A method for predicting nitrogen oxide emission of a thermal power plant based on big data

Abstract

The present application relates to big data processing technology, specifically relates to a kind of based on big data's thermal power plant nitrogen oxide emission prediction method, including constructing and training nitrogen oxide emission prediction model;The data to be measured are input into the trained nitrogen oxide emission prediction model, and the prediction result is output;The nitrogen oxide emission prediction model includes time series encoding module, time series position encoding module, image encoding module, image position encoding module and fusion module;The present application is combined by various coding techniques and deep learning architecture, model can not only improve the fitting accuracy of known data, but also enhance the generalization ability to new situation, that is, unseen data can also make more accurate prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to big data processing technology, in particular to a power plant nitrogen oxide emission prediction method based on big data. BACKGROUND

[0002] The mainstream method of power plant nitrogen oxide emission prediction relies on statistical models and early machine learning algorithms, such as linear regression, decision tree, and random forest tools. These methods try to establish a mathematical model to depict the correlation between nitrogen oxide emissions and many factors (such as fuel type, combustion temperature, air-fuel ratio, load fluctuation, etc.) by deeply studying historical operation data. However, with the passage of time and technological progress, these traditional methods have exposed some significant limitations, including:

[0003] 1) Inability to deal with complex nonlinear relationships and dynamic changes: the interaction between power plant operating parameters is complex, and simple linear or fixed structure models cannot accurately simulate these deep relationships, affecting the accuracy of the prediction.

[0004] 2) Inadequate consideration of the continuity and dependence of nitrogen oxide emissions over time when dealing with time series data: in actual production processes, emission levels are dynamically adjusted in real time with various operating conditions, while traditional prediction models often lack the ability to effectively capture this long-term dependence and sudden changes.

[0005] 3) Ignoring the value of multi-element data: unstructured data such as infrared images can directly reflect the internal combustion state of the boiler, but it has not been fully utilized, which undoubtedly limits the understanding of the nitrogen oxide generation mechanism and the prediction accuracy of the model. SUMMARY

[0006] Considering the limitations of traditional methods in predicting complex, dynamic, and nonlinear data characteristics in power systems, and the inability to meet the precise prediction needs of modern power plants under the requirements of fine management and strict environmental protection, the present application provides a power plant nitrogen oxide emission prediction method based on big data, which constructs and trains a nitrogen oxide emission prediction model; input the test data into the trained nitrogen oxide emission prediction model, and output the prediction result; the nitrogen oxide emission prediction model includes a time series encoding module, a time series position encoding module, an image encoding module, an image position encoding module, and a fusion module;

[0007] The training process of the nitrogen oxide emission prediction model includes the following steps:

[0008] S1. Collect power plant equipment state data, including time series data and infrared image data;

[0009] S2. Divide the power plant equipment state data into multiple groups of sample data using 6 hours as a time window and 2 hours as a sliding step; for each group of sample data, use the previous 4 hours as feature extraction data and the last 2 hours as prediction data; each piece of feature extraction data includes a time series and an image sequence; the image sequence includes a combustion area infrared image sequence and a denitration reactor infrared image sequence;

[0010] S3. For each piece of feature extraction data, input the time series into a time series encoding module to obtain time series encoding features, and input the image sequence into an image encoding module to obtain image encoding features; the image encoding features include combustion area image encoding features and denitration reactor image encoding features;

[0011] S4. Input the time series encoding features and the image encoding features into a time series position encoding module and an image position encoding module respectively to obtain time series position encoding features and image position encoding features; the image position encoding features include combustion area image position encoding features and denitration reactor image position encoding features;

[0012] S5. Input the time series position encoding features and the image position encoding features into a fusion module to output prediction results;

[0013] S6. Calculate the training loss and back-propagate the training model parameters until the model parameters converge and the training is completed.

[0014] Further, in the data collection process of step S1:

[0015] S11. Collect a time series data every 5 minutes; each piece of time series data is composed of 24 fields, including a thermal power plant field, a power generation load field, a fuel type field, a fuel consumption field, a boiler operation state field, a combustion efficiency field, a flue gas temperature field, a flue gas flow rate field, a excess air coefficient field, an oxygen content field, a furnace negative pressure field, a NOx initial concentration in flue gas field, a denitration system inlet flue gas temperature field, a denitration system ammonia injection amount field, a denitration system operation mode field, a denitration system outlet flue gas temperature field, a denitration system outlet NOx concentration field, a chimney height field, a local wind speed field, a local wind direction field, an atmospheric stability category field, an environmental temperature field, an environmental humidity field, and a solar radiation intensity field;

[0016] S12. Collect an infrared image data every 10 minutes, and each piece of infrared image data includes a combustion area infrared image and a denitration reactor infrared image.

