Thermal power denitration nitrogen oxide concentration dynamic prediction method and device and electronic equipment

Abstract

The invention relates to a thermal power denitration nitrogen oxide concentration dynamic prediction method, a thermal power denitration nitrogen oxide concentration dynamic prediction device and electronic equipment. The method comprises the following steps: acquiring real-time process parameters of boiler operation; the real-time process parameters are preprocessed to form input data suitable for the prediction model, and preprocessing at least comprises the step of conducting time sequence delay adjustment on the process parameters based on the flue gas flow speed and instrument monitoring delay; identifying the preprocessed input data as a first working condition type or a second working condition type according to a working condition factor representing a unit load change state; inputting the input data identified as the first working condition type into a first prediction model trained by a first data set, and inputting the input data identified as the second working condition type into a second prediction model trained by a second data set; and outputting a nitrogen oxide concentration prediction value calculated by the first prediction model or the second prediction model. The problems of ammonia spraying control misalignment and environmental protection risks caused by detection lag and severe concentration fluctuation are solved.

Description

Technical Field

[0001] This invention relates to the field of flue gas treatment and control technology in thermal power plants, and in particular to a method, device, and electronic equipment for dynamic prediction of nitrogen oxide concentration in thermal power plant denitrification. Background Technology

[0002] Against the backdrop of ultra-low emissions for environmental protection and flexible peak shaving for thermal power, the operating load of thermal power units needs to change frequently to adapt to the grid demand. At the same time, there are also local changes in the quality of coal fed into the furnace and the boiler combustion conditions. The combination of these factors leads to drastic fluctuations in the concentration of nitrogen oxides at the inlet of the selective catalytic reduction denitrification system.

[0003] At the process control level, nitrogen oxide (NOx) concentration detection commonly employs extraction-type continuous flue gas monitoring systems. However, these systems have a detection cycle of approximately 120 seconds, meaning the concentration signal received by the control system lags significantly behind the actual conditions at the boiler combustion site. Due to this detection lag, traditional distributed control system strategies struggle to respond promptly and accurately to drastically fluctuating NOx concentrations. To meet stringent environmental emission standards, a conservative control approach involving excessive ammonia injection is often adopted—injecting excessive amounts of reducing agent to ensure the outlet concentration meets standards. This not only increases reducing agent consumption but also leads to excessive ammonia injection, subsequently causing blockages in downstream equipment of the denitrification system and increased system operating power consumption. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, device, and electronic equipment for dynamic prediction of nitrogen oxide concentration in thermal power plant denitrification, addressing the aforementioned technical problem of difficulty in responding to drastic fluctuations in nitrogen oxide concentration in a timely and accurate manner.

[0005] This invention provides a method for dynamically predicting the concentration of nitrogen oxides in denitrification processes at thermal power plants, the method comprising: Obtain real-time process parameters of boiler operation; The real-time process parameters are preprocessed to form input data suitable for the prediction model, wherein the preprocessing includes at least adjusting the time delay of the process parameters based on flue gas velocity and instrument monitoring delay; Based on the operating condition factors that characterize the load change state of the unit, the preprocessed input data is identified as either a first operating condition type or a second operating condition type. Input data identified as the first working condition type is input into a first prediction model trained on a first dataset, and input data identified as the second working condition type is input into a second prediction model trained on a second dataset, wherein the first prediction model and the second prediction model are machine learning models of different types; Output the predicted value of nitrogen oxide concentration calculated by the first prediction model or the second prediction model.

[0006] In one embodiment, the preprocessing further includes a data cleaning process, specifically comprising: Time-stamped filtering removes duplicate data sequences caused by communication delays or sampling frequencies; Remove data sequences containing null values; Remove data sequences containing negative values, data sequences within the instrument calibration preset period, data sequences within the backflush preset period, paradox signal data sequences, or data sequences that repeat within the preset time window; Perform normalization on the data to eliminate dimensional differences between feature dimensions.

