A machine learning-based combustion optimization control method
By employing a machine learning-based combustion optimization control method, which utilizes GMM and CNN-LSTM models to identify operating conditions and combines dynamic weights and safety constraints, multi-objective collaborative optimization of coal-fired boilers under different operating conditions is achieved. This solves the problem of unstable control parameters in traditional methods and improves the stability and safety of the combustion process.
Patent Information
- Application Number
- CN202611520479.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-29
- Publication Date
- 2026-11-17
AI Technical Summary
Existing combustion optimization control methods for coal-fired boilers are difficult to achieve multi-objective collaborative optimization under wide load and variable operating conditions. They lack operating condition identification and safety constraints, resulting in unstable control parameters and insufficient model adaptability.
A machine learning-based combustion optimization control method is adopted. The operating conditions are identified by GMM, a CNN-LSTM combustion state prediction model is constructed, and dynamic weights and safety constraints are combined to achieve multi-objective collaborative optimization. Online closed-loop optimization and model update are used to achieve adaptive control.
It achieves adaptive, multi-objective optimization control under different operating conditions, improves the stability and safety of the combustion process, and enhances the adaptability and reliability of the model.
Smart Images

Figure CN123113390A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of combustion optimization control of coal-fired boilers and intelligent control technology of thermal power units, and in particular to a combustion optimization control method based on machine learning. Background Technology
[0002] With the development of machine learning, big data, and artificial intelligence technologies, intelligent power generation technology is becoming an important development direction for the thermal power generation industry. The combustion process in a coal-fired boiler is characterized by a large space, high temperature, strong coupling, strong nonlinearity, and large time lag. The combustion state is influenced by a variety of factors, including unit load, coal quality, air volume distribution, burnout air opening, secondary air ratio, oxygen content, and feedwater flow. Traditional combustion optimization control methods typically rely on operational experience or fixed control logic, making it difficult to accurately describe the nonlinear mapping relationships between multiple inputs and outputs, and also difficult to adapt to load fluctuations and changes in operating conditions in a timely manner.
[0003] Boiler combustion optimization is essentially a multi-objective collaborative optimization problem. It requires ensuring the safe and stable operation of the boiler while simultaneously considering objectives such as reducing NOx emissions, controlling incomplete CO combustion, stabilizing furnace outlet temperature, and smoothly adjusting control parameters. These objectives are interdependent; for example, reducing NOx emissions may affect burnout, excessively reducing airflow may lead to increased CO levels, and rapid airflow adjustments may cause combustion fluctuations. Therefore, single-objective optimization or fixed-weight optimization methods are insufficient to meet the actual control requirements under wide load and variable operating conditions.
[0004] Existing boiler combustion optimization methods based on neural networks or deep learning can establish predictive relationships between operating parameters and combustion states. However, most methods employ a single predictive model and a fixed optimization strategy, lacking differentiation between low-load, medium-load, high-load, and variable-load conditions, resulting in insufficient adaptability of the models under complex operating conditions. Furthermore, existing optimization methods often directly output optimized control parameters, lacking safety constraints such as furnace outlet temperature, oxygen content, main steam temperature, and the rate of change of control parameters, potentially posing implementation risks in engineering applications. In addition, some methods lack effective online feedback and update mechanisms, making it difficult to continuously correct the model based on changes in boiler fuel characteristics, equipment status, and operating environment.
[0005] Therefore, it is necessary to propose a closed-loop optimization control method for low-emission combustion in coal-fired boilers. Based on the establishment of a combustion state prediction model, this method introduces operating condition identification, dynamic weight decision-making, safety constraint optimization, and online model update mechanisms to achieve adaptive, multi-objective, and closed-loop optimization control of boiler combustion control parameters under different operating conditions. Summary of the Invention
[0006] The purpose of this invention is to provide a combustion optimization control method based on machine learning to solve the above-mentioned technical problems.