[0017] Further, the time sequence encoding module is a three-layer transformer-encoder model, and the input dimension and the output dimension thereof are 23 and 125 respectively; the image encoding module comprises a picture segmentation layer, a convolution layer, a fine-grained position encoding layer, a self-attention mechanism layer and a feedforward neural network.

[0018] Further, the process of inputting the image sequence into the image encoding module for processing comprises:

[0019] S31. For each combustion zone infrared image in the combustion zone infrared image sequence, the combustion zone infrared image is segmented into 16 sub-pictures by the picture segmentation layer and then input into the convolution layer for convolution processing, to output a sub-picture sequence; the sub-picture sequence is input into the fine-grained position encoding layer to obtain a sub-picture position vector sequence; the sub-picture position vector sequence is input into the self-attention mechanism layer to obtain a sub-picture attention vector sequence, and the feedforward neural network maps the sub-picture attention vector sequence to obtain a 256-dimensional combustion zone infrared image feature;

[0020] S32. The combustion zone infrared image features of all the combustion zone infrared images in the combustion zone infrared image sequence are combined to obtain a combustion zone image encoding feature;

[0021] S33. For each denitration reactor infrared image in the denitration reactor infrared image sequence, the denitration reactor infrared image is segmented into 16 sub-pictures by the picture segmentation layer and then input into the convolution layer for convolution processing, to output a sub-picture sequence; the sub-picture sequence is input into the fine-grained position encoding layer to obtain a sub-picture position vector sequence; the sub-picture position vector sequence is input into the self-attention mechanism layer to obtain a sub-picture attention vector sequence, and the feedforward neural network maps the sub-picture attention vector sequence to obtain a 256-dimensional denitration reactor infrared image feature;

[0022] S34. The denitration reactor infrared image features of all the denitration reactor infrared images in the denitration reactor infrared image sequence are combined to obtain a denitration reactor image encoding feature.

[0023] Further, the processing process of the time sequence position encoding module is represented as:

[0024]

[0025]

[0026] wherein PE (pos,i) represents the position encoding of the i-th parameter of the vector of the position pos in the time sequence encoding feature, f i represents the learnable frequency of the i-th parameter, and d represents the vector dimension.

[0027] Further, the calculation expression of the image position encoding module is:

[0028]

[0029] wherein P (pos,i) denotes the position encoding of the i-th parameter of the vector of the position pos in the image encoding feature, PPE denotes the final position encoding, P is the same as P (pos,i)

[0030] Further, the fusion module comprises three cascaded sub-structures, each of which comprises a first cross-attention layer, a second cross-attention layer and a unidirectional attention layer; wherein the processing flow of the first sub-structure comprises:

[0031] S51. Add the time sequence position encoding feature to the time sequence to obtain a time sequence feature;

[0032] S52. Add the combustion area image position encoding feature to the combustion area infrared image sequence to obtain a combustion area image sequence feature, and add the denitration reactor image position encoding feature to the denitration reactor infrared image sequence to obtain a denitration reactor image sequence feature;

[0033] S53. Input the combustion area image sequence feature and the time sequence feature into the first cross-attention layer to obtain a first time sequence fused with the combustion area image sequence feature;

[0034] S54. Input the first time sequence and the denitration reactor image sequence feature into the second cross-attention layer to obtain a second time sequence;

[0035] S55. Input the second time sequence into the unidirectional attention layer to obtain a sequence feature.

