[0053] A correlation index is established between the parameters obtained from principal component analysis and the optimal operating parameter set from historical optimization cases. All system parameters, including coal type data and unit operation data, are determined as the core operating variables for optimization control after a dual screening strategy. The optimization objective for the combustion process of the thermal power unit is defined, and a corresponding optimization objective function is established. The optimization objective is to maximize boiler thermal efficiency and minimize NO₂. X Emission concentration. The mathematical model expressions for the optimization objectives are as follows:

[0054]

[0055]

[0056] Where, η boiler The boiler thermal efficiency is calculated using the inverse balance method; Q loss (x) represents the total heat loss of the boiler during combustion; m fuel LHV is the fuel mass flow rate. fuel For the lower heating value of fuel, NO for SCR exports X Concentration (mg / Nm 3 (or ppm).

[0057] The optimization problem is subject to the following constraints, including: 1) Safety and equipment constraints (e.g., furnace temperature, steam pressure, furnace negative pressure, etc. must not exceed the equipment's permissible limits), steam temperature, damper opening, and flue gas oxygen content boundaries; 2) Environmental constraints. : Meets emission standards; 3) Operational variable boundary x q 4) Combustion controllability constraints: Oxygen content range. Specific constraints are as follows:

[0058]

[0059] The resulting multi-objective optimization problem is:

[0060]

[0061] in: The objective function vector contains and Set X is a feasible set of decision vectors. and Representing combustion efficiency and NO respectively X Both emissions were predicted using a dynamic heterogeneous proxy model. Let be the decision variable vector, representing the operating variables of the boiler combustion system; Let i be the i-th constraint function, representing the operating condition constraints (furnace temperature, pressure, oxygen content, etc.). Number the constraint functions; This represents the total number of constraint functions; This serves as the lower bound for the decision variables, limiting the minimum permissible value for each operational variable. For operation variables; The upper bound of the decision variables limits the maximum allowable value for each operational variable. The index is used to determine the index of each operational variable in vector x; The number of constraints is equivalent to .

[0062] Furthermore, to facilitate evolutionary search and comparison with the target, the target is normalized before optimization. The normalization process linearly scales the data to a specified minimum and maximum value range. The normalized dataset is defined as D, expressed by the formula:

[0063]

[0064] in, It is the raw data. and These are the maximum and minimum values ​​in the dataset. and These are the minimum and maximum boundary values ​​of the normalized data range.

[0065] Specifically, the constructed multi-stage dynamic heterogeneous agent model includes the following three stages:

[0066] 1) Collaborative global exploration modeling:

[0067] Based on the aforementioned sample data and optimization objectives, a collaborative global exploration and preliminary fitting were performed using the Kriging model and least squares support vector machine (LSSVM). The Kriging model was used to describe the global trend of combustion performance, and its predicted output can be expressed as:

[0068]

[0069] in, For deterministic trend functions, The mean is zero and the covariance is σ. 2 R(θ) is a Gaussian process, σ 2 Let R(θ) be the process variance, and let θ be the correlation function, with the parameter θ controlling the shape of the correlation function.

[0070] Meanwhile, the LSSVM model is used to enhance the fitting accuracy to highly nonlinear regions, and its expression is:

[0071]

[0072] in, For kernel function, For Lagrange multipliers, For bias terms, This represents the number of training samples.

[0073] Through joint training and complementary correction, a preliminary solution space for combustion performance is obtained.

[0074] 2) Dynamic characteristic refinement modeling:

[0075] To address the time lag and dynamic coupling characteristics of the combustion process, an Elman Network is introduced to construct a dynamic refinement model. This model utilizes its internal state feedback mechanism to capture the dynamic and time-lag characteristics of the boiler combustion process, achieving localized prediction refinement for key areas. The hidden layer output state of the Elman Network is updated by the following formula:

[0076]

[0077]

[0078] in, Given the current input feature vector, This is a hidden layer state. , , These are the input, recursion, and output weight matrices, respectively. , For bias vectors, This represents the activation function.

