Method for predicting NOx concentration at SCR inlet of coal-fired power plant based on lightweight Transform
By using a lightweight Transformer structure to retain multi-head self-attention and position coding at the encoder end, the nonlinearity and robustness issues of NOx concentration prediction at the SCR inlet of coal-fired power plants are solved, achieving high-precision and low-complexity prediction results.
Patent Information
- Application Number
- CN202511663760.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-13
- Publication Date
- 2026-02-10
AI Technical Summary
Existing methods for predicting NOx concentration at the SCR inlet of coal-fired power plants are insufficient to accurately describe the strongly nonlinear, multivariate coupled, and time-varying large lag processes of boiler systems. Furthermore, traditional Transformer models are less robust under high sampling rates and small sample or high noise conditions.
We adopt a lightweight Transformer structure, retain multi-head self-attention and positional encoding at the encoder end, remove the decoder, and project the output directly onto the encoder end, simplifying the model to sequence-to-vector, reducing computational complexity and memory requirements, while improving prediction accuracy through data preprocessing and feature selection.
It achieves high-precision prediction of NOx concentration at the SCR inlet under different operating conditions, reduces model size and computational complexity, and improves the robustness and adaptability of the prediction model.
Smart Images
Figure CN121503256A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of gas concentration detection technology, specifically to a lightweight Transformer-based NO inlet detector for SCR in coal-fired power plants. x Concentration prediction methods. Background Technology
[0002] Coal-fired power plants generate large amounts of nitrogen oxides during operation, which are a major source of air pollutants and seriously harm the environment. Controlling nitrogen oxide emissions from coal-fired power plants has been listed as a top priority for emission reduction. Accurate prediction of SCR inlet NO... x Emissions reduction helps improve denitrification efficiency and is of great significance for pollutant emission control and safe operation of coal-fired power plants.
[0003] However, the boiler system generates NO x The process is complex, involving strong nonlinearity, multivariate coupling, and time-varying large lag, making it difficult to accurately describe using specific mathematical models.
[0004] With the maturation of big data, intelligent computing power, and algorithm iteration, deep learning modeling and prediction has become a new research hotspot and mainstream trend. Currently, commonly used flue gas emission prediction schemes mainly rely on numerical simulation, engineering design, and deep learning methods for SCR systems. Examples include Long Short-Term Memory (LSTM) networks and Transformers. However, these prediction methods have drawbacks: the recursive structure limits their ability to capture long-term dependencies, and the serial computation mechanism restricts training speed, resulting in limited adaptability to complex dynamic conditions. Traditional Transformers, on the other hand, suffer from quadratic complexity, leading to memory bottlenecks in high sampling rate and ultra-long window scenarios, and exhibiting weak robustness to small sample sizes or high-noise conditions.
[0005] Currently, there is a lack of effective NOx at the SCR inlet of coal-fired power plants. x Concentration prediction methods can solve the above problems.
[0006] The information disclosed in this background section is intended only to enhance the understanding of the overall background of the invention and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention
[0007] The technical problem to be solved by this invention is to overcome the above-mentioned defects and provide a lightweight Transformer-based SCR inlet NOx for coal-fired power plants. x Concentration prediction methods.
[0008] To solve the above-mentioned technical problems, the technical solution provided by the present invention is as follows:
[0009] A lightweight Transformer-based SCR inlet NOx for coal-fired power plants x The concentration prediction method includes the following steps: S100, obtaining NO from the SCR inlet of a coal-fired power plant. x Concentration-related operating data of coal-fired boilers, the coal-fired power plant and SCR inlet NO x The concentration-related operating data of coal-fired boilers includes at least input variables and target variables; S200, using a preset data processing method, the acquired data of the coal-fired power plant and SCR inlet NO... x Preprocessing of concentration-related coal-fired boiler operating data yields preprocessed data; S300: Based on preset dataset partitioning rules, the preprocessed data is divided into training dataset, validation dataset, and test dataset; S400: Based on the partitioned training dataset, validation dataset, and test dataset, a preset lightweight Transformer model is trained; S500: The trained lightweight Transformer model is deployed as the SCR inlet NO of the coal-fired power plant. x Concentration prediction models to improve NO concentration at SCR inlet of coal-fired power plants x Accuracy of concentration prediction.
