Underground powerhouse cavern group air pollutant concentration prediction method based on improved space-time diagram convolutional network

Abstract

The invention discloses an underground powerhouse cavern group air pollutant concentration prediction method based on an improved space-time diagram convolutional network, and belongs to the technical field of air pollutant concentration prediction. Comprising the steps of obtaining initial monitoring data; acquiring a sensor arrangement diagram, constructing a normalized self-loop adjacent matrix, and describing PM2.5 interaction difficulty between monitoring stations by using a simplified diffusion distance; constructing a multi-scale spatio-temporal feature fusion model, and fusing PM2.5 concentration spatio-temporal information under different scales; and constructing a space-time diagram convolutional neural network, taking the output result of the self-loop adjacent matrix and the multi-scale space-time feature fusion model as the input of the space-time diagram convolutional neural network, splicing and outputting by adopting a multi-space-time block output fusion method, then accessing the multi-head self-attention model, and outputting a prediction result. According to the method, different information embedded in the time data under different scales can be considered, so that the model can learn the correlation between different sequences on each scale, and the prediction accuracy and generalization ability are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of air pollutant concentration prediction, and particularly to a method for predicting the air pollutant concentration of an underground powerhouse cavern group based on an improved spatio-temporal graph convolutional network. Background Art

[0002] During the excavation of an underground power plant cavern group, a large amount of PM2.5 is generated, which seriously affects the health of construction workers and pollutes the construction environment. Blasting operations, excavation work, and the exhaust emissions of engineering vehicles are the main sources of PM2.5. In the underground powerhouse cavern group of a water conservancy project, the construction unit starts the ventilation facilities to reduce PM2.5 to ensure the safety of personnel. However, the energy consumption of the ventilation system is very high, and long-term operation will not only reduce the ventilation efficiency but also cause energy waste and economic losses. At present, in the related research field of intelligent ventilation systems, experts and scholars have emphasized the necessity of improving the prediction accuracy of PM2.5 concentration. Accurately predicting the change of PM2.5 concentration is of great significance for the construction of an underground power plant cavern group. It can enable managers to start the ventilation system before a dangerous event occurs or minimize the ventilation energy consumption under acceptable environmental requirements.

[0003] Traditional models rely on predefined models and theories, usually based on the understanding of phenomena, physical laws, or mathematical rules, highly dependent on prior knowledge, and often ignore the limitations of existing theories under different task conditions. This makes it challenging to obtain a high-dimensional non-linear mapping between environmental features and target features in an unknown domain.

[0004] Although the spatio-temporal graph convolutional model (STGCN) can mine key information in data and process non-grid structured data that can be represented as a graph to achieve spatio-temporal prediction, however, as a time series data, PM2.5 concentration contains different characteristic details at different scales of time series. The existing STGCN has a fixed receptive field and is prone to ignoring global or local characteristic information. In addition, due to the action of the wind field, PM2.5 can migrate within the underground powerhouse cavern group, which leads to differences in the interaction difficulty of PM2.5 between different monitoring stations. The existing STGCN model only takes the connectivity of stations as the spatial relationship and does not consider the interaction difficulty of PM2.5 between different stations. Moreover, the existing STGCN model has a fixed structure. When dealing with different engineering tasks, relying solely on optimizing hyperparameters may ignore more effective structures. The disadvantages of existing technologies result in problems such as low prediction accuracy, low result credibility, and insufficient cross-project generalization ability in the task of predicting the air pollutant concentration of an underground powerhouse cavern group.

