Method and system for predicting carbon dioxide emission concentration of thermal power plant

Abstract

The embodiment of the invention provides a method and system for predicting the carbon dioxide emission concentration of a thermal power plant, and the method comprises the steps: collecting the fuel data and environment data of the thermal power plant, and constructing a training data set and a test data set; a carbon dioxide emission concentration prediction model is constructed, the front half part of the model comprises a feature preprocessing module, a regional feature extractor, a channel feature extractor and a feature fusion module, and the rear half part of the model comprises a multi-scale attention module, a parallel attention module and a self-adaptive feature fusion module; defining a loss function, training the constructed carbon dioxide emission concentration prediction model through the training data set, updating the model until a convergence condition is reached, and obtaining a trained carbon dioxide emission concentration prediction model; and inputting data in the test data set into a trained carbon dioxide emission concentration prediction model to predict a corresponding real-time carbon dioxide concentration value.

Description

Technical Field

[0001] This document relates to the field of carbon emission detection technology, and in particular to a method and system for predicting carbon dioxide emission concentration in a thermal power plant. Background Art

[0002] In the process of modern industrialization, thermal power plants are one of the main energy production methods, and their carbon dioxide (CO2) emissions have increasingly attracted people's attention. With the intensification of global climate change issues, countries have formulated strict environmental protection laws and regulations, requiring effective control and real-time monitoring of industrial emissions. Among them, monitoring the concentration of carbon dioxide in the flue of thermal power plants is one of the key steps to ensure that emissions meet environmental standards.

[0003] Traditional carbon dioxide emission monitoring methods usually rely on manual sampling and offline analysis. Although this method can reflect the emission situation to a certain extent, it is difficult to meet the needs of modern thermal power plants for high-precision and real-time monitoring due to its inability to achieve real-time monitoring, low monitoring frequency and cumbersome operation. Non-dispersive infrared (NDIR) technology has been widely used in the field of carbon dioxide concentration monitoring in recent years due to its fast, accurate and stable characteristics. NDIR sensors can measure the concentration of carbon dioxide based on the difference in the infrared light absorption characteristics of gases without contacting the measured gas. This technology provides important technical support for the online monitoring of carbon dioxide in the flue of thermal power plants. However, the existing carbon dioxide monitoring system based on NDIR technology still has some problems in practical applications. On the one hand, the current system mainly relies on statistical models and simple regression analysis, such as time series analysis and multivariate regression models. These methods can predict the trend of carbon dioxide concentration to a certain extent, but fail to fully consider the complex interaction between multiple influencing factors (such as fuel type and quality, air excess coefficient, temperature, humidity, atmospheric pressure and combustion efficiency, etc.), resulting in low accuracy of the prediction results. On the other hand, existing systems usually only use a single infrared band for measurement and fail to fully utilize multi-band information to correct and compensate for measurement errors, thus affecting the system's adaptability and anti-interference ability under complex working conditions.

[0004] Therefore, in order to solve the above problems, it is necessary to develop a more comprehensive monitoring method that can integrate variables that affect the measurement, such as fuel type and quality, air excess coefficient, temperature, humidity, atmospheric pressure and combustion efficiency, and use multiple infrared bands for measurement at the same time to enhance the system's anti-interference ability and environmental adaptability, and build a high-precision carbon dioxide concentration prediction model to provide strong support for environmental protection and climate change response. Summary of the invention

[0005] One or more embodiments of this specification provide a method for predicting carbon dioxide emission concentration in a thermal power plant, including:

[0006] Collect fuel data and environmental data of thermal power plants to build training and test data sets;

[0007] Construct a carbon dioxide emission concentration prediction model. The first half of the model includes a feature preprocessing module, a regional feature extractor, a channel feature extractor, and a feature fusion module. The second half includes a multi-scale attention module, a parallel attention module, and an adaptive feature fusion module.

[0008] Define the loss function, train the constructed carbon dioxide emission concentration prediction model through the training data set, update the model until the convergence condition is reached, and obtain the trained carbon dioxide emission concentration prediction model;

[0009] The data in the test data set is input into the trained carbon dioxide emission concentration prediction model to predict the corresponding real-time carbon dioxide concentration value.

[0010] Furthermore, the fuel data includes fuel type, fuel quality and combustion efficiency; the environmental data includes air excess coefficient, temperature, humidity, atmospheric pressure and light intensity absorption values ​​in multiple bands.

[0011] Furthermore, the feature preprocessing module uses a one-dimensional convolution layer to perform a convolution operation on the input one-dimensional time series data to extract local features; uses a reshaping layer to rearrange the local features extracted by the one-dimensional convolution layer into a feature matrix; processes and analyzes the rearranged feature matrix through a two-dimensional convolution layer and a Sigmoid activation function to extract higher-level features and generate an initial feature map.