[0036] Further, the process of calculating the training loss comprises:

[0037] S61. Perform averaging operation on the time sequence position encoding feature, the combustion area image position encoding feature and the denitration reactor image position encoding feature respectively to obtain a time sequence sample representation vector, a combustion area image sample representation vector and a denitration reactor image sample representation vector;

[0038] S62. Add the time sequence sample representation vector, the combustion area image sample representation vector and the denitration reactor image sample representation vector after mapping respectively, calculate the loss between the addition result and the label by using the MSE loss function, and record it as the first;

[0039] S63. Calculate the similarity between the time sequence sample representation vector and the combustion area image sample representation vector to obtain a first similarity value; calculate the similarity between the time sequence sample representation vector and the denitration reactor image sample representation vector to obtain a second similarity value; add the first similarity value and the second similarity value to obtain a second loss;

[0040] S64. Use the MSE loss function to calculate the loss between the prediction result and the label, and record it as the third loss;

[0041] S65. Add the first loss, the second loss, and the third loss to obtain the training loss.

[0042] Beneficial effects of the present invention:

[0043] The present invention provides a method for predicting nitrogen oxide emissions from thermal power plants based on big data, which addresses the problem that traditional nitrogen oxide emission prediction models rely on a single type of data (such as time series data); the present invention integrates a time series coding module and an image coding module to simultaneously process and analyze multiple types of data, thereby providing richer and more comprehensive information to make predictions. The addition of the time series position coding module and the image position coding module enables the model to capture subtle differences in time and space in the data, which is crucial for accurately predicting nitrogen oxide emissions. In addition, nitrogen oxide emissions from thermal power plants are affected by many factors, including fuel type, combustion efficiency, equipment aging, etc. These factors may cause the data to present complex patterns. By fusing modules, especially the addition of intuitive image data, the model of the present invention is able to integrate information from different coding modules to better identify and learn these complex patterns.

[0044] Traditional models may have limitations in real-time monitoring and adapting to new data. By using deep learning technology, such models can continuously learn and adapt to new data patterns, thereby improving their effectiveness in real-time monitoring scenarios.

[0045] By combining multiple coding technologies and deep learning architectures, the model can not only improve the fitting accuracy of known data, but also enhance the generalization ability to new situations, that is, it can make more accurate predictions for unseen data.

[0046] Traditional prediction methods sometimes require a lot of manual intervention to adjust model parameters or select features. The present invention reduces the need for manual intervention by automatically learning and fusing different data sources. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a flow chart of the method of the present invention;

[0048] Figure 2 This is a structural diagram of the image encoding module of the present invention;

[0049] Figure 3 This is a schematic diagram of the fusion module flow of the present invention. DETAILED DESCRIPTION

[0050] With reference to the accompanying drawings, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all the other embodiments obtained by those skilled in the art without creative efforts belong to the scope of the present application.

[0051] The present application provides a thermal power plant nitrogen oxide emission prediction method based on big data, comprising constructing and training a nitrogen oxide emission prediction model; inputting the to-be-tested data into the trained nitrogen oxide emission prediction model, and outputting a prediction result; the nitrogen oxide emission prediction model comprises a time sequence encoding module, a time sequence position encoding module, an image encoding module, an image position encoding module and a fusion module.

[0052] As shown in the formula (1), the training process of the nitrogen oxide emission prediction model comprises the following steps: Figure 1

[0053] S1. Collecting power plant equipment state data, which comprises time sequence data and infrared image data.

[0054] Specifically, in the data collection process of step S1:

