Boiler combustion optimization control method based on data driving

Abstract

The invention relates to the field of power equipment control data processing, in particular to a boiler combustion optimization control method based on data driving, which comprises the following steps of: acquiring multi-source data such as temperature field distribution, air and smoke pressure, smoke components and coal quality characteristics, and eliminating noise interference by adopting sliding window mean filtering; generating a standardized feature matrix in combination with principal component analysis and a dynamic time warping algorithm; constructing a dynamic coupling model fusing a gradient boosting decision tree and a long short-term memory network, analyzing a nonlinear relationship between pulverized coal particle size distribution and a wind-coal ratio, and predicting combustion efficiency, pollutant concentration and temperature field uniformity; and model parameter self-correction and weight dynamic adjustment are triggered through actual combustion data feedback, and a closed-loop control link is formed. According to the method, accurate modeling of the multi-physical field coupling characteristic of the combustion system is achieved, the time sequence generalization ability under the dynamic working condition is improved, and the purposes of heat efficiency improvement and pollutant emission reduction are effectively balanced.

Description

Technical Field

[0001] The present invention relates to the field of power equipment control data processing, and particularly to a data-driven boiler combustion optimization control method. Background Art

[0002] During the combustion process of a thermal power plant boiler, the data-driven technical path mainly relies on the multi-dimensional acquisition and analysis of combustion system operation data, and realizes dynamic regulation by constructing a non-linear correlation model between combustion efficiency, pollutant emissions, and operation parameters. The combustion process involves complex variables such as pulverized coal particle size distribution, primary air-secondary air ratio, and oxygen concentration field distribution. Traditional mechanism models are difficult to accurately describe the coupling characteristics of multi-physical fields. By deploying a sensor network for temperature, pressure, gas composition, etc. to collect combustion thermodynamics parameters in real time and establishing a machine learning model in combination with historical operating condition data, sensitive factors of combustion efficiency and NOx, SO2 emissions can be identified. Based on gradient boosting decision tree or deep neural network algorithms, a functional model of combustion state can be fitted, and control target values such as the optimal air-coal ratio and excess air coefficient under a specific load can be derived, and then an optimization strategy set for burner swing angle and secondary air damper opening can be generated. Finally, on the premise of ensuring stable steam parameters, boiler thermal efficiency can be improved and coordinated pollutant emissions reduction can be achieved.

[0003] The reason why traditional static data modeling methods are difficult to construct time series generalization ability lies in the double mismatch between the time-varying parameter coupling mechanism of the combustion process and the spatio-temporal correlation characteristics of data: coal type switching and load fluctuations lead to the dynamic non-stationary characteristics of the combustion system, and there are minute-level time-varying coupling relationships between key parameters such as air-coal ratio and oxygen concentration. However, the static model parameters are fixed and cannot represent the dynamic transfer function between combustion efficiency and operating variables; there is a time lag effect in the combustion state migration, and the response of the furnace temperature field distribution to the secondary air damper opening has an inertial delay of ten-minute magnitude. The alignment of input and output data of the static model does not consider the time delay accumulation characteristics of the combustion thermodynamics process; the time series data collected by the sensor network contains multi-scale correlation characteristics of combustion efficiency and pollutant emissions. Conventional regression models ignore the matching degree between the data autocorrelation structure and the state space expression of the combustion system, resulting in a decline in prediction performance when the model is extrapolated to new operating conditions. Summary of the Invention

[0004] Aiming at the deficiencies of the prior art, the present invention provides a data-driven boiler combustion optimization control method to solve the problems of inaccurate modeling of multi-physical field coupling characteristics and lack of time series generalization ability of the combustion efficiency prediction model under dynamic operating conditions.

[0005] To solve the above technical problems, the specific technical solutions of the present invention are as follows: The data-driven boiler combustion optimization control method provided by the present invention includes: Obtain multi-source sensor data of the boiler combustion system, where the multi-source sensor data includes temperature field distribution, wind and flue gas system pressure, flue gas component concentration, air volume parameters, and coal quality characteristics; Perform fusion processing on the multi-source sensor data, including eliminating noise through sliding window mean filtering, extracting sensitive factors through principal component analysis, and aligning time series data through dynamic time warping to generate a standardized feature matrix containing combustion efficiency sensitive factors; Construct a dynamic coupling model based on the standardized feature matrix, and model the standardized feature matrix through a hybrid machine learning algorithm. The hybrid machine learning algorithm includes gradient boosting decision tree and long short-term memory network to predict combustion state parameters, where the combustion state parameters include combustion efficiency, pollutant concentration, and temperature field uniformity; Based on the combustion state parameters output by the dynamic coupling model, adopt a multi-objective optimization algorithm with dynamic decomposition to synergistically optimize combustion efficiency, pollutant concentration, and temperature field uniformity, and generate a hierarchical control instruction set including air-coal ratio adjustment amount, damper opening correction amount, and burner tilt compensation amount; Send the hierarchical control instruction set to the distributed control system to drive the adjustment of the secondary damper and the burner tilt actuator, and collect the actual combustion data after execution in real time; Based on the deviation between the actual combustion data and the predicted combustion state parameters output by the dynamic coupling model, trigger parameter self-correction of the dynamic coupling model and weight dynamic adjustment of the multi-objective optimization algorithm, update the splitting node threshold of the gradient boosting decision tree and the hidden layer weights of the long short-term memory network, and feedback the adjusted parameters to the dynamic coupling model and the multi-objective optimization algorithm to form a closed-loop control link.

[0006] Further, for the data-driven boiler combustion optimization control method of the present invention, the fusion processing of multi-source sensor data includes: Perform denoising processing on the fly ash carbon content signal through sliding window mean filtering to generate denoised time series data; Perform principal component analysis on the air-coal ratio and oxygen concentration parameters in the denoised time series data to extract combustion efficiency sensitive factors and generate a low-dimensional feature vector; Use the dynamic time warping algorithm to align the minute-level sampled data of the pulverized coal flow rate with the second-level sampled data of the air volume to generate a standardized feature matrix containing coal quality characteristics, air volume ratio, and temperature distribution, and input the standardized feature matrix into the construction process of the dynamic coupling model.

[0007] Further, for the data-driven boiler combustion optimization control method of the present invention, the construction of the dynamic coupling model includes: analyzing the non-linear relationship between the pulverized coal particle size distribution and the air-coal ratio in the standardized feature matrix based on the gradient boosting decision tree algorithm, screening key control variables, and generating a list of sensitive factor weights; Performing temporal correlation learning on the oxygen concentration field and the furnace temperature in the standardized feature matrix through a long short-term memory network, combining the time-delay compensation layer to match the delay time of the damper adjustment and the temperature response, and outputting the combustion state prediction value; When it is detected that the coal type is switched or the load fluctuation amplitude exceeds the preset threshold, trigger sliding window incremental training, and update the splitting node threshold of the gradient boosting decision tree and the hidden layer weights of the long short-term memory network.

