A multi-objective optimization method and system for unit adaptive combustion and control real-time optimization

Abstract

This invention discloses a multi-objective optimization method and system for real-time adaptive combustion and control of generating units, comprising: acquiring unit operating and decision parameters; constructing a system with net power as the objective function; employing an online hybrid Gaussian process regression and uncertainty quantification network (OH-GPR-UQN) to identify the dynamic relationship between the decision parameters and the net power objective function online, estimating the gradient and quantifying its uncertainty; based on the gradient and its uncertainty, employing a Bayesian optimization-guided adaptive exploration and utilization constraint hill-climbing algorithm (BOA-CEHC) to determine the adjustment step size and direction disturbance of the decision parameters by maximizing an acquisition function that simultaneously considers "utilizing" the gradient and "exploring" uncertainty, and combining gradient projection and penalty functions to process operating constraints to obtain optimized parameters; the optimized parameters are used for closed-loop control, and operating data is fed back to update OH-GPR-UQN online. This invention can accurately identify, intelligently optimize, and robustly handle constraints, significantly improving the overall efficiency of the generating unit.

Description

Technical Field

[0001] This invention relates to the technical field of predictive control of thermal power generating units, and in particular to a real-time optimization method and system for adaptive combustion and control of generating units with multi-objective optimization. Background Technology

[0002] With the transformation of the power energy structure and the deepening of market-oriented reforms, the thermal power industry faces severe challenges. The increasing proportion of renewable energy has led to greater fluctuations in grid load. Coupled with the pilot programs for the electricity spot market and ancillary services marketization, thermal power units urgently need to improve their peak-shaving flexibility and operational efficiency to maintain competitiveness. Taking Zhejiang Province as an example, as of May 2024, the installed capacity of new energy reached 18,478,900 kilowatts (accounting for 23.35% of the province's total). The volatility of its power generation forces thermal power units to frequently and deeply shave peak loads, operating under low load and significantly varying operating conditions for extended periods, leading to increased energy consumption, accelerated equipment wear, and safety hazards. At the same time, the proportion of electricity traded through the electricity market has exceeded 50%, placing higher demands on unit energy efficiency and load response speed.

[0003] Currently, coal-fired power units face multiple bottlenecks: high coal prices force companies to use complex coal sources, and frequent fluctuations in coal quality lead to decreased combustion stability, abnormal steam temperature and pressure, unstable NOx emissions, increased ash accumulation on heated surfaces, and intensified high-temperature corrosion. Traditional extensive combustion control modes are ill-suited to the variability of coal types and the need for rapid load adjustments, becoming the core bottleneck restricting unit performance.

[0004] Optimization control of combustion processes in thermal power units faces complex technical challenges. Unit operation involves multiple coupled variables, such as fuel supply, air ratio, and steam parameters, which exhibit nonlinear relationships, making accurate modeling difficult. Simultaneously, external factors such as coal quality fluctuations and load changes also affect combustion efficiency, increasing optimization difficulty. Traditional fixed-parameter control methods are ill-suited to such dynamic and complex operating conditions, often resulting in low overall unit efficiency. Achieving real-time adaptive optimization of the combustion process while ensuring the quality of control over key operating parameters has become a critical issue. This requires the control system to quickly sense changes in operating conditions, accurately identify the coupling relationships between variables, and generate reasonable adjustment strategies accordingly. However, due to the inherent nonlinear characteristics of the system and model errors, the optimization process is prone to getting trapped in local optima or oscillations, affecting control effectiveness. Furthermore, maintaining consistency in optimization results under different load and coal quality conditions is also a key consideration. Solving these technical challenges is of great significance for improving the economic and environmental performance of thermal power units.

[0005] Currently, boiler combustion optimization methods used in the power generation industry are mainly based on the power plant performance test procedure GB / T10184-2015 and the ASME PTC 4.1 standard. This method employs a "single-factor rotation method," where only one controlled factor is changed while other factors remain constant, and the impact of this change on boiler operation is observed. After determining the optimal value for the current factor, other adjustment factors are rotated. This method does not consider the coupling effect between factors, resulting in a "best" operating condition that is not truly optimal. Furthermore, because some variables affecting boiler efficiency calculations, such as fly ash carbon content, cannot be measured in real time, the direction of combustion optimization cannot be determined online. This necessitates online testing followed by offline analysis and optimization, increasing the number of operating points to be traversed and raising time costs. In addition, units undergoing performance testing need to maintain stable operating conditions and cannot respond to automatic generation control (AGC) commands, further increasing operating costs. Therefore, there is an urgent need for a method that can perform real-time, closed-loop, and adaptive multi-variable collaborative optimization of the combustion process to improve the flexibility, economy, and environmental friendliness of the unit. Summary of the Invention

[0006] To address the problems of traditional combustion optimization methods in the background art, such as failure to consider multivariate coupling, difficulty in online measurement of optimization objectives, long optimization cycle, high cost, and inability to adapt to rapid changes in coal quality and load, this invention aims to provide a multi-objective optimization method and system for unit adaptive combustion and control in real time, which can realize real-time, closed-loop, and adaptive optimization of the combustion process.