[0007] To achieve the above objectives, this invention provides a combustion optimization control method based on machine learning, comprising the following steps: S1. Operational Data Acquisition and Database Construction: Establish a data communication connection with the DCS (Distributed Control System) of the thermal power unit, collect operating parameters during boiler operation online, and construct an operational database for combustion state prediction model training and online optimization after preprocessing the operating data; S2. Operating condition identification and combustion state prediction modeling: Based on the operating database, operating condition characteristic parameters are extracted, and GMM (Gaussian Mixture Model) is used to perform cluster analysis on the operating data. The posterior probability distribution of the current operating state belonging to each typical operating condition is calculated to identify the current operating condition. A combustion state prediction model is established to predict the combustion state and pollutant emission levels under different operating conditions. S3. Multi-objective collaborative optimization based on dynamic weights and safety constraints: Based on the current operating conditions of the boiler, a multi-objective optimization evaluation function is constructed. Combined with the boiler safety operation constraints, the three types of control parameters, namely, burnout air opening degree, secondary air ratio and total air volume, are collaboratively optimized to obtain the optimized control parameters under the current operating conditions. S4. Online Closed-Loop Optimization and Model Update: The optimized control parameters are fed back to the DCS for execution, and the operation feedback data is collected at the same time to perform online incremental updates to the combustion state prediction model, so as to realize continuous correction and adaptive optimization of the combustion state prediction model.
[0008] Preferably, S1 specifically includes: Communication is established with DCS through industrial communication interface, and operating parameters are collected at a preset sampling period. The operating parameters include at least the unit load, main steam pressure, main steam temperature, fuel quantity, feedwater quantity, burnout air opening, secondary air ratio, total air volume, furnace outlet temperature, CO content, and NOx content. The collected operating parameters are timestamped and aligned with the time, and then preprocessed by filtering and smoothing, outlier identification and removal. Finally, the original operating parameters and the preprocessed operating parameters are stored separately and organized into a dataset according to the input and output requirements of the combustion state prediction model.
[0009] Preferably, the typical operating conditions in S2 include four categories: low load, medium load, high load, and variable load.
[0010] Preferably, the expression for GMM in S2 is: ; in, This is the feature vector of the working condition; Number of operating condition categories; For the first Gaussian component weights for each working condition category; For the first The mean vector of Gaussian components for each working condition category; For the first The covariance matrix of Gaussian components for each working condition category; It is a Gaussian probability density function; Working condition feature vector The corresponding joint probability density value.
[0011] Preferably, the combustion state prediction model in S2 is constructed based on a CNN-LSTM (Convolutional Neural Network - Long Short-Term Memory) model, whose input parameter vector Set as: ; in, For unit load, Main steam pressure Main steam temperature For fuel quantity, For water supply, To burn out the wind opening, For secondary air mixing ratio, Total air volume; Output parameter vector of combustion state prediction model Set as: ; in, The furnace outlet temperature, CO content, NOx content; Mapping relationship of boiler combustion state prediction model Represented as: ; in, For the first Deep learning prediction mapping functions corresponding to each working condition category; For the first The model parameter set corresponding to each working condition category.
[0012] Preferably, the combustion state prediction model consists of a convolutional feature extraction layer, an LSTM temporal modeling layer, and a fully connected output layer. The convolutional feature extraction layer is used to extract local coupling features between multiple parameters, the LSTM temporal modeling layer is used to learn the dynamic temporal characteristics of the combustion process, and the fully connected output layer is used to output multi-objective prediction results.
[0013] Preferably, the multi-objective optimization evaluation function in S3 is: ; ; ; ; ; in, The overall optimization objective value; , , , The dynamic weights are for four optimization objectives: NOx emissions, CO content control, furnace outlet temperature stability, and control parameter stability. For NOx emission assessment items, This is an evaluation item for incomplete combustion of CO. This is an evaluation item for furnace outlet temperature deviation. To control stability evaluation items; To predict NOx content for the model, This is the baseline value for NOx emission control. This is the penalty coefficient; To predict CO content for the model, This represents the maximum allowed limit for CO. To predict the furnace outlet temperature using the model, This is a reference value for the furnace outlet temperature; For the current moment, the first One adjustable parameter, For the previous moment One adjustable parameter, , The first The upper and lower limits of an adjustable parameter. The time corresponds to the burnout air opening, secondary air ratio, and total air volume in sequence.
[0014] Preferably, the dynamic weights are generated by weighting the posterior probabilities of each working condition output by the GMM, and the calculation formula is as follows: ; in, , , , These represent the posterior probabilities of the operating conditions being low load, medium load, high load, and variable load, respectively. , , , The first The preset weights of each optimization objective under the corresponding working conditions. .