[0008] Further, for the data-driven boiler combustion optimization control method of the present invention, the dynamic decomposition multi-objective optimization algorithm includes: Based on the combustion state prediction value output by the dynamic coupling model, using the Chebyshev decomposition method to decompose the multi-objective optimization problem of combustion efficiency, pollutant concentration, and temperature field uniformity into scalar sub-problems, and dynamically adjusting the weight distribution of the sub-problems according to the real-time coal quality characteristics; Optimizing the reference point according to the deviation between the flue gas emission data and the combustion efficiency prediction value, and generating a dynamic Pareto front; Defining the air-coal ratio and the opening degree of the secondary damper as adjacent sub-problems, and generating a Pareto optimal solution set through local search; Based on the fuzzy membership function, screening the feasible solution domain that satisfies the steam pressure constraint, and combining the sensitivity backtracking analysis to quantify the gradient influence of the burner tilt angle on the temperature field uniformity, and generating a hierarchical control instruction set including the air-coal ratio adjustment amount, the damper opening correction amount, and the burner tilt angle compensation amount.

[0009] Further, for the data-driven boiler combustion optimization control method of the present invention, the model self-correction includes: When the deviation between the actual thermal efficiency and the prediction value output by the dynamic coupling model exceeds the preset threshold, use the adaptive moment estimation algorithm to update the weight parameters of the long short-term memory network; Adjust the splitting threshold of the gradient boosting decision tree according to the change in the pollutant concentration in the actual combustion data, generate updated hybrid model parameters, and feedback the updated parameters to the dynamic coupling model.

[0010] Further, for the data-driven boiler combustion optimization control method of the present invention, the weight dynamic adjustment includes: based on the change in the actual emission concentration and combustion efficiency, reversely correcting the sub-problem weight distribution generated based on the real-time coal quality characteristics in the dynamic decomposition multi-objective optimization algorithm; Retain the operating condition data within the preset time window, rollingly update the training sample set of the dynamic coupling model to suppress model degradation, and input the updated sample set into the incremental training processes of the gradient boosting decision tree and the long short-term memory network.

[0011] Furthermore, for the data-driven boiler combustion optimization control method of the present invention, the execution of the hierarchical control instruction set includes: Send the rough adjustment instruction and the fine adjustment instruction of the secondary air damper generated by the multi-objective optimization algorithm based on dynamic decomposition to the actuator in stages through the distributed control system; Drive the actuator to perform angle correction according to the burner tilt compensation amount in the hierarchical control instruction set, and synchronously feedback the temperature field distribution uniformity monitoring data to the dynamic coupling model.

[0012] Furthermore, for the data-driven boiler combustion optimization control method of the present invention, the time series alignment process further includes: Perform multi-resolution interpolation on the minute-level sampling data of the pulverized coal flow rate and the second-level sampling data of the air volume to generate a time series feature matrix with a unified time baseline; Dynamically adjust the alignment path constraint conditions of the dynamic time warping algorithm based on the coal quality characteristics, optimize the spatio-temporal alignment accuracy, and input the aligned time series feature matrix into the construction process of the dynamic coupling model.

[0013] Furthermore, for the data-driven boiler combustion optimization control method of the present invention, the sensitivity backtracking analysis includes: quantifying the gradient influence of the burner tilt on the temperature field uniformity, and generating a priority ranking of the tilt compensation amount; Combine the coupling relationship between the air damper opening adjustment amount and the steam pressure constraint in the multi-objective optimization algorithm based on dynamic decomposition to generate a time series of multi-level control instructions, and send the time series to the distributed control system.

[0014] Furthermore, for the data-driven boiler combustion optimization control method of the present invention, the closed-loop control link includes: feeding back the actual combustion data to the dynamic coupling model to trigger the incremental learning update of the split node threshold of the gradient boosting decision tree and the weights of the long short-term memory network; Feed back the change amount of the emission concentration to the multi-objective optimization algorithm based on dynamic decomposition to dynamically adjust the weight distribution of the sub-problems; Delete the expired operating condition data through the data rolling update mechanism, and maintain the matching of the dynamic coupling model with the real-time operating conditions based on the updated training sample set.

[0015] Advantages of the present invention; The beneficial effects of the present invention are reflected in that through the spatio-temporal alignment and dynamic coupling modeling of multi-source data, the prediction accuracy of combustion efficiency and the synergy of pollutant control are effectively improved. The gradient boosting decision tree is used to analyze the non-linear relationship between pulverized coal particle size distribution and air-coal ratio, and combined with the time-delay compensation mechanism of the long short-term memory network, the inertial delay characteristics of the combustion thermodynamics process are accurately characterized, and the problem of inaccurate multi-physical field coupling modeling is solved; the multi-objective optimization algorithm based on dynamic decomposition adjusts the sub-problem weights based on the real-time coal quality characteristics, generates a hierarchical control instruction set, and realizes the dynamic balance of combustion efficiency, emission indexes and temperature field uniformity; the closed-loop control link eliminates the model mismatch caused by coal type switching and load fluctuation in real time through the model self-correction and data rolling update mechanism, and enhances the self-adaptability of the system to time-varying working conditions. The above technical solutions form a complete closed loop from data acquisition, model iteration to optimization execution, significantly reducing the frequency of manual parameter adjustment and improving the economy and environmental protection of boiler operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the present invention, the accompanying drawings required for the embodiments will be briefly introduced below. Obviously, for those of ordinary skill in the art, without creative efforts, other accompanying drawings can also be obtained based on the accompanying drawings.

[0017] Figure 1 It is a flowchart of the adjustment method of the data-driven boiler combustion optimization control method provided by the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] To make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present invention and the corresponding accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the protection scope of the present invention. The technical solutions provided by the embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings. To better understand the objectives of the present invention, the present invention will be further described in detail below.

[0019] Please refer to Figure 1 , the present invention provides a data-driven boiler combustion optimization control method, including: Step S101, obtaining multi-source sensor data of the boiler combustion system, where the multi-source sensor data includes temperature field distribution, air and flue gas system pressure, flue gas component concentration, air volume parameter and coal quality characteristics; Step S102: Perform fusion processing on the multi-source sensor data, including eliminating noise through sliding window mean filtering, extracting sensitive factors through principal component analysis, and aligning time-series data through dynamic time warping to generate a standardized feature matrix containing combustion efficiency sensitive factors; Step S103: Build a dynamic coupling model based on the standardized feature matrix, and model the standardized feature matrix through a hybrid machine learning algorithm, where the hybrid machine learning algorithm includes gradient boosting decision tree and long short-term memory network, to predict combustion state parameters, and the combustion state parameters include combustion efficiency, pollutant concentration, and temperature field uniformity; Step S104: Based on the combustion state parameters output by the dynamic coupling model, use a multi-objective optimization algorithm with dynamic decomposition to synergistically optimize combustion efficiency, pollutant concentration, and temperature field uniformity, and generate a hierarchical control instruction set containing air-coal ratio adjustment amount, damper opening correction amount, and burner tilt compensation amount; Step S105: Send the hierarchical control instruction set to the distributed control system to drive the adjustment of the secondary damper and burner tilt actuator, and collect the actual combustion data after execution in real time; Step S106: Based on the deviation between the actual combustion data and the predicted combustion state parameters output by the dynamic coupling model, trigger the parameter self-correction of the dynamic coupling model and the weight dynamic adjustment of the multi-objective optimization algorithm, update the splitting node threshold of the gradient boosting decision tree and the hidden layer weights of the long short-term memory network, and feedback the adjusted parameters to the dynamic coupling model and the multi-objective optimization algorithm to form a closed-loop control link.