[0007] In a first aspect, the present invention provides a multi-objective optimization method for real-time adaptive combustion and control of generator units, which mainly includes the following steps:

[0008] S1: Data Acquisition and Preprocessing Steps: Real-time acquisition of multiple operating parameters and current decision parameters during the operation of the thermal power unit. The operating parameters include at least coal feed rate and main steam parameters, and the decision parameters include at least furnace target oxygen content, primary air pressure, and opening degree of each secondary air damper. The operating parameters and decision parameters are then preprocessed.

[0009] S2: Objective function construction and uncertainty quantification model initialization steps:

[0010] S2.1: The difference between the actual generated power of the unit and the power used by the unit is used as the real-time measurable net power objective function;

[0011] S2.2: Initialize and run an online Gaussian mixture process regression and uncertainty quantification network (OH-GPR-UQN) to establish the dynamic input-output relationship between the decision parameters, operating parameters and the net power objective function, and quantify the prediction uncertainty of the relationship;

[0012] S3: Gradient estimation and uncertainty assessment steps based on OH-GPR-UQN:

[0013] S3.1: Using the OH-GPR-UQN model, based on the current operating parameters and decision parameters obtained in S1, predict the current net power objective function value, and estimate the real-time gradient of the net power objective function relative to each of the decision parameters;

[0014] S3.2: Simultaneously, using the OH-GPR-UQN model, evaluate the confidence level or uncertainty level of the gradient estimate obtained in step S3.1;

[0015] S4: Constrained hill-climbing optimization steps based on Bayesian optimization-guided adaptive exploration and utilization:

[0016] S4.1: Construct a data acquisition function that takes into account both the optimization direction indicated by the gradient obtained in S3.1 and the exploration requirements indicated by the gradient uncertainty obtained in S3.2.

[0017] S4.2: By maximizing the acquisition function, determine the adjustment step size and adjustment direction perturbation for each of the decision parameters;

[0018] S4.3: Combining the preset unit safety and environmental protection operation constraints, the constraint processing mechanism is used to correct the adjustment step size and direction disturbance determined in S4.2, so as to ensure that the adjusted decision parameters are within the feasible domain and obtain an optimized set of decision parameters;

[0019] S5: Closed-loop control and online model update steps: The optimized decision parameters obtained in step S4.3 are sent to the distributed control system (DCS) of the thermal power unit for closed-loop control; and the controlled unit operating parameters, net power objective function value and optimization exploration point data are fed back to step S2.2 for online update and parameter adaptive adjustment of the OH-GPR-UQN model.

[0020] Secondly, the present invention also discloses a multi-objective optimization unit adaptive combustion and control real-time optimization system, including: a data acquisition and preprocessing module (10), used to acquire multiple operating parameters and current decision parameters during the operation of the thermal power unit in real time, wherein the operating parameters include at least coal feed rate and main steam parameters, and the decision parameters include at least furnace target oxygen content, primary air pressure, and opening degree of each damper of secondary air; and to preprocess the operating parameters and decision parameters;

[0021] The objective function construction and uncertainty quantification model initialization module (20) is connected to the data acquisition and preprocessing module (10) and is used for objective function construction and uncertainty quantification model initialization. It is also configured with computing resources to run the online Gaussian mixture process regression and uncertainty quantification network (OH-GPR-UQN).

[0022] The gradient estimation and uncertainty assessment module (30) is connected to the objective function construction and uncertainty quantification model initialization module (20), and is based on the gradient estimation and uncertainty assessment of OH-GPR-UQN.

[0023] The Bayesian optimization-guided constraint hill-climbing optimization module (40) is connected to the gradient estimation and uncertainty assessment module (30) and is used for Bayesian optimization-guided adaptive exploration and utilization constraint hill-climbing optimization. It contains a data acquisition function construction unit and a constraint processing unit.

[0024] The closed-loop control and online model update module (50) is connected to the Bayesian optimization-guided constraint hill climbing optimization module (40) and the distributed control system (DCS) of the thermal power unit. It is used for closed-loop control and online model update, and feeds back the data to the objective function construction and uncertainty quantification model initialization module (20) to update OH-GPR-UQN.

