A thermal deviation self-adaptive control method for a boiler heating surface of a thermal power unit based on multi-modal dynamic optimization

Abstract

The application discloses a thermal deviation self-adaptive control method for a boiler heating surface of a thermal power unit based on multi-modal dynamic optimization, and belongs to the technical field of intelligent combustion optimization control of the thermal power unit. The method comprises the following steps: a dynamic double-target optimization model is constructed, taking the minimization of a thermal deviation coefficient and the maximization of boiler efficiency as targets; a multi-modal multi-target dynamic optimization algorithm based on a response mechanism is proposed; a double-space environment drift sensing quantifies a load disturbance intensity, triggers optimal transmission matching of an external file Pareto branch, and samples and reconstructs an anisotropic credible domain; a three-partition collaborative search strategy driven by potential energy is adopted to distribute population evolution tasks, and target space front convergence and decision space multi-air distribution mode diversity are synchronously maintained. The application can quickly adapt to load fluctuation, significantly suppresses the thermal deviation of the heating surface while improving the combustion efficiency, and guarantees safe and economic operation of the unit.

Description

Technical Field

[0001] This invention relates to the field of intelligent combustion optimization control technology for thermal power units, and in particular to an adaptive control method for thermal deviation of boiler heating surfaces in thermal power units based on multimodal dynamic optimization. Background Technology

[0002] With the deepening of the construction of the new power system, thermal power units are accelerating their transformation from the traditional base load operation mode to a flexible operation mode of deep peak shaving and rapid load change. A report from the China Electricity Council indicates that 189 million kilowatts of coal-fired power units nationwide have completed flexibility retrofitting. The average minimum output load of the retrofitted units has decreased from 43.7% before the retrofit to 29.8%, and deep peak shaving has become the norm for thermal power operation. However, when the units operate under low load conditions, the furnace temperature drops sharply from approximately 1400℃ under rated conditions to approximately 1100℃, the ignition stability of pulverized coal decreases, the flame center position shifts, and the distortion of the flue gas temperature field distribution along the width direction at the furnace outlet intensifies. Taking a 1050MW ultra-supercritical boiler as an example, during deep peak shaving and load reduction, the wall temperature of the screen-type superheater rises rapidly from 570℃ to 640℃ within 30 minutes, and the wall temperature deviation can reach 100℃ within minutes. The existence of this thermal deviation causes a sharp increase in the risk of local tube wall overheating. The wall temperature of the screen-type superheater can reach up to 696°C, which can lead to tube rupture accidents and forced shutdown in severe cases, directly restricting the safety of the unit's deep peak-shaving operation.

[0003] Regarding the formation mechanism of thermal deviation in heating surfaces, numerous universities and research institutes in China have conducted extensive numerical simulation studies. Wang Tao et al. from Tsinghua University, in their paper published in the *Journal of Coal Science and Technology*, pointed out that as coal-fired power units develop towards higher parameters and deeper peak shaving, preventing tube wall overheating caused by thermal deviation in boiler heating surfaces is one of the key issues. This study used Fluent to numerically simulate a 660MW coal-fired boiler with opposing front and rear walls, establishing a coupled heat transfer model inside and outside the screen-type superheater tubes. The results showed that when the external secondary air swirl angle increased from 15° to 60°, the temperature distribution of the screen-type superheater changed from "high at both ends and low in the middle" to "low at both ends and high in the middle." When the swirl angles were 45° and 60°, respectively, the maximum wall temperature of the screen-type superheater exceeded the maximum allowable temperature of the heating surface material. Research published in the *Journal of North China Electric Power University (Natural Science Edition)* in 2018 by Wang Dongfeng et al. from North China Electric Power University indicated that the boiler combustion efficiency is optimal when both the burnout air rate and the additional air rate are around 15%. The above research shows that there is a significant contradiction between the safety of the heating surface and the economy of the boiler. Although adjusting the air distribution method can improve the wall temperature distribution, it often affects the combustion efficiency. It is difficult to balance both by relying solely on experience-based air distribution or single-objective optimization.

[0004] Regarding the monitoring and optimization of heated surface wall temperatures, existing distributed control systems in power plants primarily rely on operator experience for adjusting secondary dampers and burnout air. Actual operating data from a 1050MW unit shows that under deep peak-shaving conditions, overheating of the screen-type superheater is often accompanied by significant deviations in wall temperature on the left and right sides, and the overheating is highly sudden; once a deviation occurs on either side, the wall temperature can rise to 100℃ within minutes, making operational adjustments extremely difficult. A field study published in 2023 in *Electric Power Technology & Environmental Protection* by Shi Pengfei et al. of the State Energy Group Science and Technology Research Institute indicates that when a 600MW supercritical unit's load is deeply adjusted from 50% to 40% of rated load, the coal consumption for power generation increases by approximately 9g / (kW·h); when adjusted from 40% to 30%, the coal consumption for power generation increases by more than 50g / (kW·h). A 2024 report on industrial technology issues released by the China Energy Research Society pointed out that when a 600MW unit is deeply adjusted to 20% of its rated load, the coal consumption for power supply increases by 1.4 to 1.7 times compared to the rated load, with an absolute increase of 70 to 100 g / (kW·h). These data illustrate that the current operation mode, which relies primarily on manual experience, sacrifices economic efficiency for safety during deep peak shaving. The root cause lies in the lack of real-time quantitative assessment and automatic optimization methods for furnace outlet thermal deviation.