[0010] Optionally, the input variables include at least: coal feed rate, primary air volume, secondary air volume, air temperature, furnace negative pressure, oxygen content, economizer inlet and outlet flue gas temperature, main steam flow rate, main steam temperature, main steam pressure, reheat steam temperature, reheat steam pressure, feedwater flow rate, feedwater temperature, pulverizer current and inlet / outlet pressure difference, and active power; the target variable includes at least: measured SCR inlet NO. x Concentration data.
[0011] Optionally, in the process of using a preset data processing method, the obtained data of the coal-fired power plant and the SCR inlet NO... x Before obtaining the preprocessed data from the concentration-related coal-fired boiler operation data, the prediction method further includes: S210, traversing the dataset and searching... S220: Missing values and outliers at time points; S230: An outlier detection method based on statistical distribution is used to detect the outliers; S24: A two-stage missing value imputation method is used to impute the missing values, taking into account the characteristics of the time series.
[0012] Optionally, in the process of using a preset data processing method, the obtained data of the coal-fired power plant and the SCR inlet NO... xBefore obtaining the preprocessed data, the prediction method for coal-fired boiler operation data related to concentration includes: S240, using Spearman's rank correlation coefficient method, maximum mutual information coefficient, random forest and extreme gradient boosting tree methods to sort importance scores by average and combine them with SCR entry NO. x The mechanism involves selecting variables with higher importance scores as feature variables to reduce the dimensionality of multi-criteria fusion features.
[0013] Optionally, the outlier detection method based on statistical distribution is used to detect outliers, including: S221, for each feature variable... Calculate its mean and standard deviation Data points outside the specified range are identified as outliers. This process is expressed by the following formula:
[0014]
[0015] in, For the original data points, For the set of outliers, This is the threshold coefficient;
[0016] Optionally, considering the characteristics of the time series, a two-stage missing value imputation method is used to impute the missing values, including: S231, the first stage uses a rolling window to fill forward, for each missing time point... In the window The nearest valid value is filled in the search area, and this process is expressed by the following formula:
[0017]
[0018] in, Distance within the window The most recent valid value at any given time;
[0019] S232. The second stage uses intraday backward filling, while maintaining date boundaries, to fill missing values by maximizing the use of valid data within the same operation day. This process is expressed by the following formula:
[0020]
[0021] in, For the same date Inner distance The most recent valid value is used; if a large segment of data is missing or abnormal, it is discarded.
[0022] Optionally, the calculation formulas for the Spearman rank correlation coefficient method are as follows:
[0023]
[0024] in, The Spearman correlation coefficient. The number of sampling points. The difference in grade between the i-th data points;
[0025] The formulas for calculating the maximum mutual information coefficient are as follows:
[0026]
[0027] in, Here, x and y represent the upper bound of the grid complexity, and x and y represent the number of segments along the x and y axes, respectively. For mutual information under the G-grid;
[0028] The calculation formulas for the random forest method are as follows:
[0029]
[0030] in, For the prediction of the original sample by the b-th tree, To randomly shuffle the values of feature j in the sample before prediction;
[0031] The calculation formulas for the extreme gradient boosting tree method are as follows:
[0032]
[0033] in, The number of splits used for feature j. The optimal gain of the objective function resulting from the splitting.
[0034] Optionally, the step of training a pre-defined lightweight Transformer model based on the partitioned training dataset, validation dataset, and test dataset includes: S410, evaluating the model using an index function based on root mean square error and goodness of fit; S420, matching the coal-fired power plants in the training dataset with the SCR inlet NO... x Using concentration-related coal-fired boiler operating data as input data, the lightweight Transformer model is trained under a preset hyperparameter combination; S430, the input features of the test set are fed into the trained lightweight Transformer network model to obtain the SCR input NO. xS440: Quantitatively evaluate the model's prediction accuracy and generalization ability by calculating the error index between the predicted and measured values; S450: Repeat steps S420~S440 until the best evaluation index of the model is found, thereby obtaining the optimal hyperparameter combination and the corresponding lightweight Transformer model.
[0035] Optionally, the calculation process of the index function based on the root mean square error evaluation model is expressed by the following formula:
[0036]
[0037] in, and Let represent the predicted value and the actual value of the i-th sample, respectively, and n be the number of samples.