[0025] The technical solution provided by this invention may include the following beneficial effects:

[0026] This invention discloses a multi-objective optimization method and system for real-time adaptive combustion and control of generator units. First, an online hybrid Gaussian process regression and uncertainty quantification network (OH-GPR-UQN) is constructed and run. This network not only identifies the complex nonlinear dynamic input-output relationship between decision parameters, operating parameters, and the net power objective function, and estimates the real-time gradient of net power with respect to decision parameters, but more importantly, it can quantify the uncertainty of these predictions and gradient estimates. Second, based on the gradient estimates and uncertainty information provided by OH-GPR-UQN, a Bayesian optimization-guided adaptive exploration and utilization constrained hill-climbing algorithm (BOA-CEHC) is used for iterative optimization of the decision parameters. Finally, the optimized decision parameters are sent to the DCS for closed-loop control. The controlled operating data and newly discovered optimization points are fed back to OH-GPR-UQN for online learning and iterative updates of its model, forming a closed-loop system of continuous learning, continuous adaptation, and continuous optimization. This invention realizes adaptive optimization control of the combustion process of thermal power units, improves the overall operating efficiency of the units, and ensures the control quality of key operating parameters, providing effective support for the economical and efficient operation of the units.

[0027] Compared with existing combustion optimization methods, this invention has at least the following outstanding substantive features and significant progress: First, this invention proposes to use 'actual unit power output minus unit auxiliary power' as the objective function for combustion optimization. This objective function not only accurately reflects the net output efficiency of the unit, covering the overall performance of the boiler, turbine, and generator, but also allows all its component variables to be obtained in real time and accurately from the existing DCS system, completely solving the fundamental problems of high experimental costs, long cycles, and inability to perform closed-loop online adjustments when optimizing based on boiler efficiency (whose key parameters, such as fly ash carbon content, cannot be measured online in real time). Second, it can accurately model the complex nonlinear dynamics of the combustion process online and quantify the uncertainty of prediction and gradient, providing a richer and more reliable information foundation for subsequent intelligent optimization. It organically combines gradient-based "utilization" and uncertainty-based "exploration," enabling more intelligent searching in the multivariate decision space, effectively avoiding getting trapped in local optima, and proactively improving the identification model. This effectively overcomes the shortcomings of the traditional 'single-factor rotation method', which does not consider the coupling effect of multiple variables and is prone to getting trapped in local optima. Third, by continuously optimizing the combustion process to maximize net power, and taking into account both operational stability and environmental protection requirements, coal consumption can be effectively reduced, and the overall economic efficiency, flexibility, and cleanliness of the unit can be improved. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0029] Figure 1 This is an overall module block diagram of the multi-objective optimization unit adaptive combustion and control real-time optimization system provided in the embodiments of the present invention.

[0030] Figure 2 This is a schematic diagram of the internal structure of the Online Hybrid Gaussian Process Regression and Uncertainty Quantification Network (OH-GPR-UQN) in an embodiment of the present invention.

[0031] Figure 3 This is a schematic diagram of the workflow of the constraint hill-climbing optimization module (BOA-CEHC) based on Bayesian optimization-guided adaptive exploration and utilization in an embodiment of the present invention.

[0032] Figure 4 This is a schematic diagram of the network architecture of the optimized control system and the DCS of the thermal power unit provided in the embodiment of the present invention.

[0033] Figure 5 This is the main flowchart of the optimization method provided in the embodiments of the present invention.

[0034] Figure 6This is a simulation effect diagram of an embodiment of the present invention, which exemplarily demonstrates the comparison of the optimization trajectory of the unit's net power using the method of the present invention and the benchmark method.

[0035] Figure 7 This is a performance comparison chart of an embodiment of the present invention, which exemplarily demonstrates the comparison between the method of the present invention and the benchmark method in terms of key parameter control. Detailed Implementation

[0036] To further understand the content of this invention, a detailed description of the invention is provided in conjunction with the accompanying drawings and embodiments. The specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0037] Example 1

[0038] Reference Figure 1 This invention provides a multi-objective optimization system for unit adaptive combustion and control in real time. The system mainly includes: a data acquisition and preprocessing module 10, an objective function construction and OH-GPR-UQN model initialization / running module 20, a gradient estimation and uncertainty assessment module 30, a BOA-CEHC optimization module 40, and a closed-loop control and online model update module 50. The optimization system operates via a network (e.g., Figure 4 (As shown) It connects to the distributed control system (DCS) of the thermal power unit for data exchange and control command issuance. Among them:

[0039] The data acquisition and preprocessing module (10) is used to acquire multiple operating parameters and current decision parameters during the operation of the thermal power unit in real time. The operating parameters include at least coal feed rate and main steam parameters, and the decision parameters include at least furnace target oxygen content, primary air pressure, and opening degree of each secondary air damper. The module also preprocesses the operating parameters and decision parameters.

