An adaptive prediction method for NOx emission of high-alkali coal boiler

Abstract

The application discloses a kind of high alkali coal boiler NOx emissions self-adapting prediction method, specifically related to boiler emission prediction field, including by collecting data and dynamically extracting key features, after optimization by integrated learning, train an adaptive prediction model that can be updated online, and utilize prediction error feedback to continuously optimize fuel procurement suggestion.A kind of high alkali coal boiler NOx emissions self-adapting prediction method effectively solves real-time prediction delay problem by dynamic sampling rate adjustment strategy, reduces the power plant resource scheduling lag and cost control improper due to delay;By modulating learning rate with comprehensive robustness index, quickly adapt to changes in boiler operating conditions, improve the fuel inventory turnover rate accounting accuracy and environmental protection cost budget accuracy, avoid waste of resources and cost overrun due to unreliable prediction;By prediction based on comprehensive robustness index, significantly improve the application value of technology in power plant operation resource optimization scheduling, strengthen the refinement and efficiency of operation management.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of boiler emission prediction, and more particularly to a self-adaptive prediction method for NOx emission of high-alkali coal boiler. BACKGROUND

[0002] With the increasing strictness of environmental protection requirements of coal-fired power plants and the dual needs of operation cost control, the accurate prediction of NOx emission of high-alkali coal boiler is increasingly important in the operation management of fuel procurement planning and environmental protection cost budgeting. Traditional methods mostly rely on empirical models with fixed parameters, and rely on preset formulas and constants for estimation. However, actual measurements show that due to large fluctuations in coal quality, the prediction accuracy of the model is greatly reduced, which not only leads to a lag in combustion adjustment, but also causes an imbalance in fuel procurement ratio due to inaccurate NOx emission prediction, which cannot support the decision-making needs of power plant operation cost control and resource optimization.

[0003] To solve the adaptability defects of traditional methods, the prior art introduces a neural network method, uses a learnable offset and a modulation factor to adaptively adjust the sampling position and feature weight of the convolution kernel, thereby enhancing the model's ability to extract complex working condition features. The NOx prediction result is coupled with power plant cost accounting, fuel inventory management, and environmental protection budgeting to provide dynamic data support for fuel procurement quantity determination and environmental protection cost allocation operation decisions.

[0004] However, in actual use, there are still some shortcomings, such as real-time prediction delay, and batch data processing dependent updating mechanism, which cannot support time-sensitive operation decisions such as fuel procurement cycle and environmental protection requirement early warning. At the same time, the release of alkali metals during high-alkali coal combustion easily produces data noise, which reduces the prediction reliability and affects the accuracy of fuel inventory turnover rate accounting and environmental protection cost budgeting, thereby reducing the application value of the technology in power plant operation resource scheduling and cost control scenarios. SUMMARY

[0005] In order to overcome the above-mentioned defects of the prior art, the present application provides a self-adaptive prediction method for NOx emission of high-alkali coal boiler, which solves the problems raised in the background art by the following scheme.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical scheme:

[0007] A self-adaptive prediction method for NOx emission of high-alkali coal boiler, comprising:

[0008] S1: Real-time acquisition of first component data in the high-alkali coal boiler through a deployed sensor array;

[0009] S2: Based on the first component data, performing alkali metal noise filtering to generate second component data;

[0010] S3: Based on the second component data, dynamically extracting first component features through an adaptive feature selection model, the first component features including combustion temperature distribution, oxygen concentration and coal quality fluctuation index;

[0011] S4: Robustly optimizing the first component features through an ensemble learning technique, and outputting second component features for adaptive prediction model training;

[0012] S5: Based on the second component features generated by historical time series accumulation, initially training the adaptive prediction model, and based on the second component features generated in real-time operation, updating the model parameters of the adaptive prediction model at a fuel procurement adjustment cycle frequency, and outputting the adaptive prediction model;

[0013] S6: Obtaining the second component features based on the current time series transmitted by S4, and inputting them into the adaptive prediction model to generate real-time NOx emission prediction results and generate fuel procurement optimization suggestions;

[0014] S7: Collecting the prediction error between the real-time NOx emission prediction results and the actual operation data; when the prediction error leads to an overspending warning of fuel procurement cost, triggering the fuel procurement optimization suggestions and the model retraining or parameter adjustment of S5.

[0015] Preferably, before S1, noise pre-marking operation needs to be performed on the edge side of each sensor node for obtaining the first component data, specifically including:

[0016] At time , a sliding time window with a window length of is maintained for temporarily storing the measurement values collected by sensors, wherein the window length is set according to the sampling frequency ;

[0017] The measurement value sequence collected by the sensors within the sliding time window is subjected to layer discrete wavelet transform, and the selected wavelet basis function needs to satisfy: ;

[0018] Focusing on the layer detail coefficient , the energy value of the detail coefficient within the sliding time window is calculated, which is specifically expressed as:

[0019] ,

[0020] wherein, is the detail coefficient is the length of the sliding time window, is the high frequency component of the transmitted measurement sequence within the frequency band is the th wavelet coefficient of the th layer;

[0021] based on the historical energy reference and its standard deviation and compared with a preset noise threshold and the result is written into the noise pre-mark field. If the center point of the sliding time window is interfered by high intensity alkali metal noise, the noise pre-mark field is set to 1.

