Regional multi-site PM2.5 prediction method based on semi-dynamic graph neural network

Abstract

The invention relates to a regional multi-site PM2.5 prediction method based on a semi-dynamic graph neural network, and the method comprises the steps: obtaining historical wind fields and corresponding PM2.5 data, and carrying out the data cleaning and preprocessing, and constructing a site PM2.5 and corresponding wind field data set; the method comprises the following steps: respectively constructing a dynamic relation graph and a static relation graph between stations by utilizing geographical and wind field information, and fusing and constructing a semi-dynamic directed graph for predicting PM2.5 of the stations; constructing a space-time convolution graph model by combining space information provided by the semi-dynamic directed graph with time information, and training and verifying the space-time convolution graph model by utilizing the station PM2.5 and the corresponding wind field data set to obtain a space-time PM2.5 prediction model; and inputting the current wind field and the PM2.5 data into the space-time PM2.5 prediction model, and outputting to obtain a station PM2.5 prediction result in a future set time period. Compared with the prior art, the method has the advantages that time-space prediction can be completed for the dynamic graph, and finally more accurate prediction of the PM2.5 concentration of multiple stations in the region is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of air quality monitoring, and in particular to a regional multi-site PM2.5 prediction method based on a semi-dynamic graph neural network. Background Art

[0002] In recent years, with the improvement of economic level and the rapid development of industrial manufacturing, air pollution has become one of the most concerned issues in the world. 2.5 It is an important component of air pollutants. With the development of economy, especially in developing countries, PM 2.5 It is becoming an increasingly important issue in air pollution. Studies have shown that exposure to ambient PM 2.5 can cause serious health problems such as heart and lung damage, cancer, cardiovascular disease, and a variety of other adverse effects. Some studies have also shown that prenatal exposure to PM 2.5 Therefore, as air pollution becomes increasingly prominent, it is necessary to 2.5 It is particularly important to effectively predict the concentration and study the diffusion trend, to control environmental pollution in advance and to prevent the occurrence of pollution incidents.

[0003] PM 2.5 Concentration prediction and diffusion trend prediction have become a hot topic in the academic community. 2.5 There are many factors that influence the generation, diffusion and dissipation of PM, including natural factors such as weather, wind and geographical conditions, and human factors such as social economy and production life. 2.5 Concentration prediction models face huge difficulties. Traditional methods are often based on researchers' understanding of aerodynamics, meteorology, and atmospheric physics and chemistry, combined with real emission data for modeling. However, such methods face problems such as high modeling difficulty, high computational cost, and difficulty in feature extraction.

[0004] In recent years, deep learning methods have shown great advantages in many fields such as image processing, natural language processing, and target detection. Due to its powerful ability to automatically learn features and process large-scale data, deep learning methods have also been used in many fields related to real physics. It has also become the most studied method in air pollution prediction and has achieved relatively good results. Considering the advantages of graphs in modeling sites in non-Euclidean space, more and more studies use graphs to model the spatial relationship of pollutants. However, graph neural network methods in recent years often have some defects: wind field algorithms do not fully utilize information on regional wind fields, ignore the geographical relationship characteristics between sites, and the dynamic graph neural network model is complex, which seriously affects the model's performance in site PM. 2.5Therefore, how to improve the model and integrate wind field and geographic information into the graph neural network model to achieve more accurate and explainable PM 2.5 Prediction is a theoretical and practical engineering problem that needs to be solved urgently. Summary of the invention

[0005] The purpose of the present invention is to provide a regional multi-site PM based on a semi-dynamic graph neural network in order to overcome the defects of the above-mentioned prior art. 2.5 Prediction method to achieve PM prediction for multiple sites in the region 2.5 Accurate prediction of concentration.