[0005] Aiming at the deficiencies existing in the prior art, the present invention proposes a method for predicting the concentration of air pollutants in the underground powerhouse cavern group based on an improved spatio-temporal graph convolutional network, introducing spatio-temporal deep learning methods based on STGCN, multi-scale feature fusion methods, multi-head self-attention models, etc. into the task of predicting the concentration of air pollutants in the underground powerhouse cavern group, fusing multi-scale information, describing the interaction difficulty of PM2.5, and optimizing the model structure to solve problems such as low prediction accuracy of air pollutant concentration, low reliability of results, and insufficient cross-project generalization ability. Summary of the Invention

[0006] The object of the present invention is to propose a method for predicting the concentration of air pollutants in the underground powerhouse cavern group based on an improved spatio-temporal graph convolutional network, including the following steps:

[0007] Step 1: Obtain the initial monitoring data of the underground powerhouse cavern group and merge it with the time data;

[0008] Step 2: Obtain the sensor layout diagram, construct a normalized self-loop adjacency matrix based on the Gaussian diffusion model, and use the simplified diffusion distance to describe the interaction difficulty of PM2.5 between monitoring stations;

[0009] Step 3: According to the initial monitoring data and time data in Step 1, construct a multi-scale spatio-temporal feature fusion model to fuse the spatio-temporal information of PM2.5 concentration at different scales;

[0010] Step 4: Construct a spatio-temporal graph convolutional neural network, use the normalized self-loop adjacency matrix in Step 2 and the output result of the multi-scale spatio-temporal feature fusion model in Step 3 as the input of the spatio-temporal graph convolutional neural network, adopt the multi-spatio-temporal block output fusion method to splice the outputs and then connect them to the multi-head self-attention model to output the prediction result.

[0011] Further, the initial monitoring data in Step 1 includes temperature data, humidity data, wind speed data, and PM2.5 concentration data.

[0012] Further, Step 2 specifically includes:

[0013] Obtain the sensor layout diagram:

[0014] G=(V, E)

[0015] where V={v1, v2, ……, v m} is the node set, v m is one of the monitoring stations where sensors are arranged, and e ij =(v i , v j )∈E is the edge set, indicating whether v i is connected to v j ;

[0016] Construct an adjacency matrix \(A\) based on the sensor layout diagram \(G\). The mathematical expression for the element \(a_{ij}\) is as follows: ij

[0017]

[0018] where \(a_{ij}\) represents the element in the adjacency matrix \(A\), and \(d_{ij}\) represents the diffusion distance; ij ij

[0019] Add self-loops to the adjacency matrix. Use the degree matrix to describe the number of edges connected to a node to obtain the self-loop adjacency matrix. The calculation formula is as follows:

[0020]

[0021]

[0022] where \(I\) represents the identity matrix, \(\widetilde{A}\) represents the self-loop adjacency matrix, \(\widetilde{a}_{ij}\) represents the element in the self-loop adjacency matrix, and \(d_{ii}\) represents the element in the degree matrix;

[0023] Normalized adjacency matrix:

[0024]

[0025] where \(D\) represents the corresponding degree matrix, and \(\widetilde{A}_{sym}\) represents the normalized self-loop adjacency matrix;

[0026] Simplify the following Gaussian diffusion model:

[0027]

[0028] where \(C_0(x,y,z,u)\) represents the pollutant concentration, \(Q\) represents the pollutant emission per unit time of the pollution source, \(x\) and \(y\) represent the downwind distance and the horizontal distance from the wind direction center line respectively, \(z\) is the height of the pollution source, \(u\) is the horizontal wind speed, \(\sigma_x(x)\) and \(\sigma_y(x)\) represent the standard deviations of diffusion in the horizontal and vertical directions respectively; y z

[0029] Obtain the simplified diffusion distance:

[0030]

[0031] where \(C_0\) represents the initial concentration of PM2.5 at the pollution source, \(C\) represents the concentration of PM2.5 at point \(B\) at the same moment, and \(\sigma\) is set to 10; B x

[0032] ​​​​​​​​​​​Using the diffusion distance d B Describe the difficulty of PM2.5 interaction between monitoring sites.

