Coal-fired unit combustion NO x Method for identifying and predicting generation patterns

Abstract

The application discloses a kind of coal-fired generating unit combustion NO x Generation law identification and prediction method, comprising the following steps: S1: data acquisition and effective sample construction;S2: boiler working condition state identification and load segmentation division;S3: NO x Generation law modeling and prediction model training;S4: NO x Fluctuation quantification and fluctuation attribution analysis;S5: unit adaptive rule base construction and incremental update;S6: output control boundary data.

Description

Technical Field

[0001] This invention relates to the field of environmental protection and combustion optimization technology for coal-fired power generating units, and more specifically, to a method for optimizing NO combustion in coal-fired power generating units. x A method for generating patterns and predicting trends is used. This method identifies boiler outlet NO based on historical and real-time operational data. x The generation patterns, influencing factors, and fluctuation characteristics under steady-state and variable load conditions were identified, and the boiler outlet NO was accurately predicted. x concentration. Background Technology

[0002] Due to boiler outlet NO x NO concentration is influenced by a combination of factors, including load, excess air (oxygen), air and coal distribution to each mill / feeder, and coal quality. Traditional empirical curves are insufficient to cover all operating conditions. Due to combustion zoning, NO concentration varies between the A and B sides of the SCR inlet. x Although highly correlated, systematic biases exist. NO₂ increases during boiler steady-state and variable load processes. x The fluctuation mechanisms differ: steady-state fluctuations are mainly characterized by minute-level oxygen fluctuations, while variable load fluctuations are mainly characterized by combustion condition reconfiguration. The dominant factors may differ for different units (boiler type, mill configuration, and measuring point layout), resulting in poor generalization ability of fixed formulas. Summary of the Invention

[0003] This invention provides a method for burning NO in coal-fired power units. x A method for generating patterns for identification and prediction is proposed to address the technical problems existing in the aforementioned technologies.

[0004] To achieve the above objectives, the present invention provides a method for burning NO in coal-fired power units. x The method for generating pattern recognition and prediction includes the following steps: S1: Data Acquisition and Effective Sample Construction S1.1: Collect minute-level time-series operating data from the DCS of the coal-fired power unit. The minute-level time-series operating data from the DCS of the coal-fired power unit includes the NO at the A-side inlet of the SCR unit. x Concentration of CNO x -A, SCR device B-side inlet NO x Concentration of CNO x -B; S1.2: Data cleaning and processing: Remove invalid data records corresponding to unit load below the preset operating threshold, boiler shutdown, and furnace purging conditions; S1.3: Construct a unified SCR entry point NO x Characteristic quantity: Calculate NO on both sides A and B of the SCR device. x Concentration correlation coefficient corr(CNO) x -A,CNO x-B) If the correlation coefficient is greater than the preset correlation threshold, then the average concentration of both sides is taken as the comprehensive NO concentration. x Characteristic CNO x =(CNO x -A+CNO x -B) / 2; If the correlation coefficient is not greater than the preset correlation threshold, then for the NO on both sides of SCR device A and B... x Model them independently; S2: Boiler Operating Condition Identification and Load Segmentation S2.1: The rate of change of computer group load dP / dt = |P(t) - P(t-1)| / Δt, and the standard deviation of computer group load σ is used in a 30-minute rolling window. P Where P(t) is the real-time active load of the unit at time t, and P(t-1) is the real-time active load of the unit at time t-1; S2.2: Based on the standard deviation of the load σ P The load change rate and duration are used to determine the operating condition type. Steady-state load condition: Load standard deviation σ P ≤Stable load threshold σ stable And the continuous operating time is greater than or equal to the steady-state duration threshold T. stable ; Variable load condition: Load change rate dP / dt ≥ Variable load rate threshold v change And the continuous operating time is greater than or equal to the variable load duration threshold T. change ; The remaining data samples are classified as transitional operating conditions; S2.3: The load quartile method is used to divide the entire operating condition into three load platforms: low load, medium load, and high load. S3: Working Condition NO x Generation pattern modeling and prediction model training S3.1: For the samples processed in step S1.2, Pearson and Spearman correlation coefficients are used to initially screen influencing factors. Then, random forest and gradient boosting tree algorithms are used to calculate the impact of each operating parameter on NO. x The characteristic importance weight of concentration; S3.2: A standardized partial regression model is constructed using multiple linear regression and ridge regression to obtain the impact of each influencing factor on NO. x The magnitude and direction of the independent contribution of concentration are characterized by the standardized partial regression model determination coefficient R², which represents the interpretability of the model. S3.3: Conduct partial correlation analysis of control load variables to eliminate the interference of unit load confounding effects on the judgment of factor correlation; S3.4: Repeat steps S3.1 to S3.3 for different load segments and different operating conditions to generate an operating condition-factor correlation priority matrix M. Matrix M stores the correlation coefficient, standardization coefficient, feature importance weight, and positive and negative direction of the corresponding influencing factors under each operating condition. S3.5: Temporal segmentation of datasets and construction of a two-layer NO x Predictive model: S3.5.1: Divide the cleaned samples into a 70% training set and a 30% validation set in chronological order. Randomly shuffling the samples is prohibited to avoid leakage of temporal information. The training set is used to construct the association model and the association priority matrix M, and the validation set is used to evaluate the generalization performance of the model. S3.5.2: Construct a two-layer coupled prediction model, including a horizontal correlation model for long-term combustion optimization and a fluctuation model for minute-level stable control; the model inputs are oxygen content, unit load, coal quantity of each pulverizer, total air volume, primary air-to-coal ratio, and fuel ratio, and the output is SCR inlet NO. x Concentration prediction values; Random forest or ridge regression algorithms are selected for modeling, and validation set R², mean absolute error MAE, and root mean square error RMSE are used as the evaluation indicators for model optimization. S4: Working Condition NO x Quantitative analysis of volatility patterns and volatility attribution analysis S4.1: Calculation of quantitative indicators for steady-state operating condition fluctuations: NO within the steady-state range x Concentration standard deviation σ_NOx stable Concentration range R_NOx stable minute-level NO x Median of absolute change |ΔNOx|; S4.2: Calculation of quantitative indicators for fluctuations under variable load conditions: NO throughout the entire variable load process x Extreme concentration difference, start and end of operating conditions NO x Change, maximum NO per minute x Jump amplitude | ΔNO x |_max; S4.3: Steady-state minute-level fluctuation attribution: Calculating single-minute NO x Change ΔNO x The minute-by-minute changes in various operating parameters Δx i The correlation coefficient identifies the causes of NO x The core control parameters for short-term fluctuations; S5: Unit Adaptive Rule Library Construction and Incremental Update S5.1: Dedicated NO units are independently constructed for each individual unit. x Generate a pattern library to store the factor correlation matrix, fluctuation quantification index, and trained two-layer prediction model for each load segment and operating condition of the corresponding unit. S5.2: Real-time acquisition of newly added operating data of the unit, rolling update of the dataset and recalculation of model fit R² and ranking of core influencing factors; when the change in R² or the change in the ranking of TOP core factors exceeds the preset update threshold, the incremental update of the unit pattern library is automatically triggered. S6: Output control boundary data The trained two-layer prediction model is embedded into the unit's DCS, APS automatic power generation control system, or boiler environmental optimization system to output the SCR inlet NO under different operating conditions in real time. x Concentration prediction boundary values ​​are used for closed-loop optimization and control of combustion air and coal distribution.