[0007] In one embodiment, the preprocessing further includes a feature selection process, specifically comprising: The input feature dimensions are determined from the real-time process parameters by combining mechanistic screening and mathematical screening based on process analysis. The mechanistic screening includes selecting features from coal parameters, wind parameters, environmental parameters and other parameters, while the mathematical screening includes determining the final features through correlation analysis, sensitivity analysis, weight ranking and cross-validation.

[0008] In one embodiment, the preprocessing further includes a data lightweighting process, specifically including: Filter out low-correlation signals and retain key signals for dimensionality reduction; The data sampling interval is adjusted from the first preset interval to a longer second preset interval to reduce computing resource requirements and speed up model inference.

[0009] In one embodiment, adjusting the timing delay of the process parameters includes: Based on the flue gas velocity model and the instrument monitoring delay model, combined with fluid dynamics simulation, the temporal phase difference of the process parameters is optimized to align the temporal relationship between the independent and dependent variables.

[0010] In one embodiment, identifying the preprocessed input data as a first operating condition type or a second operating condition type based on the operating condition factor characterizing the unit load change state includes: Based on at least one of the factors of furnace temperature, load band, and load change rate, the data is identified as a first operating condition type corresponding to stable load or gradually changing load, or a second operating condition type corresponding to rapidly changing load; wherein, the load change rate is compared with a preset change rate threshold to assist in identification.

[0011] In one embodiment, the first prediction model is a causal characteristic model used to process stable or slowly changing load data, and the second prediction model is a time-sensitive model used to process rapidly changing load data. The method further includes: The hyperparameters of the first and second prediction models were tuned using Bayesian optimization to accelerate model convergence.

[0012] In one embodiment, the method further includes: When the model prediction deviation continuously exceeds a preset deviation threshold over a continuous period of time, the model self-learning is initiated. During periods of load fluctuation, incremental training is performed using downsampled data to achieve rapid learning, while during periods of stable load, instrument measurements are used to replace the output and the model is retrained to achieve refined learning.

[0013] The present invention also provides a device for dynamic prediction of nitrogen oxide concentration in denitrification of thermal power plants, the device comprising: The parameter acquisition module is used to acquire real-time process parameters of the boiler operation. A parameter preprocessing module is used to preprocess the real-time process parameters to form input data suitable for the prediction model, wherein the preprocessing includes at least adjusting the time delay of the process parameters based on flue gas velocity and instrument monitoring delay; The type identification module is used to identify the preprocessed input data as a first operating condition type or a second operating condition type based on the operating condition factors that characterize the load change state of the unit. The data input module is used to input the input data identified as the first working condition type into the first prediction model trained on the first dataset, and to input the input data identified as the second working condition type into the second prediction model trained on the second dataset, wherein the first prediction model and the second prediction model are machine learning models of different types. The prediction output module is used to output the predicted value of nitrogen oxide concentration calculated by the first prediction model or the second prediction model.

[0014] The present invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the dynamic prediction method for nitrogen oxide concentration in thermal power plant denitrification as described above.

[0015] The aforementioned method, device, and electronic equipment for dynamic prediction of nitrogen oxide concentration in thermal power plant denitrification systems effectively overcome the detection delay problem caused by the dispersed location of measuring points and the lag in instrument response by acquiring real-time process parameters of boiler operation and performing preprocessing including time-series delay adjustment. This lays a data foundation for time-series alignment for accurate prediction. Furthermore, by introducing operating condition factors characterizing the unit load change state to identify the operating condition type of the preprocessed data, and selecting the data into different types, a hybrid prediction architecture dynamically matched with the operating conditions is constructed. This architecture can adapt to different states of unit load, such as smooth and gradual changes and rapid and abrupt changes, thereby significantly improving the prediction accuracy and real-time performance of nitrogen oxide concentration at the denitrification inlet under complex and variable operating conditions, especially when facing coal quality fluctuations and load command changes. Ultimately, this provides a direct and reliable decision-making basis for implementing forward-looking and precise ammonia injection control, thus solving the problems of inaccurate ammonia injection control and environmental risks caused by detection lag and drastic concentration fluctuations. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 A flowchart of a method for dynamically predicting nitrogen oxide concentrations in denitrification of thermal power plants, as shown in one embodiment; Figure 2 Flowchart of a method for dynamic prediction of nitrogen oxide concentration in denitrification of thermal power plants, as shown in another embodiment; Figure 3 A schematic diagram of a dynamic prediction device for nitrogen oxide concentration in denitrification of thermal power plants, according to one embodiment; Figure 4 This is an internal structural diagram of an electronic device according to one embodiment. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] The following is combined Figures 1-4The present invention describes a method, apparatus, and electronic device for dynamic prediction of nitrogen oxide concentration in thermal power plant denitrification processes.