[0012] Furthermore, the regional feature extractor is used to encode the spatial correlation in the input initial feature map, specifically: firstly, extract features through two blocks consisting of a two-dimensional convolutional layer, a Sigmoid activation function and an average pooling layer; convolve all input bands through the two-dimensional convolutional layer to extract features in the local neighborhood; use the flattening layer to flatten the output features into a spatial feature vector; wherein the average pooling layer kernel is 2×2 and the stride is 2; the kernel size of all convolutional layers in the regional feature extractor is 3×3, the stride is 1, and a padding method of 1 unit is adopted;

[0013] The channel feature extractor includes two one-dimensional convolutional layers, a Sigmoid activation function and a flattening layer, and the features extracted by the one-dimensional convolutional layer are flattened by the flattening layer to obtain a channel feature vector;

[0014] The feature fusion module mixes the channel feature vector and the spatial feature vector through weight coefficients, and on the basis of predicting the mixing weight, forms the final combined feature through channel dimension splicing.

[0015] Furthermore, the multi-scale attention module adopts multiple 1×1 filter groups, which are interconnected in a hierarchical manner; the outputs of all filters are integrated to obtain filter output features, the filter output features are subjected to a maximum pooling, and then the pooled features are subjected to two-dimensional convolution and Sigmoid activation function operations, the number of channels of the pooled features are reduced and restored respectively, and finally the output features are multiplied by the filter output features and added to the original input features of the multi-scale attention module to obtain the output features of the multi-scale attention module.

[0016] Furthermore, the parallel self-attention module includes a spatial attention module and a channel attention module; the spatial attention module is used to calculate the contextual information of spatial features in the horizontal and vertical directions; the channel attention module is used to enhance the feature representation by enhancing the interdependence between feature channels; local features are separated by the spatial attention module and the channel attention module, and feature fusion is performed by adaptively learning different weights.

[0017] Furthermore, the loss function is defined, and the constructed carbon dioxide emission concentration prediction model is trained through the training data set, specifically including:

[0018] Use the flattening layer, fully connected layer and ReLU activation function to transform the obtained features into the final prediction results

[0019] Calculate the final prediction result The MAE loss value of the true sequence y is as follows:

[0020]

[0021] in and q Sequence and the value of y at time q, P represents the total time of the data set;

[0022] Backpropagation is performed based on the loss value to train and update the model.

[0023] One or more embodiments of this specification provide a system for predicting carbon dioxide emission concentration in a thermal power plant, including:

[0024] Data acquisition module: used to collect fuel data and environmental data of thermal power plants and build training data sets and test data sets;

[0025] Model building module: used to build a carbon dioxide emission concentration prediction model. The first half of the model includes a feature preprocessing module, a regional feature extractor, a channel feature extractor, and a feature fusion module. The second half includes a multi-scale attention module, a parallel attention module, and an adaptive feature fusion module.

[0026] Model training module: used to define the loss function, train the constructed carbon dioxide emission concentration prediction model through the training data set, update the model until the convergence condition is reached, and obtain the trained carbon dioxide emission concentration prediction model;

[0027] Concentration prediction module: used to input the data in the test data set into the trained carbon dioxide emission concentration prediction model to predict the corresponding real-time carbon dioxide concentration value.

[0028] One or more embodiments of the present specification provide an electronic device, including:

[0029] processor; and,

[0030] A memory arranged to store computer executable instructions, which, when executed, enable the processor to implement the steps of the above-mentioned method for predicting carbon dioxide emission concentration in a thermal power plant.

[0031] One or more embodiments of the present specification provide a storage medium for storing computer-executable instructions, which, when executed, implement the steps of the above-mentioned method for predicting carbon dioxide emission concentration in a thermal power plant.

[0032] By adopting the embodiment of the present invention, the model's ability to capture information at different scales is enhanced by extracting features in a hierarchical manner, thereby improving the performance and generalization ability of the model in monitoring carbon dioxide emission concentrations. Feature representation is enhanced by combining the global dependencies of space and channels in parallel, thereby improving the effect of carbon dioxide emission concentration prediction. The present invention can adaptively predict the contribution of each feature to the carbon dioxide emission concentration prediction task, thereby improving the prediction accuracy of the model.