[0020] In the data-driven boiler combustion optimization control method provided by the present invention, first, by deploying temperature sensors, pressure transmitters, flue gas composition analyzers, air volume measuring devices, and coal quality detectors, the temperature field distribution, wind and flue gas system pressure, flue gas composition concentration, air volume parameters, and coal quality characteristic data of the boiler combustion system are collected in real time. The multi-source sensor data is processed through sliding window mean filtering to eliminate the random noise in the fly ash carbon content signal and generate denoised time-series data; further perform principal component analysis on the air-coal ratio and oxygen concentration parameters to extract sensitive factors characterizing combustion efficiency, including the weight distribution of the excess air coefficient and the secondary damper opening, and output a low-dimensional feature vector; at the same time, use the dynamic time warping algorithm to perform multi-resolution interpolation on the minute-level sampling data of the pulverized coal flow rate and the second-level sampling data of the air volume, align the data with different time baselines to a unified time-series feature matrix, and generate a standardized feature matrix containing coal quality characteristics, air volume ratio, and temperature distribution. The above data processing process provides high-quality input for subsequent modeling by eliminating noise interference, reducing the feature dimension, and aligning time-series differences.

[0021] Based on the standardized feature matrix, a gradient boosting decision tree is used to analyze the non-linear relationship between the pulverized coal particle size distribution and the air-coal ratio, screen key control variables and generate a list of sensitive factor weights. At the same time, through a long short-term memory network, the temporal correlation characteristics between the oxygen concentration field and the furnace temperature are learned, and a minute-level delay between the damper adjustment action and the temperature response is matched through a time-delay compensation layer, and the predicted values of combustion efficiency, pollutant concentration and temperature field uniformity are output. When it is detected that the coal type is switched or the load fluctuation amplitude exceeds the preset threshold, a sliding window incremental training mechanism is triggered, and the splitting node threshold of the gradient boosting decision tree and the hidden layer weights of the long short-term memory network are updated based on the latest operating condition data to ensure the adaptability of the dynamic coupling model to the time-varying characteristics of the combustion system.

[0022] The predicted combustion state values output by the dynamic coupling model are input into a multi-objective optimization algorithm. The multi-objective problem of maximizing combustion efficiency, minimizing pollutant concentration and optimizing temperature field uniformity is decomposed into scalar sub-problems by using the Chebyshev decomposition method, and the weight allocation of each sub-problem is dynamically adjusted according to the real-time coal quality characteristics. Based on the deviation between the flue gas emission data and the predicted combustion efficiency value, a reference point is optimized to generate a dynamic Pareto front. By defining the air-coal ratio and the opening of the secondary damper as adjacent sub-problems, a local search is performed to generate a set of Pareto optimal solutions. Further, a fuzzy membership function is combined to screen the feasible solution domain that satisfies the steam pressure constraint, and the gradient influence of the burner tilt angle on the temperature field uniformity is quantified based on sensitivity backtracking analysis, and a hierarchical control instruction set including the adjustment amount of the air-coal ratio, the coarse and fine adjustment instructions of the secondary damper opening and the compensation amount of the burner tilt angle is generated.

[0023] The hierarchical control instruction set is sent to the actuator in stages through a distributed control system. First, the secondary damper is driven to execute the coarse adjustment instruction to complete the basic adjustment of the air-coal ratio. Subsequently, the fine adjustment instruction is sent, and the damper opening is corrected based on the real-time temperature field distribution uniformity monitoring data. At the same time, the burner angle is adjusted according to the burner tilt angle compensation amount to reduce the thermal stress in the local high-temperature area. The actual combustion data after execution, including the furnace temperature field distribution, flue gas emission concentration and steam pressure parameters, are collected in real time and transmitted to the closed-loop correction module.

[0024] When the deviation between the actual thermal efficiency and the predicted value of the dynamic coupling model exceeds the preset threshold, the model self-correction mechanism is triggered. The weight parameters of the long short-term memory network are updated using the adaptive moment estimation algorithm. At the same time, the splitting threshold of the gradient boosting decision tree is adjusted according to the change in pollutant concentration to generate updated hybrid model parameters. The actual combustion data is used to roll and update the training sample set according to a preset time window, and the expired operating condition data is deleted to inhibit model degradation. Then the updated sample set is input into the incremental training process. At the same time, based on the changes in emission concentration and combustion efficiency, the weight distribution of the sub-problems of the multi-objective optimization algorithm is corrected in reverse, and the reference point of the Pareto front is dynamically adjusted. By feeding the parameter update signal back to the dynamic coupling model and the multi-objective optimization algorithm, a closed-loop control link of "data acquisition - model prediction - instruction optimization - execution feedback" is formed to achieve the dynamic balance of combustion efficiency, pollutant emissions and equipment stability.

[0025] Specifically, for the data-driven boiler combustion optimization control method of the present invention, the fusion processing of multi-source sensor data includes: The carbon content in fly ash signal is denoised by sliding window mean filtering to generate denoised time series data; Principal component analysis is performed on the air-coal ratio and oxygen concentration parameters in the denoised time series data to extract combustion efficiency sensitive factors and generate a low-dimensional feature vector; The dynamic time warping algorithm is used to align the minute-level sampled data of the pulverized coal flow rate and the second-level sampled data of the air volume in time series to generate a standardized feature matrix including coal quality characteristics, air volume ratio and temperature distribution, and the standardized feature matrix is input into the construction process of the dynamic coupling model.

[0026] In the multi-source sensor data fusion processing process of the present invention, first, the carbon content in fly ash signal is processed by sliding window mean filtering. The window length is set to an integer multiple of the pulverized coal conveying period of the combustion system. The data mean within the window is calculated by frame-by-frame sliding to eliminate high-frequency random noise interference and generate the denoised time series data of the carbon content in fly ash. The denoised time series data is further input into the feature dimension reduction link. A multi-dimensional feature space is constructed for the air-coal ratio and oxygen concentration parameters. The eigenvectors of the covariance matrix are calculated by principal component analysis, and the feature dimensions with variance contribution rate exceeding the set threshold are screened. The excess air coefficient and the weight distribution of the secondary air damper opening are extracted as combustion efficiency sensitive factors to generate a low-dimensional feature vector to reduce the computational complexity of subsequent modeling.

[0027] After completing feature dimensionality reduction, time series alignment processing is performed on the pulverized coal flow rate and air volume data. Since the pulverized coal flow rate data uses a minute-level sampling frequency, while the air volume data is high-frequency sampled at the second level, the minimum alignment path of the two time series is calculated through the dynamic time warping algorithm. On the premise of retaining the time stamp benchmark of the pulverized coal flow rate data, multi-resolution interpolation is performed on the high-frequency air volume data to generate a standardized feature matrix with a unified time baseline and including coal quality characteristics, air volume ratio, and temperature distribution. The standardized feature matrix is optimized through the path constraint conditions of the dynamic time warping algorithm to correct the influence of the sudden change in the pulverized coal particle size distribution caused by coal type switching on the time series alignment accuracy, providing spatio-temporal consistent input data for the dynamic coupling model.