[0020] like Figure 1 As shown, in one embodiment, a method for dynamically predicting the concentration of nitrogen oxides in denitrification systems of thermal power plants includes the following steps: Step S110: Obtain the real-time process parameters of the boiler operation.

[0021] The process parameters of boiler operation are collected in real time from the distributed control system, programmable logic controller, or plant-level monitoring information system via a data interface protocol. These parameters include coal-related parameters (such as feeder speed and coal feed rate), air-related parameters (such as primary air volume, secondary air volume, total air volume, primary air pressure, secondary air pressure, furnace pressure, forced draft fan opening, and induced draft fan opening), environmental parameters (such as atmospheric pressure, ambient temperature, and ambient humidity), and other parameters including real-time unit load, main steam flow rate, main steam temperature, main steam pressure, furnace temperature, economizer outlet oxygen content, denitrification reactor inlet nitrogen oxide concentration, and ammonia injection flow rate. This acquisition process synchronously acquires the above parameters through multi-threaded parallel acquisition and uses a timestamp alignment mechanism to ensure data consistency, thereby ensuring the comprehensiveness and real-time nature of the input data. This provides a complete and synchronous data foundation for subsequent accurate predictions. Furthermore, acquiring data directly from the control system avoids data silos and guarantees the authority and accuracy of the data source.

[0022] Step S120: Preprocess the real-time process parameters to form input data suitable for the prediction model. The preprocessing includes at least adjusting the time delay of the process parameters based on the flue gas velocity and instrument monitoring delay.

[0023] Preprocessing also includes a data cleaning process, specifically: filtering out duplicate data sequences caused by communication delays or abnormal sampling frequencies based on time labels to eliminate redundant information; removing data sequences containing null values ​​to ensure data integrity; further removing data sequences containing negative values, abnormal data sequences within a preset instrument calibration period, interference data sequences within a preset backflush period, paradoxical signal data sequences, and duplicate data sequences appearing within a preset time window to eliminate invalid and abnormal data interference; and simultaneously performing normalization processing on the data to eliminate dimensional differences between various feature dimensions and achieve data scale uniformity. This data cleaning process significantly improves the quality and consistency of the input data, enhances the input reliability of subsequent prediction models by systematically removing outliers and noisy data, and effectively avoids the problem of certain features dominating model training due to excessively large dimensions, providing standardized, high-quality input data for machine learning models.

[0024] Preprocessing also includes a feature selection process, specifically a combination of mechanistic and mathematical selection based on process analysis. This involves determining the input feature dimensions from real-time process parameters. Mechanistic selection includes choosing features from coal type parameters, wind type parameters, environmental parameters, and other parameters. Mathematical selection involves determining the final features through correlation analysis, sensitivity analysis, weight ranking, and cross-validation. Feature selection prioritizes wind over coal, limiting the input feature dimensions to a preset number to reduce the risk of model overfitting and enhance generalization ability. Specifically, the process begins with mechanistic screening based on process analysis. Coal feed rate and coal quality analysis data are selected from coal parameters; primary air volume, secondary air volume, total air volume, furnace pressure, and forced and induced draft fan opening are selected from air parameters; atmospheric pressure and ambient temperature are selected from environmental parameters; and unit load and furnace temperature are selected from other parameters as the initial feature set. Next, correlation analysis in mathematical screening calculates the Pearson correlation coefficient between each feature and nitrogen oxide concentration to eliminate low-correlation features. Sensitivity analysis uses a perturbation method to assess the sensitivity of feature changes to the output results. Weighted ranking is used to calculate feature importance scores based on random forest or gradient boosting algorithms, and cross-validation is combined to ensure the stability and generalization ability of the selected features. The entire screening process adheres to the basic principle of prioritizing air-related features over coal-related features, i.e., prioritizing the retention of air-related features and limiting the total number of features to a preset limit. This feature selection process ensures that the input features not only conform to the physicochemical principles of the combustion process but also have significant statistical correlation through a dual selection mechanism of mechanism and mathematics. It effectively reduces the feature dimensionality and eliminates redundant information, thereby significantly mitigating the risk of model overfitting and enhancing the model's generalization ability under different operating conditions. At the same time, by focusing on key wind parameters, it improves the model's response characteristics to changes in the combustion field and provides the optimal feature input combination for subsequent prediction models.