[0006] The purpose of the present invention can be achieved by the following technical solution: A regional multi-site PM based on semi-dynamic graph neural network 2.5 The prediction method comprises the following steps:

[0007] S1. Obtain historical wind fields and corresponding PM 2.5 Data, through data cleaning and preprocessing, construct site PM 2.5 and the corresponding wind field dataset;

[0008] S2. Use geographic and wind farm information to construct dynamic and static relationship diagrams between sites, and integrate them to construct site PM 2.5 Predicted semi-dynamic directed graph;

[0009] S3, using the spatial information provided by the semi-dynamic directed graph, combined with the temporal information, to build a spatiotemporal convolutional graph model, using the site PM 2.5 The spatiotemporal convolutional graph model is trained and verified with the corresponding wind field dataset to obtain the spatiotemporal PM 2.5 Predictive models;

[0010] S4. Current wind field and PM 2.5 Data input space-time PM 2.5 In the prediction model, the output is the site PM within the future set time period 2.5 Predict the results.

[0011] Furthermore, the step S1 specifically crawls the data published on the meteorological website to obtain the historical wind field and the corresponding PM 2.5 data.

[0012] Furthermore, the specific process of step S2 is as follows:

[0013] S21, constructing a static directed graph, which is a geographic information graph and is initialized to the distance information between sites;

[0014] S22, constructing a dynamic directed graph, which is a wind field graph, using a wind field aggregation algorithm, first dividing the wind zone, and then aggregating the wind data, wherein the wind zone division area is a rectangular area with the site connection line as the center line, and the aggregated wind data is weighted addition of the u and v components (i.e., the east-west component and the north-south component) of each grid point along the site connection direction;

[0015] S23. The static directed graph and the dynamic directed graph are superimposed by using the Hadamard product method to obtain a semi-dynamic directed graph that integrates the effects of geographical and meteorological factors on spatial diffusion.

[0016] Furthermore, the specific process of step S21 is: using a directed graph G = (V, E) to describe the geographical topology of the site, where V = (V1, V2, ..., V n ) indicates the PM in the study 2.5 Monitoring station, E=(E1,E2,…E m ) indicates how geographical conditions favor PM 2.5 A measure of the spread between sites, with edges represented by the adjacency matrix A g ∈R N×N Indicates that A g The elements in are obtained through learning, ij is initialized to the inverse of the distance and will be reset to zero if it is less than a preset threshold.

[0017] Furthermore, the specific process of aggregating wind data in step S22 is as follows:

[0018] Considering wind speed and direction, the wind volume in the straight line between two stations is calculated, and the wind field relationship between stations is obtained by weighted aggregation based on the distance of grid points:

[0019]

[0020] Among them, d AB d Ai and d Bi represents the distance between (A,B), (A,i) and (B,i), is the grid point p i The wind field, is the direction vector from station A to station B. Through the dot product of the vectors, we get the wind components in the directions of the two nodes, which are calculated from the u and v components of each grid point:

[0021]

[0022] Among them, D AB Indicates the azimuth from A to B.

[0023] Furthermore, the semi-dynamic directed graph in step S23 is specifically:

[0024]

[0025] Among them, A t is the site adjacency matrix at time t, which is composed of the wind field matrix and the geographic matrix A g For a specific time node, both are of size N×N, where N is the number of sites, and then a self-loop term is added:

[0026]

[0027] Among them, I is the N×N identity matrix, λ is a hyperparameter, and λI is the self-loop term in the graph convolutional network.

[0028] Furthermore, the spatiotemporal convolutional graph model in step S3 includes three modules, namely a spatial convolution module, a temporal aggregation module and an output module, wherein the spatial convolution module is composed of a semi-dynamic directed graph convolutional network, the temporal aggregation module is composed of a long short-term memory neural network, and the output module is composed of a fully connected layer.