[0040] The objective function construction and uncertainty quantification model initialization module (20) is connected to the data acquisition and preprocessing module (10) and is used for objective function construction and uncertainty quantification model initialization. It is also configured with computing resources to run the online Gaussian mixture process regression and uncertainty quantification network (OH-GPR-UQN).

[0041] The gradient estimation and uncertainty assessment module (30) is connected to the objective function construction and uncertainty quantification model initialization module (20), and is based on the gradient estimation and uncertainty assessment of OH-GPR-UQN.

[0042] The Bayesian optimization-guided constraint hill-climbing optimization module (40) is connected to the gradient estimation and uncertainty assessment module (30) and is used for Bayesian optimization-guided adaptive exploration and utilization constraint hill-climbing optimization. It contains a data acquisition function construction unit and a constraint processing unit.

[0043] The closed-loop control and online model update module (50) is connected to the Bayesian optimization-guided constraint hill climbing optimization module (40) and the distributed control system (DCS) of the thermal power unit. It is used for closed-loop control and online model update, and feeds back the data to the objective function construction and uncertainty quantification model initialization module (20) to update OH-GPR-UQN.

[0044] Specifically, the OH-GPR-UQN in the objective function construction and uncertainty quantification model initialization module (20) includes: multiple parallel local Gaussian process regression (GPR) model processing units, a gating network decision unit, and an uncertainty quantification auxiliary network processing unit.

[0045] Specifically, the acquisition function construction unit in the Bayesian optimization-guided constrained hill climbing optimization module (40) dynamically constructs and optimizes the acquisition function described in claim 4 based on the gradient and uncertainty output by the gradient estimation and uncertainty assessment module (30) to determine the adjustment step size and direction perturbation of the decision parameters.

[0046] Specifically, the constraint processing unit in the Bayesian optimization-guided constraint hill-climbing optimization module (40) integrates the gradient projection algorithm to process hard constraints and handles soft constraints by modifying the evaluation value of the acquisition function or introducing a penalty term.

[0047] Example 2

[0048] This embodiment focuses on combustion optimization for 600MW supercritical coal-fired power units, using a typical 600MW supercritical coal-fired generating unit as an example. It applies the multi-objective optimization method and system for adaptive combustion and control real-time optimization described in this invention. (Refer to...) Figure 1 System module block diagram and Figure 5 The main flowchart of the method.

[0049] S1: Data Acquisition and Preprocessing (executed by Data Acquisition and Preprocessing Module 10)

[0050] Data source: The following parameters are collected in real time from the DCS and SIS of this 600MW unit via OPC interface or dedicated data gateway at a sampling period of 1 second:

[0051] Operating parameters: Unit load command, actual load, total coal feed, instantaneous coal feed of five coal mills (A / B / C / D / E), pulverized coal temperature at each mill outlet, primary air main pressure and flow rate, primary air regulating valve opening and air volume of each mill, secondary air main pressure and flow rate, secondary air damper opening and corresponding air volume of each layer (e.g., upper, middle, and lower layers), burnout air (OFA) damper opening and air volume, average oxygen content in the furnace, steam temperature and pressure at the outlet of each superheater / reheater, main steam temperature, main steam pressure, reheat steam temperature, economizer outlet water temperature, flue gas temperature, NOx / SO2 / CO emission concentrations, total unit power generation (P_gross), and total plant power consumption (P_aux).

[0052] Decision parameters: Key adjustable parameters set by the current operator or lower-level control system, mainly including: furnace target oxygen setpoint, primary air main pressure setpoint, target opening of secondary air dampers on each floor (or the distribution ratio of dampers on each floor under total air volume), and OFA damper target opening.

[0053] Preprocessing: Data cleaning (removal of bad points, outliers, and missing values), filtering (moving average, low-pass filtering), feature engineering (such as constructing the wind-coal ratio), and normalization / standardization (elimination of dimensions).

[0054] S2: Objective function construction and OH-GPR-UQN model initialization (executed by module 20)

[0055] S2.1: Objective function construction: Define the net power objective function J(t) = P_gross(t) - P_aux(t), where t is time.

[0056] The objective function is maximized in the subsequent optimization process. Note that J(t) is a real-time function and a scalar. Since this objective function can be accurately measured in real time under normal power generation conditions, the problem of not being able to measure boiler efficiency in real time is avoided. Furthermore, because this objective function covers the boiler, turbine, and generator, it can be used to optimize these three components simultaneously. Choosing this new objective function plays a crucial role in the optimized control of this invention.