[0038] Optionally, the process of expressing the index function based on the goodness-of-fit evaluation model by the following formula is characterized by:
[0039]
[0040] in, and Let these represent the predicted value and the actual value of the i-th sample, respectively. is the actual mean, and n is the sample size.
[0041] The advantages of this invention are as follows: This invention proposes to use a lightweight Transformer structure, which retains multi-head self-attention and positional encoding to capture global long-range dependencies at the encoder end, while removing the decoder and directly projecting the output to the encoder end, simplifying the model from sequence to sequence to sequence to sequence to vector. This reduces the model size, memory usage, and computational complexity while ensuring prediction accuracy, and has great practical application value. Attached Figure Description
[0042] Figure 1 This is a flowchart of the overall steps provided in one embodiment of the present invention;
[0043] Figure 2 This is a flowchart of the sub-steps of S200 provided in one embodiment of the present invention;
[0044] Figure 3 This is a flowchart of the sub-steps of S220 provided in one embodiment of the present invention.
[0045] Figure 4 This is a flowchart of the sub-steps of S230 provided in one embodiment of the present invention.
[0046] Figure 5 This is a flowchart of the sub-steps of S400 provided in one embodiment of the present invention.
[0047] Figure 6 This is a block diagram of the lightweight Transformer network provided in one embodiment of the present invention.
[0048] Figure 7 This is one embodiment of the SCR inlet NO of a coal-fired power plant provided by the present invention. x Concentration prediction results graph. Detailed Implementation
[0049] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses consistent with some aspects of this disclosure as detailed in the appended claims.
[0050] As mentioned earlier, controlling nitrogen oxide emissions from coal-fired power plants has been listed as a top priority for emission reduction. Nitrogen oxides are one of the main sources of air pollutants and seriously harm the environment. Large amounts of nitrogen oxides are generated during the operation of coal-fired power plants. Currently, my country's coal-fired power plants generally use selective catalytic reduction (SCR) technology for flue gas denitrification treatment, accurately predicting the NO inlet temperature at the SCR inlet. x Emissions reduction helps improve denitrification efficiency and is of great significance for pollutant emission control and safe operation of coal-fired power plants.
[0051] Currently, commonly used flue gas emission prediction schemes mainly involve research on SCR systems based on methods such as numerical simulation, engineering design, and deep learning. Traditional methods based on numerical simulation and engineering design rely on explicit characterization of physical mechanisms and the establishment of equations for their modeling and solution processes.
[0052] However, the boiler system generates NO. x The process is complex, involving strong nonlinearity, multivariate coupling, and significant time-varying lag, making it difficult to accurately describe using specific mathematical models. With the maturation of big data, intelligent computing power, and algorithm iteration, using deep learning for modeling and prediction has become a new research hotspot and mainstream trend.
[0053] Traditional neural networks, such as Long Short-Term Memory (LSTM) networks, maintain a constant error during temporal backpropagation through gating mechanisms, significantly mitigating gradient vanishing and enabling effective modeling of long-range dependencies.
[0054] Compared to the temporal recursive paradigm of LSTM, Transformer establishes direct connections between arbitrary time steps through multi-head self-attention, capturing the global context in parallel. It has both high parallelism and scalability, which can alleviate the risks of gradient truncation and overfitting caused by the expansion of the hidden state dimension of LSTM, and has long-range modeling capabilities.
[0055] However, the quadratic complexity makes high sampling rate and ultra-long window scenarios face memory bottlenecks, and its robustness to small sample or strong noise conditions is weak.
[0056] To address the aforementioned problems, this invention proposes a lightweight Transformer-based SCR inlet NOx for coal-fired power plants. x A concentration prediction method is proposed, employing a lightweight Transformer structure. This method retains multi-head self-attention and positional encoding at the encoder end to capture global long-range dependencies, while removing the decoder and directly projecting the output onto the encoder end. This simplifies the model from sequence-to-sequence to sequence-to-vector, reducing model size, memory usage, and computational complexity while maintaining prediction accuracy, thus possessing significant practical application value. This invention addresses this issue through the following approach.