[0029] Furthermore, the spatial convolution module takes the semi-dynamic graph adjacency matrix as input and uses random walk normalization to complete the matrix normalization, which is expressed as:

[0030]

[0031] Among them, X t Is the site history PM 2.5 Input, W t is the weight matrix of the graph neural network, and the input data of the spatial convolution module is X t and X t The size is N×t, where N is the number of sites and t is the length of the input historical information. The size of A is N×N×t, g Use the distance between sites to initialize. When the distance between sites i and j exceeds the preset threshold, a ij is 0, otherwise it is initialized to the inverse of the distance.

[0032] Furthermore, the process of training the spatiotemporal convolutional graph model in step S3 includes:

[0033] S31. Use a semi-dynamic graph as the adjacency matrix, input into a graph convolutional neural network, and use random walk normalization to normalize the adjacency matrix to maintain the utility of a directed graph with edge weights.

[0034] For a sequence input of length l, there is also a dynamic graph of length l and its corresponding adjacency matrix. The input and adjacency matrix are matched one-to-one based on the time step, processed separately by the spatial convolution module, and concatenated into a complete output of the same length:

[0035] H=[h t-l+1 ,…,h t ]

[0036] S32, input the obtained hidden layer H into the long short-term memory neural network to extract time information, and finally obtain the site PM through the fully connected layer 2.5 Prediction output of concentration;

[0037] S33, use the Pytorch framework to build the model, set the neural network parameters of the spatiotemporal convolutional graph model, and set the site PM 2.5 The corresponding wind field data set is input into the spatiotemporal convolutional graph model and trained with the loss function to convert PM 2.5 The parameters with the highest prediction accuracy are retained as the ideal parameters of the spatiotemporal convolutional graph model, and the spatiotemporal PM is obtained. 2.5 Predictive models.

[0038] Furthermore, in step S32, the number of layers of the long short-term memory neural network is set to 2, the batch_first parameter is set to true, and the input size is N is the batch size, L is the length of the hidden layer H sequence, H in is the hidden layer embedding size, and the output size is H out is the size of the output layer;

[0039] In the fully connected layer, the end of the sequence output by the long short-term memory neural network is extracted for input, and a one-dimensional output Y is obtained through the linear layer, whose size is the product of the number of sites and the number of prediction steps, and then the one-dimensional output is reshaped into an output Y of the same shape (N×pred) , N is the number of sites, and pred is the number of prediction steps.

[0040] Compared with the prior art, the present invention has the following advantages:

[0041] The present invention obtains historical wind fields and corresponding PM 2.5 Data to build site PM 2.5 and the corresponding wind farm data set; then, using geographic and wind farm information, the dynamic relationship diagram and static relationship diagram between sites are constructed respectively, and the PM 2.5 The predicted semi-dynamic directed graph is then used to combine the spatial information provided by the semi-dynamic directed graph with the temporal information to construct a spatiotemporal convolutional graph model. 2.5The spatiotemporal PM 2.5 Prediction model. On the one hand, a new wind field calculation method is proposed to predict the possible impact of PM2.5 on the site. 2.5 All wind farm data in the propagation area are aggregated, which solves the problem of incomplete utilization of regional wind farm information by traditional methods and can better capture the spatial dynamic relationship between sites. On the other hand, the geographic features between sites are extracted by constructing a static geographic map and added to the PM. 2.5 In the site relationship of the propagation process, the problem of traditional methods ignoring geographical information other than distance is overcome; in addition, a simple semi-dynamic graph neural network architecture is constructed, which can complete spatiotemporal prediction for dynamic graphs, and finally realize PM prediction for multiple sites in the region. 2.5 More accurate prediction of concentration.

[0042] The present invention uses geographic and wind farm information to construct dynamic and static relationship graphs between sites, respectively. The dynamic directed graph is a wind farm graph, which is implemented by a wind farm aggregation algorithm. The area is divided into rectangular areas with the site connection line as the center line. The aggregation method is to perform weighted addition of the u and v components of each grid point along the site connection line direction. The static directed graph is a geographic information graph, which is initialized with the distance information between sites, and the geographic information is further automatically trained during the model training process. The dynamic directed graph and the static directed graph are then superimposed by the Hadamard product to obtain a fusion for site PM. 2.5 The predicted semi-dynamic directed graph fully considers the impact of geographical and meteorological factors on spatial diffusion, and can ensure the accuracy of the semi-dynamic directed graph.