[0079] 3) Adaptive fusion stage:

[0080] A heterogeneous model fusion function is constructed, and an adaptive weight adjustment mechanism is designed for each sub-model. The outputs of the Kriging model, LSSVM model, and Elman network are dynamically fused to form a high-precision fusion proxy model for mid-to-late-stage optimization search. The fusion expression is:

[0081]

[0082] Where, ω i The adaptive weight coefficients for each sub-model satisfy ∑ω i =1.

[0083] The weights are dynamically adjusted based on the prediction error (RMSE) and confidence interval width (CI) of each model on the validation set, achieving self-balancing and adaptive fusion of accuracy among models to construct a high-precision dynamic surrogate model. Initially, the weights of each model are set according to their structural complexity and sample adaptability. =0.35, =0.35, =0.30, and the subsequent weights are dynamically adjusted based on the prediction error (RMSE) and confidence interval width (CI) of each model on the validation set to achieve self-balancing of accuracy and time-varying adaptability among models.

[0084] Furthermore, a multi-dimensional knowledge base of "load-coal quality-operating condition" is constructed based on historical optimization cases. Through similarity matching and transfer parameter adjustment, model knowledge transfer and retraining are achieved to accelerate the optimization process of the surrogate model under new operating conditions. This process can be divided into the following steps:

[0085] 1) Environmental mutation detection and migration triggering mechanism:

[0086] To address load fluctuations, coal quality disturbances, and abrupt changes in combustion characteristics during deep peak shaving of thermal power units, environmental abrupt change detection indicators are defined. This method quantifies the prediction bias of different surrogate models under the same input, and utilizes the inconsistency of heterogeneous models to quantify the intensity of environmental changes. The larger the value, the more drastic the environmental changes. Normalization is then performed to account for the uncertainty of the Kriging model.

[0087]

[0088] in, Let represent the solution set of the population in generation t. This is a vector of decision variables.

[0089] To improve robustness, the system normalizes the uncertainty of the Kriging model, as follows:

[0090]

[0091] When the population for k consecutive generations satisfies >ηδ (ηδ is 1.96 corresponding to a 95% confidence level), the system determines that the environment has changed abruptly, the current combustion conditions have changed abruptly, and the knowledge transfer module is triggered; For the Kriging model at the input point The standard deviation of the prediction at that location.

[0092] 2) Two-space similarity matching:

[0093] Before knowledge transfer, the system uses a dual-space similarity matching mechanism (feature space and performance space) to determine whether historical models have transfer value. First, it extracts the deep embedding feature vector φ(x) of the samples in the feature space and calculates the current task samples... Samples of historical missions Cosine similarity:

[0094]

[0095] Among them, φ(x) can be obtained by a feature extraction network, reflecting the similarity of combustion conditions in feature dimensions; The similarity in the feature space is calculated using cosine similarity.

[0096] Secondly, within the performance space, the relationship between boiler thermal efficiency and NO X After normalizing the emission targets, the performance response similarity is defined as follows:

[0097]

[0098] Among them, F t =[η, t, C] NOx [t] represents the performance vector of the current task, F s =[η, s, C] NOx [s] represents the performance vector of the historical task. It is composed of boiler combustion thermal efficiency and flue gas NO₂, respectively. X The emission index composition is used to characterize the degree of difference in combustion performance between the two operating conditions; For performance space similarity, the relationship between combustion thermal efficiency and NO x The performance vector after normalizing the two emission targets is used for calculation.

[0099] Finally, by fusing the similarity indices of the two spaces, the overall similarity is defined as follows:

[0100]

[0101] Among them weighting factors =0.6. When Sim>0.85, the system considers the historical model to have migration potential and enters the migration stage.

[0102] 3) Progressive Structure Transfer and Retraining: To reduce the risk of erroneous transfer, the system adopts a progressive structure transfer strategy. In the initial stage, only the core feature extraction structure of the historical model is retained, and the output layer is reinitialized to adapt to the new operating condition data distribution of the current thermal power unit boiler combustion. Subsequently, fine-tuning training is performed under the guidance of the new environment samples, and the changing trends of validation error and prediction confidence are evaluated after each round of training. The optimization objective of the transfer model is defined as:

[0103]

[0104] in, This represents the prediction error for the current task. , These are the parameter vectors for the current and historical models, respectively. The transfer regularization coefficient is used to balance the degree of retention of old knowledge with the degree of adaptation to new tasks; This represents the current prediction error.