[0005] In one embodiment of the present invention, preferably, in step S1.1, the DCS minute-level time-series operation data of the coal-fired unit further includes the unit load, oxygen content at both inlets of the SCR, total furnace air volume, total coal volume, instantaneous coal volume of each coal feeder, primary inlet air volume of each coal mill, main steam flow rate, total feedwater flow rate, and coal quality test data entering the furnace, wherein the coal quality test data entering the furnace includes volatile matter in the coal, fuel nitrogen content, and fixed carbon index.

[0006] In one embodiment of the present invention, preferably, the threshold values ​​in step S2.2 are: stable load threshold σ stable =1.0MW, steady-state duration threshold T stable =30min, variable load rate threshold v change =2MW / min, variable load duration threshold T change =3min.

[0007] In one embodiment of the present invention, preferably, the load segmentation criteria in step S2.3 are: low load < 40% of rated load, medium load 40%~70% of rated load, and high load > 70% of rated load.

[0008] In one embodiment of the present invention, preferably, after the steady-state fluctuation attribution in step S4.3 is completed, the identified main cause of short-term fluctuations is used as the priority operating variable for the fine-tuning of boiler steady-state control, and NO is suppressed by fine-tuning this variable. x Minute-level fluctuations.

[0009] In one embodiment of the present invention, preferably, under variable load conditions, the core impact parameter of the variable load impact is determined based on the correlation priority matrix M, and the parameter is intervened in advance during the unit's load increase and decrease process to reduce NO during the variable load phase. x Maximum jump amplitude.

[0010] In one embodiment of the present invention, preferably, double-layer NO x The prediction model has a coefficient of determination R² ≥ 0.95 on the validation set and a mean absolute error MAE ≤ 6 mg / m³.

[0011] In one embodiment of the present invention, preferably, the incremental update triggering condition in step S5.2 is: the fluctuation amplitude of the newly added data window model R² exceeds 0.05, or the top 3 core influencing factors are replaced.

[0012] In one embodiment of the present invention, preferably, in step S1.3, the preset correlation threshold is 0.85, when NO on sides A and B... x When the correlation coefficient is ≤0.85, construct independent two-layer prediction models for side A and side B respectively, and output the NO values ​​for both sides separately. x Predict concentration to achieve precise air distribution optimization on each side.

[0013] In one embodiment of the present invention, preferably, the NO output by the prediction model... x The boundary concentration values ​​are sent to the SCR ammonia injection optimization control system to achieve coordinated operation between the boiler combustion front-end control and the tail-end denitrification ammonia injection.

[0014] The present invention has the following beneficial technical effects: This invention addresses the NOx issue in existing coal-fired power units. x The control technology suffers from poor adaptability to all operating conditions and NO in steady-state / variable load conditions. x Addressing pain points such as undifferentiated fluctuation mechanisms, inability to quantify influencing factors, weak model generalization ability, and difficulty in supporting adaptive control of units, this paper provides a data-driven, condition-specific adaptive NO model for coal-fired units. x Compared to traditional empirical curves, single global regression models, and prediction schemes without working condition differentiation, the generation pattern identification and prediction method has the following multi-dimensional outstanding technical advantages: I. The data preprocessing scheme is more closely aligned with the actual conditions at the boiler site. x The characterization results are accurate and reliable. Based on minute-level, full-dimensional runtime sequence data from the unit's DCS, covering all combustion-related variables such as load, coal and air distribution, oxygen content, coal quality, and steam-water parameters, a complete NOx analysis is performed. x The generated multi-factor coupling mechanism compensates for the information loss caused by traditional modeling which only selects a small number of parameters.

[0015] The system automatically cleans invalid samples through dual filtering based on load threshold and shutdown purging conditions, avoiding interference from abnormal operating conditions with model fitting accuracy; it also addresses SCR bilateral NO. x To address the common industry-wide issue of systematic bias, a new logic for determining the correlation between two-sided concentrations is added: when the two-sided correlation is higher than 0.85, the mean value is used to represent the overall inlet NO. x When the correlation is insufficient, the model is independently modeled on both sides, which balances the simplicity of modeling with the ability to accurately represent the one-sided bias. This solves the problem of prediction distortion caused by the traditional approach of directly taking the mean and ignoring the A / B side bias.

[0016] By introducing coal quality test data such as volatile matter, fuel nitrogen, and fixed carbon as fixed inputs to the model, the impact of frequent fluctuations in coal quality on NO can be eliminated. x The interference with the prediction results significantly improves the prediction stability of units with varying coal types.