[0043] The present invention uses the spatial information provided by the semi-dynamic directed graph and combines it with the time information to construct a spatiotemporal convolutional graph model, which includes three modules, namely, a spatial convolution module, a temporal aggregation module and an output module. The spatial convolution module is composed of a semi-dynamic directed graph convolutional network, and the obtained semi-dynamic graph adjacency matrix is ​​used as input, and the matrix is ​​normalized using random walk normalization; the temporal aggregation module is composed of a long short-term memory neural network, and the long short-term memory network is used to extract the time information, and the output module is composed of a fully connected layer. In this way, spatiotemporal prediction can be completed for dynamic graphs, and PM of multiple sites in the region can be predicted. 2.5 Accurately predict the concentration. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a schematic diagram of the method flow of the present invention;

[0045] Figure 2 It is a schematic diagram of the construction process of a semi-dynamic directed graph;

[0046] Figure 3This is a schematic diagram of wind zone division;

[0047] Figure 4 is the spatiotemporal PM established in the embodiment 2.5 Schematic diagram of the prediction model;

[0048] Figure 5 PM for Spacetime 2.5 Schematic diagram of the semi-dynamic directed graph convolution module in the prediction model;

[0049] Figure 6 PM for Spacetime 2.5 Schematic diagram of the long short-term memory network in the prediction model. DETAILED DESCRIPTION

[0050] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0051] Example

[0052] like Figure 1 As shown in the figure, a regional multi-site PM based on semi-dynamic graph neural network 2.5 The prediction method comprises the following steps:

[0053] S1. Obtain historical wind fields and corresponding PM 2.5 Data, through data cleaning and preprocessing, construct site PM 2.5 and the corresponding wind field dataset;

[0054] S2. Use geographic and wind farm information to construct dynamic and static relationship diagrams between sites, and integrate them to construct site PM 2.5 Predicted semi-dynamic directed graph;

[0055] S3, using the spatial information provided by the semi-dynamic directed graph, combined with the temporal information, to build a spatiotemporal convolutional graph model, using the site PM 2.5 The spatiotemporal convolutional graph model is trained and verified with the corresponding wind field dataset to obtain the spatiotemporal PM 2.5 Predictive models;

[0056] S4. Current wind field and PM 2.5 Data input space-time PM 2.5 In the prediction model, the output is the site PM within the future set time period 2.5 Predict the results.

[0057] This embodiment applies the above solution, and the main contents are:

[0058] Step 1: Use the wind field and PM data published on the meteorological website to 2.5Data is preprocessed. First, the prediction area is divided according to longitude and latitude, and the stations that meet the conditions (stations with fewer missing historical observations) are selected. The original data is organized into matrix data that can be used as neural network input, and the missing values ​​are linearly filled. Wind field data is downloaded from the public meteorological data website, and the missing wind field data is also linearly filled. Finally, the original wind field and PM 2.5 Data set, in this example, PM 2.5 The dataset contains data from 2020 to 2022, with 50 sites.

[0059] Step 2, reference Figure 2 , using geographic and wind farm information, we construct dynamic and static relationship diagrams between sites, and further integrate them to construct site PM 2.5 Semi-dynamic directed graph for prediction.

[0060] The first step is to construct a static directed graph. Use a directed graph G = (V, E) to describe the geographical topology of the site. Where V = V = (V1, V2, ..., V n ) represents the PM in our study 2.5 Monitoring station, E=(E1,E2,…E m ) indicates how geographical conditions favor PM 2.5 A measure of the spread between sites. The edges are represented by the adjacency matrix A g ∈R N×N Indicates. g The elements in are obtained through learning. ij is initialized to the inverse of the distance, and will be reset to zero if it is less than some value.