[0105] As the model gradually stabilizes, the migration intensity coefficient λ decays dynamically in the following manner:

[0106]

[0107] The constraints of old knowledge will be gradually relaxed to make the model more consistent with the distribution of new working conditions; The migration regularization coefficient for generation t+1 is gradually reduced by the decay factor β. is the migration regularization coefficient for generation t, used to control the impact of the previous generation model on the current task.

[0108] When the transfer model's validation error shows no significant fluctuations across multiple iterations, i.e., the rate of change of the validation error satisfies:

[0109]

[0110] in, The rate of change of the verification error is the relative change of the verification error between two iterations (t and t-1); Let be the verification error of the t-th generation; This is the preset error threshold.

[0111] Furthermore, if the confidence interval width CI remains stable in multiple iterations, the system will incorporate the model into the current heterogeneous agent system for integrated prediction; if the error diverges or the confidence level decreases, the migration path will be terminated and the system will revert to the original model training process to ensure the overall modeling accuracy and stability of the system.

[0112] refer to Figure 3 Driven by a surrogate model, a multi-objective evolutionary optimization algorithm is executed to globally optimize the fused surrogate model. Once the convergence condition is met, the optimal combination of process setpoints for the boiler combustion system of the thermal power unit is output for actual combustion control. In the initialization phase, the system first generates a uniformly distributed initial population and evaluates it using the original objective function. Simultaneously, it characterizes the convergence and solution set distribution features of the optimization process. In the evolutionary phase, the parent generation produces offspring through crossover and mutation, which are predicted by a heterogeneous surrogate using feature similarity. Performance similarity The algorithm uses a weighted fusion criterion to select candidate solutions. If a significant change in the environment is detected, the historical model is updated and sample replenishment is triggered to maintain the predictive stability of the surrogate model under the new conditions. The algorithm repeats this process until the termination condition is met.

[0113] Specifically, this embodiment takes a 660MW supercritical opposed-fired boiler as the research object, and selects operating variables closely related to combustion efficiency and pollutant emissions as input parameters, including raw coal conveying rate (Rm), primary / secondary air volume (V1, V2), and damper opening (D). w ), burnout wind ratio (R) b ), oxygen content in flue gas (O set The output parameters are combustion efficiency η and NO. X Emission concentration This embodiment uses operational data acquired from the power plant's distributed control system (DCS) and online monitoring platform, with a total sample size of 720 sets. Of these, 540 sets are used for model training, and 180 sets are used for testing and verification. A combustion optimization model for the boiler system of a thermal power unit is established using a multi-stage dynamic heterogeneous proxy model.

[0114] Furthermore, data preprocessing and optimization objectives are established. Step 1.1: Data preprocessing is performed on the originally collected operating parameters, and the optimization objectives are specified as: maximizing boiler thermal efficiency and minimizing NO. X Emission concentration. The multi-objective optimization problem is expressed as follows:

[0115]

[0116] A normalization process is used to linearly map each eigenvalue to the interval [0, 1]. The formula is expressed as:

[0117]

[0118] The mean of the input variables after normalization is between 0.40 and 0.75.

[0119] Preferably, the constructed multi-stage dynamic heterogeneous agent model: First, in the collaborative global exploration modeling stage, the Kriging model and least squares support vector machine (LSSVM) are used for collaborative global exploration and preliminary fitting during the early algorithm optimization process. The predicted output of the Kriging model can be expressed as:

[0120]

[0121] Meanwhile, the expression for the LSSVM model is:

[0122]

[0123] Through joint training and complementary correction, a preliminary solution space for combustion performance is obtained. In the dynamic characteristic refinement modeling stage, that is, in the middle of the algorithm's operation, an Elman network is introduced to construct a dynamic refinement model, achieving local prediction refinement for key regions; the hidden layer output state of the Elman network is updated by the following formula:

[0124]

[0125]