[0033] Furthermore, the specific steps of constructing the multi-scale spatio-temporal feature fusion model in step 3 are as follows:

[0034] Adopt the Encoder-Decoder structure. The Encoder consists of convolution and downsampling. The convolution structure uses a kernel of size 3×3, the padding of the convolution operation is set to 1, and the max-pooling layer is used for downsampling. The max-pooling layer selects a window size of 2×2 and moves on the feature map with a stride of 2.

[0035] The feature map is restored to the shape of the input through the Decoder. Transposed convolution and skip connections are performed in the Decoder. A transposed convolution of size 2×2 with a stride of 2 is used. The feature map of the corresponding layer is obtained through skip connection and splicing, and the output feature map is activated by ReLU after convolution.

[0036] Furthermore, step 4 includes the following sub-steps:

[0037] Step S401, use the temporal gated convolutional neural network to extract temporal features. For a sequence of length T at a vertex, with the time step as the x-axis, the vertex as the y-axis, and the feature as the channel, a feature map is constructed. The input has C channels, and the output is generated through the gated linear unit. Residual connections are used to prevent gradient disappearance; use the graph convolutional neural network to extract spatial features, and input the output of the temporal gated convolutional neural network and the normalized self-loop adjacency matrix into the graph convolutional neural network.

[0038] Step S402, input data into the spatio-temporal graph convolutional model. After passing through the first spatio-temporal block, the output results are directly connected to the second spatio-temporal block and the third spatio-temporal block respectively. The second spatio-temporal block and the third spatio-temporal block are connected in parallel, and an output is obtained respectively. The outputs are concatenated on the Channel scale and merged into one output, and then input into the multi-head self-attention model for operation.

[0039] Step S403, in the multi-head self-attention model, a query vector, a key vector, and a value vector are obtained through the learned weight matrix and the linear transformation of the input elements, and calculations are performed; for each pair of elements x i and x j , calculate an attention score to represent x i for x jRegarding the degree of attention, the softmax function is used to convert the attention scores into values between 0 and 1 and make their sum equal to 1. Multiply the value vector of each element by its corresponding attention weight and then sum them up. Finally, in the multi-head attention mechanism, the input data is split into multiple heads containing query vectors, key vectors, and value vectors, and independent self-attention calculations are performed. The calculated results are concatenated to form the final output.

[0040] Furthermore, the calculation formula of the gated linear unit is as follows:

[0041]

[0042] where X l and X l+1 represent the input and output of the gated convolutional neural network respectively, Γ1, Γ2, and Γ3 represent different convolutional kernels, and b1, b2, and b3 represent different biases. is the Hadamard product, and σ is the Sigmoid function.

[0043] Furthermore, the convolution formula of the graph convolutional neural network is as follows:

[0044]

[0045] where σ represents the activation function, X (l) represents the input, X (l+1) represents the output, and W (l) is the linear transformation matrix.

[0046] The beneficial effects of the present invention are as follows:

[0047] 1. The present invention can consider different information embedded in time data at different scales, enabling the model to learn the correlations between different sequences at each scale, thereby improving the prediction accuracy and generalization ability.

[0048] 2. The present invention uses the Gaussian diffusion model to describe the diffusion distance of air pollutants and optimizes the adjacency matrix, accurately reflecting the interaction difficulty between air pollutants at different monitoring points, providing more reasonable and accurate spatial information for the model. The proposed model changes the arrangement and connection methods of multiple spatio-temporal blocks, constructs a more flexible model, and uses the multi-head self-attention model to process rich features, breaking the traditional idea of improving model performance through hyperparameter tuning and providing a practical basis for constructing a more general model.

[0049] 3. The prediction performance of the present invention has been greatly improved compared with that of a single prediction model. It can achieve fast and accurate prediction of PM2.5 concentration, thereby obtaining more accurate and reliable prediction results, providing guidance for decision-making, and ensuring the safety of personnel during underground engineering construction. At the same time, the model proposed in the present invention also provides a new idea for the prediction of other engineering parameters and has good prospects for engineering application. Description of the Drawings

[0050] Figure 1 It is a flowchart of the method for predicting the concentration of air pollutants in the underground powerhouse cavern group based on the improved spatio-temporal graph convolutional network of the present invention.