[0017] II. A refined chemical condition classification mechanism to analyze the differentiated patterns of steady-state / variable load / transient states. It accurately distinguishes steady-state, variable load, and transient operating conditions using three indicators: load change rate, 30-minute rolling load standard deviation, and duration. Simultaneously, it employs the quartile method to classify low / medium / high load platforms, overcoming the limitations of existing technologies that do not differentiate between operating conditions and cannot distinguish NO. x Limitations of the wave mechanism: steady-state NO x Fluctuations are dominated by minute-level fine-tuning of oxygen levels, and variable load NO x The mutation originates from the overall reconstruction of combustion conditions. This invention can identify the core influencing factors for the two types of conditions respectively.

[0018] Real-world case studies validate the value of differentiating operating conditions: steady-state samples account for 52.2%, while variable load samples account for only 4%, and the dominant control parameters for the two are completely different. In steady-state conditions, the core parameters are the primary air supply, coal feed rate, and oxygen supply of each mill, while in variable load conditions, the primary influencing factor is the increase or decrease in coal feed rate of the coal feeder during load increases or decreases. A unified global model would confuse the action patterns of the two types of operating conditions. This invention, which models different operating conditions, can accurately locate the control levers in different operating ranges.

[0019] III. Multi-algorithm fusion to quantify influencing factors, enabling quantitative analysis of NO x Generation mechanism to guide on-site optimization A multi-level analysis system was employed, using correlation coefficients for initial screening, random forest feature importance, ridge regression standardized partial coefficients, and control load partial correlation, to identify the relationship between each factor and NO. x The correlation direction is determined, and the independent contribution weight is quantified. At the same time, the load confounding effect is eliminated, which solves the problem that traditional experience can only qualitatively judge the "impact of air volume on NO". x The shortcomings of being unable to quantify the adjustment range; generating a standardized operating condition-factor priority matrix, intuitively outputting the TOP core control parameters under each load and operating condition, providing operators with clear operating priorities.

[0020] The time-series dataset is split (70% for training and 30% for validation), abandoning the random shuffling method to prevent leakage of time-series information. The model's generalization ability is more in line with the real continuous operation scenario of the unit, avoiding the overfitting and accuracy degradation problems caused by traditional random dataset splitting.

[0021] An innovative dual-layer coupled prediction model was constructed: a horizontal model adapted to long-term combustion benchmark optimization, and a fluctuation model specifically designed to capture minute-level NO. xDynamic changes, balancing macroscopic steady-state indicator control with instantaneous fluctuation suppression; such as Figure 2 As shown, the experimental validation model has a validation set determination coefficient R² of 0.95 and a MAE of only 5.9 mg / m³, with a prediction accuracy far exceeding that of the traditional single linear model (the global unified model has an R² of only 0.61).

[0022] IV. Quantification of NO by Working Condition x Fluctuations and attribution of fluctuations, targeted suppression of NO x Excessive fluctuations Establish two complete quantitative index systems for fluctuations: steady-state and variable load. The steady-state segment calculates NO. x Standard deviation, range, and median minute fluctuation; statistical analysis of the entire range and maximum instantaneous jump amplitude during variable load periods, allowing for intuitive quantification of NO under different operating conditions. x The degree of fluctuation allows for precise identification of load ranges prone to over-discharge and severe fluctuations.

[0023] NO under steady-state conditions x Incremental attribution analysis was performed on minute-level fluctuations, using ΔNO x The correlation with each parameter Δx identifies the root cause of short-term fluctuations (actual measurements show that minute fluctuations in oxygen content are consistent with steady-state NO). x (The primary cause of fluctuations) On-site, this main factor can be prioritized for fine-tuning to achieve rapid stabilization of nitrogen oxides; under variable load conditions, early intervention based on the factor matrix guides the coal feeder's coal distribution, significantly reducing NO during load increases and decreases. x Instantaneous peak value reduces the risk of instantaneous exceedance of SCR input limits.

[0024] V. Unit self-adaptation + incremental update rule library, adaptable to multiple boiler types and long-term variable operating conditions. Establish an independent and dedicated NO for each unit x Generate a rule library, and store correlation matrices, fluctuation indicators, and two-layer prediction models for each load segment and operating condition in a differentiated manner. This solves the problem of poor generalization of existing fixed formulas and general models to different furnace types and mill configurations, and achieves exclusive adaptation for a single unit.

[0025] An automatic incremental update mechanism is set up: new operating data is collected on a rolling basis. When the model fit R² fluctuates or the ranking of core influencing factors changes beyond the threshold, the pattern library is automatically updated. There is no need for manual remodeling. It adapts to changes in combustion characteristics caused by aging combustion equipment, coal blending scheme adjustments, and equipment upgrades after long-term operation of the unit. It maintains the accuracy of prediction and pattern recognition in the long term and reduces manual maintenance costs.

[0026] VI. The model can be directly integrated with the existing control system of the power plant to achieve combustion-denitrification synergistic closed-loop optimization. The trained prediction model can be seamlessly embedded into DCS, APS automatic power generation control systems, and boiler environmental optimization platforms, outputting NO under various operating conditions in real time. x Predicted NO concentration boundary values ​​directly guide the refined adjustment of air and coal distribution, reducing NO from the combustion source. x Generation amount.

[0027] Predictive output data can be synchronously sent to the SCR ammonia injection optimization system, realizing the linkage between front-end combustion control and tail-end denitrification ammonia injection: predicting NO in advance. x Peak levels can be reached by adjusting the ammonia injection rate in advance to reduce NO. x It can prevent instantaneous exceedances while avoiding excessive ammonia injection that could lead to ammonia escape and ammonium bisulfate clogging of the catalyst, thus simultaneously reducing denitrification operating costs and equipment wear.

[0028] VII. Comprehensive Project Implementation Benefits Environmental benefits: Under steady-state operating conditions, NO can be significantly reduced. x Minute-level fluctuation amplitude, suppressing instantaneous NO during load changes and load increases / decreases. x The peak value jump stabilizes the SCR inlet concentration and significantly reduces NO. x The frequency of emissions exceeding standards must meet the requirements for ultra-low emission control. Economic benefits: It provides precise guidance for optimizing air and coal distribution, while taking into account boiler combustion efficiency under the premise of nitrogen control; it also optimizes ammonia injection to reduce liquid ammonia consumption, reduce the risk of catalyst blockage and corrosion, and extend the service life of denitrification equipment. Operation and maintenance benefits: Full-process automated data processing, operating condition identification, and pattern updates eliminate the need for extensive manual statistics and curve fitting, reducing the workload of operators in data analysis; quantitative output of control priorities reduces reliance on operator experience, and can be quickly implemented for both new and old units. Attached Figure Description

[0029] The above and other objects, features, and advantages of the present invention will become more apparent from the more detailed description of the embodiments of the invention in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same parts or steps.