[0126] In the adaptive fusion stage, a heterogeneous model fusion function is constructed. An adaptive weight adjustment mechanism is designed for each sub-model, dynamically fusing the outputs of the Kriging model, LSSVM model, and Elman network to form a high-precision fusion proxy model for mid-to-late-stage optimization search. The fusion expression is:

[0127]

[0128] The weights are dynamically adjusted based on the prediction error (RMSE) and confidence interval width (CI) of each model on the validation set, achieving self-balancing and adaptive fusion of accuracy among models to construct a high-precision dynamic surrogate model. Initially, the weights of each model are set according to their structural complexity and sample adaptability. =0.35, =0.35, =0.30, and the subsequent weights are dynamically adjusted based on the prediction error (RMSE) and confidence interval width (CI) of each model on the validation set to achieve self-balancing of accuracy and time-varying adaptability among models.

[0129] Specifically, an environmental mutation detection index is defined and normalized by combining it with the uncertainty of the Kriging model:

[0130]

[0131] The system normalizes the uncertainty of the Kriging model, as follows:

[0132]

[0133] When the population for k consecutive generations satisfies >ηδ, the system determines that a sudden change has occurred in the environment and the current combustion condition, triggering the knowledge transfer module. Before knowledge transfer, the system uses a dual-space similarity matching mechanism based on feature space and performance space to determine whether the historical model has transfer value. First, deep embedding feature vectors of samples are extracted in the feature space, and the cosine similarity between the current task sample and the historical task sample is calculated:

[0134]

[0135] Among them, φ(x) can be obtained by the feature extraction network, reflecting the similarity of combustion conditions in the feature dimension.

[0136] Secondly, within the performance space, the relationship between boiler thermal efficiency and NO X After normalizing the emission targets, the performance response similarity is defined as follows:

[0137]

[0138] in, , The boiler combustion thermal efficiency and flue gas NO2 are respectively determined by the boiler combustion thermal efficiency and flue gas NO2. X The emission indexes are used to characterize the degree of difference in combustion performance between the two operating conditions.

[0139] Finally, by fusing the similarity indices of the two spaces, the overall similarity is defined as follows:

[0140]

[0141] Among them weighting factors =0.6, when When the value is greater than 0.85, the system considers the historical model to have migration potential and enters the migration stage.

[0142] Furthermore, a multi-objective evolutionary optimization algorithm is executed under the drive of the surrogate model to perform global optimization of the fused surrogate model. Once the convergence condition is met, the optimal combination of process setpoints for the boiler combustion system of the thermal power unit is output for actual combustion control. In the algorithm initialization phase, the system first generates a uniformly distributed initial population and evaluates it using the original objective function. Simultaneously, it is used to characterize the convergence and solution set distribution features of the optimization process. In the evolutionary phase, the parent generation produces offspring through crossover and mutation. The heterogeneous surrogate model predicts these offspring and uses a weighted fusion criterion of feature similarity and performance similarity to screen candidate solutions. If a significant change in the environment is detected, the historical model is updated and sample replenishment is triggered to maintain the predictive stability of the surrogate model under new operating conditions. To control computational costs, when the number of individuals in the convergence profile (CA) exceeds the maximum capacity, the system uses an adaptive sampling mechanism to limit the number of new samples evaluated. The number of sampled individuals is linearly related to the population size, defined as follows:

[0143]

[0144] in, This represents the total population size. When the initial total population size is selected as 100, the corresponding number of adaptively sampled individuals is 10.

[0145] CA and DA are dynamically updated in each iteration to maintain a balance between convergence and diversity. Both prediction targets are rapidly evaluated using the dynamic heterogeneous surrogate model constructed in step two. In the algorithm settings, the population size is 100, the number of iterations is 200, the crossover factor is 0.25, and the mutation factor is 0.02. Experimental data are from the thermal power unit operation database. After processing, 540 sets of typical combustion operation data were selected as the training set, covering typical operating conditions such as deep peak shaving, load fluctuation, and coal quality changes. The remaining 180 sets of data were used as the test set to verify the model's generalization ability under different operating condition disturbances. The model is considered to converge when the convergence condition is met (i.e., the average rate of change of the objective function is less than the threshold ε for several consecutive generations). n When the system verifies the baseline function and outputs the final Pareto solution set based on actual operating conditions, the optimal process set output includes: primary air volume V1 ★ Secondary air volume V2 ★ Damper opening degree w ★ Burnout wind ratio R b ★ Coal feeding rate R m ★ With the set value of oxygen content in flue gas O set ★ The data is then transmitted to the actuators via the unit-level stratified control system (DCS) to achieve online optimization and closed-loop control of boiler combustion. (See the algorithm test results diagram for reference.) Figure 4 .