[0055] S11. Collecting time sequence data for recording the state of various equipment in the power generation process, which comprises a plurality of fields, including: collecting one piece of time sequence data every 5 minutes; each piece of time sequence data is composed of 24 fields, including a thermal power plant field (representing the ID of the thermal power plant), a power generation load field (representing the power generation load in MW), a fuel type field (representing the fuel type), a fuel consumption field (representing the fuel consumption in tons / hour), a boiler operation state field (representing the boiler operation state, having two states of on and off), a combustion efficiency field (representing the combustion efficiency, %), a flue gas temperature field (representing the flue gas temperature in ℃), a flue gas flow rate field (representing the flue gas flow rate in m 3 / s), an excess air coefficient field (representing the excess air coefficient), an oxygen content field (representing the oxygen content, %), a furnace negative pressure field (representing the furnace negative pressure in Pa), a NOx initial concentration field in flue gas (representing the NOx initial concentration in flue gas in mg / m 3 ​a denitration system outlet NOx concentration field (indicating the denitration system outlet NOx concentration in mg / m3); a chimney height field (indicating the chimney height in m); a local wind speed field (indicating the local wind speed in m / s); a local wind direction field (indicating the local wind direction angle); an atmospheric stability category field (indicating the atmospheric stability category, having 5 categories of A / B / C / D / E); an ambient temperature field (indicating the ambient temperature in °C); an ambient humidity field (indicating the ambient humidity, %); and a solar radiation intensity field (indicating the solar radiation intensity in W / m2). 3 a denitration system outlet NOx concentration field (indicating the denitration system outlet NOx concentration in mg / m3); a chimney height field (indicating the chimney height in m); a local wind speed field (indicating the local wind speed in m / s); a local wind direction field (indicating the local wind direction angle); an atmospheric stability category field (indicating the atmospheric stability category, having 5 categories of A / B / C / D / E); an ambient temperature field (indicating the ambient temperature in °C); an ambient humidity field (indicating the ambient humidity, %); and a solar radiation intensity field (indicating the solar radiation intensity in W / m2). 2 a denitration system outlet NOx concentration field (indicating the denitration system outlet NOx concentration in mg / m3); a chimney height field (indicating the chimney height in m); a local wind speed field (indicating the local wind speed in m / s); a local wind direction field (indicating the local wind direction angle); an atmospheric stability category field (indicating the atmospheric stability category, having 5 categories of A / B / C / D / E); an ambient temperature field (indicating the ambient temperature in °C); an ambient humidity field (indicating the ambient humidity, %); and a solar radiation intensity field (indicating the solar radiation intensity in W / m2).

[0056] S12. Collect one infrared image data every 10 minutes, and each infrared image data includes one combustion area infrared image and one denitration reactor infrared image.

[0057] S2. Divide the power plant equipment state data into multiple groups of sample data using 6 hours as a time window and 2 hours as a sliding step; for each group of sample data, use the previous 4 hours as feature extraction data and the next 2 hours as prediction data. Each piece of feature extraction data includes a time sequence and an image sequence; the image sequence includes a combustion area infrared image sequence and a denitration reactor infrared image sequence.

[0058] Specifically, each group of sample data is composed of two sequences, one of which is a time sequence and the other of which is an image sequence; the time sequence is a 24xM matrix formed by arranging all time sequence data in the time interval of 6 hours corresponding to the group of sample data in time order, and M represents the number of time sequence data; the image sequence is a 2xN matrix formed by arranging all infrared image data in the time interval of 6 hours corresponding to the group of sample data in time order, and N represents the number of infrared image data; specifically, the image sequence is composed of a 1xN size combustion area infrared image sequence and a 1xN size denitration reactor infrared image sequence. When further divided into a feature extraction interval of the previous 4 hours and a prediction interval of the next 2 hours, the time sequence is divided into two time sequences according to the time boundary of the feature extraction interval and the prediction interval, and the image sequence is also divided into two image sequences according to the time boundary of the feature extraction interval and the prediction interval.

[0059] For example, if the start time of a group of sample data is 1:00, then its end time is 7:00. The time interval corresponding to this group of sample data is 1:00 to 7:00, of which the time interval corresponding to the feature extraction data is 1:00 to 5:00, and the time interval corresponding to the prediction data is 5:00 to 7:00; and the start time of the next sample data after this group of sample data is 3:00, and so on.

[0060] S3. For each feature extraction data, the time series is input into the time series coding module to obtain a time series coding feature, and the external image sequence is input into the image coding module to obtain an image coding feature; the image coding feature includes the combustion area image coding feature and the denitrification reactor image coding feature.

[0061] Specifically, the temporal encoding module is a three-layer transformer-encoder model, whose input dimension and output dimension are 23 and 125 respectively; Figure 2 As shown in the figure, the image encoding module includes a picture segmentation layer, a convolutional encoding layer, a fine-grained position encoding layer, a self-attention mechanism layer, and a feedforward neural network. The input dimension of the time series encoding module is 23 because the thermal power plant field in the time series data is deleted before input.