[0022] Preferably, the S2, performing alkali metal noise filtering, adopts a two-stage filtering specifically including:

[0023] The first stage is transient noise detection and elimination based on wavelet transform;

[0024] The second stage is state smoothing and estimation by a multivariate Kalman filter based on online adaptive adjustment of process noise and observation noise covariance matrices.

[0025] Preferably, the S2, in the multivariate Kalman filter of the second stage, the online adaptive adjustment of the observation noise covariance matrix and the process noise covariance matrix specifically includes:

[0026] calculating the innovation sequence at the current time, the innovation sequence being the difference between the current observation value and the one-step prediction value of the filter;

[0027] within the sliding time window, calculating the actual covariance matrix of the innovation sequence;

[0028] comparing the actual covariance matrix with a theoretical covariance matrix, the calculation formula of the theoretical covariance matrix being wherein is the observation matrix, is the state prediction covariance matrix, and T represents transposition calculation;

[0029] when the ratio of the trace of the actual covariance matrix to the trace of the theoretical covariance matrix continuously exceeds a preset range for a specified period of time, it is determined that a significant deviation occurs, triggering recursive update of the matrices and ​​​​​​

[0030] When the ratio is consistently greater than a preset upper limit, the process noise covariance matrix is ​​increased by a first proportion. ;

[0031] When the ratio remains below a preset lower limit, the observation noise covariance matrix is ​​reduced by a second ratio. .

[0032] Preferably, in step S3, the adaptive feature selection model employs an integrated filtering selection method, specifically including:

[0033] Primary feature ranking based on mutual information;

[0034] Intermediate screening is performed using recursive feature elimination with random forest as the base learner;

[0035] A stability selection mechanism is introduced, which calculates the stability score of a feature by multiple subsampling fittings. The stability score is defined as the total number of times a feature is selected divided by the total number of repetitions.

[0036] Preferably, in step S4, the second component features include meta-features synthesized based on feature stability scores and dynamic importance weights, specifically including:

[0037] For the Fusion weights of individual integrators The calculation is performed, specifically as follows:

[0038] ,

[0039] in, This represents the total number of base integrators. , Represented as the first The, the The average stability score of the first component feature upon which each base integrator depends. 、 Represented as the first The, the The normalized short-term mutual information value of the first component feature upon which each base integrator depends;

[0040] Each base integrator independently outputs a pair of preliminary predictions of NOx emission concentrations based on the first component characteristic. ;

[0041] The dynamically allocated fusion weights and the corresponding predicted value Multiply and sum to generate element features.

[0042] Preferably, the S4, the second component feature comprises a comprehensive robustness index The calculation formula is specifically represented as:

[0043] ,

[0044] Wherein, is the maximum weight of the feature, is a preset balance coefficient;

[0045] For the meta-feature generated by synthesis, that is, the fusion weight .

[0046] Preferably, the S6, based on the comprehensive robustness index value attached in the second component feature

[0047] If the comprehensive robustness index is higher than the first threshold value, the second component feature is input to the main prediction model output by S5 for rapid inference;

[0048] If the comprehensive robustness index is lower than the first threshold value but higher than the second threshold value, a lightweight calibration model is triggered to correct the output of the main prediction model, and the lightweight calibration model is a model copy fine-tuned based on recent low-noise data;

[0049] If the robustness index is lower than the second threshold value, a simplified regression model is enabled for prediction, and the prediction result is marked as a degraded mode output.

[0050] Technical effects and advantages of the present application:

[0051] 1. The present application reduces the amount of redundant data collection, shortens the time consumption of data processing and model training, effectively solves the problem of real-time prediction delay, ensures that the operation decision with strong timeliness can obtain timely data support, reduces the delay of power plant resource scheduling and the inappropriateness of cost control caused by delay, and improves the response efficiency of operation decision;

[0052] 2. The present application modulates the learning rate by the comprehensive robustness index, breaks the dependence on batch data processing, realizes real-time online updating of model parameters, quickly adapts to changes in boiler operating conditions, guarantees the reliability of prediction results in the environment of data noise caused by alkali metal precipitation in high-alkali coal combustion, improves the accuracy of fuel inventory turnover rate accounting and the accuracy of environmental protection cost budget, provides stable data basis for power plant operation resource scheduling and cost control, and avoids resource waste and cost overrun caused by unreliable prediction;

[0053] 3、The application avoids single model full-scene inference delay based on comprehensive robustness index prediction, further guarantees prediction response speed in dynamic working conditions, significantly improves the application value of the technology in power plant operation resource optimization scheduling and whole-process cost control scenes, and strengthens the refinement and efficiency of operation management. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 A step block diagram of a high-alkali coal boiler NOx emission self-adaptive prediction method according to an embodiment of the application is provided. DETAILED DESCRIPTION

[0055] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the protection scope of the application.

[0056] The terms used in the following embodiments of the application are only for the purpose of describing the specific embodiments, and are not intended to be limiting on the application. As used in the specification of the application, the singular forms "a", "an" and "the" are intended to include the plural forms, unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used in the application means and includes any or all possible combinations of one or more listed items.

[0057] Hereinafter, the terms "first" and "second" are only for the purpose of description, and cannot be understood as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Therefore, the features defined with "first" and "second" can explicitly or implicitly include one or more features, and in the description of the embodiments of the application, the meaning of "multiple" is two or more, unless otherwise specified.