[0033] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented according to the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate one or more embodiments of this specification or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0035] Figure 1 A flow chart of a method for predicting carbon dioxide emission concentration in a thermal power plant provided in one or more embodiments of this specification;

[0036] Figure 2 A schematic diagram of a network model for monitoring carbon dioxide emission concentration in a method for predicting carbon dioxide emission concentration in a thermal power plant provided in one or more embodiments of this specification;

[0037] Figure 3 A schematic diagram of a characteristic weight prediction mechanism of a method for predicting carbon dioxide emission concentration in a thermal power plant provided in one or more embodiments of this specification;

[0038] Figure 4 A schematic diagram of a feature fusion module of a method for predicting carbon dioxide emission concentration in a thermal power plant provided in one or more embodiments of this specification;

[0039] Figure 5 A schematic diagram of a multi-scale attention module of a method for predicting carbon dioxide emission concentration in a thermal power plant provided by one or more embodiments of this specification;

[0040] Figure 6 A schematic diagram of a parallel self-attention module of a method for predicting carbon dioxide emission concentration in a thermal power plant provided by one or more embodiments of this specification;

[0041] Figure 7 A schematic diagram of the composition of a system for predicting carbon dioxide emission concentration in a thermal power plant provided for one or more embodiments of this specification;

[0042] Figure 8 A schematic diagram of the structure of an electronic device provided for one or more embodiments of this specification. DETAILED DESCRIPTION

[0043] In order to enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the following will be combined with the drawings in one or more embodiments of this specification to clearly and completely describe the technical solutions in one or more embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this specification, not all of the embodiments. Based on one or more embodiments of this specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this document.

[0044] Method Embodiment

[0045] According to an embodiment of the present invention, a method for predicting carbon dioxide emission concentration in a thermal power plant is provided. Figure 1 One or more embodiments of this specification provide a flow chart of a method for predicting carbon dioxide emission concentration in a thermal power plant, such as Figure 1 As shown, the method for predicting carbon dioxide emission concentration of a thermal power plant according to an embodiment of the present invention specifically includes:

[0046] S1. Collect fuel data and environmental data of thermal power plants and build training data sets and test data sets.

[0047] Prepare training data sets and test data sets to ensure the diversity of training samples so that the model can learn the characteristics and potential changes of the data. The input variables include but are not limited to fuel data and environmental data: fuel data includes fuel type, fuel quality and combustion efficiency, etc.; the environmental data includes air excess coefficient, temperature, humidity, atmospheric pressure and light intensity absorption values ​​in multiple bands, etc. The output variable is carbon dioxide concentration; each input variable and output variable here is a time series.

[0048] S2. Construct a carbon dioxide emission concentration prediction model. The first half of the model includes a feature preprocessing module, a regional feature extractor, a channel feature extractor and a feature fusion module, and the second half includes a multi-scale attention module, a parallel attention module and an adaptive feature fusion module.

[0049] Figure 2The schematic diagram of the network model for monitoring carbon dioxide emission concentration in the present invention is shown in Figure 1. The first half of the network consists of four core components: Feature preprocessing module, Regional Feature Extractor, Channel Feature Extractor and Feature Fusion Module (FFM). First, the feature preprocessing module generates an initial feature map. Subsequently, the regional feature extractor and the channel feature extractor extract the spatial and channel features of the input feature map along different paths, respectively, and provide input data for the feature fusion module FFM. In the FFM module, the extracted features are fused to adaptively predict the contribution of each feature to the target task. The second half of the network is mainly composed of three core components: Multi-Scale Attention Module (MSA), Parallel Self-Attention Module (PSA) and Adaptive Feature Fusion Module (AFF). The multi-scale attention module is used to improve the performance and generalization ability of the model. The parallel self-attention module enhances the feature representation by combining the global dependencies of space and channels in parallel. Then, the adaptive feature fusion module fuses the features of the downsampling layer and the upsampling layer to obtain the enhanced features. Finally, the flatten layer, the fully connected layer (Dense) and the ReLU activation function are used to transform the obtained features into the final prediction results. (time series value).

[0050] The feature preprocessing module takes multiple time series as input. First, a one-dimensional convolution layer (Conv1D) is used to perform a convolution operation on the input one-dimensional time series data to extract local features. Then, a reshape layer (Reshape) is used to rearrange the features extracted by the one-dimensional convolution layer into a two-dimensional matrix. Finally, the reshaped feature matrix is ​​processed and analyzed by a two-dimensional convolution layer (Conv2D) and a Sigmoid activation function to extract higher-level features and generate an initial feature map. The calculation process is as follows:

[0051] F1=Reshape(Conv1d(input));

[0052] F2 = σ(Conv2d(F1));

[0053] Among them, input represents the input feature of the feature preprocessing module, Conv1d represents the one-dimensional convolution layer, Reshape represents the reshaping layer, σ represents the Sigmoid activation function, Conv2d represents the two-dimensional convolution layer, F1 represents the intermediate output of the feature preprocessing module, and F2 represents the final output of the feature preprocessing module.