[0062] Specifically, the process of inputting the image sequence into the image encoding module for processing includes:

[0063] S31. For each combustion area infrared image in the combustion area infrared image sequence, perform the following operations:

[0064] The convolutional layer of the image segmentation layer divides the infrared image of the burning area into 16 sub-images; for example, a 256×256 image is divided into 16 16×16 sub-images;

[0065] The 16 sub-graphs are convolutionally encoded through the convolutional coding layer to obtain the sub-graph sequence X = (x1, x2, ..., x 16 ), x i represents the i=1, 2,…, 16th subgraph vector;

[0066] The sub-image sequence is input into the fine-grained position encoding layer, and the fine-grained position encoding layer randomly initializes an embedding sequence E = (e1, e2, ..., e 16 ), add the embedding sequence and the sub-graph encoding vector sequence to obtain the sub-graph position vector sequence Z = (z1, z2, ..., z 16 ) and then output; e i represents the i=1, 2, ..., 16th embedding, z i represents the position vector of the i=1, 2, ..., 16th subgraph;

[0067] The subgraph position vector sequence is obtained through a self-attention mechanism layer to obtain a subgraph attention vector sequence, and a feedforward neural network is used to map the subgraph attention vector sequence to obtain a 256-dimensional combustion region infrared image feature;

[0068] S32. Combining the combustion region infrared image features of all combustion region infrared images in the combustion region infrared image sequence to obtain a combustion region image coding feature represents the combustion region infrared image feature of the jth combustion region infrared image, and N1 represents the number of images in the combustion region infrared image sequence.

[0069] S33. For each denitration reactor infrared image in the denitration reactor infrared image sequence, the same operation as step S31 is performed to obtain the corresponding denitration reactor infrared image feature; and the denitration reactor infrared image features of all denitration reactor infrared images in the denitration reactor infrared image sequence are combined to obtain a denitration reactor coding feature represents the denitration reactor infrared image feature of the jth denitration reactor infrared image, and N1 represents the number of images in the denitration reactor infrared image sequence.

[0070] S4. The time sequence coding feature and the image coding feature are respectively input into a time sequence position coding module and an image position coding module to obtain a time sequence position coding feature and an image position coding feature; the image position coding feature includes a combustion region image position coding feature and a denitration reactor image position coding feature.

[0071] Specifically, the processing process of the time sequence position coding module is represented as:

[0072]

[0073] wherein, PPE (pos,i) represents the position coding of the i-th parameter of the vector at position pos in the time sequence coding feature, pos represents the position (i.e. the actual position of the vector in the time sequence coding feature), f i represents the learnable frequency of the i-th parameter, and d represents the vector dimension (i.e. the dimension of the vector).

[0074] Specifically, the calculation expression of the image position coding module is:

[0075]

[0076]

[0077] wherein, PPE (pos,i) represents the position coding of the i-th parameter of the vector at position pos in the image coding feature, PPE represents the image position coding feature, and P represents the image coding feature position coding.

[0078] S5. inputting the time sequence position coding feature and the image position coding feature into a fusion module to output a prediction result.

[0079] Specifically, the fusion module includes three cascaded substructures, each of which includes a first cross-attention layer, a second cross-attention layer and a unidirectional attention layer; as shown in the figure, the processing flow of the first substructure includes: Figure 3

[0080] S51. adding the time sequence position coding feature to the time sequence to obtain a time sequence feature containing position information;

[0081] S52. adding the combustion area image position coding feature to the combustion area infrared image sequence to obtain a combustion area image sequence feature containing position information; adding the denitration reactor image position coding feature to the denitration reactor infrared image sequence to obtain a denitration reactor image sequence feature containing position information;

[0082] S53. inputting the combustion area image sequence feature and the time sequence feature into the first cross-attention layer to obtain a first time sequence fused with the combustion area image sequence feature;

[0083] S54. inputting the first time sequence obtained in S53 and the denitration reactor image sequence feature into the second cross-attention layer to obtain a second time sequence fused with the two image features;

[0084] S55. inputting the second time sequence obtained in S54 into the unidirectional attention layer to obtain a sequence feature.

[0085] Since image information is difficult to be fused only through one structure, two identical substructures are added subsequently for image information fusion.

[0086] S6. calculating a training loss and back-propagating a training model parameter until the model parameter converges and the training is completed.