[0005] In the research of combustion optimization algorithms, existing methods are mainly divided into two categories: mechanism model-driven and data-driven. Regarding mechanism models, while numerical simulation methods based on computational fluid dynamics can accurately simulate the combustion and heat transfer processes in the furnace, each simulation calculation is time-consuming and requires precise furnace geometry and coal quality parameters, making it difficult to embed real-time optimization closed loops. A 2021 study by Nanchang University showed that using Fluent to perform numerical simulations of combustion in a 660MW boiler requires establishing multi-physics coupled models, including a gas-phase turbulence k-ε two-equation model, a non-premixed combustion model, and a particulate phase random trajectory model, resulting in high computational costs and long iteration cycles. Regarding data-driven methods, existing research mainly uses neural networks and support vector regression to establish surrogate models of the combustion system. However, these models are limited by the drift characteristics of the input-output mapping relationship of the combustion system under deep peak-shaving variable operating conditions, resulting in insufficient model adaptability. A 2014 study by Song Jinghui et al. in the *Journal of Power Engineering* showed that the burnout air rate significantly affects the furnace temperature field and NOx emission concentration field. From the perspective of boiler safety operation, the burnout air rate should not exceed 0.30; from the perspective of emission control, it should be controlled above 0.23; and considering both factors, it should be controlled within the range of 0.23 to 0.30. The above parameter range drifts with changes in load and coal quality, making it difficult for static optimization methods to track such dynamic changes.

[0006] Currently, the control of boiler heating surface thermal deviation under deep peak-shaving conditions in thermal power units still faces several technical challenges:

[0007] Firstly, thermal deviation is coupled with multiple factors such as secondary air distribution method, burnout air rate, and burner tilt angle, and it is difficult to find the globally optimal air distribution configuration by adjusting a single variable one by one.

[0008] Secondly, the safety of the heating surface and the economy of the boiler constitute a typical conflicting objective, and existing single-objective or weighted summation optimization methods cannot provide an effective trade-off boundary between the two.

[0009] Third, fluctuations in load and coal quality cause the optimal air distribution scheme to drift continuously, and static optimization methods cannot track this dynamic change;

[0010] Fourth, multiple equivalent air distribution schemes with similar objectives are difficult to identify and maintain, and operators lack flexible alternative strategies under different safety margin preferences. These problems restrict the coordinated improvement of unit safety and economy under deep peak-shaving conditions.

[0011] Based on the above background, this invention proposes an adaptive control method for thermal deviation of boiler heating surface in thermal power units based on multimodal dynamic optimization. Summary of the Invention

[0012] To overcome the problems of severely uneven boiler heating surface wall temperature distribution, dramatic increase in thermal deviation, and lag in response to load changes under deep peak shaving conditions in existing technologies, this invention provides an adaptive control method for thermal deviation of boiler heating surface in thermal power units based on multimodal dynamic optimization. This method can quickly adapt to load fluctuations, significantly reduce thermal deviation of heating surface while improving combustion efficiency, and ensure safe and economical operation of the unit.

[0013] To achieve the above objectives, the present invention provides the following solution:

[0014] An adaptive control method for thermal deviation of boiler heating surface in thermal power units based on multimodal dynamic optimization includes the following steps:

[0015] Step S1: Collect the operating parameters in the distributed control system of the thermal power unit, and based on the operating parameters, construct a dynamic bi-objective optimization model with the objectives of minimizing the comprehensive thermal deviation coefficient of the heating surface and maximizing the boiler thermal efficiency.

[0016] Step S2: Real-time sensing of changes in thermal power unit load and coal quality, calculation of comprehensive drift response intensity, and triggering dual-space response reconstruction based on the comprehensive drift response intensity;

[0017] Step S3: Based on the comprehensive drift response intensity, perform branch matching and anisotropic trust domain migration on the historical Pareto wind distribution model branches in the external archive to generate a multimodal initial population;

[0018] Step S4: Introduce homotopy functions to asymptotically recover wall temperature overheat and vapor temperature constraints, and dynamically divide the population into elite development zone, cross-modal bridging zone and sparse compensation zone according to dual-space search potential energy, and perform tangential fine search, inter-branch bridging migration and sparse region directional exploration respectively.

[0019] Step S5: Maintain the external archive using a branch-keeping strategy based on geodesic special congestion distance, output the optimal Pareto solution set and send it to the distributed control system.

[0020] Preferably, in step S1, the operating parameters include: the array data of wall temperature measuring points along the width of the furnace of the outlet header of the screen-type superheater and the high-temperature reheater, the reheat steam temperature, the unit load command and coal quality parameters, and the collected data are subjected to outlier detection and missing value completion.

[0021] The data of the boiler heating surface wall temperature measuring point array comes from the multi-point thermocouple array arranged along the width direction of the outlet header of the screen-type superheater and high-temperature reheater.

[0022] The opening degree of the secondary air auxiliary damper, the opening degree of the burnout damper, and the burner swing angle of each layer were determined as the optimization decision variables.

[0023] Preferably, in step S1, the comprehensive thermal deviation coefficient of the heating surface is calculated based on the array data of wall temperature measuring points along the width of the furnace from the outlet header of the screen-type superheater and the high-temperature reheater; the boiler thermal efficiency is calculated based on the inverse balance method, and the following dynamic bi-objective optimization model is established:

[0024]

[0025] in, Indicates environmental moment The feasible decision space is defined below; x∈Ω(t) are the optimization decision variables. Let f1(x, t) be the objective function for the comprehensive thermal deviation coefficient of the heating surface, and f2(x, t) be the objective function for the negative value of boiler thermal efficiency. Let g be the environmental time. i (x, t) is the i-th dynamic inequality constraint function, h j (x, t) is the th A dynamic equality constraint function.

[0026] Preferably, in step S2, the normalized drift of the external file in the target space, decision space, and constraint space is calculated, and the comprehensive drift response intensity is obtained by adaptive weighting. When the comprehensive drift response intensity exceeds a preset threshold, it is determined that the environment has changed significantly, triggering dual-space response reconstruction.

[0027] Preferably, in step S2, the quantification method of the comprehensive drift response intensity is as follows: the target space drift ΔF, the decision space drift ΔX, and the constraint space drift ΔG are defined as follows:

[0028]

[0029] in, For the environment at all times External archives, Let be the center vector of all individuals in the external archive in the target space. For the Euclidean norm, Use a small constant to prevent the denominator from being zero;

[0030]

[0031] in, Let be the central vector of all individuals in the external archive in the decision space. Use a small constant to prevent the denominator from being zero;

[0032]

[0033] in, Let be the number of inequality constraints. For all individuals in the external archives in the first The average degree of violation on each constraint, Use a small constant to prevent the denominator from being zero;

[0034] The overall response intensity is:

[0035]

[0036] in, , , For adaptive weighting coefficients, satisfying + + =1, dynamically adjusted according to the unit load change rate; when > The system is identified as experiencing a significant environmental change, triggering branch-level migration and trusted domain reconstruction. This is a preset threshold.