[0057] S2.2: Initialization of the OH-GPR-UQN model (refer to...) Figure 2 ):

[0058] Operating Condition Zone Division and Local GPR Model: Based on historical data and using clustering algorithms (such as K-Means) or expert experience, the unit's operating conditions are divided into several typical zones (such as low / medium / high load zones, and zones affected by different coal types). A local GPR model is initialized for each operating condition zone. The squared exponential kernel (RBF kernel) or Matern kernel is selected, and the initial kernel parameters are set based on the statistical characteristics of historical data for that operating condition zone or obtained through offline optimization. A set of induction points is maintained using a sparse GPR method (such as FITC).

[0059] Gated Network: Design a small multilayer perceptron (MLP) as the gated network. The inputs are the current unit load and main operating parameters, and the outputs are the activation weights or selection signals of each local GPR model.

[0060] Uncertainty Quantization Auxiliary Network (UQN_aux): Initializes a small feedforward neural network. Inputs are the predicted mean μ_i and variance σ_i^2 of the activated local GPR model output, and the current decision parameters u. Output is a comprehensive uncertainty measure σ_∇J of the gradient ∇J.

[0061] Initially, historical data was used to perform preliminary offline training and parameter tuning for each component of OH-GPR-UQN.

[0062] S3: Gradient estimation and uncertainty assessment based on OH-GPR-UQN (performed by module 30)

[0063] In each optimization iteration cycle:

[0064] S3.1: Gradient Estimation: The gated network determines the current operating region and outputs the local GPR model weights w_i. Each activated local GPR model i predicts the net power J under the current decision parameters u_t, obtaining μ_i(J|u_t) and σ_i^2(J|u_t), and calculates the gradient components ∂μ_i(J|u_t) / ∂u_{t,j}. The gradient is fused as: ∇J_hat = Σ w_i[∂μ_i(J|u_t) / ∂u_{t,1}, ..., ∂μ_i(J|u_t) / ∂u_{t,M}]^T.

[0065] S3.2: Uncertainty Assessment: The UQN_aux network receives μ_i, σ_i^2 of each activated GPR and the current u_t as input, and outputs the uncertainty measure σ_∇J of the gradient vector ∇J_hat as a whole or for each component.

[0066] S4: Bayesian Optimization-Guided Adaptive Exploration and Utilization Constrained Hill Climbing Optimization (BOA-CEHC) (Executed by Module 40) (Refer to) Figure 3 )

[0067] S4.1: Constructing the Acquisition Function: An improved UCB acquisition function is adopted: A(u_candidate) = MeanImprovement(u_candidate | ∇J_hat, J_current) + κ UncertaintyTerm(u_candidate | σ_∇J, σ^2_J_fused). MeanImprovement is the net power improvement estimated at the candidate point u_candidate based on the current gradient. UncertaintyTerm quantifies the value of exploring near the u_candidate point, based on gradient uncertainty σ_∇J or model prediction uncertainty. κ is a dynamically adjusted balancing factor.

[0068] S4.2: Maximizing the acquisition function to determine the step size and direction perturbation: Within the neighborhood of the current decision point u_t, find the u_candidate_opt that maximizes the acquisition function A(u_candidate) through numerical optimization. Adjust the vector Δu_raw = u_candidate_opt - u_t.

[0069] S4.3: Constraint Handling Mechanism: u_t + Δu_raw is checked. Hard constraints (such as furnace oxygen content 2%-4%, damper opening 0-100%, main steam temperature <571℃) are handled using gradient projection or a repair strategy. Soft constraints (such as NOx emission expectation around 80mg / Nm³) are handled by converting them into a penalty term P_NOx(u_candidate) on the acquisition function A(u_candidate). The final allowable adjustment amount Δu_final is obtained, and the optimized decision parameter u_opt = u_t + Δu_final.

[0070] S5: Closed-loop control and online model updates (executed by module 50)

[0071] The u_opt is sent to the corresponding control loop of the unit's DCS.

[0072] The system continuously collects and distributes unit operation data.

[0073] Newly acquired data points (u_opt, runtime parameters, J_actual) are used to update OH-GPR-UQN online. The activated local GPR model updates its posterior or induced points using its online learning algorithm. Gated network parameters are adjusted based on recent prediction performance. The UQN_aux network also undergoes incremental learning.

[0074] The S1-S5 processes are executed cyclically.