[0028] In the above fusion processing steps, the sliding window mean filtering eliminates sensor noise and improves the signal-to-noise ratio of the data; the principal component analysis eliminates redundant features through orthogonal transformation and focuses on the key influencing factors of combustion efficiency; the dynamic time warping algorithm solves the problem of different sampling frequencies of multi-source data and constructs a feature expression with a unified time benchmark. Each step progresses sequentially to form a complete preprocessing link from raw data cleaning, feature dimensionality reduction to time series alignment, providing a data basis for the high-precision prediction of the dynamic coupling model.

[0029] Specifically, for the data-driven boiler combustion optimization control method of the present invention, the construction of the dynamic coupling model includes: based on the gradient boosting decision tree algorithm, non-linear relationship analysis is performed on the pulverized coal particle size distribution and the air-to-coal ratio in the standardized feature matrix, key control variables are screened, and a sensitive factor weight list is generated; Through the long short-term memory network, time series correlation learning is performed on the oxygen concentration field and the furnace temperature in the standardized feature matrix, and the time delay compensation layer is combined to match the delay time of the damper adjustment and the temperature response, and the combustion state prediction value is output; When it is detected that the coal type switches or the load fluctuation amplitude exceeds the preset threshold, sliding window incremental training is triggered to update the split node threshold of the gradient boosting decision tree and the hidden layer weights of the long short-term memory network.

[0030] In the construction process of the dynamic coupling model of the present invention, first, based on the gradient boosting decision tree algorithm, non-linear relationship analysis is performed on the pulverized coal particle size distribution and the air-to-coal ratio in the standardized feature matrix. By traversing the distribution histogram of the pulverized coal particle size in the feature matrix, the mutual information entropy between each bin interval and the air-to-coal ratio parameter is calculated, and the particle size intervals with mutual information entropy exceeding the set threshold are selected as key control variables; further, the Gini coefficient is used to evaluate the contribution degree of each control variable to the combustion efficiency, and a sensitive factor weight list including the secondary damper opening and the excess air coefficient is generated, providing a basis for the feature importance ranking for subsequent multi-objective optimization. The weight list is dynamically updated through the feature importance ranking algorithm to reflect the migration characteristics of the key control variables under the coal type switching condition.

[0031] After completing the analysis of the non-linear relationship, a long short-term memory network is used to process the oxygen concentration field and the time series data of the furnace temperature in the standardized feature matrix. By designing the network time step based on the thermodynamic response period of the combustion system, an input sequence including the historical oxygen concentration field distribution and the real-time air-coal ratio parameter is constructed to learn the time series correlation pattern between the oxygen concentration fluctuation and the temperature field distribution. A time delay compensation layer is embedded in the hidden layer of the long short-term memory network, and the forgetting gate weight of the network memory unit is dynamically adjusted according to the delay time between the damper adjustment action and the temperature response, and the combustion state parameters including the predicted value of the combustion efficiency, the pollutant concentration and the temperature field uniformity are output. The time delay compensation layer corrects the matching accuracy between the network output and the inertial delay of the combustion system by introducing the difference matrix between the adjustment action time stamp and the temperature response time stamp.

[0032] When it is detected that the coal type is switched or the load fluctuation amplitude exceeds the preset threshold, a sliding window incremental training mechanism is triggered. Based on the sliding time window, the latest operating condition data is intercepted, and the incremental gradient boosting decision tree algorithm is used to locally fit the pulverized coal particle size distribution characteristics, and the splitting node threshold is updated to adapt to the combustion characteristics of the current coal type. At the same time, the weights of the hidden layer of the long short-term memory network are fine-tuned online, and the adaptive learning rate adjustment algorithm is used to balance the weight update amplitude of the historical data and the new data, and maintain the prediction stability of the model under dynamic operating conditions. During the incremental training process, the expired data is removed as the window slides, preventing the prediction performance decay of the model caused by the data distribution drift. The updated model parameters are fed back to the dynamic coupling model in real time to form a closed-loop parameter iteration link.

[0033] Specifically, for the data-driven boiler combustion optimization control method of the present invention, the dynamic decomposition multi-objective optimization algorithm includes: Based on the combustion state prediction value output by the dynamic coupling model, the multi-objective optimization problem of combustion efficiency, pollutant concentration and temperature field uniformity is decomposed into scalar sub-problems by using the Chebyshev decomposition method, and the weight distribution of the sub-problems is dynamically adjusted according to the real-time coal quality characteristics; Optimize the reference point according to the deviation between the flue gas emission data and the predicted value of the combustion efficiency, and generate a dynamic Pareto front; Define the air-coal ratio and the opening of the secondary damper as adjacent sub-problems, and generate a Pareto optimal solution set through local search; Based on the fuzzy membership function, the feasible solution domain satisfying the steam pressure constraint is screened, and the gradient influence of the burner tilt angle on the temperature field uniformity is quantified by combining the sensitivity backtracking analysis, and a hierarchical control instruction set including the air-coal ratio adjustment amount, the damper opening correction amount and the burner tilt angle compensation amount is generated.

[0034] The multi-objective optimization algorithm described in the present invention is based on the predicted combustion state values output by the dynamic coupling model. The multi-objective problem of maximizing combustion efficiency, minimizing pollutant concentration, and optimizing temperature field uniformity is decomposed into multiple scalar sub-problems by using the Chebyshev decomposition method. According to the calorific value of coal type, volatile content, and ash fusion point parameters in the real-time coal quality characteristics, the weight distribution strategy of each sub-problem is dynamically adjusted. The weight of optimizing temperature field uniformity is increased for coal types with high ash fusion point, and the optimization priority of combustion efficiency is enhanced for coal types with high volatile content, forming a weight distribution matrix that matches the current combustion condition. The weight distribution matrix is periodically updated through a sliding time window mechanism to respond to the combustion characteristic migration caused by coal type switching.

[0035] Based on the deviation between the measured values of NOx and SO2 concentrations in the flue gas emission data and the predicted values of the dynamic coupling model, the reference point of the multi-objective optimization problem is dynamically corrected. The deviation is mapped to the offset coefficient of the Pareto front through normalization processing to generate a dynamic Pareto front solution space that includes the current condition constraints. The air-coal ratio and the opening degree of the secondary air damper are defined as adjacent sub-problems within the Pareto front solution space. A neighborhood search space is constructed based on the physical coupling relationship of the combustion system control variables, and an adaptive step size strategy is used to perform local search within the neighborhood to screen the Pareto optimal solution set that satisfies the dynamic balance constraints of the air-coal ratio and oxygen concentration.

[0036] In the Pareto optimal solution set, the membership degree of the steam pressure constraint condition is calculated through a fuzzy membership function, and the solution vectors with membership degrees lower than the set threshold are eliminated, retaining the feasible solution domain that satisfies the main steam pressure fluctuation range of the boiler. Further, a sensitivity backtracking analysis algorithm is used to quantify the gradient influence of the burner tilt angle adjustment on the temperature field uniformity index based on historical operation data, generating a priority list of tilt angle compensation amounts. Combining the time series relationship between the coarse adjustment and fine adjustment commands of the damper opening degree, the tilt angle compensation amount with the highest priority is encoded with the air-coal ratio adjustment amount and the correction amount of the secondary air damper opening degree to generate a hierarchical control instruction set that includes execution timing and amplitude parameters. The instruction set is issued in stages through a distributed control system, and the tilt angle compensation amount is preferentially executed to quickly improve the temperature field distribution, and then the combustion efficiency and emission index are refined through the correction of the damper opening degree.