[0037] Preferably, in step S3, based on the comprehensive drift response intensity, the historical Pareto wind distribution model branches in the external archives are identified and divided. The soft correspondence matching relationship between the historical branches and the current candidate branches is established through the entropy regularized optimal transmission model. Regularized migration estimation is performed on each matching branch to determine the prediction center of each branch in the new environment. With the center of the migrated branch as the center, an anisotropic ellipsoidal credibility domain is spanned by the tangential basis and normal complement space. The tangential and normal sampling radii are adaptively adjusted according to the response intensity to reconstruct the population and generate a multimodal initial population adapted to the new working conditions.

[0038] Preferably, in step S3, the branch matching is achieved by constructing a migration cost matrix between the historical Pareto branch and the current candidate branch. The first historical branch and the first The migration cost for each candidate branch is:

[0039]

[0040] in, , The first The target values ​​and decision variable values ​​of each historical branch center , The first The target value and decision variable value of each candidate branch center , These are the local tangent basis vectors for the historical branch and the candidate branch, respectively. , , These are the weighting coefficients. It is the absolute value symbol;

[0041] Matching relationships are solved by... The optimal transmission model is obtained at the cost of entropy regularization, which allows each historical branch to be matched to the current candidate branch in a soft correspondence manner.

[0042] Preferably, in step S3, the specific method of the anisotropic trusted domain migration is as follows: for each Pareto branch that has completed the migration, the migration branch center is used as the reference point. Centered on the tangential base With normal space Zhang Cheng's ellipsoid credibility region, the first The first branch generated The sampling method for the initial individuals is as follows:

[0043]

[0044] in, Represents element-wise product. For random perturbation vectors, Let the radius be the tangential trust region. Let the radius be the normal region of confidence. Based on the radius, The scaling factor is determined by the eigenvalues ​​of the branch covariance matrix;

[0045] The number of particles allocated to each branch is:

[0046]

[0047] in, The total population size The total number of branches, For the first The quality of a branch is determined by a combination of its convergence index, average constraint violation rate, and individual information entropy within the branch.

[0048] Preferably, in step S4, a homotopy function is introduced to progressively recover the wall temperature overheating constraint and the steam temperature deviation constraint. The individuals in the population are arranged in ascending order of dual-space search potential energy and dynamically divided into an elite development zone, a cross-modal bridging zone, and a sparse compensation zone. In the elite development zone, a local fine search is performed along the tangent of its branch. In the cross-modal bridging zone, branches with similar target spaces and different decision spaces are selected to perform bridging migration. In the sparse compensation zone, directional compensation exploration is performed based on the branch scarcity and constraint boundary information to achieve co-evolution of convergence and multimodal preservation.

[0049] Preferably, in step S4, a homotopy function is introduced to progressively recover the wall temperature overheat constraint and the steam temperature deviation constraint, specifically as follows:

[0050] No. The relaxed feasible region is defined as follows:

[0051]

[0052] in, For the first Inequality constraint functions, For the first The generation A relaxed boundary of constraints, , This is the initial relaxation amount. The contraction time constant;

[0053] The allowable temperature limit for the wall temperature constraint is determined based on the material of the heated surface pipe.

[0054] The equality constraint tolerance is:

[0055]

[0056] in, For the initial equality tolerance, G max The maximum number of iterations, The contraction index;

[0057] Introduce a homotopy penalty function:

[0058]

[0059] in, The original objective function vector, This is a dynamic penalty coefficient. The initial penalty value, To penalize the rate of growth, moderate constraint violations are allowed in the early stages of iteration to preserve historical information, and the strict feasible region is gradually restored as the iteration progresses.

[0060] Preferably, in step S4, the dual-space search potential is constructed by: taking into account the convergence error of each individual. Constraint Violation Target space congestion Decision-making space geodesy congestion The scarcity of its branches and historical contributions H contrib (x), construct the dual-space search potential energy:

[0061]

[0062] in, , , , , , The weighting coefficients for each indicator; the population is calculated as follows: Sort in ascending order, first One was designated as an elite development zone, in the middle One is included in the cross-modal bridging region, the remaining ones Included in the sparse compensation zone, among which, For population size, , , The partitioning ratio coefficient is determined by the response intensity. Adaptive adjustment with population branch entropy, and .

[0063] Preferably, in step 5, a branch-keeping strategy is used to maintain the external archives. The archives are pruned based on the geodesic special congestion distance of individuals within the branch. Multiple equivalent solution sets with similar target space performance but significantly different air distribution methods are retained. The optimal Pareto solution set is sent to the distributed control system for execution, thereby realizing online dynamic closed-loop control of the boiler heating surface thermal deviation.

[0064] Compared with the prior art, the beneficial technical effects of the present invention are as follows:

[0065] This invention achieves rapid inheritance and relocation of historical effective air distribution mode branches after environmental changes through dual-space environmental drift perception and response intensity quantification, avoiding the waste of convergence information caused by completely random restarts. Through entropy regularization optimal transmission branch matching and anisotropic credibility domain reconstruction, it simultaneously maintains the convergence of the target space and the diversity of multiple air distribution modes in the decision space, alleviating the problem of branch collapse caused by convergence pressure in traditional methods. Through potential energy-driven three-zone collaborative search and homotopy dynamic constraint processing, it balances boiler thermal efficiency while ensuring that wall temperature does not exceed limits and steam temperature deviation is controllable, overcoming the deficiency of existing methods in providing a safe-economic-effective trade-off boundary. Through an external archive maintenance strategy for branch retention, it provides retention capability for equivalent operating schemes with similar target performance but significantly different air distribution methods, enhancing the flexibility and robustness of combustion regulation under deep peak-shaving conditions. This invention realizes online dynamic optimization and closed-loop control of boiler heating surface thermal deviation during deep peak-shaving load changes, maintaining high boiler thermal efficiency while reducing peak wall temperature and thermal deviation coefficient, providing an innovative solution for the coordinated improvement of safety and economy in deep peak-shaving operation of thermal power units.