[0025] The process parameters are adjusted for time delays, including optimizing the time phase difference based on flue gas velocity models and instrument monitoring delay models, combined with fluid dynamics simulations, to align the time relationship between independent and dependent variables. Specifically, the transmission delay time of flue gas from the furnace to the monitoring point of the denitrification reactor is calculated based on the established flue gas velocity model, which comprehensively considers boiler load, flue structure dimensions, and real-time flue gas temperature and pressure parameters. Simultaneously, the response lag time of the nitrogen oxide analyzer is quantified based on the instrument monitoring delay model, which integrates the instrument sampling cycle, gas pretreatment time, and electrochemical sensor response characteristics. Furthermore, combined with computational fluid dynamics simulations, a three-dimensional flow field and concentration field simulation of the boiler flue system is performed to accurately obtain the gas phase transmission characteristics between each monitoring point. Finally, through the above multi-model fusion calculations, the time phase difference of the collected process parameters is dynamically compensated and optimized, ensuring precise alignment of the time relationship between all input independent variables and the output dependent variable, i.e., nitrogen oxide concentration. This timing delay adjustment process eliminates the timing mismatch problem caused by the physical delay of flue gas transmission and the lag in instrument response, establishing an accurate temporal causal relationship between the model input features and output labels. It significantly improves the timing consistency of the prediction model input data under dynamic operating conditions, especially when the load changes, and provides highly synchronized training samples and prediction inputs for subsequent machine learning models, thereby solving the prediction bias problem caused by timing misalignment.

[0026] Step S130: Based on the operating condition factors characterizing the load change state of the unit, the preprocessed input data is identified as either the first operating condition type or the second operating condition type.

[0027] Based on at least one factor among furnace temperature, load band, and load change rate, the data is identified as either a first operating condition type corresponding to stable or gradually changing load, or a second operating condition type corresponding to rapidly changing load. The load change rate is compared with a preset change rate threshold to aid in identification. Specifically, furnace temperature signals, unit load band parameters, and load change rate values ​​are acquired in real time. The accurate load change rate is obtained by calculating the load change gradient within a certain time window and compared in real time with a preset change rate threshold. When the load change rate is below the preset threshold and furnace temperature fluctuations are within a stable range, the current operating condition is identified as the first operating condition type corresponding to stable or gradually changing load. When the load change rate exceeds the preset threshold or furnace temperature fluctuates drastically, it is identified as the second operating condition type corresponding to rapidly changing load. This identification process enables accurate dynamic classification of boiler operating conditions. By using multi-parameter fusion judgment, it significantly improves the accuracy and reliability of operating condition identification, providing a precise basis for selecting an appropriate prediction model. This ensures that a causal characteristic model is used to guarantee prediction stability under stable operating conditions, while a time-sensitive model is used to capture dynamic characteristics under rapidly changing operating conditions, thereby improving the adaptability and prediction accuracy of nitrogen oxide concentration prediction under different operating conditions.

[0028] Step S140: Input data identified as the first working condition type is input into the first prediction model trained on the first dataset, and input data identified as the second working condition type is input into the second prediction model trained on the second dataset, wherein the first prediction model and the second prediction model are machine learning models of different types.