[0061] The second step is to construct a dynamic directed graph. 2.5 The wind direction and intensity are constantly changing, compared to the short-term invariance of geography. In order to extract the wind relationship between sites, this scheme uses a method to calculate the wind relationship from the original grid data. The algorithm can be divided into two steps. For two interacting sites, the reference Figure 3 ,First, the wind zones are divided according to geographical locations, and then the wind data of the nodes in the wind zones are aggregated.,The wind volume on the straight line between two stations is calculated by considering the wind speed and wind direction.,The wind field relationship between stations is obtained by weighted aggregation based on the distance,of grid points.

[0062]

[0063] where d AB ,d Ai and d Bi represents the distance between (A,B), (A,i) and (B,i), is the grid point pi The wind field, is the direction vector from station A to station B. By taking the dot product of the vectors, we get the wind components in the directions of the two nodes. It can be calculated from the u and v components:

[0064]

[0065] Where D AB Indicates the azimuth from A to B.

[0066] In the third step, the static and dynamic site spatial relationship diagrams are superimposed to obtain a semi-dynamic directed graph that integrates the effects of geographical and meteorological factors on spatial diffusion.

[0067]

[0068] Among them A t is the site adjacency matrix at time t, which is composed of the wind field matrix and the geographic matrix A g The combination of . The input is obtained from the real-world wind data in the second step. g is obtained through training, A g The elements in are used to measure the weakening or strengthening effect of the inter-station geography on the wind field. The final adjacency matrix is ​​obtained by the following formula:

[0069]

[0070] λI is a self-loop term commonly used in graph convolutional networks, where I is the identity matrix.

[0071] Step 3, reference Figure 4 , Figure 5 , Figure 6 , using the spatial information provided by the semi-dynamic directed graph and combining it with the temporal information to construct a spatiotemporal PM 2.5 The prediction model (consists of three modules, namely, spatial convolution module, temporal aggregation module and output module, wherein the spatial convolution module is composed of a semi-dynamic directed graph convolution network, the temporal aggregation module is composed of a long short-term memory neural network, and the output module is composed of a fully connected layer) is trained using the training set data to obtain the optimal weight.

[0072] First, we build a spatial convolution module, use a semi-dynamic graph as the adjacency matrix, input it into the improved graph convolutional neural network, and use random walk normalization to normalize the adjacency matrix to maintain the utility of the directed graph with edge weights.

[0073]

[0074] in and A g is the adjacency matrix obtained in the previous step that represents the spatial relationship between sites, λI is the self-loop term, X t Is the site history PM 2.5 The input of W t is the weight matrix of the graph neural network. The input data of this module is X t and Where X t The size of is N×t, where N is the number of sites and t is the length of the input history information. The size of A is N×N×t. g Use the distance between sites to initialize. When the distance between sites i and j exceeds 200, a ij is 0, otherwise it is initialized to the inverse of the distance. In this embodiment, N is the number of sites, 50, and the length of the historical information t is 12.

[0075] The historical data of a certain length are convolved and concatenated separately to obtain the complete hidden layer output of the same length.

[0076] H=[h t-l+1 ,…,h t ]

[0077] The obtained hidden layer H is then input into the long short-term memory (LSTM) neural network to extract time information, and finally the site PM is obtained through the fully connected layer. 2.5 The predicted output of the concentration. The number of LSTM layers is set to 2, the batch_first parameter is set to true, and the input size is X (N×L×Hin) , where N is the batch size, L is the length of the hidden layer sequence obtained previously, 12, and H in is the hidden layer embedding size, which is 50 in the model. The output size is X (N×L×Hout) , where H out is the size of the output layer, which is set to 128 in the model. The data obtained by LSTM is then input into the fully connected layer and reshaped into an output Y of the same shape (N×pred) , where N is the number of stations and pred is the number of prediction steps. In the model, the number of stations is 50 and pred is 36, which means the PM of the stations in the next 36 hours is predicted. 2.5 concentration.