[0066] 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

[0067] 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.

[0068] Figure 1 The flowchart of an adaptive control method for thermal deviation of boiler heating surface in thermal power unit based on multimodal dynamic optimization is provided by the present invention. Detailed Implementation

[0069] 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.

[0070] 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.

[0071] like Figure 1 As shown, this invention provides an adaptive control method for thermal deviation of boiler heating surface in thermal power units based on multimodal dynamic optimization. It utilizes a multimodal multi-objective dynamic optimization algorithm to dynamically optimize the boiler secondary damper opening, burnout air rate, and burner swing angle under deep peak shaving conditions in thermal power units. This method aims to smooth the thermal deviation distribution of the screen-type superheater and high-temperature reheater along the width direction, ensuring that the heating surface wall temperature does not exceed the limit while also taking into account the boiler thermal efficiency, thereby achieving a synergistic improvement in the safety and economy of deep peak shaving operation of thermal power units.

[0072] This invention uses the control of heat transfer surface thermal deviation in a 660MW ultra-supercritical coal-fired boiler with opposed front and rear walls under deep peak-shaving conditions as a specific embodiment. It employs a opposed-front-and-rear-wall combustion method, with wall temperature measuring points arranged along the width of the furnace in the outlet headers of the screen-type superheater and high-temperature reheater, forming a wall temperature measuring point array. The unit's distributed control system (DCS) collects wall temperature measuring point data, reheat steam temperature, unit load commands, and coal quality parameters in real time via the OPC protocol. The population size N is set to 100, the external file capacity to 50, and the maximum number of iterations Tmax to 100 generations. The specific scheme is as follows:

[0073] Step S1: Data acquisition and preprocessing.

[0074] Specifically, a multi-source online sensor network deployed on the boiler body and its auxiliary equipment is used to collect real-time operating data of the thermal power unit under different load and coal quality conditions. The collected data is transmitted to the plant-level monitoring information system (MIS) via the OPC protocol and is spatiotemporally aligned based on the unit's operating characteristics.

[0075] The collected data includes: array data of wall temperature measuring points along the width of the furnace from the outlet headers of the screen-type superheater and high-temperature reheater (12 measuring points each); reheat steam temperature (1 measuring point on each side); unit load command (AGC command); coal quality parameters (lower heating value, volatile matter, ash content); current values ​​of operating variables such as the opening of the secondary air auxiliary damper, the opening of the burnout damper, and the burner swing angle at each level; and parameters required for thermal efficiency calculation such as flue gas temperature and oxygen content in the flue gas.

[0076] Furthermore, outlier detection and missing value completion are performed during the data preprocessing stage.

[0077] Outlier detection employs the 3σ criterion: for each measurement point's time-series data, the mean value within a sliding window is calculated. Data points exceeding the range of μ±3σ are marked as outliers and removed.

[0078] Missing value completion uses a linear interpolation method. For data segments with more than 5 consecutive missing sampling points, historical data under similar operating conditions are used for replacement.

[0079] The opening degrees of the secondary air auxiliary dampers and burnout dampers on each floor, as well as the vertical swing angle of the burner, were determined as the optimization decision variables. In this embodiment, the optimization decision variables are... The decision variables include: the opening of the secondary air auxiliary dampers on the front and rear walls (3 layers each, 6 dimensions); the opening of the burnout dampers on the front and rear walls (1 layer each, 2 dimensions); and the burner's vertical swing angle (1 dimension), for a total of d=9 dimensions. The value range of each variable is determined according to the boiler design specifications, with damper openings ranging from 0 to 100% and burner swing angles ranging from -30° to +5°.

[0080] Furthermore, a dynamic bi-objective optimization model is constructed.

[0081] A dynamic bi-objective optimization problem is established, aiming to minimize the comprehensive thermal deviation coefficient of the heated surface and maximize the boiler thermal efficiency. Let the real-time wall temperature at the i-th wall temperature measuring point be _i_. Then the average wall temperature of the heated surface is:

[0082]

[0083] Define the instantaneous thermal deviation coefficient of the i-th measuring point as:

[0084]

[0085] The combined thermal deviation coefficient f1 between the screen-type superheater and the high-temperature reheater is:

[0086]

[0087] Where, δ sp δ represents the maximum absolute value of the instantaneous thermal deviation coefficient among all wall temperature measuring points of the screen-type superheater. rh This represents the maximum absolute value of the instantaneous thermal deviation coefficient among all wall temperature measuring points of the high-temperature reheater. α and β are the weighting coefficients of the screen-type superheater and the high-temperature reheater, respectively. α+β=1. In this embodiment, based on the allowable temperature margins of the two types of heat-receiving surface metal materials, α=0.45 and β=0.55, respectively.

[0088] Boiler thermal efficiency η is calculated based on the inverse balance method:

[0089]

[0090] in, For the heat loss of flue gas, Heat loss due to incomplete combustion of chemicals, This represents the heat loss due to incomplete combustion in the machine. The optimization objective is to maximize η, which is equivalent to minimizing f2 = -η.

[0091] Establish the following dynamic bi-objective optimization model:

[0092]

[0093] in, Indicates environmental moment The feasible decision space below; To optimize decision variables, For environmental moments, The comprehensive thermal deviation coefficient of the heated surface. For negative boiler thermal efficiency, dynamic inequality constraints. Including: Screen-type superheater wall temperature T wall Upper limit constraint (T) wall ≤ T max In this embodiment, the allowable temperature of T91 material is 610°C, and the upper limit of the high-temperature reheater wall temperature is constrained, with the reheat steam temperature T... reheat Deviation constraint is |T reheat - T set | ≤ 5℃, where T set This is the setpoint for the reheat steam temperature. Dynamic equality constraints. This includes the balance between unit load and coal consumption.

[0094] Step S2: Environmental drift identification and response intensity quantification.