[0075] In actual implementation, performance is tested and compared with benchmark methods in existing technologies, such as... Figure 6 The results show that, compared with the traditional hill-climbing method, the BOA-CEHC of the present invention can perform a more effective global search in terms of maximizing net power J(t), escape local optima, and reach a higher global optimum. Figure 7 The two sub-figures compare the advantages of the present invention's method over the benchmark method in terms of NOx emission control (compliance with constraints) and main steam temperature stability (control quality). The top figure shows that NOx emissions are stably controlled below the limit and close to the expected value. The bottom figure shows that the main steam temperature fluctuation is effectively suppressed, and the temperature control of the present invention is more stable.

[0076] Example 3

[0077] This embodiment focuses on the synergistic optimization of combustion and denitrification in a 350MW subcritical CFB boiler unit. It is applied to a 350MW circulating fluidized bed (CFB) boiler unit equipped with a selective non-catalytic reduction (SNCR) denitrification system.

[0078] S1: Data Acquisition and Preprocessing

[0079] Additional operating parameters: In addition to standard parameters, these include: bed temperature (dense phase zone, dilute phase zone), bed pressure, return feeder temperature, and limestone feed rate. SNCR system parameters: ammonia injection rate, spray gun position, and dilution air volume.

[0080] Decision parameter adjustments: Combustion side (total air volume, primary / secondary air ratio, seeding air volume, target bed temperature) and SNCR side (total ammonia injection volume, ammonia distribution ratio of each spray gun).

[0081] S2: Objective function construction and OH-GPR-UQN model initialization

[0082] S2.1: Objective function: Net power J(t) = P_gross(t) - P_aux(t). NOx emissions are used as a hard constraint and a penalty term in the optimization process.

[0083] S2.2: OH-GPR-UQN Model Initialization: Operating zone division considers CFB characteristics (load, coal type, bed temperature pattern). LRA-Layer (part of OH-GPR-UQN) will explicitly model and predict both the original NOx formation concentration (combustion-side effects) and the final emission concentration after SNCR removal.

[0084] S3: Gradient Estimation and Uncertainty Assessment Based on OH-GPR-UQN

[0085] The OH-GPR-UQN estimates the gradient of J relative to the combustion-side decision parameters, and the gradient (or influence model) of the final NOx emission concentration relative to the combustion-side and SNCR-side decision parameters.

[0086] S4: Bayesian Optimization-Guided Adaptive Exploration and Utilization-Based Constrained Hill Climbing Optimization (BOA-CEHC)

[0087] The acquisition function A(u_candidate) primarily aims to maximize J, but is strongly constrained by NOx emissions. The decision vector u_candidate contains combustion-side and SNCR-side parameters.

[0088] Constraint handling mechanism: 1. When BOA-CEHC provides a combustion-side adjustment suggestion Δu_combustion, OH-GPR-UQN predicts the new original NOx formation concentration.

[0089] 2. For this initial NOx, optimize the SNCR parameters (ammonia injection rate, distribution ratio) to ensure that the final NOx meets the standard and the possibility of ammonia depletion is small.

[0090] 3. If NOx still exceeds the limit under the optimal SNCR operation, then call back or modify Δu_combustion, or the collection function has already suppressed the direction of high original NOx generation through the penalty term during construction.

[0091] The combustion-side and SNCR-side decision parameters u_opt are obtained after co-optimization.

[0092] S5: Closed-loop control and online model updates

[0093] u_opt sends the combustion and SNCR control loops to the DCS. The section on SNCR system behavior in OH-GPR-UQN has also been updated.

[0094] This third embodiment demonstrates that the method of the present invention can achieve more complex integrated intelligent optimization of "combustion-denitrification", reflecting the characteristics of "multi-objective optimization" and "adaptive control".

[0095] In summary, the core idea of ​​this invention is to use the real-time measurable net output power of the unit as the direct optimization target. By applying small, safe disturbances to the control system and observing their impact on the objective function (i.e., online system identification), the gradient of the objective function with respect to each adjustable parameter is estimated. Then, using this gradient information, combined with an adaptive step-size strategy that considers model accuracy and convergence speed, multi-variable optimization is performed. Constraint processing ensures the safety of the optimization process, ultimately achieving continuous adaptive optimization of the unit combustion process. This overcomes the shortcomings of existing thermal power unit combustion optimization methods, such as the weak adaptability of the system identification model to complex dynamics, the lack of reliability assessment of gradient estimation, the easy trapping of optimization algorithms in local optima, and the simplistic exploration and utilization balance strategy. It provides a multi-objective adaptive real-time optimization method and system that can achieve high-precision identification, intelligent optimization, and robust constraint processing of the unit combustion process.

[0096] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0097] 1. Real-time online global optimization: By using a real-time measurable objective function J(t), combined with online system identification and multivariate optimization, real-time, closed-loop optimization of the combustion process is achieved, overcoming the shortcomings of traditional methods that rely on offline tests and cannot fully reflect the unit's net output.