[0087] Specifically, the process of calculating the training loss includes:

[0088] S61. respectively performing an averaging operation on the time sequence position coding feature, the combustion area image position coding feature and the denitration reactor image position coding feature to obtain a time sequence sample feature vector, a combustion area image sample feature vector and a denitration reactor image sample feature vector;

[0089] S62. respectively mapping the time sequence sample feature vector, the combustion area image sample feature vector and the denitration reactor image sample feature vector to the dimension of the prediction step number, then adding them, calculating the loss between the added result and the label by using an MSE loss function, and recording it as the first; ​

[0090] S63. Calculate the similarity between the time sequence sample feature vector and the combustion area image sample feature vector to obtain a first similarity value; calculate the similarity between the time sequence sample feature vector and the denitration reactor image sample feature vector to obtain a second similarity value; add the first similarity value and the second similarity value to obtain a second loss;

[0091] S64. Calculate the loss between the prediction result and the label by using the MSE loss function, and record it as a third loss;

[0092] S65. Add the first loss, the second loss and the third loss to obtain a training loss.

[0093] In the present application, unless otherwise explicitly specified and limited, the terms "mounting", "setting", "connecting", "fixing", "rotating" and the like should be understood in a broad sense, for example, can be fixedly connected, or can be detachably connected, or can be integrated; can be mechanically connected, or can be electrically connected; can be directly connected, or can be indirectly connected through an intermediate medium; can be the internal communication of two elements or the interaction relationship between two elements, unless otherwise explicitly limited, the above-mentioned terms in the present application can be understood according to the specific meaning of the above-mentioned terms in the present application by those skilled in the art according to the specific circumstances.

[0094] Although the embodiments of the present application have been shown and described, it can be understood by those skilled in the art that various changes, modifications, replacements and variations can be made to the embodiments without departing from the principles and spirits of the present application, the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for predicting nitrogen oxide emissions of a thermal power plant based on big data, the method comprising: collecting data of the thermal power plant; and predicting nitrogen oxide emissions of the thermal power plant based on the collected data. The nitrogen oxide emission prediction model is constructed and trained; the to-be-tested data is input into the trained nitrogen oxide emission prediction model, and a prediction result is output; the nitrogen oxide emission prediction model comprises a time sequence coding module, a time sequence position coding module, an image coding module, an image position coding module and a fusion module; The training process of the nitrogen oxide emission prediction model comprises the following steps: S1. Collecting power plant equipment state data, which comprises time sequence data and infrared image data; S2. Dividing the power plant equipment state data into multiple groups of sample data by adopting 6 hours as a time window and 2 hours as a sliding step; for each group of sample data, the previous 4 hours are used as feature extraction data, and the last 2 hours are used as prediction data; each piece of feature extraction data comprises a time sequence and an image sequence; the image sequence comprises a combustion area infrared image sequence and a denitration reactor infrared image sequence; S3. For each piece of feature extraction data, the time sequence is input into the time sequence coding module to obtain time sequence coding features, and the image sequence is input into the image coding module to obtain image coding features; the image coding features comprise combustion area image coding features and denitration reactor image coding features; S4. The time sequence coding features and the image coding features are input into the time sequence position coding module and the image position coding module respectively to obtain time sequence position coding features and image position coding features; the image position coding features comprise combustion area image position coding features and denitration reactor image position coding features; S5. The time sequence position coding features and the image position coding features are input into the fusion module to output prediction results; The fusion module comprises three cascaded substructures, each substructure comprising a first cross-attention layer, a second cross-attention layer and a one-way attention layer; wherein the processing procedure of the first substructure comprises: S51. Adding the time sequence position coding features to the time sequence to obtain time sequence features; S52. Adding the combustion area image position coding features to the combustion area infrared image sequence to obtain combustion area image sequence features, and adding the denitration reactor image position coding features to the denitration reactor infrared image sequence to obtain denitration reactor image sequence features; S53. Inputting the combustion area image sequence features and the time sequence features into the first cross-attention layer to obtain a first time sequence fused with the combustion area image sequence features; S54. Inputting the first time sequence and the denitration reactor image sequence features into the second cross-attention layer to obtain a second time sequence; S55. Inputting the second time sequence into the one-way attention layer to obtain a sequence feature; S6. Calculating the training loss and back-propagating the training model parameters until the model parameters converge and the training is completed.