[0095] Specifically, the system senses changes in unit load and coal quality fluctuations in real time, calculates the normalized drift of external files in the target space, decision space, and constraint space, obtains the comprehensive drift response intensity through adaptive weighting, and determines that the environment has changed significantly when the response intensity exceeds a preset threshold, triggering dual-space response reconstruction.

[0096] set up For the environment at all times The external archive contains the non-dominated solution set in the current environment. Define the center of the target space:

[0097]

[0098] Where F(x, t) is the objective function vector of individual x at environmental time t.

[0099] Decision Space Center:

[0100]

[0101] The target space normalized drift ΔF is:

[0102]

[0103] in, Let be the center vector of all individuals in the external archive in the target space. For the Euclidean norm, It is a small constant, and in this example, it is taken as... .

[0104] The decision space drift ΔX is:

[0105]

[0106] in, It is a small constant, and in this example, it is taken as... .

[0107] The constrained space normalized drift ΔG is:

[0108]

[0109] in, For small constants, in this embodiment, we take... , used to prevent the denominator from being zero; p is the number of inequality constraints; Let be the average violation rate of all individuals in the external archive on the i-th constraint.

[0110] The overall drift response intensity is:

[0111]

[0112] in, , , For adaptive weighting coefficients, satisfying + + =1, based on the unit load change rate Dynamically adjust weights: when When ≤1% Pe / min, ( , , =(0.4,0.3,0.3); when 1% < When ≤ 3% Pe / min, ( , , ) = (0.3, 0.4, 0.3); when When > 3% Pe / min, ( , , = (0.3, 0.3, 0.4).

[0113] when > When a strong environmental change is detected, branch-level migration and trusted domain reconstruction are triggered. This embodiment uses a preset threshold. Take 0.60. Taking the scenario where the unit load command decreases from 660MW to 330MW at a rate of 3% / min as an example, the comprehensive response strength... It jumped from approximately 0.2 to 0.72, exceeding the threshold. This triggers a response refactoring.

[0114] Step S3: Pareto memory, branch matching, and regularized transfer estimation.

[0115] Specifically, based on the comprehensive drift response intensity, the branches of historical Pareto wind distribution patterns in the external archives are identified and divided. In this step, the non-dominated solutions in the external archive A(t-1) are first identified and divided into branches. Based on the Euclidean distance and geodesic-specific congestion distance in the decision space, spectral clustering is used to divide the external archives into K Pareto branches. (k=1,2,…,K). Each branch The local structure consists of branch centers Local covariance matrix Σk and tangential basis V k Characterization:

[0116]

[0117]

[0118] Wherein, tangential basis V k It consists of the first m principal eigenvectors of Σk. In this embodiment, m=3, meaning that the three eigenvectors with the largest eigenvalues ​​in the covariance matrix are retained as the tangent basis of this branch. (Normal space complement) It consists of the remaining (dm) eigenvectors of Σk, where d is the total dimension of the decision variables (d=9 in this embodiment). The tangential basis matrix V k With the normal complement space matrix Satisfy: V k any column vector and Any column vectors are orthogonal.

[0119] In the current environment, after re-evaluating the external file A(t-1) from the previous environment under the current objective function, the migration cost between the k-th historical branch and the l-th candidate branch is:

[0120]

[0121] in, , Let be the target value and the decision variable value of the k-th historical branch center, respectively. , Let be the objective value and the decision variable value of the j-th candidate branch center, respectively. , These are the local tangent basis vectors for the historical branch and the candidate branch, respectively. , , The weighting coefficients satisfy α+β+γ=1. In this embodiment, we take the following values ​​respectively: =0.4、 =0.4、 =0.2.

[0122] Branch Matching Matrix The following entropy-regularized optimal transmission model was obtained by solving:

[0123]

[0124]

[0125] Where Γ is the branch matching matrix with dimension K. h ×K c K h K represents the number of Pareto branches identified in the historical context. c This represents the number of candidate branches in the current environment. Let Γ be the migration cost between the k-th historical branch and the l-th candidate branch, and let Γ be the element in the matrix. kl This indicates the degree of matching between the k-th historical branch and the l-th candidate branch. This indicates that the sum of each row of matrix Γ equals p. h The value of the corresponding row in the middle, that is, the total matching quality of each historical branch, is assigned to each candidate branch; This indicates that the sum of each column of matrix Γ equals p. c The values ​​in the corresponding columns, i.e., the total matching quality of each candidate branch, are allocated from each historical branch; p h With p c η represents the quality distribution of the historical branch and the candidate branch, respectively, where η is the entropy regularization coefficient, and in this embodiment, η = 0.2.

[0126] After completing branch matching, estimate the migration vector Δμ of each historical branch in the new environment. k The following regularized transfer model is solved:

[0127]

[0128] in, Let be the migration vector to be solved; The current branch target center, This is the transpose of the gradient matrix of the objective function with respect to the decision variables. For the matched target branch center, As a branch tangential basis, I is the transpose of the tangential basis matrix; I is the identity matrix; To constrain the degree of violation, The transpose of the gradient matrix of the constraint violation degree with respect to the decision variables; λ X With λ G λ is the regularization coefficient, which is taken as λ in this embodiment. X =0.5、λ G =0.3.

[0129] After the migration, the branch center was updated to:

[0130]

[0131] in, For the first The central position of each branch in the previous environment; The first result obtained by solving the regularized transfer model The migration vectors of each branch;

[0132] Furthermore, anisotropic trust domain sampling is performed. For each Pareto branch that has completed the migration, the branch center after the migration is used as the reference point. Centered on the tangential base With normal space Zhang Cheng's ellipsoid credibility region

[0133] No. The sampling form of the q-th initial individual generated by each branch is:

[0134]

[0135] in, It represents the element-wise product (Hadamard product). For random perturbation vectors, Let the radius be the tangential trust region. Let the radius be the normal region of confidence. The base radius is taken as 8% of the decision variable range in this embodiment, that is, the base radius of the damper opening is taken as 8%, and the base radius of the burner swing angle is taken as 4.8°. The scaling factor is determined by the eigenvalues ​​of the branch covariance matrix;

[0136] The number of particles allocated to each branch is:

[0137]

[0138] Where, N total The total population size Q represents the total number of branches. k For the first The quality of each branch.