[0030] Figure 1 This invention relates to an embodiment of a coal-fired power unit for burning NO. x Flowchart of the method for generating pattern recognition and prediction; Figure 2 NO x A curve comparing predicted and actual concentration values. Detailed Implementation

[0031] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0032] Unless otherwise specifically stated, the technical or scientific terms used in the embodiments of this invention should be understood in their ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains. The terms "comprising" or "including," as used in the embodiments of this invention, do not limit the shapes, numbers, steps, actions, operations, components, elements, and / or groups thereof mentioned, nor do they exclude the appearance or addition of one or more other different shapes, numbers, steps, actions, operations, components, elements, and / or groups thereof, or the inclusion of these.

[0033] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention. It should also be understood that, for ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale, and techniques, methods, and apparatus known to those skilled in the art may not be discussed in detail; however, where appropriate, the illustrated techniques, methods, and apparatus should be considered part of the specification. In all the examples shown and discussed herein, any other specific example may have different values. It should be noted that similar symbols and letters in the following figures denote similar items; therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.

[0034] In the description of the embodiments of the present invention, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In the embodiments of the present invention, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in a suitable manner in any one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in the embodiments of the present invention, as well as the features of different embodiments or examples.

[0035] Hereinafter, exemplary embodiments according to the present invention will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments of the present invention. It should be understood that the present invention is not limited to the exemplary embodiments described herein.

[0036] This invention provides a data-driven, unit-adaptive NO combustion method. x Generate a pattern recognition method, and implement: (1) Automatically filter valid running samples from DCS data and construct entry point NO. x Characteristic quantity; (2) Identify the dominant factors and directions by load segment and operating condition (steady state / variable load); (3) Quantify steady-state fluctuations and the impact of variable loads, and locate operable control levers; (4) Output of boundary NO that can be embedded in DCS / APS or environmental optimization system x Concentration value.

[0037] First embodiment: Figure 1 This invention relates to an embodiment of a coal-fired power unit for burning NO. x A flowchart of the method for generating pattern recognition and prediction, such as... Figure 1 As shown, the coal-fired power unit provided by this invention burns NO x The method for generating pattern recognition and prediction includes the following steps: S1: Data Acquisition and Effective Sample Construction S1.1: Collect minute-level time-series operating data from the DCS of the coal-fired power unit. The minute-level time-series operating data from the DCS of the coal-fired power unit includes the NO at the A-side inlet of the SCR unit. x Concentration of CNO x -A, SCR device B-side inlet NO x Concentration of CNO x -B; In this embodiment, in step S1.1, the DCS minute-level time-series operation data of the coal-fired unit also includes the unit load, oxygen content at both inlets of the SCR, total furnace air volume, total coal volume, instantaneous coal volume of each coal feeder, primary inlet air volume of each coal mill, main steam flow rate, total feedwater flow rate, and coal quality test data entering the furnace. Among them, the coal quality test data entering the furnace includes volatile matter in coal, fuel nitrogen content, and fixed carbon index.

[0038] S1.2: Data cleaning and processing: Remove invalid data records corresponding to unit load below the preset operating threshold, boiler shutdown, and furnace purging conditions; S1.3: Construct a unified SCR entry point NO x Characteristic quantity: Calculate NO on both sides A and B of the SCR device. x Concentration correlation coefficient corr(CNO) x -A,CNO x -B) If the correlation coefficient is greater than the preset correlation threshold, then the average concentration of both sides is taken as the comprehensive NO concentration. xCharacteristic CNO x =(CNO x -A+CNO x -B) / 2; If the correlation coefficient is not greater than the preset correlation threshold, then for the NO on both sides of SCR device A and B... x Model them independently; S2: Boiler Operating Condition Identification and Load Segmentation S2.1: The rate of change of computer group load dP / dt = |P(t) - P(t-1)| / Δt, and the standard deviation of computer group load σ is used in a 30-minute rolling window. P Where P(t) is the real-time active load of the unit at time t, and P(t-1) is the real-time active load of the unit at time t-1; S2.2: Based on the standard deviation of the load σ P The load change rate and duration are used to determine the operating condition type. Steady-state load condition: Load standard deviation σ P ≤Stable load threshold σ stable And the continuous operating time is greater than or equal to the steady-state duration threshold T. stable ; Variable load condition: Load change rate dP / dt ≥ Variable load rate threshold v change And the continuous operating time is greater than or equal to the variable load duration threshold T. change ; The remaining data samples are classified as transitional operating conditions; S2.3: The load quartile method is used to divide the entire operating condition into three load platforms: low load, medium load, and high load. S3: Working Condition NO x Generation pattern modeling and prediction model training S3.1: For the samples processed in step S1.2, Pearson and Spearman correlation coefficients are used to initially screen influencing factors. Then, random forest and gradient boosting tree algorithms are used to calculate the impact of each operating parameter on NO. x The characteristic importance weight of concentration; S3.2: A standardized partial regression model is constructed using multiple linear regression and ridge regression to obtain the impact of each influencing factor on NO. x The magnitude and direction of the independent contribution of concentration are characterized by the standardized partial regression model determination coefficient R², which represents the interpretability of the model. S3.3: Conduct partial correlation analysis of control load variables to eliminate the interference of unit load confounding effects on the judgment of factor correlation; S3.4: Repeat steps S3.1 to S3.3 for different load segments and different operating conditions to generate an operating condition-factor correlation priority matrix M. Matrix M stores the correlation coefficient, standardization coefficient, feature importance weight, and positive and negative direction of the corresponding influencing factors under each operating condition. S3.5: Temporal segmentation of datasets and construction of a two-layer NO x Predictive model: S3.5.1: Divide the cleaned samples into a 70% training set and a 30% validation set in chronological order. Randomly shuffling the samples is prohibited to avoid leakage of temporal information. The training set is used to construct the association model and the association priority matrix M, and the validation set is used to evaluate the generalization performance of the model. S3.5.2: Construct a two-layer coupled prediction model, including a horizontal correlation model (NO) for long-term combustion optimization. x (Baseline concentration - combustion impact factor) and fluctuation models for minute-level stable control (NO) x Change ΔNO x -Change in operating parameters Δx i The model inputs are oxygen content, unit load, coal quantity of each pulverizer, total air volume, primary air-to-coal ratio, and fuel ratio. The output is SCR inlet NO. x Concentration prediction values; Random forest or ridge regression algorithms are selected for modeling, and validation set R², mean absolute error MAE, and root mean square error RMSE are used as the evaluation indicators for model optimization. S4: Working Condition NO x Quantitative analysis of volatility patterns and attribution analysis of volatility S4.1: Calculation of quantitative indicators for steady-state operating condition fluctuations: NO within the steady-state range x Concentration standard deviation σ_NOx stable Concentration range R_NOx stable minute-level NO x Median of absolute change |ΔNOx|; S4.2: Calculation of quantitative indicators for fluctuations under variable load conditions: NO throughout the entire variable load process x Extreme concentration difference, start and end of operating conditions NO x Change, maximum NO per minute x Jump amplitude | ΔNO x |_max; S4.3: Steady-state minute-level fluctuation attribution: Calculating single-minute NO x Change ΔNO x The minute-by-minute changes in various operating parameters Δx i The correlation coefficient identifies the causes of NO x The core control parameters for short-term fluctuations; S5: Unit Adaptive Rule Library Construction and Incremental Update S5.1: Dedicated NO units are independently constructed for each individual unit. x Generate a pattern library to store the factor correlation matrix, fluctuation quantification index, and trained two-layer prediction model for each load segment and operating condition of the corresponding unit. S5.2: Real-time acquisition of newly added operating data of the unit, rolling update of the dataset and recalculation of model fit R² and ranking of core influencing factors; when the change in R² or the change in the ranking of TOP core factors exceeds the preset update threshold, the incremental update of the unit pattern library is automatically triggered. S6: Output control boundary data The trained two-layer prediction model is embedded into the unit's DCS, APS automatic power generation control system, or boiler environmental optimization system to output the SCR inlet NO under different operating conditions in real time. x Concentration prediction boundary values ​​are used for closed-loop optimization and control of combustion air and coal distribution.