[0037] Specifically, for the data-driven boiler combustion optimization control method described in the present invention, the model self-correction includes: When the deviation between the actual thermal efficiency and the predicted value output by the dynamic coupling model exceeds the preset threshold, the weight parameters of the long short-term memory network are updated using the adaptive moment estimation algorithm; According to the change in pollutant concentration in the actual combustion data, the splitting threshold of the gradient boosting decision tree is adjusted to generate updated hybrid model parameters, and the updated parameters are fed back to the dynamic coupling model.

[0038] During the model self - calibration process of the present invention, when the deviation between the actual thermal efficiency and the predicted value output by the dynamic coupling model exceeds the preset threshold, the adaptive moment estimation algorithm is triggered to update the weight parameters of the long - short - term memory network. Based on the residual sequence between the measured value and the predicted value of the furnace temperature field distribution in the actual combustion data, the gradient direction of the output layer weight of the long - short - term memory network is calculated. By dynamically adjusting the learning rate and the momentum factor, the connection weights of the hidden - layer neurons are updated in batches to reduce the temperature field prediction error. The adaptive moment estimation algorithm introduces an exponentially - weighted moving average of the historical gradient squares during the weight update process to suppress the gradient oscillation caused by sudden changes in the combustion conditions and improve the network convergence stability.

[0039] After the weight update of the long - short - term memory network is completed, according to the real - time change trends of NOx and SO2 concentrations in the flue gas emission data, the splitting threshold parameters of the gradient - boosting decision tree are adjusted. By calculating the correlation coefficient between the pollutant concentration change rate and the combustion efficiency sensitivity factor, the feature dimension and the splitting point of the splitting node are dynamically selected, and the feature branches that have a significant impact on the pollutant concentration fluctuation are preferentially retained. During the splitting threshold adjustment process, the decision tree branch weights are reconstructed based on the pollutant concentration gradient values within the sliding time window to generate the hybrid model parameters that integrate the latest operating conditions characteristics.

[0040] The updated weights of the long - short - term memory network and the splitting threshold parameters of the gradient - boosting decision tree are fed back to the dynamic coupling model in real - time, and the original model configuration is overwritten through the parameter replacement mechanism. The feedback process adopts a version control strategy, retaining copies of the model parameters for the last three iterations. When the new parameters cause the prediction error to increase, it automatically rolls back to the historical stable version. Synchronously trigger the data rolling update mechanism, delete the operating condition data in the training sample set whose time stamps exceed 48 hours, and inject the latest collected actual combustion data into the incremental training queue to maintain the dynamic matching between the model parameters and the real - time combustion state.

[0041] Specifically, for the data - driven boiler combustion optimization control method of the present invention, the dynamic weight adjustment includes: based on the change amounts of the actual emission concentration and the combustion efficiency, reversely correcting the sub - problem weight allocation generated based on the real - time coal quality characteristics in the dynamically decomposed multi - objective optimization algorithm; Retain the operating condition data within the preset time window, and roll - update the training sample set of the dynamic coupling model to suppress model degradation, and input the updated sample set into the incremental training processes of the gradient - boosting decision tree and the long - short - term memory network.

[0042] During the dynamic weight adjustment process of the present invention, based on the change amounts of the actual emission concentration and the combustion efficiency, the sub-problem weight allocation of the multi-objective optimization algorithm with dynamic decomposition is corrected in reverse. By calculating the deviation rates of the measured values and the predicted values of the NOx and SO2 concentrations, and combining the proportion of the volatile matter and the ash content in the coal quality characteristics, the weight coefficients of the sub-problems of combustion efficiency optimization, pollutant reduction, and temperature field uniformity are dynamically adjusted. For coal types with high ash content, the weight proportion of the sub-problem of pollutant reduction is increased, and for coal types with high volatile matter, the priority of the sub-problem of combustion efficiency is enhanced, generating a weight allocation matrix matching the current coal quality characteristics. The weight coefficients are iteratively updated through the exponential smoothing algorithm to balance the correlation between the historical weight distribution and the real-time operating conditions.

[0043] After completing the weight allocation correction, based on a preset time window, the operating condition data of the most recent operating cycle is intercepted, and the training sample set of the dynamic coupling model is updated in a first-in, first-out manner. The expired data samples with time stamps exceeding the window range are deleted, and the real-time data containing key operating conditions such as coal type switching and load fluctuation is retained, suppressing the risk of model degradation caused by data distribution deviation. The updated sample set is input into the incremental training module of the gradient boosting decision tree and the long short-term memory network. By using the local data resampling technique to balance the proportion of new and old samples, the splitting node threshold of the gradient boosting decision tree is updated using the mini-batch gradient descent algorithm. At the same time, based on the time series segmented training strategy, the weights of the hidden layer of the long short-term memory network are fine-tuned to improve the generalization ability of the model for dynamic combustion conditions. The incremental training results are synchronized to the dynamic coupling model in real time, forming a collaborative iterative mechanism for weight allocation optimization and model parameter update.

[0044] Specifically, for the data-driven boiler combustion optimization control method of the present invention, the execution of the hierarchical control instruction set includes: The rough adjustment instruction and the fine adjustment instruction of the secondary air damper generated by the multi-objective optimization algorithm with dynamic decomposition are sent to the actuator in stages through the distributed control system; According to the burner tilt compensation amount in the hierarchical control instruction set, the actuator is driven to perform angle correction, and the temperature field distribution uniformity monitoring data is synchronously fed back to the dynamic coupling model.

[0045] During the execution of the hierarchical control instruction set described in the present invention, first, the distributed control system analyzes the rough adjustment instruction and fine adjustment instruction of the secondary air damper generated by the dynamically decomposed multi-objective optimization algorithm. The rough adjustment instruction is generated based on the reference adjustment amount of the air-coal ratio, and is sent to the secondary air damper actuator in stages through a preset priority queue to complete the rapid reference positioning of the damper opening; the fine adjustment instruction calculates the opening correction amount using a proportional-integral control algorithm according to the real-time temperature field distribution uniformity monitoring data, and is sent at a second-level interval after the rough adjustment is completed to achieve the refined adjustment of the damper opening. The staged sending mechanism isolates the execution timings of the rough adjustment and fine adjustment instructions through a time window control strategy to suppress the action conflict of the actuator.

[0046] During the adjustment of the secondary air damper, the burner tilt compensation amount in the hierarchical control instruction set is synchronously analyzed. Based on the priority sorting of the burner tilt compensation amount, the compensation amount is converted into an actuator drive signal, and the steering and speed of the burner swing angle motor are controlled through pulse width modulation to complete the tilt compensation action. During the generation of the drive signal, a swing angle position closed-loop feedback mechanism is introduced to compare the deviation between the target tilt angle and the actual angle, and dynamically adjust the pulse modulation frequency to improve the angle correction accuracy. The burner tilt angle data and the temperature field distribution uniformity monitoring data after execution are collected in real time and synchronously transmitted to the dynamic coupling model through the data bus to trigger the online verification and parameter update of the combustion state prediction value.