[0029] The first prediction model is a causal characteristic model, used to handle stable or slowly varying load data, while the second prediction model is a time-sensitive model, used to handle rapidly varying load data. Based on the load condition identification results, the corresponding model prediction channels are automatically activated. For the first load condition type corresponding to stable or slowly varying loads, the preprocessed data is input into the causal characteristic model, which is pre-trained using steady-state operating data. This model employs random forest or gradient boosting decision tree algorithms, focusing on learning the steady-state mapping relationship between input features and nitrogen oxide concentration. For the second load condition type corresponding to rapidly varying loads, the time-adjusted data is input into the time-sensitive model, which is pre-trained using dynamic load data. This model employs Long Short-Term Memory (LSTM) or Gated Recurrent Unit (GRU) algorithms, specifically capturing the lag correlation between time-series dynamic characteristics and concentration changes. Both models undergo hyperparameter tuning through Bayesian optimization, and the first and second datasets are independently constructed from historical data of different load conditions to ensure model specificity. This process achieves accurate matching between the prediction model and the operating conditions. By designing specialized models, it fully leverages the interpretability of causal models under steady-state conditions and the tracking capability of time series models under dynamic conditions. It resolves the technical contradiction that a single model cannot simultaneously achieve steady-state accuracy and dynamic response speed, enabling the system to obtain more stable prediction output when the load is stable and maintain higher prediction and tracking accuracy when the load changes drastically.

[0030] Bayesian optimization was used to tune the hyperparameters of the first and second prediction models to accelerate model convergence. Specifically, for the causal model and the time-sensitive model, their hyperparameter optimization objective functions were established respectively. A Gaussian process was constructed based on the Bayesian optimization framework as a surrogate model to fit the nonlinear relationship between the performance of the prediction model and the hyperparameters. The next set of hyperparameter combinations to be evaluated was iteratively selected by improving the expected acquisition function, and the posterior distribution of the surrogate model was updated after each iteration. This optimization process continued until the preset number of iterations or performance convergence threshold was reached, thereby obtaining the globally optimal hyperparameter configurations for the causal model and the time-sensitive model, respectively. This hyperparameter tuning process significantly improved the efficiency of hyperparameter optimization through the intelligent search strategy of Bayesian optimization, avoiding the waste of computational resources in traditional grid search and random search. It not only accelerated the training and convergence speed of the two prediction models, but also ensured that the models achieved optimal prediction performance under their respective applicable conditions. The causal model obtained better generalization ability under stable conditions, and the time-sensitive model showed more accurate time-series tracking characteristics under dynamic conditions.

[0031] Step S150: Output the predicted value of nitrogen oxide concentration calculated by the first prediction model or the second prediction model.

[0032] Based on the operating condition identification results, the system automatically selects the output of either the causal characteristic model or the time-sensitive model as the final predicted value. This predicted value is transmitted in real-time to the human-machine interface of the distributed control system for dynamic display via a data interface protocol. Simultaneously, the predicted data is written to a real-time database and a trend curve is generated. When the predicted value exceeds a preset alarm threshold, an audible and visual alarm signal is automatically triggered, and the predicted value is sent as a feedforward signal to the ammonia injection control system for denitrification optimization control. Under model self-learning trigger conditions, the system automatically compares the deviation between the predicted value and the actual instrument measurement value. When the deviation exceeds the limit continuously, the instrument measurement value is temporarily used to replace the output to ensure control reliability. This output process realizes real-time visualization of the prediction results and closed-loop control application. Direct access to the distributed control system ensures deep integration of predicted data and production control. The alarm mechanism effectively improves operational safety, and the feedforward signal significantly improves the accuracy and response speed of ammonia injection control. The intelligent output replacement mechanism ensures the continuous and stable operation of the system under abnormal model conditions, ultimately constructing a complete output closed loop integrating prediction, display, alarm, and control.