[0054] The structure of the region feature extractor is as follows: Figure 2 As shown in the lower left branch, its main function is to encode the spatial correlation in the input features to improve the accuracy of carbon dioxide emission concentration prediction. To incorporate spatial context information, the regional feature extractor first extracts features through two blocks consisting of a two-dimensional convolutional layer (Conv2D), a Sigmoid activation function, and an average pooling layer (Average Pooling). The kernel size of all convolutional layers in the regional feature extractor is 3×3, the stride is 1, and a padding method of 1 unit is used (that is, padding around the input boundary). The two-dimensional convolutional layer extracts features within the local neighborhood by convolving all input bands. In order to obtain position-invariant features, the regional feature extractor applies an average pooling layer with a kernel of 2×2 and a stride of 2 in two spatial dimensions. Finally, the output is flattened into a spatial feature vector f using a flattening layer (Flatten). spa The calculation process is as follows:

[0055] f spa =Flatten(AP(σ(Conv2d(F2))));

[0056] Among them, F2 represents the output of the feature preprocessing module, Conv2d represents the two-dimensional convolutional layer, σ represents the Sigmoid activation function, AP represents the average pooling layer, Flatten represents the flattening layer, and f spa represents the spatial feature vector.

[0057] The structure of the channel feature extractor is as follows Figure 2 The upper left branch shows that in order to extract local channel features, the channel feature extractor applies a one-dimensional convolutional layer (Conv1D) with a unit step along the channel dimension. The channel feature extractor consists of two blocks consisting of a one-dimensional convolutional layer and a Sigmoid activation function, and a flattening layer (Flatten), resulting in a channel feature vector f cha The calculation process is as follows:

[0058] f cha = Flatten(σ(Conv1d(F2)));

[0059] Among them, F2 represents the output of the feature preprocessing module, Conv1d represents the one-dimensional convolutional layer, σ represents the Sigmoid activation function, Flatten represents the flattening layer, and f cha represents the channel feature vector.

[0060] Figure 3 This is a schematic diagram of the feature weight prediction mechanism of the present invention, which is designed as a two-layer neural network. In the first layer, the input spatial feature vector f spa and channel feature vector f cha They are projected into two vectors of the same dimension by separate linear operators. Let U spa and b spa The weight matrix and bias vector of the linear layer used to process spatial features, U cha and b cha Then they represent the corresponding parameters for processing channel features. Then, the projected features are processed by the nonlinear activation function σ (Sigmoid activation function). The output of the first layer can be expressed as σ(U spa f spa +b spa ) andσ(U cha f cha +b cha ). In the second layer, the input vector is passed through the ELM extreme learning machine (ELM) to obtain a new feature vector. The scalar score of the feature weight prediction can be expressed as:

[0061] s spa =ELM(σ(U spa f spa +b spa ))

[0062] s cha =ELM(σ(U cha f cha +b cha ))

[0063] Among them, U spa and b spa Represents the weight matrix and bias vector of the linear layer used to process spatial features respectively; U cha and b cha represents the weight matrix and bias vector of the linear layer used to process channel features respectively; σ represents the Sigmoid activation function; ELM represents the extreme learning machine; f spa and f cha Represent the spatial eigenvector and channel eigenvector respectively; s spa and chaThey represent the scalar scores of spatial features and channel features of the feature weight prediction mechanism respectively.

[0064] Finally, a softmax operation is used to convert the scalar score of each feature into a weight coefficient, calculated as follows:

[0065]

[0066]

[0067] Among them, s spa and cha They represent the scalar scores of spatial features and channel features of the feature weight prediction mechanism, respectively, and p spa and p cha They respectively represent the weight coefficients of the spatial features and channel features of the feature weight prediction mechanism, and exp represents an exponential function with the natural constant e as the base.

[0068] Figure 4 FIG. 1 is a schematic diagram of the feature fusion module of the present invention. In this structure, the fusion module combines channel and spatial features through weight coefficient p cha and p spa Mixing, based on the predicted mixing weights, forms the final combined features through channel dimension splicing.

[0069] First, the present invention converts the input spatial feature fspa and channel feature f cha Converted into output features g respectively spa and g cha The middle layer between the input and output features contains a block consisting of a two-dimensional convolutional layer (Conv2D), a batch normalization layer (BatchNormalization, BN) and a Sigmoid activation function, and a block consisting of a two-dimensional convolutional layer and a batch normalization layer. The Conv2D+BN+Sigmoid block converts the input f spa and f cha The purpose of the Conv2D+BN block is to normalize the two features to the same value range before connecting them, preventing one feature from dominating and encouraging the contribution of both features. spa and g cha Connect along the channel dimension and form the final spatial channel feature through a fully connected layer (Dense). The combined spatial channel feature is expressed as:

[0070] f comb =Dense([p spa g spa ;p cha g cha ])

[0071] Among them; represents the feature vector connection representation along the channel dimension, Dense represents the fully connected layer, and g spa and g cha Represent the output features of space and channel respectively, p spa and p cha They represent the weight coefficients of spatial features and channel features of the feature weight prediction mechanism, respectively, and f comb Represents the final spatial channel features of the feature fusion module.