[0098] 2. Effective handling of coupling effects: By employing multivariate system identification and multivariate optimization algorithms, the complex coupling relationships between various combustion control variables can be effectively captured and utilized to find a better cooperative operating point.

[0099] 3. Strong adaptability: Through periodic or event-triggered online system identification, the model can adapt to changes in operating conditions such as coal quality changes, load fluctuations, and equipment characteristic drift, maintaining the robustness of the optimization effect.

[0100] 4. A Bayesian optimization-guided adaptive exploration and exploitation constrained hill-climbing algorithm (BOA-CEHC) is employed for iterative optimization of decision parameters. Instead of simply adjusting the step size based on gradient magnitude and fixed rules, a data acquisition function (such as an improved UCB or EI) is constructed. This function simultaneously evaluates two aspects: the expected net power improvement from moving along the optimal direction indicated by the current gradient estimate; and moving towards regions with higher gradient estimation uncertainty or model prediction uncertainty to obtain more information, improve the model, or discover unexpected better solutions. By maximizing this data acquisition function in the decision parameter space, the "optimal" step size for the next hill-climbing operation and whether some exploratory perturbation of the gradient direction is needed are determined. If the "exploitation" value of the gradient direction is significantly higher than its "exploration" value, a larger step is tended along the gradient direction; conversely, a smaller step size may be chosen, and the exploration may be biased towards directions with higher uncertainty. The balance factor κ in the data acquisition function can be dynamically adjusted according to the overall optimization progress and model confidence.

[0101] 5. Improved economy and flexibility: By optimizing combustion, it is expected to effectively reduce coal consumption for power supply, improve the unit's load response rate and peak shaving flexibility, adapt to the needs of the electricity market, and reduce the unit's coal consumption for power supply by 0.9g / (kw•h), resulting in significant fuel cost savings and CO2 emission reduction.

[0102] 6. Reduced optimization costs and time: It avoids the large amount of manpower, material resources and time costs required for traditional performance testing, and the optimization process can be carried out during normal unit operation (including response AGC).

[0103] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A multi-objective optimization method for unit adaptive combustion and control real-time optimization, characterized in that, Includes the following steps: S1: Data Acquisition and Preprocessing Steps: Real-time acquisition of multiple operating parameters and current decision parameters during the operation of the thermal power unit. The operating parameters include at least coal feed rate and main steam parameters, and the decision parameters include at least furnace target oxygen content, primary air pressure, and secondary air damper opening. The operating parameters and decision parameters are then preprocessed. S2: Objective function construction and uncertainty quantification model initialization steps: S2.1: The difference between the actual generated power of the unit and the power used by the unit is used as the real-time measurable net power objective function; S2.2: Initialize and run an online Gaussian mixture process regression and uncertainty quantification network OH-GPR-UQN to establish the dynamic input-output relationship between the decision parameters, operating parameters and the net power objective function, and quantify the prediction uncertainty of the relationship; The OH-GPR-UQN includes: multiple local Gaussian process regression GPR models, each of which is responsible for identifying a specific working condition region; A gating network dynamically determines the activation weights of each local GPR model or selects one of the dominant models based on the current operating parameters and decision parameters; and an uncertainty quantification auxiliary network is used to further refine the uncertainty assessment of gradient estimation based on the predicted mean, variance, and input characteristics of each local GPR model. S3: Gradient estimation and uncertainty assessment steps based on OH-GPR-UQN: S3.1: Using the OH-GPR-UQN model, based on the current operating parameters and decision parameters obtained in S1, predict the current net power objective function value, and estimate the real-time gradient of the net power objective function relative to each of the decision parameters; S3.2: Simultaneously, using the OH-GPR-UQN model, evaluate the confidence level or uncertainty level of the gradient estimate obtained in step S3.1; S4: Constrained hill-climbing optimization steps based on Bayesian optimization-guided adaptive exploration and utilization: S4.1: Construct a data acquisition function that takes into account both the optimization direction indicated by the gradient obtained in S3.1 and the exploration requirements indicated by the gradient uncertainty obtained in S3.

2. S4.2: By maximizing the acquisition function, determine the adjustment step size and adjustment direction perturbation for each of the decision parameters; S4.3: Combining the preset unit safety and environmental protection operation constraints, the constraint processing mechanism is used to correct the adjustment step size and direction disturbance determined in S4.2, so as to ensure that the adjusted decision parameters are within the feasible domain and obtain an optimized set of decision parameters; S5: Closed-loop control and online model update steps: The optimized decision parameters obtained in step S4.3 are sent to the distributed control system (DCS) of the thermal power unit for closed-loop control; and the controlled unit operating parameters, net power objective function value, and optimization exploration point data are fed back to step S2.2 for online update and parameter adaptive adjustment of the OH-GPR-UQN model.