[0146] The beneficial effects of this invention are as follows:

[0147] (1) This invention proposes a multi-stage dynamic heterogeneous proxy modeling method, which organically integrates the Kriging model, support vector machine and recurrent neural network, taking into account both global fitting ability and time response characteristics, and improving the prediction accuracy and stability of the model in complex combustion systems.

[0148] (2) The present invention introduces an adaptive weighting mechanism and a transfer learning strategy, enabling the model to automatically adjust the parameter distribution according to different loads, coal quality and working conditions, realizing cross-scenario knowledge transfer and self-learning capabilities, and significantly reducing the cost of remodeling and calibration.

[0149] (3) The present invention adopts a multi-objective optimization framework driven by a proxy model, which effectively balances the contradiction between combustion efficiency and pollutant emissions, and improves the operation economy and cleanliness of thermal power units in the deep peak shaving stage.

[0150] (4) By coupling with the component layered control system (DCS), this invention realizes online updating and closed-loop control of optimization results, and has good engineering feasibility and promotion and application value.

[0151] 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.

[0152] 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 combustion optimization method using a multi-stage dynamic heterogeneous proxy model for boiler systems in thermal power units, characterized in that, include: The operation data of the target thermal power unit under different load and coal quality conditions are collected, and the operation data are sequentially subjected to outlier detection, missing value completion, correlation screening, and kernel principal component analysis to obtain the core operation variables. Build a system to maximize boiler thermal efficiency and minimize NO X A multi-objective optimization problem with emissions as the objective; The core operational variables are normalized. Based on the multi-objective optimization problem, the core operational variables are explored globally and initially fitted using the Kriging model and least squares support vector machine to obtain a preliminary solution space. The dynamic time delay characteristics of the preliminary solution space are captured and locally refined using an Ullman recurrent neural network to obtain a dynamically refined solution space. A high-precision surrogate model is obtained by fusing the preliminary solution space and the dynamic refined solution space using an adaptive weighting mechanism. A knowledge base of load, coal quality, and operating conditions was constructed based on historical optimization cases. Based on the load-coal quality-operating condition knowledge base, similarity matching and transfer parameter adjustment are used to perform model knowledge transfer and retraining on the high-precision proxy model to obtain a transfer proxy model. The migration surrogate model is globally optimized using a multi-objective evolutionary optimization algorithm until the convergence condition is met, and the optimal combination of process settings is obtained. The optimal combination of process settings is then updated in the combustion controller of the target thermal power unit.

2. The combustion optimization method for a multi-stage dynamic heterogeneous proxy model in a thermal power unit boiler system according to claim 1, characterized in that, Operating data of the target thermal power unit under different load and coal quality conditions were collected. Outlier detection, missing value completion, correlation screening, and kernel principal component analysis were then performed on the operating data to obtain the core operating variables, including: Operational data is collected using a multi-source online sensor network deployed on and around the boiler body. This operational data includes: raw coal characteristic data and unit operating parameters. The raw coal characteristic data includes: moisture, ash content, and volatile matter. The unit operating parameters include: boiler temperature field distribution, air-to-coal ratio, oxygen content, damper opening, burnout air ratio, and SCR outlet NO. X concentration; The running data is spatiotemporally aligned based on the alignment time window width; the expression for the alignment time window width is: ;in, The width of the alignment time window; The characteristic combustion reaction time of coal type; The aligned running data is then subjected to outlier detection and missing value completion. Based on the completed running data, the Spearman rank correlation coefficient is calculated, and variables with an absolute correlation coefficient with the target variable that is not less than a set threshold are selected. Kernel principal component analysis is then used to extract low-dimensional features from the selected results to obtain the core operational variables.