[0058] As shown in the accompanying drawings Figure 1 A self-adaptive prediction method for NOx emission of a high-alkali coal boiler, by collecting data and dynamically extracting key features, after optimization by integrated learning, training an adaptive prediction model that can be updated online, and continuously optimizing the model using prediction error feedback, specifically including the following steps:

[0059] S1: Real-time collection of first component data in the high-alkali coal boiler through the deployed sensor array;

[0060] S2: Based on the first component data, alkali metal noise filtering is performed to generate second component data;

[0061] S3: dynamically extracting first component features by an adaptive feature selection model based on the second component data, the first component features including combustion temperature distribution, oxygen concentration and coal quality fluctuation index;

[0062] S4: robustly optimizing the first component features by an ensemble learning technique, and outputting second component features for adaptive prediction model training;

[0063] S5: initially training the adaptive prediction model based on the generated second component features accumulated in a historical time sequence, and updating model parameters of the adaptive prediction model based on the second component features generated in real-time operation at a fuel procurement adjustment cycle as the update frequency, and outputting the adaptive prediction model;

[0064] S6: obtaining the second component features based on the current time sequence transmitted by S4, and inputting the second component features into the adaptive prediction model to generate real-time NOx emission prediction results and generate fuel procurement optimization suggestions;

[0065] S7: collecting prediction errors between the real-time NOx emission prediction results and actual operation data; when the prediction errors cause fuel procurement cost overbudget warning, triggering the fuel procurement optimization suggestions and model retraining or parameter adjustment of S5.

[0066] Specifically, in S1, a pre-deployed sensor array adopts a hierarchical heterogeneous topology structure, including a combustion core layer, a furnace main body layer and a tail flue layer; the first component data are all packaged as data packets on the edge side, and the contents include a sensor unique identifier, a time stamp, a type identifier of a measurement parameter, a topology layer identifier to which the sensor belongs, a sensor read value and its unit, and a noise pre-mark field; the topology layer identifier can set the combustion core layer as "CC", the furnace main body layer as "FMB" and the tail flue layer as "RFL".

[0067] In the embodiment, the combustion core layer is arranged in the burner outlet and the near combustion zone within 1-3 meters from the burner nozzle, with a density not greater than 0.5 meters, a high-density deployment of a high-speed K-type thermocouple with a response time less than 100 milliseconds, a high-temperature extraction type NOx probe with a sampling period of 500 milliseconds, and a fly ash carbon content laser detector capable of measuring unburned carbon online, and instantaneous signals related to alkali metal precipitation and initial NOx generation, including at least instantaneous gas temperature, NOx generation concentration and unburned carbon content; the furnace main body layer covers the middle and upper space of the boiler furnace, specifically the area from 3 meters from the burner nozzle to the furnace outlet, and at least covers flame detectors in visible and infrared bands, sub-zone NOx concentration detectors using laser absorption spectroscopy technology, and high-temperature resistance cameras in the area from 3 meters from the burner nozzle to the furnace outlet, according to a uniform grid of 2 meters x 2 meters; the tail flue layer covers the area from the furnace outlet to the boiler outlet, and at least covers a high-temperature resistance camera and a high-temperature resistance camera in the area from the furnace outlet to the boiler outlet, according to a uniform grid of 2 meters x 2 meters. The detection sensor and the infrared thermal imager are used to obtain the macroscopic distribution information of the combustion products and the temperature field in the furnace, including at least the flame existence signal and intensity, the carbon monoxide concentration, the oxygen concentration, and the temperature field of the furnace wall and the flue gas; the tail flue layer is arranged in the vertical flue before the economizer or the air preheater inlet, and the final measurement NOx concentration sensor, the soot concentration monitor based on the laser backscattering principle, and the capacitive humidity sensor are arranged at the position 3 times of the flue hydraulic diameter upstream of the center of the flue section, so as to capture the final emission data after sufficient reaction and mixing, including at least the NOx emission concentration, the soot concentration in the flue gas, and the flue gas humidity; the data collection of each layer sensor is uniformly scheduled by a central FPGA data collection controller, each sensor node is time-synchronized through the hardware clock synchronization mechanism based on the IEEE 1588 precision clock protocol, the time stamp error of the whole array in different physical positions is less than 1 microsecond, and the FPGA data collection controller is configured with a multi-channel parallel ADC, so as to synchronously trigger and read the data of all sensor channels at a basic sampling frequency of not less than 100 Hz .

[0068] It should be noted that the collection of the first component data introduces a dynamic sampling rate adjustment strategy based on feedback of the adaptive prediction model, and the implementation includes: when the short-term prediction confidence output by the adaptive prediction model in S5 is higher than 0.9 and the boiler load change rate is lower than 2% / min, the baseline sampling frequency of 100 Hz is run; when the prediction confidence is continuously lower than 0.85 or the load change rate is continuously higher than 5% / min for 500 ms, the high-frequency collection mode is automatically triggered, the sampling frequency is increased to 500 Hz, and lasts at least 2 seconds to capture the complete dynamic process when the working condition changes dramatically; when the prediction confidence is continuously higher than 0.88 and the load change rate is continuously lower than 2% / min for 2 seconds, the baseline sampling frequency mode is switched to.