[0072] In order to better predict the concentration of carbon dioxide emissions, the model needs to have stronger multi-scale feature representation capabilities. This paper proposes a multi-scale attention module (MSA). Figure 5 This is a schematic diagram of the multi-scale attention module of the present invention, which constructs hierarchical residual connections in a single residual block, represents multi-scale features at a granular level, and increases the receptive field range of each network layer, thereby enhancing the feature representation capability.

[0073] MSA uses different structures to represent multi-scale features. Compared with traditional methods, MSA uses more 1×1 filter groups instead of a single 3×3 filter to extract multi-scale features. These different filter groups are connected to each other in a hierarchical manner, such as Figure 5 As shown in Figure 2, MSA first divides the 1×1 convolution feature map evenly into 5 feature subsets, and the feature subsets are denoted as x i , i = [1, 2, 3, 4, 5]. Except for x1, each feature subset corresponds to a filter f i . Subset the feature x i With the output y of the previous layer i-1 Add together as filter f i The filter f i The output y i The definition is as follows:

[0074]

[0075] Each filter f of the MSA i All information of the i-1th layer features before segmentation will be received to obtain a larger receptive field. Finally, by integrating all y i , i = [1, 2, 3, 4, 5] to obtain the output feature C. The process is expressed as follows:

[0076] C=Concat(y1, y2, y3, y4, y5);

[0077] Among them, y i , i = [1, 2, 3, 4, 5] represents the filter f iThe output of y, Concat represents the integration operation along the channel dimension, and C represents the integration of all y i The output features are obtained. This multi-scale processing method can not only extract local and global information, but also improve the efficiency of convolution processing by segmenting and integrating channels.

[0078] In the prediction of carbon dioxide emission concentration, the attention mechanism captures information by locating key areas in the features. SENet (Squeeze-and-Excitation Networks) uses a compression-expansion strategy to generate channel weight mappings to select channel features that need to be paid attention to. This strategy is called a channel attention mechanism. The channel attention mechanism can more effectively guide the network to focus on key areas. The present invention combines the channel attention mechanism with MSA, so that MSA has both powerful multi-scale feature extraction capabilities and feature selection capabilities of the channel attention mechanism. Specifically, first, feature C is subjected to a maximum pooling operation to obtain feature Z, and then two Conv2D+Sigmoid operations are performed to reduce and restore the number of channels of Z respectively. Finally, the output feature is multiplied by C to enhance its representation capability, and the original input features of the MSA module are added to obtain the output features of the MSA module. The process is expressed as follows:

[0079] Z = MaxPool(C);

[0080] MSA output =σ[H i σ(H r Z)]×C+MSA input ;

[0081] Where C represents the integration of all y along the channel dimension i , i = [1, 2, 3, 4, 5], MaxPool represents the maximum pooling layer, Z represents the output feature obtained by performing a maximum pooling on feature C, σ represents the Sigmoid activation function, H r represents a 1×1 2D convolutional layer with reduction ratio r, H i Represents a 1×1 two-dimensional convolutional layer that restores the feature dimension, × represents element multiplication, + represents element addition, MSA input , MSA output Represent the input and output features of the MSA module respectively.

[0082] One challenge in predicting CO2 emission concentration is the uneven distribution of local features across different channels. Most CNN-based models use the same processing method for local features across different channels, which does not work well for complex data patterns. To address these issues, this paper proposes a parallel self-attention module that separates local features by space and channel, and adaptively learns different weights, so that the network can better handle the complexity in the data. Figure 6 This is a schematic diagram of the parallel self-attention module of the present invention, which is mainly composed of a spatial attention module and a channel attention module in parallel.

[0083] First, define the spatial attention module:

[0084] In the scenario of predicting CO2 emission concentration, understanding the global features of the scene is crucial for accurate prediction, and long-range context information helps capture these features. Therefore, the present invention introduces a spatial attention module to calculate the context information of spatial features, which enhances the pixel-level representation through lightweight calculations. Unlike traditional self-attention, spatial attention calculates context information in the horizontal and vertical directions, rather than context information between all pixels.