[0047] Specifically, for the data-driven boiler combustion optimization control method described in the present invention, the time series alignment process further includes: Perform multi-resolution interpolation on the minute-level sampling data of the pulverized coal flow rate and the second-level sampling data of the air volume to generate a time series feature matrix with a unified time baseline; Dynamically adjust the alignment path constraint conditions of the dynamic time warping algorithm based on the coal quality characteristics to optimize the spatio-temporal alignment accuracy, and input the aligned time series feature matrix into the construction process of the dynamic coupling model.

[0048] During the time series alignment process described in the present invention, first, for the minute-level sampling data of the pulverized coal flow rate and the second-level sampling data of the air volume, a linear interpolation method is used to perform multi-resolution interpolation reconstruction on the low-frequency pulverized coal flow rate data. Based on the second-level time stamp, equally spaced interpolation data points are generated between two adjacent pulverized coal flow rate sampling points to form a continuous sequence that matches the time resolution of the air volume data. A phase calibration mechanism for the pulverized coal conveying cycle of the combustion system is introduced during the interpolation process, and the interpolation step size is dynamically adjusted according to the opening and closing states of the coal bunker feeding valve to avoid the accumulation of interpolation errors caused by periodic material interruption, and generate a second-level sequence of the pulverized coal flow rate with a unified time baseline.

[0049] After the interpolation reconstruction is completed, the pulverized coal flow and air volume data are aligned in time and space based on the dynamic time warping algorithm. According to the pulverized coal particle size distribution range and volatile content ratio in the coal quality characteristics, the slope constraint of the warping path is dynamically set. For high-volatile coal, the lateral offset restriction of the warping path is relaxed to match its rapid combustion characteristics; for large-particle pulverized coal, the longitudinal constraint of the warping path is increased to reflect its combustion lag effect. The optimal warping path is calculated through the dynamically adjusted path constraint conditions, and the pulverized coal flow and air volume data of different sampling frequencies are mapped to a unified space-time coordinate system to generate a time series feature matrix containing coal quality characteristics, air volume ratio and temperature distribution.

[0050] The aligned time series feature matrix is ​​processed by data standardization to normalize the coal powder flow, air volume and temperature parameters to the same dimensional range, eliminating the influence of sensor range differences on model input. The standardized feature matrix is ​​input into the construction process of the dynamic coupling model as the joint input data of the gradient boosting decision tree and the long short-term memory network. The feature matrix triggers the data validity verification mechanism during the model construction process, eliminates abnormal data points due to interpolation errors exceeding the threshold range, and maintains the temporal and spatial consistency of the model input data.

[0051] Specifically, in the data-driven boiler combustion optimization control method of the present invention, the sensitivity backtracking analysis includes: quantifying the gradient effect of the burner inclination angle on the uniformity of the temperature field, generating a priority ranking of the inclination angle compensation amount; Combined with the coupling relationship between the damper opening adjustment amount and the steam pressure constraint in the dynamic decomposition multi-objective optimization algorithm, a time series of multi-level control instructions is generated, and the time series is sent to the distributed control system.

[0052] In the sensitivity retrospective analysis process described in the present invention, firstly, based on the correspondence between the burner inclination adjustment record and the temperature field uniformity index in the historical operation data, the gradient influence coefficient of different inclination compensation amounts on the change of the temperature field standard deviation is calculated. The adjustment frequency of each inclination compensation direction and amplitude is counted through a sliding time window, and a priority ranking list of inclination compensation amounts is generated in combination with the ranking results of the temperature field uniformity improvement efficiency. The priority ranking is arranged in descending order according to the absolute value of the gradient influence coefficient, and the inclination compensation instructions with high contribution to the improvement of the temperature field uniformity are executed first.

[0053] After completing the priority sorting, a wind-coal ratio-steam pressure transfer function model is constructed by combining the coupling relationship between the damper opening adjustment amount and the steam pressure constraint in the multi-objective optimization algorithm with dynamic decomposition. By analyzing the influence amplitude of the rough adjustment and fine adjustment commands of the damper opening on the main steam pressure fluctuation, the maximum allowable adjustment amount threshold under the steam pressure constraint condition is calculated. Based on the threshold, the damper opening adjustment command is segmented by amplitude, and a multi-level control command time series including a rough adjustment stage, a pressure stabilization stage, and a fine adjustment stage is generated.

[0054] The multi-level control command time series is converted into a protocol format that can be parsed by the distributed control system through an instruction encoder. The damper opening adjustment command in the rough adjustment stage is preferentially issued. After the steam pressure fluctuation enters the stable interval, the inclination compensation command and the damper fine adjustment command are sequentially executed according to the time series. During the execution process, the steam pressure change rate and the temperature field uniformity monitoring data are synchronously collected and real-time fed back to the dynamic coupling model for online verification of the combustion state prediction value, forming a closed-loop control link for adjustment command issuance, execution effect verification, and model parameter update.

[0055] Specifically, for the data-driven boiler combustion optimization control method of the present invention, the closed-loop control link includes: feeding back the actual combustion data to the dynamic coupling model to trigger the incremental learning update of the splitting node threshold of the gradient boosting decision tree and the weights of the long short-term memory network; feeding back the change amount of the emission concentration to the multi-objective optimization algorithm with dynamic decomposition to dynamically adjust the weight distribution of sub-problems; Deleting expired working condition data through a data rolling update mechanism, and maintaining the matching of the dynamic coupling model with the real-time working condition based on the updated training sample set.

[0056] During the operation of the closed-loop control link of the present invention, the measured values of the furnace temperature field distribution, flue gas emission concentration, and steam pressure parameters in the actual combustion data are real-time fed back to the dynamic coupling model through the data bus. The data triggers the incremental learning update process of the splitting node threshold of the gradient boosting decision tree. Based on a sliding time window, the latest working condition data is intercepted, and an incremental feature importance evaluation algorithm is used to calculate the change in the correlation degree between the pulverized coal particle size distribution and the combustion efficiency sensitive factor, dynamically adjusting the feature selection priority of the splitting node; simultaneously triggering the online fine-tuning of the weights of the long short-term memory network, calculating the gradient direction of the hidden layer based on the temperature field prediction residual sequence, and adjusting the weight update step size through an adaptive momentum factor to suppress the network oscillation caused by sudden changes in the combustion working condition and improve the temperature field prediction accuracy.

[0057] The change in emission concentration is converted into the deviation rate of NOx and SO2 concentrations through normalization processing and fed back to the weight allocation module of the multi-objective optimization algorithm with dynamic decomposition. According to the comparison result between the deviation rate and the preset threshold, the weight coefficient allocation ratio of the sub-problems of combustion efficiency optimization, pollutant reduction, and temperature field uniformity is dynamically adjusted. For pollutant indicators with high deviation rates, the weight proportion of the corresponding sub-problems is increased; combined with the sulfur content and nitrogen content parameters in the real-time coal quality characteristics, the coal type adaptation coefficient of the weight distribution matrix is reconstructed to generate a multi-objective optimization weight configuration matching the current combustion state.

[0058] Based on the preset time window length, the data rolling update mechanism executes the first-in, first-out strategy on the training sample set, deleting expired operating condition data outside the window range. Retain the real-time data samples containing coal type switching marks and load fluctuation intervals, and balance the distribution ratio of new and old data through local resampling. The updated training sample set is input into the incremental training module of the gradient boosting decision tree and the long short-term memory network, and the small batch data chunk loading technology is used to update the model parameters segment by segment according to the time series. The incremental training results are synchronized to the online model through the version control protocol. When the prediction error caused by the new parameters exceeds the tolerance, it automatically rolls back to the historical stable version to maintain the matching of the dynamic coupling model and the real-time operating conditions.