[0069] In a possible implementation, on the edge side of each sensor node transmission, the noise pre-marking operation includes: at time , a sliding time window with a window length of is maintained for temporarily storing the measurement values of sensors, wherein the window length is set according to the sampling frequency ; in this embodiment, when , the value of corresponds to a data window of 0.1 seconds; the measurement value sequence of the sensor collected in the sliding time window is subjected to layer discrete wavelet transform, and the selected wavelet basis function needs to satisfy: And decomposed into approximate coefficients and multiple levels of detail coefficients Focusing on the first Layer detail factor , Represented as a sequence of transmitted measurements in the frequency band Calculate the detail factor for high-frequency components within the range. In the sliding time window Internal energy value Specifically, it is expressed as:

[0070] ,

[0071] in, Represented as detail coefficients In the sliding time window The length inside, Represented as the first The first layer Wavelet coefficients; based on historical energy benchmarks and its standard deviation and with a preset noise threshold Comparison: The determination result is then written into the noise pre-label field of the corresponding data point; at time... The sliding time window is determined when the energy of the transmitted measurement sequence in the target noise frequency band significantly exceeds the fluctuation range of historical normal operating conditions. The center point is affected by high-intensity alkali metal noise, so the noise pre-labeling field is set to 1.

[0072] It should be noted that the energy value mentioned Used to quantize the current moment The intensity of the transmitted measurement sequence within the target noise frequency band, specifically corresponding to the wavelet decomposition of the [missing value]. Layer detail factor The specific frequency band it represents ;when At that time, the corresponding frequency band is approximately This frequency band can effectively capture the transient characteristics of alkali metal precipitation. After the boiler is installed and commissioned, and confirmed to be operating under stable, low-noise rated conditions, sensor data should be continuously collected for at least 30 minutes, and the energy value of the target frequency band should be calculated as a historical energy benchmark. The preset noise threshold This is a configurable parameter, set to 3 in this embodiment. When the judgment result significantly exceeds 3 standard deviations, the sliding time window is determined. The center point is subject to high-intensity alkali metal noise interference.

[0073] Specifically, in S2, performing alkali metal noise filtering on the first component data carrying the noise pre-mark field based on S1 transmission specifically includes: a first stage is transient noise detection and elimination based on wavelet transform; a second stage is state smoothing and estimation by a multivariate Kalman filter based on online adaptive adjustment of process noise and observation noise covariance matrix, finally generating clean data generated after double-stage filtering pipeline deep processing, i.e. the second component data.

[0074] Further, the transient noise detection and elimination based on wavelet transform includes: dynamically selecting a wavelet base function, which is adaptively selected according to the noise pre-mark field in the first component data and the topology layer identifier to which the sensor belongs: for data with a topology layer identifier of "CC", a dbN wavelet base with good tight support is pre-selected by default to better capture and eliminate transient noise; for data with a topology layer identifier of "FMB" or "RFL", a symN wavelet base with good symmetry is pre-selected by default; regardless of the pre-selected wavelet base, if the noise pre-mark field is 1, a strengthened filtering mode is immediately started; otherwise, a standard filtering mode is used.

[0075] In the embodiment, uniform 3-layer discrete wavelet decomposition is performed.

[0076] It should be noted that the threshold processing of the wavelet transform is strongly associated with the filtering mode: in the strengthened filtering mode, a hard threshold function is used, and the adaptive threshold is calculated as follows: wherein is a general threshold value, is a window length, is a standard deviation of the first layer wavelet decomposition coefficients; in the standard filtering mode, a soft threshold function is used, and the threshold is a general threshold value .

[0077] It should be noted that when the noise pre-mark field is 1, the strengthened filtering mode started not only uses the threshold, but also includes a reconstruction verification: after reconstructing the data with the wavelet coefficients, the residual error between the reconstructed data and the original data is calculated; if there are still peaks exceeding the threshold in the residual error, the data segment is marked as "extreme abnormality", and the predicted value based on historical trends by the Kalman filter is output as the temporary data at this moment instead of the wavelet reconstruction value, so as to ensure the continuity of the output data stream.

[0078] In a possible implementation, the observation noise covariance matrix and the process noise covariance matrix online adaptive adjustment, including: calculating a new sequence of innovation at the current time, the new sequence of innovation being the difference between the current observation value and the filter one-step prediction value; calculating the actual covariance matrix of the new sequence of innovation within a set sliding time window; comparing the actual covariance matrix with the theoretical covariance matrix, the calculation formula of the theoretical covariance matrix being , wherein is an observation matrix, is a state prediction covariance matrix, and T represents transposition calculation; when the ratio of the trace of the actual covariance matrix to the trace of the theoretical covariance matrix continuously exceeds a preset range for a specified number of periods, it is determined that a significant deviation occurs, triggering recursive update of the matrices and : when the ratio continuously exceeds a preset upper limit, the process noise covariance matrix is increased by a first proportion; when the ratio continuously is less than a preset lower limit, the observation noise covariance matrix is decreased by a second proportion.

[0079] It should be noted that the current observation value is the output data after the first-stage transient noise detection and elimination processing based on wavelet transform in S2; the filter one-step prediction value is calculated by a state transition equation based on the state estimation value at the previous time by the Kalman filter; the observation matrix is determined according to the physical relationship between the state vector and the observation vector, i.e. the data transmitted by the sensor; the state vector is an n×1 column vector, and the elements thereof are determined according to the boiler combustion mechanism, and the initial state vector is set as the average value of each sensor data in the first effective sampling window after startup.