[0051] Figure 2 It is a schematic diagram of the multi-scale spatio-temporal feature fusion model of the present invention.

[0052] Figure 3 It is a structural diagram of the spatio-temporal graph convolutional neural network model of the present invention.

[0053] Figure 4 It is a schematic diagram of the PM2.5 concentration prediction result. Detailed Embodiments

[0054] The present invention proposes a method for predicting the concentration of air pollutants in the underground powerhouse cavern group based on the improved spatio-temporal graph convolutional network. The following further describes the present invention with reference to the drawings and specific embodiments.

[0055] Figure 1 It is a flowchart of the method for predicting the concentration of air pollutants in the underground powerhouse cavern group based on the improved spatio-temporal graph convolutional network of the present invention, including the following steps:

[0056] Step 1: Obtain initial monitoring data; the initial monitoring data includes environmental temperature data, humidity data, wind speed data, and PM2.5 concentration data of the underground powerhouse cavern group, and merge the environmental temperature data, humidity data, wind speed data, and PM2.5 concentration data with time data.

[0057] Step 2: Obtain the sensor layout diagram, construct an adjacency matrix based on the Gaussian diffusion model, and describe the difficulty of PM2.5 interaction between monitoring sites with the diffusion distance;

[0058] Obtain the sensor layout diagram:

[0059] G=(V, E)

[0060] where V={v1, v2, ……, v m} is the node set, v m is one of the monitoring sites where sensors are arranged, and e ij =(v i , v j )∈E is the edge set, indicating that v i and vj Connected or not;

[0061] Construct an adjacency matrix A based on the sensor layout diagram G, and the mathematical expression of element a ij is as follows:

[0062]

[0063] where a ij represents the element in the adjacency matrix A, and d ij represents the diffusion distance;

[0064] Add self-loops to the adjacency matrix, describe the number of edges connected to a node with the degree matrix, and obtain the self-loop adjacency matrix. The calculation formula is as follows:

[0065]

[0066] where I represents the identity matrix, represents the self-loop adjacency matrix, represents the element in the self-loop adjacency matrix, represents the element in the degree matrix;

[0067] Normalized adjacency matrix:

[0068]

[0069] where represents the corresponding degree matrix, represents the normalized self-loop adjacency matrix;

[0070] Simplify the following Gaussian diffusion model:

[0071]

[0072] where C0(x,y,z,u) represents the pollutant concentration, Q represents the pollutant emission per unit time of the pollution source, x and y represent the downwind distance and the horizontal distance from the wind direction center line respectively, z is the height of the pollution source, u is the horizontal wind speed, and σ y (x) and σ z (x) represent the standard deviations of diffusion in the horizontal and vertical directions respectively;

[0073] Obtain the simplified diffusion distance:

[0074]

[0075] where C0 represents the initial concentration of PM2.5 at the pollution source, C B represents the concentration of PM2.5 at point B at the same moment, and σ x is set to 10;

[0076] Using the diffusion distance d B Describe the difficulty of PM2.5 interaction between monitoring stations.

[0077] Step 3: Construct a multi-scale spatio-temporal feature fusion model, as Figure 2 shown. Integrate the spatio-temporal information of PM2.5 concentration at different scales; the constructed multi-scale spatio-temporal feature fusion model specifically adopts an Encoder-Decoder structure:

[0078] The Encoder consists of convolution and downsampling operations, responsible for feature extraction. In this embodiment, the convolution structure uses a kernel of size 3×3, and the padding of the convolution operation is set to 1 to avoid dimension mismatch problems. After the convolution operation, a max-pooling layer is used for downsampling. The max-pooling layer selects a window size of 2×2 and moves on the feature map with a stride of 2.