[0085] The input feature of the feature map is recorded as F∈R C×W×H , C×W×H represent the number of channels, width and height of the input features respectively, and then perform 1×1 two-dimensional convolution to generate {Q, K}∈R C′×W×H , where Q and K represent the feature maps of the compressed channels, The definition of spatial matrix operation is as follows:

[0086]

[0087] Among them, j i.p ∈R (H+W-1)×W×H is the feature map Q p With S i，p The correlation coefficient between p Represents the value of feature map Q at position p. S i，p ∈R C′ YesS p The i-th element of S p ∈R (H+W-1)×C′ It is a set of vectors in the same row or column at position p in K. Q and K represent the feature mapping of the compressed channel of the input feature F. T represents the matrix transpose operation, and × represents the matrix multiplication operation. Then, j is multiplied in the channel dimension. i.p Apply a softmax layer to compute the attention map A.

[0088] Next, a 1×1 convolution filter is performed on the input feature F to obtain the feature map V. p ∈R C Represents the value of the position p of the feature map V in the spatial dimension. Through the aggregation operation, we can obtain feature dependencies over longer distances and enhance the local features F p Pixel-level representation of:

[0089]

[0090] Among them, E p is the output feature vector at position p, E∈R C×W×H . A i，p represents the value at position p and channel i in the attention map A, and A represents the attention to j in the channel dimension i.p The attention map calculated by applying the softmax layer. W i，p W p The value at channel i, W p ∈R (H+W-1)×C represents the set of vectors in V that are in the same row or column as position p. V represents the feature map obtained by performing a 1×1 convolution filter on the input feature F. F p is the value of position p in the input feature F, + is the element addition operation, and × is the matrix multiplication operation. Finally, E p The output of the spatial attention module is obtained by performing a Conv2D operation as input. The spatial attention map collects interactive information between different feature regions, namely context information, which helps to improve the adaptability in the prediction of carbon dioxide emission concentration.

[0091] Then define the channel attention module (Channel Attention):

[0092] In the process of predicting CO2 emission concentration, different feature channels may contain unbalanced information, which may be weakened or distorted due to the complexity of the data. Therefore, it is key to implicitly enhance the feature representation by enhancing the interdependence between feature channels. Different from the spatial attention module, the channel attention module first converts the input feature F∈R C×W×H Reshape to H∈R C×N , and then calculate the channel dependency matrix B. The following is the definition of the channel matrix operation:

[0093] B=H×H T ;

[0094] Among them, H represents the feature obtained by reshaping the input feature F of the feature map, × is a matrix multiplication operation, T represents a matrix transpose operation, and B represents a channel dependency matrix. Then, B calculates the channel attention map feature M∈R by using a softmax layer C×C Finally, we get the output feature E∈R C×W×H :

[0095] M = softmax(B);

[0096] E = β(M × H) + F;

[0097] Among them, B represents the channel dependency matrix, softmax represents the calculation using the softmax layer, M represents the channel attention mapping feature, β is the learnable scale parameter, H represents the feature obtained after the input feature F of the feature map is reshaped, F represents the input feature of the feature map, × is the matrix multiplication operation, + is the element addition operation, and E represents the output feature. Finally, a Conv2D operation is performed to obtain the output of the channel attention module. In this way, the interdependence between feature channels is enhanced, thereby improving the feature representation.

[0098] Finally, define the adaptive feature fusion module (Adaptive Feature Fusion):

[0099] In order to more effectively train deeper convolutional networks, the present invention adopts shortcut connections and different layer feature fusion methods to alleviate the problem of feature degradation. In the present invention, in order to make up for the deficiency of feature fusion, an adaptive hybrid operation is used to fuse the features of the downsampling layer and the upsampling layer. The specific formula is as follows:

[0100] f ↑i+1 =σ(λ i )*f ↓i +σ(1-λ i )*f ↑i ;

[0101] Among them, f ↓i represents the downsampled feature map, f ↑i represents the upsampled feature map, f ↑i+1 Represents the downsampling feature map f ↓i And the upsampled feature map f ↑i The output features are obtained by weighted fusion, i represents the i-th layer of the feature map (i=1, 2, ...), σ is a sigmoid operator constrained in the range of (0, 1), * represents element multiplication, + represents element addition, and λ i is a learnable parameter. This adaptive hybrid operation can more flexibly combine features of different scales, making feature fusion more complete, thereby improving the expressiveness and performance of the model.

[0102] S3. Define the loss function, train the constructed carbon dioxide emission concentration prediction model through the training data set, update the model until the convergence condition is reached, and obtain the trained carbon dioxide emission concentration prediction model.

[0103] Use the flattening layer, fully connected layer and ReLU activation function to transform the obtained features into the final prediction results

[0104] Calculate the final prediction result The MAE loss value of the true sequence y is as follows:

[0105]

[0106] in and q Sequence and the value of y at time q, P represents the total time of the data set;

[0107] The model is trained and updated by backpropagation according to the loss value until the convergence condition is reached to obtain a trained carbon dioxide emission concentration prediction model.