[0078] Finally, the constructed spatiotemporal semi-dynamic graph model is trained. The training set is input into the neural network, the results are output through the neural network, and the weights of the neural network are updated using the gradient descent method.

[0079] In this embodiment, the Loss function is defined as:

[0080] Loss=LossRMSE +Loss MAE

[0081] Among them, RMSE and MAE represent the root mean square error and absolute average error respectively, and their calculation formulas are:

[0082]

[0083] Among them, O i is the site observation value, P i is the model prediction value.

[0084] Step 4: PM for regional sites 2.5 The obtained model is used to input the test set or the latest acquired data into the trained spatiotemporal PM 2.5 In the prediction model, the site PM 2.5 This example verifies through experiments that the dataset is put into the spatiotemporal PM 2.5 The prediction performance of the prediction model was higher than that of the baseline model.

[0085] In summary, this proposal designs a new wind field calculation method to 2.5 All wind field data in the propagation area are aggregated to construct a dynamic directed graph as a wind field representation; a geographical relationship graph between sites is also constructed to 2.5 In addition, based on the constructed semi-dynamic directed site graph, a semi-dynamic graph neural network model is proposed. 2.5 The data and its corresponding graph at the same time point are convolved to extract the spatial relationship, and the output is spliced ​​to obtain a hidden layer containing time series information. The hidden layer is input into the time convolution module to extract the time relationship, and finally the output prediction is obtained through the fully connected layer.

[0086] This proposal designs a new wind field calculation method to analyze the possible impacts of PM2.5 on the site 2.5 All wind farm data in the propagation area are aggregated, which solves the problem of incomplete utilization of regional wind farm information in the past methods and can better capture the spatial dynamic relationship between sites. The geographic features between sites are extracted by constructing a static geographic map and added to the PM. 2.5 In the site relationship of the propagation process, it can overcome the problem of ignoring geographical information other than distance in the past methods. Constructing a simple semi-dynamic graph neural network architecture can complete spatiotemporal prediction for dynamic graphs, and finally realize PM for multiple sites in the region. 2.5 More accurate prediction of concentration.

Claims

1. A regional multi-site PM based on semi-dynamic graph neural network 2.5 The prediction method is characterized in that The following steps are involved: S1. Obtain historical wind fields and corresponding PM 2.5 Data, through data cleaning and preprocessing, construct site PM 2.5 and the corresponding wind field dataset; S2. Use geographic and wind farm information to construct dynamic and static relationship diagrams between sites, and integrate them to construct site PM 2.5 Predicted semi-dynamic directed graph; S3, using the spatial information provided by the semi-dynamic directed graph, combined with the temporal information, to build a spatiotemporal convolutional graph model, using the site PM 2.5 The spatiotemporal convolutional graph model is trained and verified with the corresponding wind field dataset to obtain the spatiotemporal PM 2.5 Predictive models; S4. Current wind field and PM 2.5 Data input space-time PM 2.5 In the prediction model, the output is the site PM within the future set time period 2.5 Predict the results.

2. A regional multi-site PM based on a semi-dynamic graph neural network according to claim 1 2.5 The prediction method is characterized in that The step S1 specifically crawls the data published on the meteorological website to obtain the historical wind field and the corresponding PM 2.5 data.

3. A regional multi-site PM based on a semi-dynamic graph neural network according to claim 1 2.5 The prediction method is characterized in that The specific process of step S2 is: S21, constructing a static directed graph, which is a geographic information graph and is initialized to the distance information between sites; S22, constructing a dynamic directed graph, which is a wind field graph, using a wind field aggregation algorithm, first dividing the wind zone, and then aggregating the wind data, wherein the wind zone division area is a rectangular area with the site connection line as the center line, and the aggregated wind data is weighted addition of the u and v components of each grid point along the site connection line direction; S23. The static directed graph and the dynamic directed graph are superimposed by using the Hadamard product method to obtain a semi-dynamic directed graph that integrates the effects of geographical and meteorological factors on spatial diffusion.