[0039] In one embodiment of the present invention, preferably, the threshold values ​​in step S2.2 are: stable load threshold σ stable =1.0MW, steady-state duration threshold T stable =30min, variable load rate threshold v change =2MW / min, variable load duration threshold T change =3min.

[0040] In one embodiment of the present invention, preferably, the load segmentation criteria in step S2.3 are: low load < 40% of rated load, medium load 40%~70% of rated load, and high load > 70% of rated load.

[0041] In one embodiment of the present invention, preferably, after the steady-state fluctuation attribution in step S4.3 is completed, the identified main cause of short-term fluctuations is used as the priority operating variable for the fine-tuning of boiler steady-state control, and NO is suppressed by fine-tuning this variable. x Minute-level fluctuations.

[0042] In one embodiment of the present invention, preferably, under variable load conditions, the core impact parameter of the variable load impact is determined based on the correlation priority matrix M, and the parameter is intervened in advance during the unit's load increase and decrease process to reduce NO during the variable load phase. x Maximum jump amplitude.

[0043] In one embodiment of the present invention, preferably, double-layer NO x The prediction model has a coefficient of determination R² ≥ 0.95 on the validation set and a mean absolute error MAE ≤ 6 mg / m³.

[0044] In one embodiment of the present invention, preferably, the incremental update triggering condition in step S5.2 is: the fluctuation amplitude of the newly added data window model R² exceeds 0.05, or the top 3 core influencing factors are replaced.

[0045] In one embodiment of the present invention, preferably, in step S1.3, the preset correlation threshold is 0.85, when NO on sides A and B... xWhen the correlation coefficient is ≤0.85, construct independent two-layer prediction models for side A and side B respectively, and output the NO values ​​for both sides separately. x Predicting concentration enables precise air distribution optimization on each side.

[0046] In one embodiment of the present invention, preferably, the NO output by the prediction model... x The boundary concentration values ​​are sent to the SCR ammonia injection optimization control system to achieve coordinated operation between the boiler combustion front-end control and the tail-end denitrification ammonia injection.

[0047] Second embodiment: This example is for a 300MW coal-fired power unit NO. x Engineering applications of generation pattern identification and prediction.

[0048] 1. Basic Overview of the Generating Unit This implementation focuses on a 300MW subcritical coal-fired power plant in a domestic power plant, equipped with five HP-type medium-speed coal mills and dual-sided SCR denitrification devices. The unit's daily operating load ranges from 50 to 300MW, with significant fluctuations in the volatile matter content of the coal and the nitrogen content of the fuel. The NO at the SCR inlet on both sides of the unit... x There is a long-term, small, systematic deviation; NO during steady-state operation x Significant minute-level fluctuations are common during load increases and decreases, and NO is prone to occur. x For instantaneous peak values, conventional manual experience curves cannot be adapted to precise control under full load and all operating conditions.

[0049] 2. S1 Data Acquisition and Sample Construction Implementation Process Data collection: Minute-level time-series operating data is retrieved from the unit's DCS system, with a sampling interval of Δt = 1 minute. The collected parameters fully cover: unit active load P, SCR-A side inlet NO. x Concentration of CNO x -A、SCR-B Side Inlet NO x Concentration of CNO x - Oxygen content at the inlet of side B and A / B, total air volume in the furnace, total coal feed, instantaneous coal feed of each coal feeder (A / B / C / D / E), primary inlet air volume of each coal mill, main steam flow, total feedwater flow, and daily coal quality test data (volatile matter, fuel nitrogen, fixed carbon).

[0050] The total data collection time was approximately 216 hours, with a total of 12,961 original records.