[0108] S4. Input the data in the test data set into the trained carbon dioxide emission concentration prediction model to predict the corresponding real-time carbon dioxide concentration value.

[0109] The beneficial effects of the present invention are as follows:

[0110] By adopting the embodiment of the present invention, the model's ability to capture information at different scales is enhanced by extracting features in a hierarchical manner, thereby improving the performance and generalization ability of the model in monitoring carbon dioxide emission concentrations. Feature representation is enhanced by combining the global dependencies of space and channels in parallel, thereby improving the effect of carbon dioxide emission concentration prediction. The present invention can adaptively predict the contribution of each feature to the carbon dioxide emission concentration prediction task, thereby improving the prediction accuracy of the model.

[0111] System Example

[0112] According to an embodiment of the present invention, a system for predicting carbon dioxide emission concentration in a thermal power plant is provided. Figure 7 A schematic diagram of a system for predicting carbon dioxide emission concentration in a thermal power plant provided in one or more embodiments of this specification, such as Figure 7 As shown, the system for predicting carbon dioxide emission concentration in a thermal power plant according to an embodiment of the present invention specifically includes:

[0113] Data collection module 70: used to collect fuel data and environmental data of the thermal power plant and construct training data sets and test data sets;

[0114] Model building module 72: used to build a carbon dioxide emission concentration prediction model. The first half of the model includes a feature preprocessing module, a regional feature extractor, a channel feature extractor and a feature fusion module, and the second half includes a multi-scale attention module, a parallel attention module and an adaptive feature fusion module;

[0115] Model training module 74: used to define a loss function, train the constructed carbon dioxide emission concentration prediction model through a training data set, update the model until a convergence condition is reached, and obtain a trained carbon dioxide emission concentration prediction model;

[0116] Concentration prediction module 76: used to input the data in the test data set into the trained carbon dioxide emission concentration prediction model to predict the corresponding real-time carbon dioxide concentration value.

[0117] The embodiment of the present invention is a system embodiment corresponding to the above-mentioned method embodiment. The specific operations of each module can be understood by referring to the description of the method embodiment, which will not be repeated here.

[0118] Device Example 1

[0119] An embodiment of the present invention provides an electronic device, such as Figure 8 As shown, it includes: a memory 80, a processor 82, and a computer program stored in the memory 80 and executable on the processor 82. When the computer program is executed by the processor 82, the following method steps are implemented:

[0120] S1. Collect fuel data and environmental data of thermal power plants and build training data sets and test data sets;

[0121] S2. Construct a carbon dioxide emission concentration prediction model. The first half of the model includes a feature preprocessing module, a regional feature extractor, a channel feature extractor, and a feature fusion module. The second half includes a multi-scale attention module, a parallel attention module, and an adaptive feature fusion module.

[0122] S3. define a loss function, train the constructed carbon dioxide emission concentration prediction model through the training data set, update the model until the convergence condition is reached, and obtain a trained carbon dioxide emission concentration prediction model;

[0123] S4. Input the data in the test data set into the trained carbon dioxide emission concentration prediction model to predict the corresponding real-time carbon dioxide concentration value.

[0124] Device Example 2

[0125] An embodiment of the present invention provides a computer-readable storage medium, on which a program for implementing information transmission is stored. When the program is executed by the processor 82, the following method steps are implemented:

[0126] S1. Collect fuel data and environmental data of thermal power plants and build training data sets and test data sets;

[0127] S2. Construct a carbon dioxide emission concentration prediction model. The first half of the model includes a feature preprocessing module, a regional feature extractor, a channel feature extractor, and a feature fusion module. The second half includes a multi-scale attention module, a parallel attention module, and an adaptive feature fusion module.

[0128] S3. define a loss function, train the constructed carbon dioxide emission concentration prediction model through the training data set, update the model until the convergence condition is reached, and obtain a trained carbon dioxide emission concentration prediction model;

[0129] S4. Input the data in the test data set into the trained carbon dioxide emission concentration prediction model to predict the corresponding real-time carbon dioxide concentration value.

[0130] The computer-readable storage medium in this embodiment includes, but is not limited to, ROM, RAM, magnetic disk or optical disk, etc.