[0080] In the embodiment, the state transition matrix is obtained based on linearization of a simplified physical model of the boiler combustion process, and a first-order autoregressive model is used for approximation, and the values on the diagonal are obtained by analyzing historical data; for the states that can be directly measured, the corresponding elements of the observation matrix are 1; for the states that need to be indirectly estimated through observation values, the corresponding observation equations are established; the state vector can be defined as , wherein represents temperature, represents actual NOx concentration, represents oxygen concentration, represents air volume; a specific embodiment of the preset range is [0.5, 2.0], and a specific embodiment of the specified number of periods is 3, i.e. when the ratio of the trace of the actual covariance matrix of the new sequence of innovation to the trace of the theoretical covariance matrix is less than 0.5 or higher than 2.0 for 3 consecutive times, it is determined that a significant deviation occurs, triggering parameter update: increasing the process noise covariance matrix by a first proportion, and the specific representation is: ; reducing the observation noise covariance matrix by a second proportion is specifically expressed as: .

[0081] Specifically, in S3, a candidate feature set is constructed based on the second component data The construction step includes but is not limited to: in a sliding window, the time domain statistics of the data obtained by each sensor are calculated in real time, including but not limited to: mean value reflecting average level, standard deviation reflecting fluctuation intensity, skewness reflecting distribution asymmetry, kurtosis reflecting distribution steepness, and linear trend slope of the last 10 seconds of data; based on the principle of combustion science, multiple sensor readings are combined, including but not limited to: based on the elemental analysis data of the coal entering the furnace and the theoretical combustion temperature estimated value calculated based on the air volume, the ratio of the coal feeder speed signal to the corresponding air volume of the burner is calculated, and the oxygen excess coefficient is calculated according to the measured oxygen concentration.

[0082] In one possible implementation, dynamically extracting the first component features by the adaptive feature selection model includes: performing primary feature sorting based on mutual information; performing intermediate screening by recursive feature elimination using a learning-based learner of random forest; and introducing a stability selection mechanism, calculating the stability score of the features by multiple subsampling fitting, and only retaining the features with a stability score exceeding a preset threshold to constitute the first component features.

[0083] In an embodiment, the primary feature sorting based on mutual information uses Kraskov estimation method to calculate the mutual information value between each candidate feature and the target variable NOx emission is specifically expressed as:

[0084] ,

[0085] wherein, is expressed as the numerical value of the candidate feature discretized in a period of time, is expressed as the value of the NOx emission discretized, respectively, the marginal frequency of individual occurrence, the marginal frequency of individual occurrence, and the frequency of joint occurrence; after the calculation, the mutual information values of all features are arranged in descending order, and according to a preset number of retained features, the features with the strongest correlation with the target variable are selected into the next screening stage; in a preferred embodiment, the number of retained features is set to 50.

[0086] ​It should be noted that the update frequency of the adaptive feature selection model in S3 is independent of and higher than the update frequency of the adaptive prediction model in S5.

[0087] In another embodiment, the recursive feature elimination with random forest as base learner is used for the intermediate screening, and the first 50 features based on mutual information are eliminated for redundancy and weak correlation features, including: the key hyperparameters of the base learner are set as: the number of decision trees in the forest is 100, the maximum depth of each tree is 15, and the minimum number of samples required for internal node redivision is 5; the feature elimination mechanism adopts a reverse elimination strategy, trains a random forest based on 50 features, and calculates the importance of each feature based on the Gini impurity reduction amount; in each iteration, the features with the last 10% of the feature importance ranking are eliminated, and the model is retrained using the remaining features; the iteration is repeatedly performed until the number of feature subsets is reduced to the preset retention number, which is 20 in a preferred embodiment.

[0088] In another embodiment, the adaptive feature selection model performs a stability selection mechanism for final selection, and the recursive feature elimination process is repeatedly verified by the bootstrap sampling method as follows: based on the remaining 20 features, 100 bootstrap samples are performed, and 70% of the samples are randomly selected with replacement to form a subsample set; the recursive feature elimination process is independently run on each subsample set, and the number of times each feature is selected in 100 subsamples is recorded to obtain a stability score, which is defined as the total number of times the feature is selected divided by the total number of repetitions; only features with a stability score exceeding a preset threshold of 0.85 are retained; if the number of features after screening with a threshold of 0.85 is less than the preset minimum feature retention number, the threshold is automatically lowered to 0.8; in a preferred embodiment, the preset minimum feature retention number is 5.

[0089] It should be noted that the adaptive feature selection model calculates a dynamic importance weight for each feature after performing the stability selection mechanism to reflect its relative importance under the current working condition, and the input parameters include: the stability score from the stability selection mechanism and the short-term mutual information value calculated based on the real-time data in the current time window , which is specifically represented as:

[0090] ,

[0091] wherein, is the weight coefficient of the stability score, and the preferred value thereof in the present embodiment is 0.6, This is represented by normalizing the short-term mutual information values ​​to the [0,1] interval using the min-max method, so that each feature of the output is associated with a dynamic weight in the [0,1] interval.