[0057] Example 1:
[0058] Please refer to the instruction manual appendix. Figures 1 to 7 As shown in the figure, this invention provides a lightweight Transformer for the NO inlet of a coal-fired power plant SCR. x Concentration prediction method, which includes the following steps:
[0059] S100, Obtain NO from coal-fired power plants and SCR inlet x Concentration-related operating data of coal-fired boilers, the coal-fired power plant and SCR inlet NO x Concentration-related operating data for coal-fired boilers should include at least input variables and target variables.
[0060] The input variables include at least: coal feed rate, primary air volume, secondary air volume, air temperature, furnace negative pressure, oxygen content, economizer inlet and outlet flue gas temperature, main steam flow rate, main steam temperature, main steam pressure, reheat steam temperature, reheat steam pressure, feedwater flow rate, feedwater temperature, pulverizer current and inlet / outlet pressure difference, and active power. The target variable includes at least: measured SCR inlet NO. x Concentration data.
[0061] For example, the dataset collected in this embodiment comes from a 600MW supercritical once-through boiler. It includes operating data of the coal-fired boiler from different months in 2021 (July, August, and September), each with a length of 10 days and a sampling interval of 5 seconds, along with measured SCR inlet NO. xConcentration data; coal-fired boiler operating data includes total coal quantity, total air volume, furnace negative pressure, SCR inlet oxygen content, SCR inlet flue gas temperature, main steam temperature and pressure, instantaneous flow rate of coal feeder (A, B, C, D, E, F) control cabinet, contact current of coal mill (A, B, C, D, E, F) motor, differential pressure of flue gas at inlet and outlet of air preheater A and primary and secondary air temperatures, pressure and air volume at inlet and outlet of blower A and active power, etc.
[0062] S200. Using a preset data processing method, the obtained data of the coal-fired power plant and SCR inlet NO... x Preprocessing of concentration-related coal-fired boiler operating data yields preprocessed data.
[0063] The preset data processing method may further include the following steps.
[0064] S210. Traverse the dataset and find missing and outlier values at any given time.
[0065] S220. An outlier detection is performed on the outlier using a statistical distribution-based outlier identification method.
[0066] In this first embodiment, step S220 may further include: S221, for each feature variable Calculate its mean and standard deviation Data points outside the specified range are identified as outliers. This process is expressed by the following formula:
[0067] (1)
[0068] in, For the original data points, For the set of outliers, This is the threshold coefficient.
[0069] S230. Based on the characteristics of time series, a two-stage missing value imputation method is used to impute the missing values.
[0070] In this first embodiment, step S230 may further include:
[0071] S231. In the first stage, a scrolling window is used for forward filling, for each missing time point. In the window Fill in the nearest valid value within the search (e.g., This process is expressed by the following formula:
[0072] (2)
[0073] in, Distance within the window The most recent valid value at any given time.
[0074] S232. The second stage uses intraday backward filling, while maintaining date boundaries, to fill missing values by maximizing the use of valid data within the same operation day. This process is expressed by the following formula:
[0075] (3)
[0076] in, For the same date Inner distance The most recent valid value at any given time (e.g., choosing) (If the data is from the same day, it will be discarded if a large amount of data is missing or abnormal).
[0077] S240 employs the Spearman rank correlation coefficient method, maximum mutual information coefficient, random forest, and extreme gradient boosting tree method to rank importance scores by average and combine them with the SCR entry NOx generation mechanism to select variables with higher importance scores as feature variables, thereby reducing the dimensionality of multi-criteria fusion features.
[0078] The calculation formulas for the Spearman rank correlation coefficient method are as follows:
[0079] (4)
[0080] in, The Spearman correlation coefficient. The number of sampling points. Let be the grade difference of the i-th data.
[0081] The formulas for calculating the maximum mutual information coefficient are as follows:
[0082] (5)
[0083] in, Here, x and y represent the upper bound of the grid complexity, and x and y represent the number of segments along the x and y axes, respectively. For mutual information under the G grid.
[0084] The calculation formulas for the random forest method are as follows:
[0085] (6)
[0086] in, For the prediction of the original sample by the b-th tree, To make predictions, the values of feature j in the sample are randomly shuffled.
[0087] The calculation formulas for the extreme gradient boosting tree method are as follows:
[0088] (7)
[0089] in, The number of splits used for feature j. The optimal gain of the objective function resulting from the splitting.