[0015] Preferably, the safe operation constraints in S3 include: The range of the burnout air opening is 0~100%; The secondary air mix ratio ranges from 0% to 100%. The optimization search range for total air volume is 95% to 105% of the baseline total air volume under the current load, and the adjustment amount of total air volume in a single instance shall not exceed 2% of the baseline total air volume. The optimized control parameter adjustment is then limited and smoothed before being output.
[0016] Preferably, S4 specifically includes: An external optimization control module is deployed on the plant's computing server and interacts with the DCS in real time through an industrial communication interface. The external optimization control module reads the DCS operating parameters in real time, identifies the operating conditions and calls the corresponding combustion state prediction model. When the predicted indicators exceed the preset range, it starts optimization calculation and generates control commands to be fed back to the DCS for execution. When the amount of newly added operational data reaches a preset threshold or the prediction error of the combustion state prediction model exceeds a set threshold, incremental training is used to correct the parameters of the combustion state prediction model. After the accuracy verification is qualified, the original combustion state prediction model is replaced.
[0017] Therefore, the above-mentioned combustion optimization control method based on machine learning has the following beneficial effects: 1. Strong adaptability to operating conditions: The GMM is used to probabilistically identify the boiler operating conditions, classify four typical operating conditions and establish a combustion state prediction model, which solves the problem of insufficient adaptability of a single model under wide load and variable operating conditions. The optimization strategy switches smoothly with the probability of operating conditions, avoiding the control abrupt changes during the transition of operating conditions in the traditional fixed weight method.
[0018] 2. Multi-objective dynamic collaborative optimization: Construct a multi-objective optimization evaluation function that includes NOx emissions, CO control, furnace temperature stability and control stability. Based on the probability of the current operating condition, dynamically adjust the weight of each objective, take into account the core control requirements under different operating conditions, and achieve balanced optimization among multiple objectives.
[0019] 3. Comprehensive safety constraint system: During the optimization process, upper and lower limit constraints of control parameters and single adjustment amount limit constraints are introduced to smooth the output command, effectively avoiding combustion instability caused by large fluctuations in control parameters, and significantly improving the safety and reliability of engineering applications.
[0020] 4. Closed-loop continuous self-optimization: An external optimization control module interacts with the DCS to achieve online closed-loop control; at the same time, an incremental model online update mechanism is established, which can continuously correct the model based on operational feedback, adapt to the slow changes in fuel characteristics, equipment status and operating environment, and ensure the long-term effectiveness of the model.
[0021] 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
[0022] Figure 1 This is a flowchart of a combustion optimization control method based on machine learning according to the present invention. Figure 2 This is a structural diagram of the multi-condition combustion state prediction model of the present invention; Figure 3 This is a flowchart of the online closed-loop optimization control of the present invention. Detailed Implementation
[0023] To make the objectives, technical solutions, and advantages disclosed in the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the embodiments of the present invention and are not intended to limit the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of this application. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout.
[0024] It should be noted that the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, such as a process, method, system, product, or server that includes a series of steps or units, not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such process, method, product, or device.
[0025] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0026] like Figures 1-3As shown, this invention provides a combustion optimization control method based on machine learning. This embodiment uses a 660MW supercritical coal-fired boiler as the application object, and specifically includes the following steps: S1. Data Acquisition and Database Construction: This embodiment uses data acquisition and preprocessing software developed based on Python as the core of data processing. It establishes a communication connection with the DCS of the thermal power unit through industrial communication boards, data acquisition boards or DCS data interfaces to obtain key operating parameters of the boiler and unit during operation online.
[0027] Data acquisition is performed according to a preset sampling period. In this embodiment, the sampling period is 3 seconds, meaning that boiler operating parameters are collected every 3 seconds. The collected operating parameters include, but are not limited to, unit load, main steam pressure, main steam temperature, fuel quantity, feedwater quantity, burnout air opening, secondary air ratio, total air volume, furnace outlet temperature, CO content, and NOx content. During the acquisition process, each operating parameter is timestamped, and data is aligned according to a unified time reference to ensure that data from different measuring points have a corresponding relationship under the same operating conditions.