3. The combustion optimization method for a multi-stage dynamic heterogeneous proxy model in a thermal power unit boiler system according to claim 2, characterized in that, Build a system to maximize boiler thermal efficiency and minimize NO X Multi-objective optimization problems with emissions as the objective include: To maximize boiler thermal efficiency and minimize NO X Emission concentration was set as the overall optimization target; Set constraints; the constraints include: furnace temperature constraints, steam pressure constraints, furnace negative pressure constraints, emission standard constraints, oxygen content constraints, and operating variable constraints; A multi-objective optimization problem is constructed based on the overall optimization objective and the constraints; the expression of the multi-objective optimization problem is: ;in, For the decision vector set; A vector of decision variables; The objective function vector; For combustion efficiency; NO X emission; Let i be the i-th constraint function; Number the constraint functions; This represents the total number of constraint functions; This serves as the lower bound for the decision variable. For operation variables; This serves as the upper bound for the decision variable; To obtain the sign for the decision variable; The number of constraints.

4. The combustion optimization method for a multi-stage dynamic heterogeneous proxy model in a thermal power unit boiler system according to claim 3, characterized in that, The expression for the high-precision proxy model is: ;in, ; ; ; ; The resulting high-precision proxy model; , , These are the first adaptive weight coefficient, the second adaptive weight coefficient, and the third adaptive weight coefficient, respectively. , , These are the prediction outputs of the Kriging model, the least squares support vector machine, and the Ullman recurrent neural network, respectively. It is a deterministic trend function; The mean is zero and the covariance is σ. 2 R(θ) is a Gaussian process; σ 2 R(θ) represents the process variance; R(θ) represents the correlation function. This represents the number of training samples; For Lagrange multipliers; For kernel functions; This is the first bias term; To output the weight matrix; This is a hidden layer state; This is the first bias vector; The input weight matrix; This is the current input feature vector; This is the recursive weight matrix; This is the second bias vector; This represents the activation function.

5. The combustion optimization method for a multi-stage dynamic heterogeneous proxy model in a thermal power unit boiler system according to claim 4, characterized in that, Based on the load-coal quality-operating condition knowledge base, similarity matching and transfer parameter adjustment are used to perform model knowledge transfer and retraining on the high-precision proxy model to obtain a transferred proxy model, including: The load-coal quality-operating condition knowledge base is constructed based on the feature space mapped by the kernel principal component analysis and the performance space output by the multi-objective optimization problem. A normalized index for detecting environmental mutations is constructed; the expression for the normalized index for detecting environmental mutations is: ;in, ; This serves as a normalized index for detecting environmental mutations. This serves as an indicator for detecting environmental mutations in generation t. For the Kriging model at the input point The standard deviation of the forecast at that location; Let be the solution set of the population in generation t; The normalized index for environmental mutation detection is calculated based on the preliminary solution space and the dynamic refined solution space to obtain the current detection index; When the current detection index is greater than the mutation threshold, the feature space and the performance space are calculated using the total similarity formula to obtain the similarity between the two spaces; the expression of the total similarity formula is: ;in, ; ; The similarity between the two spaces; As a weighting factor; The similarity of the feature space; The similarity of the performance space; For the current task sample Samples of historical missions Cosine similarity; For the current task sample The deep embedding feature vector; Sample of historical missions The deep embedding feature vector; This is the performance vector for the current task; This represents the performance vector of historical tasks. When the similarity between the two spaces is greater than a similarity threshold, the high-precision surrogate model is subjected to transfer iterative learning using a transfer optimization objective until the prediction error between two adjacent iterations meets the condition of having no significant fluctuation range, thus obtaining the transfer surrogate model; the expression for the transfer optimization objective is: ;in, ; This represents the prediction error for the current task. This represents the current prediction error; For migration regularization coefficients; This is the parameter vector of the current model; This is the parameter vector of the historical model; The migration regularization coefficient for the (t+1)th generation; Let be the migration regularization coefficient for generation t; The expression for the range of error with no significant fluctuation is: ;in, To verify the rate of change of error; Let be the verification error of the t-th generation; This is the preset error threshold.

Images