[0090] For example, in this embodiment one, the feature variables with the highest average importance scores among the four methods mentioned above and an average score greater than 0.3 are selected, combined with the SCR inlet NO of the coal-fired boiler system. x The generation mechanism determines the final characteristic variables.
[0091] S300. Based on the preset dataset partitioning rules, the obtained preprocessed data is divided into training dataset, validation dataset and test dataset.
[0092] For example, the preset dataset partitioning rules include: the training dataset accounts for 80% of the normalized data, the validation dataset accounts for 15% of the normalized data, and the remaining data is the test dataset (for example, the data from the first 8 days of the flight crew is selected as the training dataset, the last hour is the test dataset, and the remaining data in between is the validation dataset).
[0093] S400: Based on the partitioned training, validation, and test datasets, train the pre-defined lightweight Transformer model. This may include the following steps.
[0094] S410, Index function based on root mean square error and goodness-of-fit evaluation model.
[0095] The calculation process of the index function based on the root mean square error evaluation model is expressed by the following formula:
[0096] (8)
[0097] in, and Let n represent the predicted value and the actual value of the i-th sample, respectively, and n be the number of samples (e.g., 870).
[0098] The process of expressing the index function based on the goodness-of-fit evaluation model using the following formula is as follows:
[0099] (9)
[0100] in, and Let these represent the predicted value and the actual value of the i-th sample, respectively. This represents the actual mean, and n is the sample size (e.g., 870).
[0101] S420, Connect the coal-fired power plants in the training dataset with the SCR entry NO. x Using concentration-related coal-fired boiler operating data as input data, and under preset hyperparameter combinations (e.g., number of network layers, number of hidden units, learning rate, etc.), the lightweight Transformer model is trained (e.g., ...). Figure 6 (As shown).
[0102] S430. Feed the input features of the test set into the trained lightweight Transformer network model to obtain the SCR input NO. x Predicted concentration values.
[0103] S440. Quantitatively evaluate the prediction accuracy and generalization ability of the model by calculating the error index between the predicted value and the measured value (i.e., the aforementioned root mean square error and goodness of fit).
[0104] For example, in this embodiment, the optimal hyperparameter combination is as follows: model dimension is 32, number of attention heads is 4, number of layers is 3, feedforward network dimension is 64, dropout rate is 0.1, learning rate is 0.0005, number of training epochs is 50, and early stopping patience value is 10.
[0105] S450. Repeat steps S420 to S440 until the best evaluation metric for the model is found, thus obtaining the optimal combination of hyperparameters and the corresponding lightweight Transformer model.
[0106] S500, Deploy the trained lightweight Transformer model as the SCR inlet NO of a coal-fired power plant. x Concentration prediction models to improve NO concentration at SCR inlet of coal-fired power plants x Accuracy of concentration prediction.
[0107] Based on the obtained lightweight Transformer network, the SCR inlet NO of the unit under different historical periods and various operating conditions was analyzed. x Concentrations were predicted one by one, and the results were compared with the measured values to prove that the model could maintain reliable predictive performance in different operating periods.
[0108] For example, such as Figure 7 As shown, the SCR inlet NO x The actual concentration value is shown by the black dashed line; the SCR inlet NO... x The concentration prediction value is shown by the black solid line. The root mean square error and goodness of fit are 0.5075 and 0.9968, respectively. The results show that the method designed in this invention can accurately predict the NO at the SCR inlet. x concentration.
[0109] Therefore, in summary, the advantages of this invention compared to existing technologies are as follows: This invention firstly collects coal-fired boiler operating data and SCR inlet NO within a certain period of time based on a coal-fired power plant data acquisition system. x Concentration data; outlier and missing data are processed, and the dataset is divided into training, validation, and test sets; multi-criteria fusion feature reduction is performed on the feature variables, and variables with higher scores are selected as feature variables based on importance; then, root mean square error and goodness of fit are defined as evaluation index functions for the model; further, a lightweight Transformer is trained to obtain the SCR input NO. x Predicted concentration values; repeatedly train the model using a grid search method until the best evaluation index is obtained, thus obtaining the optimal hyperparameter combination and the corresponding lightweight Transformer network; based on the obtained lightweight Transformer network, analyze the SCR inlet NO under different historical periods and various operating conditions of the unit. x The concentration was predicted one by one, and the predictive performance of the model was demonstrated by comparing it with the measured values; this invention predicts the NO concentration at the SCR inlet. x The predicted concentration values closely match the actual values, meeting the accuracy requirements of closed-loop control and possessing significant practical application value.