[0028] The collected operational data is first preprocessed, including filtering, smoothing, outlier identification, and outlier removal, in order to reduce the impact of measurement point noise, instantaneous disturbances, and instrument anomalies on model training. Among these, filtering and smoothing can employ moving average filtering, median filtering, or other digital filtering methods, while outliers can be identified and removed based on the reasonable range of operational parameters, statistical distribution characteristics, or the variation amplitude of adjacent sampling points.
[0029] Simultaneously, the monitoring system performs synchronous backup of the raw data, storing the original operating data in a local or server database before preprocessing to facilitate subsequent data traceability, model validation, and algorithm optimization. The preprocessed operating data is organized according to model input / output requirements, forming a dataset for operating condition identification, combustion state prediction model training, validation, and online optimization. Furthermore, two sets of feature extractions are performed on the preprocessed data: the first set extracts unit load, oxygen content, air-coal ratio, main steam flow, and total air volume, forming an operating condition feature vector. The first group only inputs GMM to complete the operating condition identification and outputs the posterior probability of each operating condition for subsequent dynamic weight calculation of multi-objective optimization; the second group selects unit load, main steam pressure, main steam temperature, fuel quantity, feedwater quantity, burnout air opening, secondary air ratio, and total air volume to form the input parameter vector. The time-series samples were then organized and fed into a combustion state prediction model built on a CNN-LSTM model.
[0030] S2. Operating Condition Identification and Combustion State Prediction Modeling: Based on the boiler operation dataset obtained in S1, the boiler operating conditions are first identified, and combustion state prediction models are established for different operating conditions. The PyTorch deep learning framework is used to complete model construction, training, and validation. The combustion state prediction model is built based on a CNN-LSTM model, and its input parameter vector... Set as: ; in, For unit load, Main steam pressure Main steam temperature For fuel quantity, For water supply, To burn out the wind opening, For secondary air mixing ratio, Total air volume; Output parameter vector of combustion state prediction model Set as: ; in, The furnace outlet temperature, CO content, NOx content; Combustion state prediction models are used to characterize the nonlinear dynamic mapping relationship between key boiler operating parameters and furnace outlet temperature, CO content, and NOx content. The mapping relationship of the boiler combustion state prediction model... Represented as: ; in, For the first Deep learning prediction mapping functions corresponding to each working condition category; For the first The model parameter set corresponding to each working condition category.
[0031] Specifically, the Gaussian Model (GMM) is first used for operating condition identification. The GMM fits the boiler's operating state space using multiple Gaussian distributions, determines the probability distribution of the current operating state belonging to each typical operating condition based on the posterior probability, and selects the operating condition with the highest probability as the dominant operating condition, then calls the corresponding combustion state prediction model. The GMM expression is: ; in, This is the feature vector of the working condition; Number of operating condition categories; For the first Gaussian component weights for each working condition category; For the first The mean vector of Gaussian components for each working condition category; For the first The covariance matrix of Gaussian components for each working condition category; It is a Gaussian probability density function; Working condition feature vector The corresponding joint probability density value. In this embodiment, the operating conditions are divided into four categories: low load, medium load, high load, and variable load. For the combustion state prediction model corresponding to each operating condition, the unit load, main steam pressure, main steam temperature, fuel quantity, feedwater quantity, burnout air opening, secondary air ratio, and total air volume are used as input parameters, and the furnace outlet temperature, CO content, and NOx content are used as output parameters.
[0032] Considering the inertia and hysteresis characteristics of boiler combustion, the historical time window of the CNN-LSTM model in this embodiment is set to 30 seconds. That is, based on a 3-second sampling period, the operating parameters of 10 consecutive sampling points are selected to form the input sequence, used to predict the furnace outlet temperature, CO content, and NOx content at the next moment. In terms of model structure, the combustion state prediction model includes a convolutional feature extraction layer, an LSTM temporal modeling layer, and a fully connected output layer. The convolutional feature extraction layer is used to extract local coupling features between different operating parameters; the LSTM temporal modeling layer is used to learn the dynamic influence of operating parameters changing over time on the combustion output during boiler combustion; and the fully connected output layer is used to output the multi-objective prediction results under the corresponding operating conditions. Figure 2 In the LSTM module, c0, c1, ..., c 10 The h0, h1, ..., h1 represent the feature vectors output by the convolutional feature extraction layer at each sampling time. 10 This represents the hidden state vector output by the LSTM temporal modeling layer at each sampling time.