[0059] In the specific implementation of the present invention, first, a temperature sensor array, a wind and flue gas pressure transmitter, a flue gas composition analyzer, and an on-line coal quality detection device are deployed in the boiler combustion system to form a multi-source data acquisition network. The temperature sensors are installed at the four corners and the middle of the furnace in a cross shape, and the temperature field distribution data is collected every 10 seconds; the air volume measurement device uses a Venturi tube combined with a differential pressure transmitter to collect the primary air and secondary air flow rates at a second-level frequency; the coal quality detector combines a belt scale and a laser particle size analyzer to output a histogram of the pulverized coal particle size distribution every minute. The collected original data is transmitted to the edge computing node through the industrial Ethernet and undergoes mean filtering noise reduction processing. The filtering window length is set to 3 times the pulverized coal transportation cycle to eliminate the pulse interference of the fly ash carbon content signal.

[0060] In the data preprocessing stage, principal component analysis is performed on the denoised air-coal ratio and oxygen concentration parameters, and the first 3 principal components with a cumulative variance contribution rate of 85% are retained, and the excess air coefficient and the adjustment sensitivity of the secondary air damper are extracted as key sensitive factors. Aiming at the time series alignment problem between the minute-level pulverized coal flow rate data and the second-level air volume data, the dynamic time warping algorithm is used to set a path constraint with a horizontal offset of no more than 5 seconds, and cubic spline interpolation is performed on the air volume data to generate a standardized feature matrix with a unified time baseline. This matrix contains 28-dimensional features such as coal quality characteristics, air volume ratio, and temperature field standard deviation, which are used as the input of the dynamic coupling model.

[0061] When constructing the dynamic coupling model, the maximum depth of the gradient boosting decision tree is set to 7 layers. Based on the Gini coefficient splitting criterion, the influence weights of each bin interval of the pulverized coal particle size distribution on the air-coal ratio are analyzed, and a feature importance ranking table is generated. The long short-term memory network adopts a 3-layer and 128-unit structure. When embedding in the second layer, a time-delay compensation layer is embedded. According to the difference matrix between the damper adjustment timestamp and the temperature response timestamp, the forgetting gate weight coefficient is dynamically adjusted to compensate for the combustion thermodynamics inertia delay. When the coal quality detector recognizes a coal type switch or a load fluctuation exceeding 15%, sliding window incremental training is triggered, the operating conditions data of the most recent 2 hours are retained, and the Adam optimizer is used to update the model parameters at a learning rate of 0.001.

[0062] When the multi-objective optimization module runs, based on the volatile content in the real-time coal quality characteristics, the weight of the combustion efficiency sub-problem is set in the dynamic range of 0 - 0.6, and the weight of the ash softening temperature parameter determining the temperature field uniformity fluctuates between 0.2 - 0.4. The three-objective problem is decomposed into 12 scalar sub-problems by the Chebyshev decomposition method, and the Pareto front solution set is generated through neighborhood search. The fuzzy membership function sets the steam pressure fluctuation threshold of ±0.5 MPa. After screening the feasible solution domain, combined with the historical operation data, a burner tilt - temperature field gradient influence matrix is constructed, and a hierarchical control instruction set with priority ranking is generated.

[0063] In the instruction execution stage, the distributed control system decomposes the coarse adjustment instruction of the secondary damper into 5% opening step lengths and issues it in 3 times, with an interval of 30 seconds each time; the fine adjustment instruction is issued at a second-level frequency with an accuracy of 0.5%. The burner tilt compensation adopts PID closed-loop control, and each 15 degrees of the stepper motor is divided into an adjustment interval, and the change of the temperature field standard deviation is monitored synchronously. The actual combustion data after execution is uploaded through the OPC protocol. When the thermal efficiency deviation exceeds 2% continuously for 3 times, the model self-correction process is triggered: the long short-term memory network updates the weights using the RMSProp optimizer, the gradient boosting decision tree adjusts the splitting threshold to more than 0.01 of the information gain rate, and the updated parameters are injected into the online model after being verified by MD5.

[0064] For data closed-loop management, a circular buffer is used to store the operating conditions data of the most recent 72 hours, and data rolling update is performed every 15 minutes. The expired data is marked and automatically archived. The incremental training module loads data in 32 batches in the mini-batch manner, and the learning rate decay coefficient is set to 0.95 to prevent model oscillation. The weight dynamic adjustment module calculates the emission concentration deviation rate every hour. When the NOx exceedance duration exceeds 10 minutes, the weight of the pollutant sub-problem is increased by 0.1, and the weight fluctuation is smoothed through the moving average algorithm to maintain the stability of multi-objective optimization.

[0065] The present invention solves the problem of inaccurate multi - physical - field coupling modeling by constructing a dynamic coupling model and a multi - objective optimization cooperation mechanism. Based on the gradient - boosting decision tree, the non - linear correlation between pulverized coal particle size distribution and air - coal ratio is analyzed, and key control variables are screened to generate a list of sensitive factor weights; the long - short - term memory network combined with a time - delay compensation layer learns the temporal dependence relationship between oxygen concentration field and furnace temperature, dynamically matches the minute - level delay between damper adjustment and temperature response, and accurately characterizes the multi - physical - field coupling characteristics of combustion efficiency, pollutant concentration, and temperature field uniformity. Through the sliding - window incremental training mechanism, model parameters are updated when the coal type changes or the load fluctuation exceeds the threshold, eliminating the mismatch problem of the static model for dynamic working conditions.

[0066] Aiming at the lack of temporal generalization ability, the present invention uses the dynamic time warping algorithm to align the multi - resolution temporal features of pulverized coal flow rate and air volume data, generating a spatio - temporally consistent standardized feature matrix. The long - short - term memory network is embedded with a time - delay compensation layer, and the weight of the network forgetting gate is corrected by adjusting the difference matrix of action and response timestamps, capturing the inertial delay effect of the combustion thermodynamics process. The data rolling - update mechanism combined with a closed - loop feedback link collects actual combustion data in real - time to trigger model self - calibration, suppresses the decay of prediction performance caused by data distribution drift, and enhances the adaptability of the model to time - varying working conditions.

[0067] The closed - loop control link realizes dynamic balanced control through the collaborative iteration of the multi - objective optimization algorithm and execution feedback. The Chebyshev decomposition method is used to dynamically adjust the weights of sub - problems of combustion efficiency, pollutant emission reduction, and temperature field uniformity, generating a Pareto - optimal solution set; sensitivity back - tracking analysis quantifies the influence gradient of burner tilt on the temperature field, and combined with the fuzzy membership function, a feasible solution domain is screened to generate a hierarchical control instruction set. The feedback of actual combustion data triggers incremental learning and weight adjustment, forming a closed - loop of "data acquisition - model prediction - instruction optimization - execution feedback" to maintain the continuous adaptation of the multi - physical - field coupling model to dynamic working conditions.