[0139] Q k The expression is:

[0140]

[0141] Among them, C k As a convergence index, H is the average constraint violation rate. k This represents the information entropy of individuals within a branch. This allocation rule allows branches with good convergence, high feasibility, and sparse internal structure to obtain more search resources.

[0142] Step S4: Dynamic constraint processing.

[0143] Specifically, a homotopy function is introduced to asymptotically recover the wall temperature overheat constraint and the steam temperature deviation constraint. The g-th generation relaxation feasible region is defined as:

[0144]

[0145] in, For the first There are inequality constraint functions, where p is the number of inequality constraints. For the first The generation A relaxed boundary condition for constraints, wherein,

[0146]

[0147] In the formula, As the initial relaxation amount, in this embodiment, the actual constraint boundary is relaxed by 10%, that is, the actual constraint of the upper limit of wall temperature is 610℃, and after the initial relaxation it is 671℃; the actual constraint of reheat steam temperature deviation is ±5℃, and after the initial relaxation it is ±5.5℃. The contraction time constant is taken in this embodiment. = Tmax / 4 = 25. As the iteration progresses, the constraint boundary gradually tightens, and in the later stage of the iteration, it recovers to the true constraint limit (i.e., wall temperature 610℃, steam temperature deviation ±5℃).

[0148] The equality constraint tolerance is:

[0149]

[0150] in, Let G be the initial equality tolerance, g be the current iteration number, and G be the initial equality tolerance. max κ represents the maximum number of iterations and the contraction exponent.

[0151] Introduce a homotopy penalty function:

[0152]

[0153] in, Let p be the original objective function vector, p be the number of inequality constraints, and ρ(g) be the dynamic penalty coefficient.

[0154]

[0155] in, As the initial penalty value, in this embodiment, we take... =5, To punish the growth rate, take =3, which allows for moderate constraint violations in the early stages of iteration to preserve historical information, and gradually restores the strictly feasible region as the iteration progresses.

[0156] Furthermore, a dual-space search potential and a three-region collaborative search are employed. Individuals in the population are arranged in ascending order of dual-space search potential and dynamically divided into an elite development zone, a cross-modal bridging zone, and a sparse compensation zone. In the elite development zone, a local fine-grained search is performed along the tangent of its branch. In the cross-modal bridging zone, branches with similar target spaces but different decision spaces are selected for bridging migration. In the sparse compensation zone, directional compensation exploration is performed based on branch scarcity and constraint boundary information, achieving a co-evolution of convergence and multimodal preservation.

[0157] For each individual x, construct a two-space search potential. First, define the following normalization index:

[0158] The convergence error is the normalized distance from the individual to the current Pareto front.

[0159] To constrain the degree of violation;

[0160] For target space congestion;

[0161] To measure the congestion level of the decision-making space;

[0162] The scarcity of the branch to which an individual belongs;

[0163] H contrib (x) represents an individual's historical contribution to external archives.

[0164] Constructing the search potential energy:

[0165]

[0166] in, , , , , , These are the weighting coefficients for each indicator; in this embodiment, they are taken as 0.25, 0.20, 0.15, 0.15, 0.15, and 0.10, respectively.

[0167] Population Sort in ascending order, first One was designated as part of the elite development zone (Zone I), in the middle One is assigned to the cross-modal bridging region (Region II), and the remaining... It is included in the sparse compensation zone (Zone III), in which, For population size, , , The partitioning ratio coefficient is determined by the response intensity. Adaptive adjustment with population branch entropy, and In this embodiment: when ≤0.3 and H contrib When ≥0.6, (θ) I , θ II , θ III ) = (0.50, 0.30, 0.20); when > 0.6 or H contrib When < 0.4, (θ I , θ II , θ III = (0.25, 0.35, 0.40).

[0168] More specifically, Elite Development Zone I: Let individual x belong to branch B. k Its tangential projection and normal projection are as follows:

[0169]

[0170]

[0171] The update strategy is as follows:

[0172]

[0173] in, For small random perturbations along the tangential direction, To minimize the deviation along the normal direction, elite individuals primarily move along the Pareto tangent, with the normal direction used only for minor deviation correction.

[0174] More specifically, in the cross-modal bridging region II: bridging branches are selected according to the principle of "similar target spaces and separate decision spaces."

[0175]

[0176] in, This indicates the branch B to which the individual in Zone II currently belongs. i Select branch B j The probability of it being a bridging target branch; , Branch B i With branch B j The target space center vector, , Branch B i With branch B j The target space center vector, The target space Gaussian kernel bandwidth controls the sensitivity of the target space distance; The decision space is defined by the Gaussian kernel bandwidth, which controls the sensitivity of the decision space distance. For branch B j The scarcity level is obtained by transforming the proportion of individuals in that branch to the total population using a negative logarithmic method; the smaller the proportion, the higher the scarcity level.

[0177] Individual update direction is:

[0178]

[0179] in, Add a random perturbation to the bridging vector from the current branch leader to the target bridging branch leader.

[0180] More specifically, in Sparse Compensation Exploration Region III: the probability of sparse branch selection is defined as follows:

[0181]

[0182] Among them, R branch,s For the first The scarcity of each branch, For the first Average constraint violation rate of each branch The scale parameter for constraint violation is used to control the degree of influence of constraint violation on the selection probability. In this embodiment, it is set to .

[0183] The individual update method is as follows:

[0184]

[0185] in, For the selected sparse branch center, Determined by the branch covariance and response intensity, r explore Let u be the radius of exploration (which increases with increasing response intensity), and let x be a standard normal random vector. opposite For reverse learning position (x) opposite =l b +u b -x,l b with u b (These are the lower and upper bounds of the decision variable, respectively).

[0186] Step S5: External file maintenance and closed-loop control for branch retention.

[0187] Specifically, a branch-and-hold strategy is adopted to maintain external archives. Archives are pruned based on the geodesic-specific congestion distance of individuals within the branch. Multiple equivalent solution sets with similar target space performance but significantly different air distribution methods are retained. The optimal Pareto solution set is then sent to the distributed control system for execution, thereby achieving online dynamic closed-loop control of boiler heating surface thermal deviation.