[0033] The dynamic prediction method for nitrogen oxide concentration in thermal power plant denitrification systems presented in this embodiment effectively improves the quality and temporal consistency of input data by adjusting the time-series delay of real-time process parameters and performing data cleaning. Feature selection reduces the risk of model overfitting and enhances generalization ability, while simultaneously reducing computational resource requirements and accelerating model inference speed. Dynamic identification based on operating condition factors and dual-model prediction significantly improves prediction accuracy and response speed under different load changes, especially under rapid load changes. Hyperparameter tuning through Bayesian optimization accelerates the model convergence process. Therefore, the overall accuracy, robustness, and practicality of nitrogen oxide concentration prediction are improved.

[0034] like Figure 2 As shown, in one embodiment, the preprocessing further includes a data lightweighting process, specifically comprising the following steps: Step S210: Filter out low-correlation signals and retain key signals for dimensionality reduction.

[0035] First, based on the feature importance scores and Pearson correlation coefficient analysis results obtained in the feature screening stage, low correlation signals with nitrogen oxide concentration prediction below a preset threshold are systematically filtered out, while all key signals are retained to achieve effective dimensionality reduction.

[0036] Step S220: Adjust the data sampling interval from the first preset interval to a longer second preset interval to reduce computing power resource requirements and speed up model inference.

[0037] The sampling interval of the data acquisition system is adjusted from the original first preset interval (usually a high-frequency acquisition of 1-5 seconds) to a longer second preset interval (usually an optimized acquisition of 10-30 seconds). This adjustment process is achieved by directly modifying the configuration parameters of the data acquisition module.

[0038] This embodiment significantly reduces data dimensionality and data flow through dual lightweighting processes. While retaining key process information, it reduces the amount of data processing by as much as 60%-80%. This not only greatly reduces the computing resources required for model training and predictive inference, enabling the system to run stably on ordinary industrial computing hardware, but also increases the model inference speed by more than 40%, meeting the timeliness requirements of the denitrification control system for real-time prediction. Furthermore, by eliminating noise signals and redundant data, it further enhances the model's generalization ability and predictive stability.

[0039] In one embodiment, when the model prediction deviation continuously exceeds a preset deviation threshold over a continuous time period, model self-learning is initiated. During periods of load fluctuation, incremental training using downsampled data is employed for rapid learning. During periods of stable load, instrument measurements are used to replace the output, and the model is retrained for refined learning. Specifically, during periods of load fluctuation, representative historical data is extracted from a real-time database using downsampling to form a training subset. Incremental training algorithms are then used to rapidly update the parameters of the prediction model for rapid learning. During periods of stable load, the model automatically switches to refined learning mode, temporarily replacing the model output with high-precision instrument measurements to ensure control loop stability. Simultaneously, high-quality operating data from this period is used to retrain the model with all parameters. The technical effect of this self-learning process is to achieve online dynamic optimization and continuous performance improvement of the prediction model. Through differentiated learning strategies, both the real-time nature of model updates under rapidly changing operating conditions and the accuracy of parameter optimization under steady-state conditions are ensured. This effectively overcomes model drift problems caused by changes in fuel characteristics, equipment aging, or environmental factors, ensuring that the prediction system always maintains the best match with the actual operating state, ultimately forming an intelligent prediction system with self-correcting capabilities.

[0040] The following describes the dynamic prediction device for nitrogen oxide concentration in thermal power plant denitrification provided by the present invention. The dynamic prediction device for nitrogen oxide concentration in thermal power plant denitrification described below can be referred to in correspondence with the dynamic prediction method for nitrogen oxide concentration in thermal power plant denitrification described above.

[0041] like Figure 3 As shown, in one embodiment, a dynamic prediction device for nitrogen oxide concentration in denitrification of thermal power plants includes a parameter acquisition module 310, a parameter preprocessing module 320, a type identification module 330, a data input module 340, and a prediction value output module 350.

[0042] The parameter acquisition module 310 is used to acquire real-time process parameters of the boiler operation.

[0043] The parameter preprocessing module 320 is used to preprocess the real-time process parameters to form input data suitable for the prediction model, wherein the preprocessing includes at least adjusting the time delay of the process parameters based on the flue gas velocity and instrument monitoring delay.

[0044] The type identification module 330 is used to identify the preprocessed input data as a first operating condition type or a second operating condition type based on the operating condition factors that characterize the load change state of the unit.