[0051] Data cleaning: The load operating threshold was set at 50MW. Records of abnormal operating conditions such as low load start-up and shutdown, furnace purging, and coal mill start-up and shutdown purging were removed. After cleaning, all samples were valid data of normal unit operation under load, with a total of 12961 minutes of valid samples.

[0052] NO x Construction of comprehensive characterization parameters: Calculate NO on both sides A and B x Concentration correlation coefficient corr (CNO) x -A, CNO x -B) = 0.9721, which is greater than the preset threshold of 0.85. Therefore, the average of both sides is used as the comprehensive NO. x Characteristic quantity: , No separate modeling is required for each side, simplifying the calculation process.

[0053] 3. Implementation of S2 operating condition identification and load segmentation Basic parameter settings: Sampling interval Δt = 1 min; 30 min rolling window calculation of load standard deviation σ P ; Operating condition judgment threshold: σ stable =1.0MW, T stable =30min; v change =2MW / min, T change =3min.

[0054] The load is divided into three segments using the quartile method: low load <101MW, medium load 101~213MW, and high load >213MW.

[0055] Statistical results of working condition classification: For each of the 12,961 valid samples, the load change rate dP / dt and the 30-minute rolling load standard deviation σ_P were calculated row by row to complete the automatic determination of the operating condition. The total duration of the steady-state load segment was 6771 minutes, accounting for 52.2% of all valid samples. The total duration of the variable load segment was 514 minutes, accounting for 4.0% of all valid samples. The total duration of the transition phase was 5676 minutes, accounting for 43.8% of all valid samples.

[0056] Samples of each load section and NO x Mean statistics:

[0057] It can be seen that NO in the low load rangex The overall generation level was the highest, with NO in the medium load range. x The lowest average value was found on different load platforms. x The difference in background concentration is significant, and the unified global model cannot distinguish this characteristic.

[0058] 4. S3 Working Condition NO x Generation pattern modeling and prediction model training 4.1 Temporal Partitioning of Datasets All 12,961 valid samples were divided according to the order of collection time: the first 70% of the time series data was used as the training set, and the last 30% of the time series data was used as the validation set. The data was not randomly shuffled to avoid leakage of time series information.

[0059] Pearson correlation and Spearman correlation were used to perform initial screening of factors. Random forest (RF) was used to calculate feature importance, and ridge regression was used to calculate standardized partial regression coefficients. Partial correlation analysis of control load was introduced to eliminate load confounding interference. Models were created separately for all time periods, steady state, and variable load, and a working condition-factor priority matrix M was generated.

[0060] 4.2 Model Fitting Effect and Core Influencing Factors under Different Working Conditions 1) Global model for all time periods (N=12961) The model's coefficient of determination R² = 0.610, indicating a poor fit. Top 5 correlation coefficients: C Coal feeder coal quantity (-0.540), C Mill inlet air flow rate (-0.539), SCR inlet oxygen quantity (+0.504), D Mill inlet air flow rate (-0.382), E Coal feeder coal quantity (-0.259); The top 3 most important features of random forest are: SCR (Inlet Oxygen Content) 49.4%, E (Coal Feeder Coal Quantity) 12.2%, and Actual Coal Quantity 11.9%.

[0061] 2) Steady-state operating condition model (N=6771) The model's coefficient of determination R² = 0.698, indicating a significant improvement in fitting accuracy; Top 5 correlation coefficients: C mill inlet air flow (-0.661), C coal feeder coal quantity (-0.583), SCR inlet oxygen content average (+0.569), D mill inlet air flow (-0.520), actual coal quantity (-0.276); The top 3 most important features of random forest are: SCR (Inlet Oxygen Content) 47.9%, C (Coal Feeder Quantity) 14.6%, and E (Coal Feeder Quantity) 11.8%.

[0062] 3) Variable load model (N=514) The model's coefficient of determination R² = 0.630; Top 5 correlation coefficients: B Coal feeder coal quantity (+0.474), total air volume (+0.245), C Mill inlet air flow (-0.191), A Coal feeder coal quantity (-0.157), actual coal quantity (+0.155); The top 3 most important features in random forest are: B feeder coal quantity 44.6%, SCR inlet oxygen quantity 24.5%, and A feeder coal quantity 22.1%.

[0063] 4.3 Core Influencing Factors of Load Division Low load (<101MW) TOP3: SCR inlet oxygen (+0.78), C mill inlet air flow (-0.68), total air volume (-0.63); Medium load (101~213MW) TOP3: C Coal feeder coal quantity (-0.50), SCR inlet oxygen quantity (+0.48), total coal quantity (-0.32); Top 3 High Load (>213MW): C Coal feeder coal quantity (-0.49), B Coal feeder coal quantity (+0.45), Total air volume (+0.40).

[0064] 4.4 Construction and Accuracy Verification of the Two-Layer Coupled Prediction Model Building a two-level prediction model: A horizontal correlation model for the long-term benchmark NO x The optimized fluctuation model is used for minute-level dynamic fluctuation prediction. The model input parameters include oxygen content, unit load, coal mill output, total air volume, primary air-to-coal ratio, and coal quality indicators. The output is the SCR inlet NO. x Predict concentration.

[0065] The optimal model was selected by using the validation set R², MAE, and RMSE, and the random forest algorithm was finally selected. The measured validation set model had a determination coefficient R² = 0.950, a mean absolute error MAE = 5.9 mg / m³, and the predicted scatter points closely matched the ideal fitted line, with prediction accuracy much higher than that of the global single-condition model.

[0066] 5. S4 NO x Quantification of fluctuation patterns and attribution of steady-state fluctuations Steady-state fluctuation quantification indicators; The steady-state effective segmentation consists of 46 segments, with NO within each segment. x The median standard deviation was 7.4 mg / m³, the median concentration range was 35.4 mg / m³, and the maximum range in a single range was 116.8 mg / m³; NO under steady-state conditions xThe median absolute value of the change per minute is 2.84 mg / m³.

[0067] Quantitative indicators of load fluctuation: The variable load is effectively divided into 84 segments, with NO within each segment. x The median standard deviation was 6.7 mg / m³, the median concentration range was 14.4 mg / m³, the maximum instantaneous jump range in a single period was 164.8 mg / m³, and the NO concentration during the variable load phase was... x The risk of peak impact is significantly higher than that of steady state.