[0033] During the training of the combustion state prediction model, datasets for each working condition are first constructed according to the classification results, and then divided into training, validation, and test sets. A CNN-LSTM network model is built using the PyTorch framework, with the error between predicted and measured values used as the loss function. The model parameters are iteratively updated using backpropagation and an optimizer until the model loss converges or the preset number of training epochs is reached. After training, the generalization ability of the model for each working condition is evaluated using the test set.
[0034] The combustion state prediction model evaluation metrics used in this embodiment include mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²). 2 ). Among them, R 2The closer the value is to 1, the better the model fit; the smaller the values of MAE, RMSE, and MAPE, the smaller the prediction error. The test results of typical combustion state prediction models are shown in Table 1.
[0035] Table 1: Test Results
[0036] S3. Multi-objective collaborative optimization based on dynamic weights and safety constraints: After obtaining the trained combustion state prediction model, a dynamic weighted Bayesian optimization algorithm is used to collaboratively optimize the burnout air opening, secondary air ratio, and total air volume based on the operating condition identification results. The optimization objective is to achieve NOx emission reduction, CO content control, furnace outlet temperature stability, and smooth adjustment of control parameters while ensuring the safe and stable operation of the boiler. Adjustable optimization control parameters are used. Represented as: ; in, This indicates the burnout air opening, used to adjust the amount of air entering the upper burnout zone of the furnace; This indicates the secondary air distribution ratio, used to characterize the proportion of secondary air relative to the total air volume in different layers and areas; This represents the total air volume, used to characterize the overall amount of air entering the boiler furnace to participate in combustion.
[0037] With the objectives of stabilizing furnace temperature, reducing NOx emissions, controlling CO content, and ensuring stable adjustment of control parameters, a multi-objective optimization evaluation function is constructed. : ; ; ; ; ; in, The overall optimization objective value; , , , The dynamic weights are for four optimization objectives: NOx emissions, CO content control, furnace outlet temperature stability, and control parameter stability. For NOx emission assessment items, This is an evaluation item for incomplete combustion of CO. This is an evaluation item for furnace outlet temperature deviation. To control stability evaluation items; To predict NOx content for the model, This is the baseline value for NOx emission control. This is the penalty coefficient; To predict CO content for the model, This represents the maximum allowed limit for CO. To predict the furnace outlet temperature using the model, This is a reference value for the furnace outlet temperature; For the current moment, the first One adjustable parameter, For the previous moment One adjustable parameter, , The first The upper and lower limits of an adjustable parameter. The time corresponds sequentially to the burnout air opening, secondary air ratio, and total air volume; the weights of each optimization objective are generated by weighting the operating condition identification results of GMM. , The calculation formula is: ; in, , , , These represent the posterior probabilities of the operating conditions being low load, medium load, high load, and variable load, respectively. , , , The first The preset weights of each optimization objective under the corresponding working conditions.
[0038] When the boiler's operating status changes, the weights of each optimization objective change continuously according to the operating condition probability, achieving a smooth switching of the optimization strategy and avoiding control abrupt changes during operating condition transitions in traditional fixed-weight methods. In this embodiment, the NOx emission control baseline value... Set as CO allowed upper limit Set as Reference value for furnace outlet temperature The temperature is determined based on the load segment of the unit: 400℃ for low load, 410℃ for medium load, and 403℃ for high load.
[0039] Regarding safety constraints, the upper and lower limits for the burnout air opening are 0-100%; the upper and lower limits for the secondary air ratio are 0-100%; and the total air volume is based on the baseline total air volume under the current load. Dynamically determined, the search scope is optimized as follows: And the total air volume adjustment in a single instance shall not exceed .
[0040] This embodiment prioritizes "low-load NOx control, high-load temperature stability, and variable-load smooth control" and sets preset weights for each operating condition as shown in Table 2.