[0045] The data input module 340 is used to input the input data identified as the first working condition type into the first prediction model trained on the first dataset, and to input the input data identified as the second working condition type into the second prediction model trained on the second dataset, wherein the first prediction model and the second prediction model are machine learning models of different types.

[0046] The prediction output module 350 is used to output the predicted value of nitrogen oxide concentration calculated by the first prediction model or the second prediction model.

[0047] Figure 4 This example illustrates a schematic diagram of the physical structure of an electronic device, which can be a smart terminal. Its internal structure diagram can be as follows: Figure 4 As shown. The electronic device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for dynamically predicting the concentration of nitrogen oxides in thermal power plant denitrification processes. This method includes: Obtain real-time process parameters of boiler operation; The real-time process parameters are preprocessed to form input data suitable for the prediction model. The preprocessing includes at least adjusting the time delay of the process parameters based on the flue gas velocity and instrument monitoring delay. Based on the operating condition factors that characterize the load change state of the unit, the preprocessed input data is identified as either the first operating condition type or the second operating condition type. Input data identified as the first working condition type is input into the first prediction model trained on the first dataset, and input data identified as the second working condition type is input into the second prediction model trained on the second dataset, wherein the first prediction model and the second prediction model are machine learning models of different types; Output the predicted nitrogen oxide concentration calculated by the first or second prediction model.

[0048] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the electronic device to which the present invention is applied. A specific electronic device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0049] On the other hand, the present invention also provides a computer storage medium storing a computer program, which, when executed by a processor, implements a method for dynamically predicting the concentration of nitrogen oxides in thermal power plant denitrification processes. This method includes: Obtain real-time process parameters of boiler operation; The real-time process parameters are preprocessed to form input data suitable for the prediction model. The preprocessing includes at least adjusting the time delay of the process parameters based on the flue gas velocity and instrument monitoring delay. Based on the operating condition factors that characterize the load change state of the unit, the preprocessed input data is identified as either the first operating condition type or the second operating condition type. Input data identified as the first working condition type is input into the first prediction model trained on the first dataset, and input data identified as the second working condition type is input into the second prediction model trained on the second dataset, wherein the first prediction model and the second prediction model are machine learning models of different types; The predicted nitrogen oxide concentration calculated by the first or second prediction model is output. In another aspect, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer instructions from the computer-readable storage medium, and when the processor executes the computer instructions, it implements a method for dynamically predicting nitrogen oxide concentration in thermal power plant denitrification, the method comprising: Obtain real-time process parameters of boiler operation; The real-time process parameters are preprocessed to form input data suitable for the prediction model. The preprocessing includes at least adjusting the time delay of the process parameters based on the flue gas velocity and instrument monitoring delay. Based on the operating condition factors that characterize the load change state of the unit, the preprocessed input data is identified as either the first operating condition type or the second operating condition type. Input data identified as the first working condition type is input into the first prediction model trained on the first dataset, and input data identified as the second working condition type is input into the second prediction model trained on the second dataset, wherein the first prediction model and the second prediction model are machine learning models of different types; Output the predicted nitrogen oxide concentration calculated by the first or second prediction model.

[0050] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory.

[0051] By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0052] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0053] The above-described embodiments are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A method for dynamically predicting the concentration of nitrogen oxides in denitrification processes at thermal power plants, characterized in that, The method includes: Obtain real-time process parameters of boiler operation; The real-time process parameters are preprocessed to form input data suitable for the prediction model, wherein the preprocessing includes at least adjusting the time delay of the process parameters based on flue gas velocity and instrument monitoring delay; Based on the operating condition factors that characterize the load change state of the unit, the preprocessed input data is identified as either a first operating condition type or a second operating condition type. Input data identified as the first working condition type is input into a first prediction model trained on a first dataset, and input data identified as the second working condition type is input into a second prediction model trained on a second dataset, wherein the first prediction model and the second prediction model are machine learning models of different types; Output the predicted value of nitrogen oxide concentration calculated by the first prediction model or the second prediction model.