[0079] Then, the feature map is restored to the shape of the input through the Decoder. A key step in the Decoder is transposed convolution and skip connection. Transposed convolution with a size of 2×2 and a stride of 2 is used. The feature map of the corresponding layer is obtained through skip connection and concatenation. Finally, the output feature map is activated through ReLU after convolution.

[0080] Step 4: Construct a spatio-temporal graph convolutional neural network, as Figure 3 shown. Use the output result of the multi-scale spatio-temporal feature fusion model and the self-loop adjacency matrix as the input of the spatio-temporal graph convolutional neural network, adopt the multi-temporal block output fusion method to splice the outputs and then connect them to a multi-head self-attention model, and the output obtains the prediction result. It includes the following sub-steps:

[0081] Step S401, the spatio-temporal graph convolution model includes three spatio-temporal blocks (ST-Block), and each spatio-temporal block includes a time-gated convolutional neural network and a graph convolutional neural network.

[0082] Use the time-gated convolutional neural network to extract time features. The size of the convolution kernel of the time-gated convolutional neural network is 1×K t , to achieve one-dimensional causal convolution. For a sequence V={v t-T+1 ,...,v t} with length T at a vertex, with the time step as the x-axis, the vertex as the y-axis, and the feature as the channel, construct a feature map. The input has C channels. Therefore, the input representation of the time convolution is X∈R T×C , and the convolution kernel is represented as 2C0 represents the number of output channels.

[0083] After the input sequence passes through the time convolution, it generates an output through a gated linear unit (GLU). Residual connection is used to prevent the problem of gradient disappearance. The calculation formula of the gated linear unit is as follows:

[0084]

[0085] where X l and X l+1 represent the input and output of the gated convolutional neural network respectively. Γ1, Γ2, and Γ3 represent different convolutional kernels. b1, b2, and b3 represent different biases. is the Hadamard product, performing element-wise multiplication. σ is the Sigmoid function.

[0086] The graph convolutional neural network is used to extract spatial features. The output of the temporal gated convolutional neural network and the adjacency matrix constructed in step 2 are input into the graph convolutional neural network. The graph convolutional neural network can fully consider the spatial correlation between geospatial data. The graph convolution formula is as follows:

[0087]

[0088] where σ represents the activation function, X (l) represents the input, X (l+1) represents the output, and W (l) is the linear transformation matrix.

[0089] Step S402, the multi-spatiotemporal block output fusion method is used to describe the process of splicing the outputs of parallel spatiotemporal blocks.

[0090] In the spatiotemporal graph convolution model, after the input data passes through the first spatiotemporal block, the output results are directly connected to the second and third spatiotemporal blocks respectively. The second and third spatiotemporal blocks are connected in parallel, and each obtains an output. The multi-spatiotemporal block output fusion method is used to splice the outputs of multiple spatiotemporal blocks on the Channel scale and merge them into one output, which is input into the multi-head self-attention model for further operations. The parallel spatiotemporal block structure developed in this embodiment can increase or decrease the number of parallel spatiotemporal blocks according to different engineering actual needs, preventing the model gradient from disappearing due to too many cascaded spatiotemporal blocks.

[0091] Step S403, the calculation process of the multi-head self-attention mechanism includes:

[0092] Calculating a query vector, a key vector, and a value vector; these vectors are obtained through the learned weight matrix and the linear transformation of the input elements. The calculation formula is as follows:

[0093]

[0094] where Q i , K i , V i represent the query vector, key vector, and value vector of the i-th element respectively, and W Q , W K , WV represent the corresponding weight matrices respectively, and x i represents an element in the input sequence.

[0095] For each pair of elements x i and x j , an attention score is calculated to represent the degree of attention of x i to x j . Using the softmax function, the attention scores are converted into values between 0 and 1 and their sum is 1. Multiply the value vectors of each element by their corresponding attention weights and then sum to obtain the final output. The calculation formula for this process is as follows:

[0096]

[0097] w ij = softmax(score(Q i , K j ))

[0098]

[0099] where d k is the dimension of the key vector, and score(Q i , K j ) represents the attention score calculated according to Q i and K j .