[0110] It should be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.
[0111] Furthermore, the terms "system" and "network" are often used interchangeably in this paper. The term "and / or" in this paper merely describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " in this paper generally indicates that the preceding and following related objects have an "or" relationship.
[0112] It should be understood that, in the embodiments of the present invention, "B corresponding to A" means that B is associated with A, and B can be determined based on A. However, it should also be understood that determining B based on A does not mean that B is determined solely based on A; B can also be determined based on A and / or other information.
[0113] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.
[0114] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.
[0115] In the embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the couplings or direct couplings or communication connections shown or discussed may be indirect couplings or communication connections through some interfaces, apparatuses, or units, or they may be electrical, mechanical, or other forms of connection.
[0116] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of the present invention, depending on actual needs.
[0117] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.
[0118] From the above description of the embodiments, those skilled in the art will clearly understand that the present invention can be implemented in hardware, firmware, or a combination thereof. When implemented in software, the above-described functions can be stored in a computer-readable medium or transmitted as one or more instructions or code on a computer-readable medium. Computer-readable media include computer storage media and communication media, wherein communication media include any medium that facilitates the transmission of a computer program from one place to another. Storage media can be any available medium accessible to a computer. For example, but not limited to, computer-readable media can include RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code having the form of instructions or data structures and accessible to a computer. Furthermore, any connection can suitably be a computer-readable medium. For example, if the software is transmitted from a website, server, or other remote source using coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL), or wireless technologies such as infrared, radio, and microwave, then coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technologies such as infrared, radio, and microwave are included in the definition of the medium. As used in this invention, disk and disc include compressed optical discs (CDs), laser discs, optical discs, digital versatile discs (DVDs), floppy disks, and Blu-ray discs, wherein disks typically magnetically copy data, while discs optically copy data using lasers. The combinations described above should also be included within the scope of protection for computer-readable media.
[0119] In summary, the above description is merely a preferred embodiment of the technical solution of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A lightweight Transformer-based SCR inlet NOx for coal-fired power plants x Concentration prediction methods are characterized by, Includes the following steps: S100, Obtain NO from coal-fired power plants and SCR inlet x Concentration-related operating data of coal-fired boilers, the coal-fired power plant and SCR inlet NO x Concentration-related operating data for coal-fired boilers should include at least input variables and target variables; S200. Using a preset data processing method, the obtained data of the coal-fired power plant and SCR inlet NO... x Preprocessing of concentration-related coal-fired boiler operating data yields preprocessed data. S300. Based on the preset dataset partitioning rules, the obtained preprocessed data is divided into a training dataset, a validation dataset, and a test dataset. S400: Based on the partitioned training dataset, validation dataset, and test dataset, train the pre-defined lightweight Transformer model; S500, Deploy the trained lightweight Transformer model as the SCR inlet NO of a coal-fired power plant. x Concentration prediction models to improve NO concentration at SCR inlet of coal-fired power plants x Accuracy of concentration prediction.
2. The NO inlet of a coal-fired power plant SCR based on a lightweight Transformer as described in claim 1 x Concentration prediction method, characterized in that, The input variables include at least: coal feed rate, primary air volume, secondary air volume, air temperature, furnace negative pressure, oxygen content, economizer inlet and outlet flue gas temperature, main steam flow rate, main steam temperature, main steam pressure, reheat steam temperature, reheat steam pressure, feedwater flow rate, feedwater temperature, coal mill current and inlet / outlet pressure difference, and active power. The target variable includes at least: measured SCR inlet NO. x Concentration data.
3. The NO inlet of a coal-fired power plant SCR based on a lightweight Transformer as described in claim 1 x Concentration prediction method, characterized in that, The preset data processing method is used to process the obtained NO values of the coal-fired power plant and the SCR inlet. x Before obtaining the preprocessed data from the concentration-related coal-fired boiler operating data, the prediction method further includes: S210. Traverse the dataset and search. Missing and outlier values at any given time; S220. An outlier detection method based on statistical distribution is used to detect the outlier. S230. Based on the characteristics of time series, a two-stage missing value imputation method is used to impute the missing values.