[0092] Specifically, in S4, the ensemble learning technique is as follows: parallel startup of random forest, gradient boosting tree and support vector regressor as base ensemblers; dynamic allocation of fusion weights to the outputs of each base ensembler based on feature stability score and dynamic importance weight, synthesizing meta-features; dynamic pruning based on the consistency variance among meta-features; the output second component features include: the feature with the highest stability score among the features inherited from S3, the synthesized meta-features, and a comprehensive robustness index attached to each feature.

[0093] In one possible implementation, dynamically assigning fusion weights to the outputs of each base integrator includes: [The text abruptly ends here, so the translation stops.] Fusion weights of individual integrators The calculations include: obtaining the fusion weights by calculating the unnormalized score of each base integrator. Specifically, it is expressed as:

[0094] ,

[0095] in, This represents the total number of base integrators. , Represented as the first The, the The average stability score of the first component feature upon which each base integrator depends. 、 Represented as the first The, the The normalized short-term mutual information values ​​of the first component features upon which each basis integrator relies; each basis integrator independently outputs a pair of preliminary predicted NOx emission concentrations based on the first component features. The dynamically allocated fusion weights and the corresponding predicted value Multiply and sum to generate eigenvalues. Specifically, it is expressed as:

[0096] .

[0097] The first component features are robustly optimized using ensemble learning techniques, and the second component features are output for training the adaptive prediction model.

[0098] Furthermore, the dynamic pruning mechanism includes: calculating the variance of the predicted values ​​of all base integrators on the current feature at each processing cycle. ; set dynamic threshold wherein is set as 0.01 in the present embodiment, is expressed as the absolute value of the boiler load change rate, is expressed as a proportionality coefficient; when it is detected that the current fusion weight will be automatically identified and temporarily removed the lowest base integrator corresponds to the meta-feature, until the next cycle is included.

[0099] Further, the comprehensive robustness index is calculated according to the formula, which is specifically expressed as:

[0100] ,

[0101] wherein, is the maximum weight of the feature, whose value depends on the source of the feature, is a preset balance coefficient for adjusting the relative influence of stability and current importance, which is defined as 0.7 in the present embodiment.

[0102] It should be noted that for the feature with the highest stability score among the features inherited from S3, its normalized dynamic importance weight is taken; for the meta-feature generated by synthesis, the fusion weight .

[0103] Specifically, in S5, the adaptive prediction model is an online learning network constructed based on a gated recurrent unit; the online updating of the adaptive prediction model adopts a weighted incremental learning algorithm based on the comprehensive robustness index , the parameter updating process of which is modulated by the comprehensive robustness index attached to the second component feature, which is used as the sample weight of the loss function to suppress the influence of low reliability data on the parameters of the adaptive prediction model; and the parameter updating frequency is strictly aligned with the fuel procurement adjustment period, i.e., if there is no coal replacement in the period, the parameters are updated according to the period, and if there is coal replacement, additional parameter calibration is performed.

[0104] Further, the second component features generated in the accumulated historical time series and their corresponding measured NOx emission values are used to fully train the online learning network using the small batch gradient descent method until the prediction accuracy of the adaptive prediction model on the validation set tends to be stable, and the initial training is completed; at least 12 complete fuel procurement adjustment periods of historical data are required during the initial training, covering different seasons and different coal batches, to ensure that the model has basic adaptation ability to the coal quality change in the procurement period.

[0105] Further, the comprehensive robustness index The second component feature of the comprehensive robustness index value is integrated into the optimization process of the adaptive prediction model, and the value is taken as a sample weight to amplify the contribution of high-reliability data to the loss and reduce the influence of low-reliability data; in the parameter updating phase, an online sequential learning strategy is adopted to calculate the gradient of the loss function on all parameters of the model based on only the current single sample, and the model parameters are updated in the opposite direction of the gradient with a dynamic learning rate modulated by the value; in this embodiment, when the value is higher than the threshold value 0.8, the learning rate is increased to 0.00015 to accelerate the learning of important knowledge; when the value is lower than the threshold value 0.3, the learning rate is reduced to 0.00005 to suppress noise interference. It should be noted that the fuel procurement adjustment cycle is updated as the frequency, including: the fuel procurement adjustment cycle takes the "coal quality batch and inventory consumption cycle" as the core basis, and preferentially uses the average turnover cycle obtained by statistical analysis of the historical procurement data of the power plant. If the procurement plan contains coal batches of different mining areas and different alkali contents, the cycle can be dynamically adjusted to the time length when the batch of coal is expected to be consumed. The starting time of the cycle is synchronized with the "issuance date of the procurement plan" of the procurement department to ensure that the model update matches the procurement decision rhythm. On the first day of each fuel procurement adjustment cycle, the following is automatically triggered: extract the key data in the last cycle, including at least the actual alkali content of the coal purchased in the cycle, the measured value of NOx emission after combustion, and the unit cost data of the fuel, and fuse them with the second component feature generated in real time in the cycle to serve as the core data set for model parameter updating; if there is a change of coal species in the cycle, an additional parameter calibration is triggered within 24 hours after the change of coal species to avoid model lag caused by sudden change of coal quality, and the data range of temporary calibration is limited to at least 4 hours of stable operation data after the change of coal species.