2. The method of claim 1, wherein, The local GPR model employs a sparse online Gaussian process regression algorithm or a recursive update algorithm to update its kernel function parameters, noise hyperparameters, and induced points online.

3. The method of claim 1, wherein, In step S4.1, the acquisition function is an improved upper confidence interval (UCB) function, which has the form A(u)=μ(u)+κσ_grad(u), where μ(u) represents the expected net power improvement based on the current gradient moving to point u in the decision parameter space, σ_grad(u) represents the uncertainty of gradient estimation near point u, and κ is a hyperparameter that balances exploration and exploitation. This hyperparameter is dynamically adjusted according to the optimization process and the overall confidence of the model.

4. The method according to claim 1 or 3, characterized in that, In step S4.3, the constraint handling mechanism includes: using gradient projection to adjust and project suggestions that exceed the hard constraint boundary back into the feasible region; and converting the deviation of the soft constraint into a penalty term for the acquisition function.

5. The method according to claim 1, characterized in that, In step S5, the online update of the OH-GPR-UQN model includes updating the posterior distribution of the Gaussian process regression using newly acquired data points, and adjusting the parameters of the gating network to optimize the division of the working condition region and the model switching / fusion logic.

6. A multi-objective optimization unit adaptive combustion and control real-time optimization system, characterized in that, include: The data acquisition and preprocessing module (10) is used to acquire multiple operating parameters and current decision parameters during the operation of the thermal power unit in real time. The operating parameters include at least coal feed rate and main steam parameters, and the decision parameters include at least furnace target oxygen content, primary air pressure, and opening degree of each secondary air damper. The module also preprocesses the operating parameters and decision parameters. The objective function construction and uncertainty quantification model initialization module (20) is connected to the data acquisition and preprocessing module (10) and is used for objective function construction and uncertainty quantification model initialization. It is also configured with computing resources to run the online Gaussian process regression and uncertainty quantification network OH-GPR-UQN. The OH-GPR-UQN includes: multiple local Gaussian process regression GPR models, each of which is responsible for identifying a specific working condition region. A gating network dynamically determines the activation weights of each local GPR model or selects one of the dominant models based on the current operating parameters and decision parameters; and an uncertainty quantification auxiliary network is used to further refine the uncertainty assessment of gradient estimation based on the predicted mean, variance, and input characteristics of each local GPR model. The gradient estimation and uncertainty assessment module (30) is connected to the objective function construction and uncertainty quantification model initialization module (20), and is based on the gradient estimation and uncertainty assessment of OH-GPR-UQN. The Bayesian optimization-guided constraint hill-climbing optimization module (40) is connected to the gradient estimation and uncertainty assessment module (30) and is used for Bayesian optimization-guided adaptive exploration and utilization constraint hill-climbing optimization. It contains a data acquisition function construction unit and a constraint processing unit. The closed-loop control and online model update module (50) is connected to the Bayesian optimization-guided constraint hill climbing optimization module (40) and the distributed control system DCS of the thermal power unit. It is used for closed-loop control and online model update, and feeds back the data to the objective function construction and uncertainty quantification model initialization module (20) to update OH-GPR-UQN.

7. The system of claim 6, wherein, The OH-GPR-UQN in the objective function construction and uncertainty quantification model initialization module (20) includes: multiple parallel local Gaussian process regression (GPR) model processing units, a gating network decision unit, and an uncertainty quantification auxiliary network processing unit.

8. The system of claim 6 or 7, wherein, The acquisition function construction unit in the Bayesian optimization-guided constrained hill-climbing optimization module (40) dynamically constructs and optimizes the acquisition function based on the gradient and uncertainty output by the gradient estimation and uncertainty assessment module (30) to determine the adjustment step size and directional perturbation of the decision parameters. The acquisition function is an improved upper confidence interval UCB function, which has the form A(u)=μ(u)+κσ_grad(u), where μ(u) represents the expected net power improvement based on the current gradient moving to the decision parameter space point u, σ_grad(u) represents the uncertainty of gradient estimation near point u, and κ is a hyperparameter that balances exploration and utilization. This hyperparameter is dynamically adjusted according to the optimization process and the overall confidence of the model.

9. The system of claim 6, wherein, The constraint processing unit in the Bayesian optimization-guided constraint hill-climbing optimization module (40) integrates gradient projection algorithm to process hard constraints and handles soft constraints by modifying the evaluation value of the acquisition function or introducing a penalty term.

Images