[0188] For individuals within the same Pareto branch, the geodesic distance d between individuals is calculated by constructing a k-nearest neighbor graph within the branch. geo (x, y). The geodesic special crowding distance for individual x is:

[0189]

[0190] Where B(x) is the Pareto branch to which individual x belongs, σ g The Gaussian kernel bandwidth (σ is taken in this embodiment) g = 0.5), F(x) and F(y) are the objective function vectors of individuals x and y, respectively.

[0191] When external files exceed the capacity limit, priority will be given to those that satisfy the condition that the number of individual branches is greater than the minimum quota Q. min Remove the individual with the smallest GSC(x) value from the branch:

[0192]

[0193] Where A is the current external file, Q minTo determine the minimum retention quota for a branch, Q is taken in this embodiment. min = 0.6×|A| / K, where |A| is the number of external file individuals, which is 50 in this embodiment, and K is the current branch number.

[0194] When selecting a leader, sampling is performed based on the joint probability P of the GSC value and branch quality. leader (x) is:

[0195]

[0196] Where GSC(x) is the geodesic-specific crowding distance of individual x, Q B (x) represents the quality of the branch B(x) to which individual x belongs.

[0197] When the convergence condition is met (in this embodiment, the rate of change of the average objective function for 10 consecutive generations is lower than a set threshold), the final Pareto optimal air distribution scheme set is output. After optimization using the method of this invention in this embodiment, the Pareto optimal air distribution scheme set can, for example, include two sets of typical equivalent air distribution schemes:

[0198] Option A prioritizes economic efficiency: its burnout damper opening is relatively small, the average opening of the secondary damper is relatively large, and the burner swing angle is positive. Under this option, the boiler thermal efficiency is high, but the peak wall temperature and thermal deviation coefficient of the screen superheater are relatively large.

[0199] Option B prioritizes safety: its burnout damper opening is relatively large, the average opening of the secondary damper is relatively small, and the burner swing angle is negative. Under this option, the peak wall temperature and thermal deviation coefficient of the screen superheater are significantly reduced, but the boiler thermal efficiency is slightly reduced.

[0200] The two schemes are independent of each other in the target space and differ significantly in the decision space, demonstrating the multimodal preservation capability of the method of this invention. Operators can flexibly choose the scheme that prioritizes economy or safety based on the current slagging status of the heated surface, the soot blowing cycle, and the grid dispatch requirements.

[0201] The final selected optimal air distribution scheme is sent from the optimization calculation server to the unit's distributed control system (DCS) via the OPC protocol. The DCS then converts the secondary damper opening command, burnout damper opening command, and burner tilt angle command in the scheme step by step and sends them to each actuator, realizing online dynamic closed-loop control of the boiler's heating surface thermal deviation. Whenever the load command changes or the coal quality fluctuation exceeds the set threshold, the optimization process from step one to step eight is automatically repeated to form a closed-loop control link.

[0202] Therefore, this invention provides an adaptive control method for thermal deviation of boiler heating surfaces in thermal power units based on multimodal dynamic optimization, effectively solving the problems of severely uneven boiler heating surface wall temperature distribution, dramatic increase in thermal deviation, and lag in response to load changes under deep peak shaving conditions. This invention constructs a dynamic bi-objective optimization model with the objectives of minimizing the thermal deviation coefficient and maximizing boiler efficiency; proposes a multimodal multi-objective dynamic optimization algorithm based on a response mechanism, quantifies the intensity of load disturbances through dual-space environment drift sensing, triggers optimal transmission matching and anisotropic reliable domain sampling reconstruction of the Pareto branch of the external archive; and employs a potential energy-driven three-zone collaborative search strategy to allocate population evolution tasks, simultaneously maintaining the convergence of the objective space front and the diversity of multiple air distribution modes in the decision space. This method can quickly adapt to load fluctuations, significantly smoothing out thermal deviation of the heating surface while improving combustion efficiency, ensuring the safe and economical operation of the unit.

[0203] Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. Furthermore, those skilled in the art will recognize that, based on the ideas of this invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A method for adaptive control of thermal deviation of boiler heating surface in thermal power unit based on multimodal dynamic optimization, characterized in that, Includes the following steps: Step S1: Collect the operating parameters in the distributed control system of the thermal power unit, and construct a dynamic bi-objective optimization model based on the operating parameters, with the objectives of minimizing the comprehensive thermal deviation coefficient of the heating surface and maximizing the boiler thermal efficiency; Step S2: Real-time sensing of changes in thermal power unit load and coal quality, calculation of comprehensive drift response intensity, and triggering dual-space response reconstruction based on the comprehensive drift response intensity; Step S3: Based on the comprehensive drift response intensity, perform branch matching and anisotropic trust domain migration on the historical Pareto wind distribution model branches in the external archive to generate a multimodal initial population; Step S4: Introduce homotopy functions to asymptotically recover wall temperature overheat and vapor temperature constraints, and dynamically divide the population into elite development zone, cross-modal bridging zone and sparse compensation zone according to dual-space search potential energy, and perform tangential fine search, inter-branch bridging migration and sparse region directional exploration respectively. Step S5: Maintain the external archive using a branch-keeping strategy based on geodesic special congestion distance, output the optimal Pareto solution set and send it to the distributed control system.

2. The adaptive control method for thermal deviation of boiler heating surface in thermal power unit based on multimodal dynamic optimization according to claim 1, characterized in that, In step S1, the operating parameters include: the array data of wall temperature measuring points along the width of the furnace of the outlet header of the screen-type superheater and high-temperature reheater, the reheat steam temperature, the unit load command and the coal quality parameters. The wall temperature measuring point array data comes from a multi-point thermocouple array arranged along the width direction in the outlet header of the screen-type superheater and high-temperature reheater. The opening degree of the secondary air auxiliary damper, the opening degree of the burnout damper, and the burner swing angle of each layer were determined as the optimization decision variables.