[0106]

[0107] ​​​​Specifically, in S6, based on the optimization logic of emissions and costs, fuel procurement optimization suggestions are generated. These suggestions include: data on the current alkali content of coal, calorific value fluctuation range, real-time NOx emission forecast results, and pre-set fuel candidate database data, which at least includes the procurement cost, transportation cycle, alkali content label, and supplier fulfillment rate for each candidate coal type; the content generated by the fuel procurement optimization suggestions covers at least: coal type ratio suggestions, based on the gap between NOx emission forecasts and environmental standards, suggesting a reduction in the proportion of high-alkali coal procurement and a specific ratio of low-alkali coal; procurement quantity suggestions, combining current fuel inventory and the coal quality stability required for NOx forecasts, suggesting the procurement quantity per batch; and procurement cycle adaptation suggestions, if environmental standards tighten during the forecast period, suggesting a shortened procurement cycle to avoid long-term inventory of high-alkali coal leading to subsequent emission exceedances.

[0108] Furthermore, based on the comprehensive robustness index attached to the second component feature... Value, dynamically select the predicted execution path: if the comprehensive robustness index If the value exceeds the first threshold, the second component feature is input into the main prediction model output by S5 for rapid inference; if the comprehensive robustness index is... If the value is below the first threshold but above the second threshold, a lightweight calibration model is triggered to correct the output of the main prediction model. This lightweight calibration model is a copy of the model fine-tuned based on recent low-noise data. If the robustness index... If the value is below the second threshold, a simplified regression model is used for prediction, and the prediction result is marked as a degraded mode output; and regardless of the prediction path, the confidence level of the fuel procurement optimization suggestion must be tied to the reliability of the prediction path.

[0109] It should be noted that the lightweight calibration model is a copy of the main prediction model after a dynamic pruning mechanism. It retains the first two layers of the original model's feedforward network and has been marked with the comprehensive robustness index within the last 24 hours. Fine-tuning training is performed on datasets exceeding a first threshold, specifically designed to learn compensatory mappings for low-quality features; the simplified regression model is a pre-defined multiple linear regression model, whose input consists of stable features selected from the second component features through mutual information, and whose parameters are periodically refitted based on operating data from the past week; when entering degraded mode output, a warning signal is simultaneously sent to the boiler control system, and the control strategy is automatically switched until... The value recovers to above the second threshold.

[0110] In one embodiment, the first threshold is set to 0.75, based on historical high-confidence prediction samples. The value was obtained by statistical analysis of the 95th percentile; the second threshold was set to 0.3, determined by the critical point at which performance drops sharply during noise injection testing.

[0111] Further, the main prediction model and the lightweight calibration model are executed in parallel to obtain first prediction value and first uncertainty interval and second prediction value and second uncertainty interval respectively; based on the width of the first uncertainty interval and the second uncertainty interval, prediction decision weights are assigned to the first prediction value and the second prediction value; according to the prediction decision weights, the first prediction value and the second prediction value are weighted and averaged to generate a final real-time NOx emission prediction result.

[0112] In another embodiment, the first uncertainty interval is calculated by the Monte Carlo Dropout mechanism built in the main prediction model after 50 forward samplings, the mean and the standard deviation of the prediction result are obtained, and the interval range is ; the second uncertainty interval is directly output by the lightweight calibration model through quantile regression, and the upper and lower bounds of the 90% confidence interval are output; the width of the first uncertainty interval is , and the width of the second uncertainty interval is , then the decision weight of the main prediction model is specifically represented as: , and the decision weight of the lightweight calibration model is ; the final NOx emission prediction result is specifically represented as:

[0113] .

[0114] Specifically, in S7, the prediction error between the real-time NOx emission prediction result and the actual operation data is collected; when the prediction error leads to a fuel procurement cost overrun warning, the warning threshold is set by the procurement department in advance, and the following is triggered simultaneously: analyzing the core reasons for the cost overrun caused by the prediction error, and triggering the model retraining or parameter adjustment of S5 based on a multi-level trigger judgment chain: instantaneous abnormality detection to generate an alarm; short-term performance degradation detection based on the weighted error of short-term moving average to trigger parameter adjustment; model degradation detection based on the long-term trend of the weighted error combined with the concept drift index to trigger model retraining; and in each level of trigger judgment, the execution deviation of the fuel procurement optimization suggestion needs to be evaluated simultaneously, if the deviation ratio exceeds 20%, the model operation is triggered, and a "suggestion execution deviation warning" is sent to the procurement department to avoid cost overrun caused by human execution factors.

[0115] ​​In the embodiment, if the error is caused by "coal quality fluctuation prediction deviation", the coal type screening threshold in the fuel procurement optimization suggestion is corrected, the alkali content upper limit is lowered, and the coal quality sampling frequency is increased; if the error is caused by "NOx treatment cost prediction deviation", the weight of cost and emission in the suggestion is adjusted, and the coal type with low treatment cost is preferentially recommended.