2. The method of claim 1, wherein the method is based on big data. In the data collection process of step S1: S11. Collecting one time series data every 5 minutes; each time series data is composed of 24 fields, including the field of thermal power plant, the field of power generation load, the field of fuel type, the field of fuel consumption, the field of boiler operation state, the field of combustion efficiency, the field of flue gas temperature, the field of flue gas flow rate, the field of excess air coefficient, the field of oxygen content, the field of furnace negative pressure, the field of initial concentration of NOx in flue gas, the field of flue gas temperature at the inlet of denitration system, the field of ammonia injection amount of denitration system, the field of operation mode of denitration system, the field of flue gas temperature at the outlet of denitration system, the field of NOx concentration at the outlet of denitration system, the field of chimney height, the field of local wind speed, the field of local wind direction, the field of atmospheric stability category, the field of environmental temperature, the field of environmental humidity, and the field of solar radiation intensity; S12. Collecting one infrared image data every 10 minutes, each infrared image data including one infrared image of combustion area and one infrared image of denitration reactor.

3. The method of claim 1, wherein the method is based on big data. The time series encoding module is a three-layer transformer-encoder model, and the input dimension and the output dimension thereof are 23 and 125 respectively; the image encoding module includes a picture segmentation layer, a convolution layer, a fine-grained position encoding layer, a self-attention mechanism layer and a feedforward neural network.

4. The method of claim 3, wherein the method is based on big data. The process of inputting the image sequence into the image encoding module for processing includes: S31. For each infrared image of combustion area in the infrared image sequence of combustion area, the infrared image of combustion area is segmented into 16 sub-images by the picture segmentation layer and then input into the convolution layer for convolution processing, outputting a sub-image sequence; the sub-image sequence is input into the fine-grained position encoding layer to obtain a sub-image position vector sequence; the sub-image position vector sequence is input into the self-attention mechanism layer to obtain a sub-image attention vector sequence, and the feedforward neural network maps the sub-image attention vector sequence to obtain a 256-dimensional infrared image feature of combustion area; S32. Combining the infrared image features of combustion area of all infrared images of combustion area in the infrared image sequence of combustion area to obtain a combustion area image encoding feature; S33. For each infrared image of denitration reactor in the infrared image sequence of denitration reactor, the infrared image of denitration reactor is segmented into 16 sub-images by the picture segmentation layer and then input into the convolution layer for convolution processing, outputting a sub-image sequence; the sub-image sequence is input into the fine-grained position encoding layer to obtain a sub-image position vector sequence; the sub-image position vector sequence is input into the self-attention mechanism layer to obtain a sub-image attention vector sequence, and the feedforward neural network maps the sub-image attention vector sequence to obtain a 256-dimensional infrared image feature of denitration reactor; S34. Combining the infrared image features of denitration reactor of all infrared images of denitration reactor in the infrared image sequence of denitration reactor to obtain a denitration reactor image encoding feature.

5. The method for predicting nitrogen oxide emissions of a thermal power plant based on big data according to claim 1, characterized in that, The processing process of the time series position encoding module is represented as: where PE (pos,i) denotes the position encoding of the i-th parameter of the vector representing the position pos in the time series encoding feature, f i denotes the learnable frequency of the i-th parameter, d denotes the vector dimension.

6. The method for predicting nitrogen oxide emissions of a thermal power plant based on big data according to claim 1, wherein, The calculation expression of the image position encoding module is: where P (pos,i) denotes the position encoding of the i-th parameter of the vector representing the position pos in the image encoding feature, PPE denotes the image position encoding feature, and P denotes the image encoding feature position encoding.

7. The method of predicting nitrogen oxide emissions from a thermal power plant based on big data according to claim 1, wherein, The process of calculating the training loss includes: S61. Performing an average operation on the time series position encoding feature, the combustion area image position encoding feature and the denitration reactor image position encoding feature respectively to obtain a time series sample representation vector, a combustion area image sample representation vector and a denitration reactor image sample representation vector; S62. Add the mapped time series sample feature vectors, combustion area image sample feature vectors, and denitration reactor image sample feature vectors, respectively, calculate the loss between the added result and the label using the MSE loss function, and record it as the first loss; S63. Calculate the similarity between the time series sample feature vectors and the combustion area image sample feature vectors to obtain a first similarity value; calculate the similarity between the time series sample feature vectors and the denitration reactor image sample feature vectors to obtain a second similarity value; add the first similarity value and the second similarity value to obtain a second loss; S64. Calculate the loss between the prediction result and the label using the MSE loss function, and record it as the third loss; S65. Add the first loss, the second loss, and the third loss to obtain the training loss.

Images