3. The adaptive control method for thermal deviation of boiler heating surface in thermal power unit based on multimodal dynamic optimization according to claim 1, characterized in that, In step S1, the comprehensive thermal deviation coefficient of the heating surface is calculated based on the wall temperature measurement point array data of the outlet header of the screen-type superheater and the high-temperature reheater along the width of the furnace; the boiler thermal efficiency is calculated based on the inverse balance method, and the following dynamic bi-objective optimization model is established: in, Indicates environmental moment The feasible decision space is defined below; x∈Ω(t) are the optimization decision variables. Let f1(x, t) be the objective function for the comprehensive thermal deviation coefficient of the heating surface, and f2(x, t) be the objective function for the negative value of boiler thermal efficiency. Let g be the environmental time. i (x, t) is the i-th dynamic inequality constraint function, h j (x, t) is the th A dynamic equality constraint function.

4. The adaptive control method for thermal deviation of boiler heating surface in thermal power unit based on multimodal dynamic optimization according to claim 1, characterized in that, In step S2, the normalized drift of the external file in the target space, decision space, and constraint space is calculated, and the comprehensive drift response intensity is obtained by adaptive weighting. When the comprehensive drift response intensity exceeds a preset threshold, it is determined that the environment has changed significantly, triggering dual-space response reconstruction.

5. The adaptive control method for thermal deviation of boiler heating surface in thermal power unit based on multimodal dynamic optimization according to claim 4, characterized in that, In step S2, the quantification method of the comprehensive drift response intensity is as follows: the target space drift ΔF, the decision space drift ΔX, and the constraint space drift ΔG are defined as follows: in, For the environment at all times External archives, Let be the center vector of all individuals in the external archive in the target space. For the Euclidean norm, Use a small constant to prevent the denominator from being zero; in, Let be the central vector of all individuals in the external archive in the decision space. Use a small constant to prevent the denominator from being zero; in, Let be the number of inequality constraints. For all individuals in the external archives in the first The average degree of violation on each constraint, Use a small constant to prevent the denominator from being zero; The overall response intensity is: in, , , For adaptive weighting coefficients, satisfying + + =1, dynamically adjusted according to the unit load change rate; when > The system is identified as experiencing a significant environmental change, triggering branch-level migration and trusted domain reconstruction. This is a preset threshold.

6. The adaptive control method for thermal deviation of boiler heating surface in thermal power unit based on multimodal dynamic optimization according to claim 5, characterized in that, In step S3, the branch matching is achieved by constructing a migration cost matrix between the historical Pareto branch and the current candidate branch. The first historical branch and the first The migration cost for each candidate branch is: in, , The first The target values ​​and decision variable values ​​of each historical branch center , The first The target value and decision variable value of each candidate branch center , These are the local tangent basis vectors for the historical branch and the candidate branch, respectively. , , These are the weighting coefficients. It is the absolute value symbol; Matching relationships are solved by... The optimal transmission model is obtained at the cost of entropy regularization, which allows each historical branch to be matched to the current candidate branch in a soft correspondence manner.

7. The adaptive control method for thermal deviation of boiler heating surface in thermal power unit based on multimodal dynamic optimization according to claim 6, characterized in that, In step S3, the specific method of anisotropic trusted domain migration is as follows: for each Pareto branch that has completed the migration, the migration branch center is used as the reference point. Centered on the tangential base With normal space Zhang Cheng's ellipsoid credibility region, the first The first branch generated The sampling method for the initial individuals is as follows: in, Represents element-wise product. For random perturbation vectors, Let the radius be the tangential trust region. Let the radius of the normal credibility region be . Based on the radius, The scaling factor is determined by the eigenvalues ​​of the branch covariance matrix; The number of particles allocated to each branch is: in, The total population size The total number of branches, For the first The quality of a branch is determined by a combination of its convergence index, average constraint violation rate, and individual information entropy within the branch.

8. The adaptive control method for thermal deviation of boiler heating surface in thermal power unit based on multimodal dynamic optimization according to claim 7, characterized in that, In step S4, a homotopy function is introduced to progressively recover the wall temperature overheating constraint and the steam temperature deviation constraint. The population individuals are arranged in ascending order of dual-space search potential energy and dynamically divided into an elite development zone, a cross-modal bridging zone, and a sparse compensation zone. In the elite development zone, a local fine search is performed along the tangent of the branch to which it belongs. In the cross-modal bridging zone, branches with similar target spaces and different decision spaces are selected to perform bridging migration. In the sparse compensation zone, directional compensation exploration is performed based on the branch scarcity degree and constraint boundary information.

9. The adaptive control method for thermal deviation of boiler heating surface in thermal power unit based on multimodal dynamic optimization according to claim 8, characterized in that, In step S4, a homotopy function is introduced to asymptotically recover the wall temperature overheat constraint and the steam temperature deviation constraint, specifically as follows: No. The relaxed feasible region is defined as follows: in, For the first Inequality constraint functions, For the first The generation A relaxed boundary of constraints, , This is the initial relaxation amount. The contraction time constant; The allowable temperature limit for wall temperature constraint is determined based on the material of the heated surface pipe; The equality constraint tolerance is: in, For the initial equality tolerance, G max The maximum number of iterations, The contraction index; Introduce a homotopy penalty function: in, The original objective function vector, This is a dynamic penalty coefficient. The initial penalty value, To penalize the rate of growth, moderate constraint violations are allowed in the early stages of iteration to preserve historical information, and the strict feasible region is gradually restored as the iteration progresses.

10. The adaptive control method for thermal deviation of boiler heating surface in thermal power unit based on multimodal dynamic optimization according to claim 9, characterized in that, In step S4, the dual-space search potential is constructed by integrating the convergence error of each individual. Constraint Violation Target space congestion Decision-making space geodesy congestion The scarcity of its branches and historical contributions H contrib (x), construct the dual-space search potential energy: in, , , , , , The weighting coefficients for each indicator; the population is calculated as follows: Sort in ascending order, first One was designated as an elite development zone, in the middle One is included in the cross-modal bridging region, the remaining ones Included in the sparse compensation zone, among which, For population size, , , The partitioning ratio coefficient is determined by the response intensity. Adaptive adjustment with population branch entropy, and .

Images