4. The NO inlet of a coal-fired power plant SCR based on a lightweight Transformer as described in claim 3. x Concentration prediction method, characterized in that, The preset data processing method is used to process the obtained NO values of the coal-fired power plant and the SCR inlet. x Before obtaining the preprocessed data from the concentration-related coal-fired boiler operating data, the prediction method further includes: S240, using Spearman's rank correlation coefficient method, maximum mutual information coefficient, random forest and extreme gradient boosting tree methods to average importance scores and combine with SCR entry NO. x The mechanism involves selecting variables with higher importance scores as feature variables to reduce the dimensionality of multi-criteria fusion features.
5. The NO inlet of a coal-fired power plant based on a lightweight Transformer as described in claim 3. x Concentration prediction method, characterized in that, The method of using outlier identification based on statistical distribution to detect outliers includes: S221. For each feature variable Calculate its mean and standard deviation Data points outside the specified range are identified as outliers. This process is expressed by the following formula: in, For the original data points, For the set of outliers, This is the threshold coefficient.
6. The NO inlet of a coal-fired power plant based on a lightweight Transformer as described in claim 3 x Concentration prediction method, characterized in that, The method for imputing missing values, tailored to the characteristics of time series data, employs a two-stage missing value imputation approach, including: S231. In the first stage, a scrolling window is used for forward filling, for each missing time point. In the window The nearest valid value is filled in the search area, and this process is expressed by the following formula: in, Distance within the window The most recent valid value at any given time; S232. The second stage uses intraday backward filling, while maintaining date boundaries, to fill missing values by maximizing the use of valid data within the same operating day. This process is expressed by the following formula: in, For the same date Inner distance The most recent valid value is used; if a large segment of data is missing or abnormal, it is discarded.
7. The NO inlet of a coal-fired power plant based on a lightweight Transformer as described in claim 4 x Concentration prediction method, characterized in that, The formulas for calculating the Spearman rank correlation coefficient method are as follows: in, The Spearman correlation coefficient. The number of sampling points. The difference in grade between the i-th data points; The formulas for calculating the maximum mutual information coefficient are as follows: in, Here, x and y represent the upper bound of the grid complexity, and x and y represent the number of segments along the x and y axes, respectively. For mutual information under the G-grid; The calculation formulas for the random forest method are as follows: in, For the prediction of the original sample by the b-th tree, To randomly shuffle the values of feature j in the sample before prediction; The calculation formulas for the extreme gradient boosting tree method are as follows: in, The number of splits used for feature j. The optimal gain of the objective function resulting from the splitting.
8. The NO inlet of a coal-fired power plant SCR based on a lightweight Transformer as described in claim 1 x Concentration prediction method, characterized in that, The training of the pre-defined lightweight Transformer model based on the partitioned training, validation, and test datasets includes: S410, Index function based on root mean square error and goodness-of-fit evaluation model; S420, Connect the coal-fired power plants in the training dataset with the SCR entry NO. x The lightweight Transformer model is trained using concentration-related coal-fired boiler operating data as input data and under a preset combination of hyperparameters. S430. Feed the input features of the test set into the trained lightweight Transformer network model to obtain the SCR input NO. x Predicted concentration values; S440. Quantitatively evaluate the prediction accuracy and generalization ability of the model by calculating the error index between the predicted value and the measured value. S450. Repeat steps S420 to S440 until the best evaluation metric for the model is found, thus obtaining the optimal combination of hyperparameters and the corresponding lightweight Transformer model.
9. The NO inlet of a coal-fired power plant based on a lightweight Transformer as described in claim 8 x Concentration prediction method, characterized in that, The calculation process of the index function based on the root mean square error evaluation model is expressed by the following formula: in, and Let represent the predicted value and the actual value of the i-th sample, respectively, and n be the number of samples.
10. The NO inlet of a coal-fired power plant based on a lightweight Transformer as described in claim 8 x Concentration prediction method, characterized in that, The process of expressing the index function based on the goodness-of-fit evaluation model using the following formula is as follows: in, and Let these represent the predicted value and the actual value of the i-th sample, respectively. is the actual mean, and n is the sample size.