[0041] Table 2: Preset weights for each operating condition
[0042] S4. Online Closed-Loop Optimization and Model Update: An external optimization control module is built to interact with the DCS of the thermal power unit. This module reads the current unit operating parameters, identifies the current operating condition, and calls the corresponding combustion state prediction model and multi-objective optimization model to calculate the optimized control parameters under the current operating condition. Control optimization commands are generated based on the difference between the optimized control parameters and the current control parameters, and the adjustment amount of the control parameters is limited. The control parameter adjustment amount can be expressed as: ; in, Indicates the adjustment amount of the control parameter. This represents the optimized control parameter vector. This represents the current control parameter vector; it feeds back control optimization instructions to the DCS for execution, realizing online closed-loop optimization of boiler combustion control parameters, and collecting operational feedback data after control; based on the operational feedback data, it updates the combustion state prediction model online, realizing continuous correction and adaptive optimization of model parameters.
[0043] In this embodiment, the external optimization control module is deployed on a high-performance computing server within the plant. The server is configured to meet the requirements of real-time data acquisition, operating condition identification, CNN-LSTM model inference, and Bayesian optimization calculation, and interacts with the DCS through an industrial communication interface.
[0044] The software integrates operating condition identification, combustion state prediction model, dynamic weighted multi-objective optimization, and DCS communication interface program to realize real-time acquisition of DCS operation data, operating condition identification, state prediction, optimization search, and control command feedback.
[0045] During boiler operation, the external optimization control module reads the DCS operating parameters in real time and updates them according to the sampling period. Specifically, it first identifies the current operating condition based on the characteristic parameters of the operating condition, and automatically calls the prediction model under the corresponding operating condition to predict the furnace outlet temperature, CO content and NOx content. Then, it judges whether the current combustion state meets the optimization target based on the prediction results. When the predicted indicators meet the operating requirements, the current control parameters remain unchanged. When the predicted indicators exceed the preset range, the dynamic weighted Bayesian optimization algorithm is called to perform collaborative optimization to obtain the optimal combination of control parameters.
[0046] To avoid large fluctuations in control parameters, the adjustment amounts are limited and smoothed before generating control optimization commands, which are then fed back to the DCS for execution, achieving online closed-loop optimization control of boiler combustion control parameters. Simultaneously, an external optimization control module continuously collects and stores real-time operating data during the closed-loop control process. When the amount of new data reaches a preset threshold or the model prediction error exceeds a set threshold, an online model update mechanism is triggered. Online updates employ incremental training, using new operating data to correct the parameters of the prediction model for the corresponding operating condition. After the updated model meets accuracy requirements, it replaces the original online model and continues to participate in subsequent optimization control processes, achieving continuous correction and adaptive optimization of the combustion optimization control model.
[0047] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.
Claims
1. A combustion optimization control method based on machine learning, characterized in that, Includes the following steps: S1. Operational data acquisition and database construction: Establish a data communication connection with the DCS of the thermal power unit, collect the operating parameters during the boiler operation process online, and construct an operational database for combustion state prediction model training and online optimization after preprocessing the operating data; S2. Operating condition identification and combustion state prediction modeling: Based on the operating database, operating condition characteristic parameters are extracted, and GMM is used to perform cluster analysis on the operating data. The posterior probability distribution of the current operating state belonging to each typical operating condition is calculated to identify the current operating condition. A combustion state prediction model is established to predict the combustion state and pollutant emission levels under different operating conditions. S3. Multi-objective collaborative optimization based on dynamic weights and safety constraints: Based on the current operating conditions of the boiler, a multi-objective optimization evaluation function is constructed. Combined with the boiler safety operation constraints, the three types of control parameters, namely, burnout air opening degree, secondary air ratio and total air volume, are collaboratively optimized to obtain the optimized control parameters under the current operating conditions. S4. Online Closed-Loop Optimization and Model Update: The optimized control parameters are fed back to the DCS for execution, and the operation feedback data is collected at the same time to perform online incremental updates to the combustion state prediction model, so as to realize continuous correction and adaptive optimization of the combustion state prediction model.
2. The combustion optimization control method based on machine learning according to claim 1, characterized in that, S1 specifically includes: Communication is established with DCS through industrial communication interface, and operating parameters are collected at a preset sampling period. The operating parameters include at least the unit load, main steam pressure, main steam temperature, fuel quantity, feedwater quantity, burnout air opening, secondary air ratio, total air volume, furnace outlet temperature, CO content, and NOx content. The collected operating parameters are timestamped and aligned with the time, and then preprocessed by filtering and smoothing, outlier identification and removal. Finally, the original operating parameters and the preprocessed operating parameters are stored separately and organized into a dataset according to the input and output requirements of the combustion state prediction model.