[0100] Finally, in the multi-head attention mechanism, the input data is split into multiple heads, each head has its own query vector, key vector and value vector, and independent self-attention calculations are performed. Then, the results of these calculations are concatenated to form the final output:

[0101] Z = W O ·Concat(z1, z2,..., z h )

[0102] where W O is the output weight matrix.

[0103] In this embodiment, Figure 4 is a schematic diagram of the PM2.5 concentration prediction result. In Figure 4 , the smaller the difference in the ordinates of the predicted value and the true value, the higher the prediction accuracy of the model.

[0104] The present invention is compared with the Transformer model, convolutional long short-term memory network (CNN-LSTM) and spatio-temporal graph convolutional network (STGCN) in the prior art, and the evaluation metrics include mean absolute error (MAE), root mean square error (RMSE) and coefficient of determination (R2 )。The results are shown in Table 1 as follows:

[0105] Table 1 Comparison between the present invention and the comparative model

[0106]

[0107] It can be seen from Table 1 that the present invention has better prediction performance.

[0108] The present invention can consider different information embedded in time data at different scales, enabling the model to learn the correlation between different sequences at each scale, thereby improving the prediction accuracy and generalization ability. Compared with the prediction performance of a single prediction model, it has a great improvement, and can achieve fast and accurate prediction of PM2.5 concentration, thus obtaining more accurate and reliable prediction results.

Claims

1. A method for predicting the concentration of air pollutants in an underground powerhouse cavern group based on an improved spatio-temporal graph convolutional network, characterized in that, It includes the following steps: Step 1: Obtain the initial monitoring data of the underground powerhouse cavern group and merge it with the time data; Step 2: Obtain the sensor layout diagram, construct a normalized self-loop adjacency matrix based on the Gaussian diffusion model, and use the simplified diffusion distance to describe the difficulty of PM2.5 interaction between monitoring stations; Step 3: According to the initial monitoring data and time data in Step 1, construct a multi-scale spatio-temporal feature fusion model to fuse the spatio-temporal information of PM2.5 concentration at different scales; Step 4: Construct a spatio-temporal graph convolutional neural network. Take the normalized self-loop adjacency matrix in Step 2 and the output result of the multi-scale spatio-temporal feature fusion model in Step 3 as the input of the spatio-temporal graph convolutional neural network. Adopt the multi-spatio-temporal block output fusion method to splice the outputs and then connect them to the multi-head self-attention model to output the prediction result.

2. The method for predicting the air pollutant concentration of the underground powerhouse cavern group based on the improved spatio-temporal graph convolutional network according to claim 1, wherein, The initial monitoring data in Step 1 includes temperature data, humidity data, wind speed data, and PM2.5 concentration data.

3. The method for predicting the air pollutant concentration of the underground powerhouse cavern group based on the improved spatio-temporal graph convolutional network according to claim 1, characterized in that, Step 2 specifically includes: Obtain the sensor layout diagram: G=(V, E) Among them, V = {v1, v2, ……, v m} is the node set, v m is one of the monitoring sites where sensors are arranged, e ij = (v i , v j ) ∈ E is the edge set, indicating whether v i is connected to v j or not; Construct an adjacency matrix A based on the sensor layout diagram G, and the mathematical expression of element a ij is as follows: Among them, a ij represents an element in the adjacency matrix A, and d ij represents the diffusion distance; Add a self-loop to the adjacency matrix, and use the degree matrix to describe the number of edges connected to the node to obtain the self-loop adjacency matrix. The calculation formula is as follows: where \(I\) represents the identity matrix, represents the self-loop adjacency matrix, represents the elements in the self-loop adjacency matrix, represents the elements in the degree matrix; Normalized adjacency matrix: wherein, represents the corresponding degree matrix, represents the normalized self-loop adjacency matrix; Simplify the following Gaussian diffusion model: Among them, C0(x, y, z, u) represents the pollutant concentration, Q represents the pollutant emission per unit time of the pollution source, x and y respectively represent the downwind distance and the horizontal distance from the center line of the wind direction, z is the height of the pollution source, u is the horizontal wind speed, σ y (x) and σ z (x) represent the standard deviations of diffusion in the horizontal and vertical directions respectively; Obtain the simplified diffusion distance: Among them, C0 represents the initial concentration of PM2.5 at the pollution source, and C B represents the concentration of PM2.5 at point B at the same moment, and σ x is set to 10; Using the diffusion distance d B Describe the difficulty of PM2.5 interaction between monitoring stations.