[0131] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting carbon dioxide emission concentration in a thermal power plant, characterized in that: include: Collect fuel data and environmental data of thermal power plants to build training and test data sets; Construct a carbon dioxide emission concentration prediction model. The first half of the model includes a feature preprocessing module, a regional feature extractor, a channel feature extractor, and a feature fusion module. The second half includes a multi-scale attention module, a parallel attention module, and an adaptive feature fusion module. Define the loss function, train the constructed carbon dioxide emission concentration prediction model through the training data set, update the model until the convergence condition is reached, and obtain the trained carbon dioxide emission concentration prediction model; The data in the test data set is input into the trained carbon dioxide emission concentration prediction model to predict the corresponding real-time carbon dioxide concentration value.

2. The method according to claim 1, characterized in that The fuel data includes fuel type, fuel quality and combustion efficiency; the environmental data includes air excess coefficient, temperature, humidity, atmospheric pressure and light intensity absorption values ​​in multiple bands.

3. The method according to claim 1, characterized in that The feature preprocessing module uses a one-dimensional convolution layer to perform a convolution operation on the input one-dimensional time series data to extract local features; The reshape layer is used to rearrange the local features extracted by the one-dimensional convolution layer into a feature matrix; the rearranged feature matrix is ​​processed and analyzed by the two-dimensional convolution layer and the Sigmoid activation function to extract higher-level features and generate the initial feature map.

4. The method according to claim 1, characterized in that The regional feature extractor is used to encode the spatial correlation in the input initial feature map. Specifically, firstly, features are extracted through two blocks consisting of a two-dimensional convolutional layer, a Sigmoid activation function and an average pooling layer; all input bands are convolved through the two-dimensional convolutional layer to extract features in the local neighborhood; the output features are flattened into a spatial feature vector using a flattening layer; wherein the average pooling layer has a kernel of 2×2 and a stride of 2; the kernel size of all convolutional layers in the regional feature extractor is 3×3, the stride is 1, and a padding method of 1 unit is adopted; The channel feature extractor includes two one-dimensional convolutional layers, a Sigmoid activation function and a flattening layer, and the features extracted by the one-dimensional convolutional layer are flattened by the flattening layer to obtain a channel feature vector; The feature fusion module mixes the channel feature vector and the spatial feature vector through weight coefficients, and on the basis of predicting the mixing weight, forms the final combined feature through channel dimension splicing.

5. The method according to claim 1, characterized in that The multi-scale attention module adopts multiple 1×1 filter groups, which are interconnected in a hierarchical manner; the outputs of all filters are integrated to obtain filter output features, the filter output features are subjected to a maximum pooling, and then the pooled features are subjected to two-dimensional convolution and Sigmoid activation function operations, the number of channels of the pooled features are reduced and restored respectively, and finally the output features are multiplied by the filter output features and added to the original input features of the multi-scale attention module to obtain the output features of the multi-scale attention module.

6. The method according to claim 1, characterized in that The parallel self-attention module includes a spatial attention module and a channel attention module; the spatial attention module is used to calculate the context information of spatial features in the horizontal direction and the vertical direction; the channel attention module is used to enhance the feature representation by enhancing the interdependence between feature channels; The local features are separated by the spatial attention module and the channel attention module, and the features are fused by adaptively learning different weights.

7. The method according to claim 1, characterized in that Define the loss function and train the constructed carbon dioxide emission concentration prediction model through the training data set. Specifically include: Use the flattening layer, fully connected layer and ReLU activation function to transform the obtained features into the final prediction results Calculate the final prediction result The MAE loss value of the true sequence y is as follows: in and q Sequence and the value of y at time q, P represents the total time of the data set; Backpropagation is performed based on the loss value to train and update the model.

8. A system for predicting carbon dioxide emission concentration in a thermal power plant, characterized in that: include: Data acquisition module: used to collect fuel data and environmental data of thermal power plants and build training data sets and test data sets; Model building module: used to build a carbon dioxide emission concentration prediction model. The first half of the model includes a feature preprocessing module, a regional feature extractor, a channel feature extractor, and a feature fusion module. The second half includes a multi-scale attention module, a parallel attention module, and an adaptive feature fusion module. Model training module: used to define the loss function, train the constructed carbon dioxide emission concentration prediction model through the training data set, update the model until the convergence condition is reached, and obtain the trained carbon dioxide emission concentration prediction model; Concentration prediction module: used to input the data in the test data set into the trained carbon dioxide emission concentration prediction model to predict the corresponding real-time carbon dioxide concentration value.

9. An electronic device, characterized in that: include: processor; as well as, A memory arranged to store computer executable instructions, wherein when the computer executable instructions are executed, the processor implements the steps of the method for predicting carbon dioxide emission concentration in a thermal power plant as claimed in any one of claims 1 to 7.

10. A storage medium, characterized in that: Used to store computer executable instructions, which, when executed, implement the steps of the method for predicting carbon dioxide emission concentration in a thermal power plant as described in any one of claims 1 to 7.

Images