3. The combustion optimization control method based on machine learning according to claim 1, characterized in that: Typical operating conditions in S2 include four categories: low load, medium load, high load, and variable load.
4. The combustion optimization control method based on machine learning according to claim 1, characterized in that: The expression for GMM in S2 is: ; in, This is the feature vector of the working condition; Number of operating condition categories; For the first Gaussian component weights for each working condition category; For the first The mean vector of Gaussian components for each working condition category; For the first The covariance matrix of Gaussian components for each working condition category; It is a Gaussian probability density function; Working condition feature vector The corresponding joint probability density value.
5. The combustion optimization control method based on machine learning according to claim 4, characterized in that: The combustion state prediction model in S2 is built based on a CNN-LSTM model, whose input parameter vector Set as: ; in, For unit load, Main steam pressure Main steam temperature For fuel quantity, For water supply, To burn out the wind opening, For secondary air mixing ratio, Total air volume; Output parameter vector of combustion state prediction model Set as: ; in, The furnace outlet temperature, CO content, NOx content; Mapping relationship of boiler combustion state prediction model Represented as: ; in, For the first Deep learning prediction mapping functions corresponding to each working condition category; For the first The model parameter set corresponding to each working condition category.
6. The combustion optimization control method based on machine learning according to claim 5, characterized in that: The combustion state prediction model consists of a convolutional feature extraction layer, an LSTM temporal modeling layer, and a fully connected output layer. The convolutional feature extraction layer is used to extract local coupling features between multiple parameters, the LSTM temporal modeling layer is used to learn the dynamic temporal characteristics of the combustion process, and the fully connected output layer is used to output multi-objective prediction results.
7. The combustion optimization control method based on machine learning according to claim 1, characterized in that: The multi-objective optimization evaluation function in S3 is: ; ; ; ; ; in, The overall optimization objective value; , , , The dynamic weights are for four optimization objectives: NOx emissions, CO content control, furnace outlet temperature stability, and control parameter stability. For NOx emission assessment items, This is an evaluation item for incomplete combustion of CO. This is an evaluation item for furnace outlet temperature deviation. To control stability evaluation items; To predict NOx content for the model, This is the baseline value for NOx emission control. This is the penalty coefficient; To predict CO content for the model, This represents the maximum allowed limit for CO. To predict the furnace outlet temperature using the model, This is a reference value for the furnace outlet temperature; For the current moment, the first One adjustable parameter, For the previous moment One adjustable parameter, , The first The upper and lower limits of an adjustable parameter. The time corresponds to the burnout air opening, secondary air ratio, and total air volume in sequence.
8. The combustion optimization control method based on machine learning according to claim 7, characterized in that: The dynamic weights are generated by weighting the posterior probabilities of each working condition output by the GMM, and the calculation formula is as follows: ; in, , , , These represent the posterior probabilities of the operating conditions being low load, medium load, high load, and variable load, respectively. , , , The first The preset weights of each optimization objective under the corresponding working conditions. .
9. The combustion optimization control method based on machine learning according to claim 1, characterized in that: The safe operating constraints in S3 include: The range of the burnout air opening is 0~100%; The secondary air mix ratio ranges from 0% to 100%. The optimization search range for total air volume is 95% to 105% of the baseline total air volume under the current load, and the adjustment amount of total air volume in a single instance shall not exceed 2% of the baseline total air volume. The optimized control parameter adjustment is then limited and smoothed before being output.
10. The combustion optimization control method based on machine learning according to claim 1, characterized in that: S4 specifically includes: An external optimization control module is deployed on the plant's computing server and interacts with the DCS in real time through an industrial communication interface. The external optimization control module reads the DCS operating parameters in real time, identifies the operating conditions and calls the corresponding combustion state prediction model. When the predicted indicators exceed the preset range, it starts optimization calculation and generates control commands to be fed back to the DCS for execution. When the amount of newly added operational data reaches a preset threshold or the prediction error of the combustion state prediction model exceeds a set threshold, incremental training is used to correct the parameters of the combustion state prediction model. After the accuracy verification is qualified, the original combustion state prediction model is replaced.