Claims

1. A data-driven boiler combustion optimization control method, characterized in that: include: Acquire multi-source sensor data of the boiler combustion system, wherein the multi-source sensor data includes temperature field distribution, air and smoke system pressure, flue gas component concentration, air volume parameters and coal quality characteristics; The multi-source sensor data is subjected to fusion processing, including eliminating noise through sliding window mean filtering, extracting sensitive factors through principal component analysis, and aligning time series data through dynamic time warping to generate a standardized feature matrix containing combustion efficiency sensitive factors; Building a dynamic coupling model based on the standardized feature matrix, modeling the standardized feature matrix through a hybrid machine learning algorithm, the hybrid machine learning algorithm including a gradient boosting decision tree and a long short-term memory network, to predict combustion state parameters, the combustion state parameters including combustion efficiency, pollutant concentration, and temperature field uniformity; Based on the combustion state parameters output by the dynamic coupling model, a dynamic decomposition multi-objective optimization algorithm is used to coordinately optimize the combustion efficiency, pollutant concentration and temperature field uniformity, and generate a hierarchical control instruction set including the air-coal ratio adjustment amount, the air door opening correction amount and the burner inclination compensation amount; The hierarchical control instruction set is sent to the distributed control system to drive the secondary air door and the burner tilt actuator to adjust, and the actual combustion data after execution is collected in real time; Based on the deviation between the actual combustion data and the predicted combustion state parameters output by the dynamic coupling model, the self-correction of the parameters of the dynamic coupling model and the dynamic adjustment of the weights of the multi-objective optimization algorithm are triggered, the split node threshold of the gradient boosting decision tree and the hidden layer weights of the long short-term memory network are updated, and the adjusted parameters are fed back to the dynamic coupling model and the multi-objective optimization algorithm to form a closed-loop control link.

2. The data-driven boiler combustion optimization control method according to claim 1 is characterized in that: The fusion processing of multi-source sensor data includes: The fly ash carbon content signal is denoised by sliding window mean filtering to generate denoised time series data; Performing principal component analysis on the air-coal ratio and oxygen concentration parameters in the denoised time series data, extracting combustion efficiency sensitive factors, and generating low-dimensional feature vectors; A dynamic time warping algorithm is used to align the minute-level sampling data of coal powder flow and the second-level sampling data of air volume to generate a standardized feature matrix containing coal quality characteristics, air volume ratio and temperature distribution, and the standardized feature matrix is ​​input into the construction process of the dynamic coupling model.

3. The data-driven boiler combustion optimization control method according to claim 1 is characterized in that: The construction of the dynamic coupling model includes: analyzing the nonlinear relationship between the coal powder particle size distribution and the air-coal ratio in the standardized feature matrix based on the gradient boosting decision tree algorithm, screening key control variables, and generating a sensitive factor weight list; The oxygen concentration field and the furnace temperature in the standardized feature matrix are subjected to time-series association learning through a long short-term memory network, and the delay time of the air door adjustment and the temperature response is matched in combination with a time-delay compensation layer to output a combustion state prediction value; When it is detected that the coal type is switched or the load fluctuation amplitude exceeds a preset threshold, the sliding window incremental training is triggered to update the split node threshold of the gradient boosting decision tree and the hidden layer weight of the long short-term memory network.

4. The data-driven boiler combustion optimization control method according to claim 3 is characterized in that: The multi-objective optimization algorithm of dynamic decomposition includes: Based on the combustion state prediction value output by the dynamic coupling model, the multi-objective optimization problem of combustion efficiency, pollutant concentration and temperature field uniformity is decomposed into scalar sub-problems by using Chebyshev decomposition method, and the weight distribution of the sub-problems is dynamically adjusted according to the real-time coal quality characteristics; Optimizing the reference point according to the deviation between the smoke emission data and the predicted value of the combustion efficiency to generate a dynamic Pareto frontier; Define the air-coal ratio and the secondary air door opening as adjacent sub-problems, and generate the Pareto optimal solution set through local search; The feasible solution domain that satisfies the steam pressure constraint is screened based on the fuzzy membership function. The gradient effect of the burner inclination angle on the temperature field uniformity is quantified by combining the sensitivity backtracking analysis. A hierarchical control instruction set including the air-coal ratio adjustment amount, the air door opening correction amount and the burner inclination angle compensation amount is generated.

5. The data-driven boiler combustion optimization control method according to claim 1 is characterized in that: The model self-correction includes: When the deviation between the actual thermal efficiency and the predicted value output by the dynamic coupling model exceeds a preset threshold, an adaptive moment estimation algorithm is used to update the weight parameters of the long short-term memory network; The splitting threshold of the gradient boosting decision tree is adjusted according to the change of pollutant concentration in the actual combustion data, updated hybrid model parameters are generated, and the updated parameters are fed back to the dynamic coupling model.

6. The data-driven boiler combustion optimization control method according to claim 1 is characterized in that: The dynamic weight adjustment includes: based on the actual emission concentration and the change in combustion efficiency, reversely correcting the weight allocation of sub-problems generated based on real-time coal quality characteristics in the multi-objective optimization algorithm of dynamic decomposition; The operating condition data within the preset time window is retained, the training sample set of the dynamic coupling model is updated in a rolling manner to suppress model degradation, and the updated sample set is input into the incremental training process of the gradient boosting decision tree and the long short-term memory network.

7. The data-driven boiler combustion optimization control method according to claim 1 is characterized in that: The execution of the hierarchical control instruction set includes: The secondary air valve coarse adjustment command and fine adjustment command generated by the multi-objective optimization algorithm based on dynamic decomposition are sent to the actuator in stages through the distributed control system; The actuator is driven to perform angle correction according to the burner inclination compensation amount in the hierarchical control instruction set, and the temperature field distribution uniformity monitoring data is synchronously fed back to the dynamic coupling model.

8. The data-driven boiler combustion optimization control method according to claim 3 is characterized in that: The timing alignment process further comprises: Perform multi-resolution interpolation on the minute-level sampling data of pulverized coal flow and the second-level sampling data of air volume to generate a time series feature matrix with a unified time baseline; The warping path constraint conditions of the dynamic time warping algorithm are dynamically adjusted based on the coal quality characteristics, the time-space alignment accuracy is optimized, and the aligned time series feature matrix is ​​input into the construction process of the dynamic coupling model.

9. The data-driven boiler combustion optimization control method according to claim 4 is characterized in that: The sensitivity backtracking analysis includes: quantifying the gradient effect of the burner inclination angle on the uniformity of the temperature field and generating a priority ranking of the inclination angle compensation amount; Combined with the coupling relationship between the damper opening adjustment amount and the steam pressure constraint in the dynamic decomposition multi-objective optimization algorithm, a time series of multi-level control instructions is generated, and the time series is sent to the distributed control system.

10. The data-driven boiler combustion optimization control method according to claim 1, characterized in that: The closed-loop control link includes: feeding back actual combustion data to the dynamic coupling model to trigger the incremental learning update of the split node threshold of the gradient boosting decision tree and the long short-term memory network weight; Feeding back the emission concentration change to the dynamically decomposed multi-objective optimization algorithm to dynamically adjust the weight distribution of sub-problems; The expired operating condition data is deleted through a data rolling update mechanism, and the matching of the dynamic coupling model with the real-time operating condition is maintained based on the updated training sample set.

Images