4. The method for predicting the air pollutant concentration of the underground powerhouse cavern group based on the improved spatio-temporal graph convolutional network according to claim 1, characterized in that, The specific steps for constructing the multi-scale spatio-temporal feature fusion model in Step 3 are as follows: Adopt the Encoder-Decoder structure. The Encoder consists of convolution and downsampling. The convolution structure uses a kernel of size 3×3, the padding of the convolution operation is set to 1, and the max-pooling layer is used for downsampling. The max-pooling layer selects a window size of 2×2 and moves on the feature map with a stride of 2; Restore the feature map to the shape of the input through the Decoder. In the Decoder, transposed convolution and skip connection are performed, and a transposed convolution of size 2×2 with a stride of 2 is used. Obtain the feature map of the corresponding layer through skip connection and splicing, and activate the output feature map through ReLU after convolution.

5. The air pollutant concentration prediction method for underground powerhouse cavern groups based on the improved spatio-temporal graph convolutional network according to claim 1, characterized in that Step 4 includes the following sub-steps: Step S401, adopt a time-gated convolutional neural network to extract time features. For a sequence of length T at a vertex, construct a feature map with the time step as the x-axis, the vertex as the y-axis, and the feature as the channel. The input has C channels, generate the output through a gated linear unit, and adopt residual connection to prevent gradient disappearance; adopt a graph convolutional neural network to extract spatial features, and input the output of the time-gated convolutional neural network and the normalized self-loop adjacency matrix into the graph convolutional neural network; Step S402, input data into the spatio-temporal graph convolutional model. After the first spatio-temporal block, the output results are directly connected to the second spatio-temporal block and the third spatio-temporal block respectively. The second spatio-temporal block and the third spatio-temporal block are connected in parallel to obtain an output respectively. Splice the outputs on the Channel scale, merge them into one output, and input them into the multi-head self-attention model for calculation; Step S403, in the multi-head self-attention model, obtain a query vector, a key vector, and a value vector through a learned weight matrix and a linear transformation of the input elements, and perform calculations; for each pair of elements x i and x j , calculate an attention score to represent the attention degree of x i to x j , use the softmax function to convert the attention scores into values between 0 and 1 and make their sum equal to 1, multiply the value vector of each element by its corresponding attention weight, and then sum; finally, in the multi-head attention mechanism, split the input data into multiple heads containing query vectors, key vectors, and value vectors, perform independent self-attention calculations, and concatenate the calculation results to form the final output.

6. The method for predicting the air pollutant concentration of the underground powerhouse cavern group based on the improved spatio-temporal graph convolutional network according to claim 5, characterized in that The calculation formula of the gated linear unit is as follows: Among them, X l and X l+1 represent the input and output of the gated convolutional neural network respectively, Γ1, Γ2, and Γ3 represent different convolutional kernels, and b1, b2, and b3 represent different biases. is the Hadamard product, and σ is the Sigmoid function.

7. The method for predicting the air pollutant concentration of the underground powerhouse cavern group based on the improved spatio-temporal graph convolutional network according to claim 5, characterized in that, The convolution formula of the graph convolutional neural network is as follows: Among them, σ represents the activation function, X (l) represents the input, X (l+1) represents the output, W (l) is the linear transformation matrix.

Images