[0068] Steady-state minute-level fluctuation attribution analysis: Calculate ΔNO x Correlation coefficients with minute-by-minute changes in various operating parameters Δxi; Top 5 causes of short-term fluctuations: d (SCR inlet oxygen content) (+0.631), d (C coal feeder coal quantity) (-0.344), d (D mill inlet air flow) (-0.180), d (E coal feeder coal quantity) (-0.179), d (B coal feeder coal quantity) (+0.092); Prove that the minute-by-minute fine adjustment of oxygen content under steady state is NO x The primary cause of minor fluctuations is to fine-tune the oxygen level in the furnace on-site to achieve stable nitrogen levels.

[0069] 6. Construction and incremental update testing of the S5 unit-specific rule library Establish an independent and dedicated NO for this 300MW unit x Generate a rule library, storing the factor priority matrix M, fluctuation quantification index, and trained two-layer prediction model for each load segment, steady-state / variable load / transitional state operating condition.

[0070] The system continuously collects new minute-level running data for 7 consecutive days, updates the dataset in real time, and recalculates the model's R² and the top 3 core influencing factors. An update threshold is set: when the change in R² is greater than 0.05 or when the top 3 core influencing factors are replaced, the system automatically triggers an incremental update of the pattern database.

[0071] During the test, due to the adjustment of the coal type fed into the furnace, the model R² dropped from 0.950 to 0.892. The change exceeded the threshold, and the system automatically completed the incremental update of the rule library. After the update, the model R² recovered to 0.947, and no manual remodeling was required.

[0072] 7. Project Implementation and Application Results The trained two-layer predictive model is embedded into the unit's DCS environmental optimization module and APS automatic power generation control system, and simultaneously connected to the SCR ammonia injection optimization system to achieve coordinated control of combustion and denitrification. Steady-state operating conditions: Fine-tuning of furnace oxygen and NO based on fluctuation attribution results. x The median minute-level fluctuation decreased from 2.84 mg / m³ to 1.1 mg / m³, and the SCR inlet NO... x The volatility decreased by 61%; Variable load adjustment: Based on the variable load condition factor matrix, intervene in advance to adjust the coal distribution of feeders B and A, and adjust NO during the variable load phase. x The maximum instantaneous jump decreased from 164.8 mg / m³ to 62.3 mg / m³, eliminating the instantaneous peak exceeding the standard; Denitrification synergistic optimization: based on real-time NO x By adjusting the ammonia injection rate in advance based on the predicted values, the average daily consumption of liquid ammonia decreased by 7.2%, and the problems of catalyst blockage and ammonia escape were significantly alleviated. Operation and maintenance optimization: The entire process of data cleaning, operating condition identification, and pattern updates is automated, eliminating the need for operators to manually compile statistics and draw experience curves, greatly reducing the workload of manual data analysis. New operators can directly rely on the factor priority matrix to optimize air and coal distribution.

[0073] 8. Advantages of Comparison of Examples Traditional single global models do not fit uniformly regardless of operating conditions, resulting in an R² of only 0.61. They cannot distinguish between steady-state and variable load dominant factors, and their guidance for on-site control is weak. This embodiment models the model according to different operating conditions and load segments. The validation set R² of the two-layer prediction model reaches 0.95, which can accurately quantify the control weight of each parameter, automatically identify the root cause of fluctuations and adaptively update the model. It adapts to long-term changes in coal quality and equipment operating conditions of the unit, and fully meets the needs of power plant ultra-low emission refined combustion optimization engineering.