2. The method for dynamic prediction of nitrogen oxide concentration in thermal power plant denitrification according to claim 1, characterized in that, The preprocessing also includes a data cleaning process, specifically including: Time-stamped filtering removes duplicate data sequences caused by communication delays or sampling frequencies; Remove data sequences containing null values; Remove data sequences containing negative values, data sequences within the instrument calibration preset period, data sequences within the backflush preset period, paradox signal data sequences, or data sequences that repeat within the preset time window; Perform normalization on the data to eliminate dimensional differences between feature dimensions.

3. The method for dynamic prediction of nitrogen oxide concentration in thermal power plant denitrification according to claim 1, characterized in that, The preprocessing also includes a feature selection process, specifically including: The input feature dimensions are determined from the real-time process parameters by combining mechanistic screening and mathematical screening based on process analysis. The mechanistic screening includes selecting features from coal parameters, wind parameters, environmental parameters and other parameters, while the mathematical screening includes determining the final features through correlation analysis, sensitivity analysis, weight ranking and cross-validation.

4. The method for dynamic prediction of nitrogen oxide concentration in thermal power plant denitrification according to claim 1, characterized in that, The preprocessing also includes a data lightweighting process, specifically including: Filter out low-correlation signals and retain key signals for dimensionality reduction; The data sampling interval is adjusted from the first preset interval to a longer second preset interval to reduce computing resource requirements and speed up model inference.

5. The method for dynamic prediction of nitrogen oxide concentration in thermal power plant denitrification according to claim 1, characterized in that, The timing delay adjustment of the process parameters includes: Based on the flue gas velocity model and the instrument monitoring delay model, combined with fluid dynamics simulation, the temporal phase difference of the process parameters is optimized to align the temporal relationship between the independent and dependent variables.

6. The method for dynamic prediction of nitrogen oxide concentration in thermal power plant denitrification according to claim 1, characterized in that, The step of identifying the preprocessed input data into a first operating condition type or a second operating condition type based on the operating condition factors characterizing the unit load change state includes: Based on at least one of the factors of furnace temperature, load band, and load change rate, the data is identified as a first operating condition type corresponding to stable load or gradually changing load, or a second operating condition type corresponding to rapidly changing load; wherein, the load change rate is compared with a preset change rate threshold to assist in identification.

7. The method for dynamic prediction of nitrogen oxide concentration in thermal power plant denitrification according to claim 1, characterized in that, The first prediction model is a causal characteristic model used to handle stable or slowly changing load data; the second prediction model is a time-sensitive model used to handle rapidly changing load data; the method further includes: The hyperparameters of the first and second prediction models were tuned using Bayesian optimization to accelerate model convergence.

8. The method for dynamic prediction of nitrogen oxide concentration in thermal power plant denitrification according to any one of claims 1 to 7, characterized in that, The method further includes: When the model prediction deviation continuously exceeds a preset deviation threshold over a continuous period of time, the model self-learning is initiated. During periods of load fluctuation, incremental training is performed using downsampled data to achieve rapid learning, while during periods of stable load, instrument measurements are used to replace the output and the model is retrained to achieve refined learning.

9. A device for dynamically predicting the concentration of nitrogen oxides in denitrification systems of thermal power plants, characterized in that, The device includes: The parameter acquisition module is used to acquire real-time process parameters of the boiler operation. A parameter preprocessing module is used to preprocess the real-time process parameters to form input data suitable for the prediction model, wherein the preprocessing includes at least adjusting the time delay of the process parameters based on flue gas velocity and instrument monitoring delay; The type identification module is used to identify the preprocessed input data as a first operating condition type or a second operating condition type based on the operating condition factors that characterize the load change state of the unit. The data input module is used to input the input data identified as the first working condition type into the first prediction model trained on the first dataset, and to input the input data identified as the second working condition type into the second prediction model trained on the second dataset, wherein the first prediction model and the second prediction model are machine learning models of different types. The prediction output module is used to output the predicted value of nitrogen oxide concentration calculated by the first prediction model or the second prediction model.

10. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the dynamic prediction method for nitrogen oxide concentration in thermal power plant denitrification as described in any one of claims 1 to 8.

Images