[0116] Secondly: the drawings of the disclosed embodiments only involve the structures involved in the disclosed embodiments, other structures can refer to the usual design, and in the case of no conflict, the same embodiment and different embodiments of the present application can be combined with each other;

[0117] Finally: the above only describes the preferred embodiments of the present application and is not used to limit the present application, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A method of adaptive prediction of NOx emissions from a high-alkali coal-fired boiler, characterized by, The method comprises the following steps: S1: collecting first component data in a high-alkali coal boiler in real time through a deployed sensor array; S2: performing alkali metal noise filtering based on the first component data to generate second component data; S3: dynamically extracting first component features, including combustion temperature distribution, oxygen concentration, and coal quality fluctuation index, through an adaptive feature selection model based on the second component data; The adaptive feature selection model adopts an integrated filtering selection method, specifically including: primary feature sorting based on mutual information; intermediate screening using recursive feature elimination with a random forest-based learning machine; and introducing a stability selection mechanism to calculate the stability score of the feature through multiple subsampling fitting, wherein the stability score is defined as the total number of times the feature is selected divided by the total number of repetitions; S4: Robustly optimizing the first component features through an ensemble learning technique, outputting second component features for adaptive prediction model training; the second component features contain meta-features synthesized based on feature stability scores and dynamic importance weights, specifically including: calculating the fusion weights of the first component features, specifically represented as: ​​ wherein, represents the total number of base integrators, , respectively represent the average stability score of the first component feature relied on by the th and the th base integrator, , respectively represent the normalized short-term mutual information value of the first component feature relied on by the th and the th base integrator; each base integrator independently outputs a pair of preliminary prediction values of NOx emission concentration based on the first component feature ; the dynamically allocated fusion weights are multiplied by and summed with the corresponding prediction values to generate a meta feature; the second component feature further comprises a comprehensive robustness index , the calculation formula of which is specifically represented as: wherein, is represented as the most significant weight of the feature, is represented as a preset balance coefficient, is represented as a stability score; for a meta-feature generated by synthesis, i.e., a fusion weight ; S5: initially training the adaptive prediction model based on the generated second component features accumulated in the history time sequence, and updating the model parameters of the adaptive prediction model online based on the second component features generated in real-time operation with a fuel procurement adjustment period as the update frequency, and outputting the adaptive prediction model; S6: obtaining the second component feature based on the current timing transmitted by S4, and inputting it into the adaptive prediction model to generate a real-time NOx emission prediction result and generate a fuel procurement optimization suggestion; based on the comprehensive robustness index attached in the second component feature value, dynamically selecting a prediction execution path: If we consider the robustness index If the value is higher than the first threshold, the second component feature is input into the main prediction model output by S5 for rapid inference; if the integrated robustness index below the first threshold but above the second threshold, triggering a correction of the output of the main prediction model by a light-weight calibration model, the light-weight calibration model being a fine-tuned copy of the model based on recent low-noise data; robustness index if the robustness index is lower than the second threshold value, a simplified regression model is enabled to make a prediction and its prediction result is marked as a degraded mode output; S7: collecting the prediction error between the real-time NOx emission prediction result and the actual operation data; when the prediction error leads to an overspending warning of the fuel procurement cost, triggering the fuel procurement optimization suggestion and the model retraining or parameter adjustment of S5.

2. The self-adaptive prediction method of NOx emission of a high-alkali coal boiler according to claim 1, characterized in that: Before S1, noise pre-marking operation needs to be performed on the edge side of each sensor node, specifically including: At time , a sliding time window with a window length of is maintained for temporarily storing the measurement values collected by the sensors, wherein the window length is set according to the sampling frequency . performing a discrete wavelet transform on a sequence of measurement values collected by the sensor within a sliding time window performing a discrete wavelet transform on a sequence of measurement values collected by the sensor within a sliding time window The selected wavelet base function needs to satisfy: ; Focus on the first Layer detail factor Calculate detail coefficients In the sliding time window Internal energy value Specifically, it is expressed as: wherein is a detail coefficient is a length within a sliding time window is a high frequency component of the transmitted measurement value sequence within a frequency band is a wavelet coefficient of the layer is a wavelet coefficient Based on historical energy benchmarks and its standard deviation and with a preset noise threshold Comparison: The determination result is written into the noise pre-labeling field to determine the sliding time window. The center point is affected by high-intensity alkali metal noise, so the noise pre-labeling field is set to 1.

3. The self-adaptive prediction method of NOx emission of a high-alkali coal boiler according to claim 1, characterized in that: In S2, the double-stage filtering for performing alkali metal noise filtering specifically includes: The first stage is transient noise detection and elimination based on wavelet transform; The second stage is state smoothing and estimation by a multivariate Kalman filter based on online adaptive adjustment of process noise and observation noise covariance matrix.

4. The self-adaptive prediction method of NOx emission of a high-alkali coal boiler according to claim 3, characterized in that: S2, an observation noise covariance matrix of the second stage multivariable Kalman filter and an online adaptive adjustment of a process noise covariance matrix specifically comprising: Calculate the innovation sequence at the current time, which is the difference between the current observation value and the one-step prediction value of the filter; Calculate the actual covariance matrix of the innovation sequence within the sliding time window; The actual covariance matrix is compared with a theoretical covariance matrix, the formula of which is wherein is an observation matrix, is a state prediction covariance matrix, T represents a transpose calculation; When the ratio of the trace of the actual covariance matrix to the trace of the theoretical covariance matrix consistently exceeds a predetermined range for a specified number of periods, a significant deviation is determined to have occurred, triggering recursive updating of the matrix and ​ when the ratio is continuously greater than a preset upper limit, increasing the process noise covariance matrix by a first proportion ; decreasing the observation noise covariance matrix by a second ratio when the ratio is continuously less than a preset lower limit .

Images