[0074] It is understood that the above embodiments are merely exemplary implementations used to illustrate the principles of the present invention, and the present invention is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A method for burning NO in coal-fired power units x The method for generating patterns for identification and prediction is characterized by, Includes the following steps: S1: Data Acquisition and Effective Sample Construction S1.1: Collect minute-level time-series operating data from the DCS of the coal-fired power unit. The minute-level time-series operating data from the DCS of the coal-fired power unit includes the NO at the A-side inlet of the SCR unit. x Concentration of CNO x -A, SCR device B-side inlet NO x Concentration of CNO x -B; S1.2: Data cleaning and processing: Remove invalid data records corresponding to unit load below the preset operating threshold, boiler shutdown, and furnace purging conditions; S1.3: Construct a unified SCR entry point NO x Characteristic quantity: Calculate NO on both sides A and B of the SCR device. x Concentration correlation coefficient corr(CNO) x -A,CNO x -B) If the correlation coefficient is greater than the preset correlation threshold, then the average concentration of both sides is taken as the comprehensive NO concentration. x Characteristic CNO x =(CNO x -A+CNO x -B) / 2; If the correlation coefficient is not greater than the preset correlation threshold, then for the NO on both sides of SCR device A and B... x Model them independently; S2: Boiler Operating Condition Identification and Load Segmentation S2.1: The rate of change of computer group load dP / dt = |P(t) - P(t-1)| / Δt, and the standard deviation of computer group load σ is used in a 30-minute rolling window. P Where P(t) is the real-time active load of the unit at time t, and P(t-1) is the real-time active load of the unit at time t-1; S2.2: Based on the standard deviation of the load σ P The load change rate and duration are used to determine the operating condition type. Steady-state load condition: Load standard deviation σ P ≤Stable load threshold σ stable And the continuous operating time is greater than or equal to the steady-state duration threshold T. stable ; Variable load condition: Load change rate dP / dt ≥ Variable load rate threshold v change And the continuous operating time is greater than or equal to the variable load duration threshold T. change ; The remaining data samples are classified as transitional operating conditions; S2.3: The load quartile method is used to divide the entire operating condition into three load platforms: low load, medium load, and high load. S3: Working Condition NO x Generation pattern modeling and prediction model training S3.1: For the samples processed in step S1.2, Pearson and Spearman correlation coefficients are used to initially screen influencing factors. Then, random forest and gradient boosting tree algorithms are used to calculate the impact of each operating parameter on NO. x The characteristic importance weight of concentration; S3.2: A standardized partial regression model is constructed using multiple linear regression and ridge regression to obtain the impact of each influencing factor on NO. x The magnitude and direction of the independent contribution of concentration are characterized by the standardized partial regression model determination coefficient R², which represents the interpretability of the model. S3.3: Conduct partial correlation analysis of control load variables to eliminate the interference of unit load confounding effects on the judgment of factor correlation; S3.4: Repeat steps S3.1 to S3.3 for different load segments and different operating conditions to generate an operating condition-factor correlation priority matrix M. Matrix M stores the correlation coefficient, standardization coefficient, feature importance weight, and positive and negative direction of the corresponding influencing factors under each operating condition. S3.5: Temporal segmentation of datasets and construction of a two-layer NO x Predictive model: S3.5.1: Divide the cleaned samples into a 70% training set and a 30% validation set in chronological order. Randomly shuffling the samples is prohibited to avoid leakage of temporal information. The training set is used to construct the association model and the association priority matrix M, and the validation set is used to evaluate the generalization performance of the model. S3.5.2: Construct a two-layer coupled prediction model, including a horizontal correlation model for long-term combustion optimization and a fluctuation model for minute-level stable control; the model inputs are oxygen content, unit load, coal quantity of each pulverizer, total air volume, primary air-to-coal ratio, and fuel ratio, and the output is SCR inlet NO. x Concentration prediction values; Random forest or ridge regression algorithms are selected for modeling, and validation set R², mean absolute error MAE, and root mean square error RMSE are used as the evaluation indicators for model optimization. S4: Working Condition NO x Quantitative analysis of volatility patterns and volatility attribution analysis S4.1: Calculation of quantitative indicators for steady-state operating condition fluctuations: NO within the steady-state range x Concentration standard deviation σ_NOx stable Concentration range R_NOx stable minute-level NO x Median of absolute change |ΔNOx|; S4.2: Calculation of quantitative indicators for fluctuations under variable load conditions: NO throughout the entire variable load process x Extreme concentration difference, start and end of operating conditions NO x Change, maximum NO per minute x Jump amplitude | ΔNO x |_max; S4.3: Steady-state minute-level fluctuation attribution: Calculating single-minute NO x Change ΔNO x The minute-by-minute changes in various operating parameters Δx i The correlation coefficient identifies the causes of NO x The core control parameters for short-term fluctuations; S5: Unit Adaptive Rule Library Construction and Incremental Update S5.1: Dedicated NO units are independently constructed for each individual unit. x Generate a pattern library to store the factor correlation matrix, fluctuation quantification index, and trained two-layer prediction model for each load segment and operating condition of the corresponding unit. S5.2: Real-time acquisition of newly added operating data of the unit, rolling update of the dataset and recalculation of model fit R² and ranking of core influencing factors; when the change in R² or the change in the ranking of TOP core factors exceeds the preset update threshold, the incremental update of the unit pattern library is automatically triggered. S6: Output control boundary data The trained two-layer prediction model is embedded into the unit's DCS, APS automatic power generation control system, or boiler environmental optimization system to output the SCR inlet NO under different operating conditions in real time. x Concentration prediction boundary values ​​are used for closed-loop optimization and control of combustion air and coal distribution.

2. The method for burning NO in coal-fired power units according to claim 1 x The method for generating patterns for identification and prediction is characterized by, In step S1.1, the DCS minute-level time-series operation data of the coal-fired unit also includes the unit load, oxygen content at both inlets of the SCR, total furnace air volume, total coal volume, instantaneous coal volume of each coal feeder, primary inlet air volume of each coal mill, main steam flow rate, total feedwater flow rate, and coal quality test data entering the furnace. Among them, the coal quality test data entering the furnace includes volatile matter in coal, fuel nitrogen content, and fixed carbon index.

3. The method for burning NO in coal-fired power units according to claim 1 x The method for generating patterns for identification and prediction is characterized by, The threshold values ​​in step S2.2 are: stable load threshold σ stable =1.0MW, steady-state duration threshold T stable =30min, variable load rate threshold v change =2MW / min, variable load duration threshold T change =3min.

4. The method for burning NO in coal-fired power units according to claim 1 x The method for generating patterns for identification and prediction is characterized by, The load segmentation criteria in step S2.3 are as follows: low load < 40% of rated load, medium load 40%~70% of rated load, and high load > 70% of rated load.

5. The method for burning NO in coal-fired power units according to claim 1 x The method for generating patterns for identification and prediction is characterized by, After completing step S4.3 of attributing steady-state fluctuations, the identified main causes of short-term fluctuations are used as the priority operating variables for refined steady-state control of the boiler. NO is suppressed by fine-tuning this variable. x Minute-level fluctuations.

6. The method for burning NO in coal-fired power units according to claim 1 x The method for generating patterns for identification and prediction is characterized by, Under variable load conditions, the core impact parameter of the load change impact is determined based on the correlation priority matrix M. This parameter is intervened in advance during the unit's load increase and decrease process to reduce NO during the load change phase. x Maximum jump amplitude.

7. The method for burning NO in coal-fired power units according to claim 1 x The method for generating patterns for identification and prediction is characterized by, Double-layer NO x The prediction model has a coefficient of determination R² ≥ 0.95 on the validation set and a mean absolute error MAE ≤ 6 mg / m³.

8. The method for burning NO in coal-fired power units according to claim 1 x The method for generating patterns for identification and prediction is characterized by, The incremental update trigger conditions in step S5.2 are: the fluctuation range of the newly added data window model R² exceeds 0.05, or the top 3 core influencing factors are replaced.

9. The method for burning NO in coal-fired power units according to claim 1 x The method for generating patterns for identification and prediction is characterized by, In step S1.3, the preset correlation threshold is 0.

85. When NO on sides A and B... x When the correlation coefficient is ≤0.85, construct independent two-layer prediction models for side A and side B respectively, and output the NO values ​​for both sides separately. x Predict concentration to achieve precise air distribution optimization on each side.

10. The method for burning NO in a coal-fired power unit according to claim 1 x The method for generating patterns for identification and prediction is characterized by, The NO output of the prediction model x The boundary concentration values ​​are sent to the SCR ammonia injection optimization control system to achieve coordinated operation between the boiler combustion front-end control and the tail-end denitrification ammonia injection.

Images