4. A regional multi-site PM based on a semi-dynamic graph neural network according to claim 3 2.5 The prediction method is characterized in that The specific process of step S21 is: using a directed graph G = (V, E) to describe the geographical topology of the site, where V = (V1, V2, ..., V n ) indicates the PM in the study 2.5 Monitoring station, E=(E1,E2,…E m ) indicates how geographical conditions favor PM 2.5 A measure of the spread between sites, with edges represented by the adjacency matrix A g ∈R N×N Indicates that A g The elements in are obtained through learning, ij is initialized to the inverse of the distance and will be reset to zero if it is less than a preset threshold.

5. A regional multi-site PM based on a semi-dynamic graph neural network according to claim 4 2.5 The prediction method is characterized in that The specific process of aggregating wind data in step S22 is as follows: Considering wind speed and direction, the wind volume in the straight line between two stations is calculated, and the wind field relationship between stations is obtained by weighted aggregation based on the distance of grid points: Among them, d AB ,d Ai and d Bi represents the distance between (A,B), (A,i) and (B,i), is the grid point p i The wind field, is the direction vector from station A to station B. Through the dot product of the vectors, we get the wind components in the directions of the two nodes, which are calculated from the u and v components of each grid point: Among them, D AB Indicates the azimuth from A to B.

6. A regional multi-site PM based on a semi-dynamic graph neural network according to claim 5 2.5 The prediction method is characterized in that The semi-dynamic directed graph in step S23 is specifically: Among them, A t is the site adjacency matrix at time t, which is composed of the wind field matrix and the geographic matrix A g For a specific time node, both are of size N×N, where N is the number of sites, and then a self-loop term is added: Among them, I is the N×N identity matrix, λ is a hyperparameter, and λI is the self-loop term in the graph convolutional network.

7. A regional multi-site PM based on a semi-dynamic graph neural network according to claim 6. 2.5 The prediction method is characterized in that The spatiotemporal convolutional graph model in step S3 includes three modules, namely a spatial convolution module, a temporal aggregation module and an output module, wherein the spatial convolution module is composed of a semi-dynamic directed graph convolutional network, the temporal aggregation module is composed of a long short-term memory neural network, and the output module is composed of a fully connected layer.

8. A regional multi-site PM based on a semi-dynamic graph neural network according to claim 7 2.5 The prediction method is characterized in that The spatial convolution module takes the semi-dynamic graph adjacency matrix as input and uses random walk normalization to complete the matrix normalization, which is expressed as: Among them, X t Is the site history PM 2.5 Input, W t is the weight matrix of the graph neural network, and the input data of the spatial convolution module is X t and X t The size is N×t, where N is the number of sites and t is the length of the input historical information. The size of A is N×N×t, g Use the distance between sites to initialize. When the distance between sites i and j exceeds the preset threshold, a ij is 0, otherwise it is initialized to the inverse of the distance.

9. A regional multi-site PM based on a semi-dynamic graph neural network according to claim 8 2.5 The prediction method is characterized in that The process of training the spatiotemporal convolutional graph model in step S3 includes: S31. Use a semi-dynamic graph as the adjacency matrix, input into a graph convolutional neural network, and use random walk normalization to normalize the adjacency matrix to maintain the utility of a directed graph with edge weights. For a sequence input of length l, there is also a dynamic graph of length l and its corresponding adjacency matrix. The input and adjacency matrix are matched one-to-one based on the time step, processed separately by the spatial convolution module, and concatenated into a complete output of the same length: H=[h t-l+1 ,…,h t ] S32, input the obtained hidden layer H into the long short-term memory neural network to extract time information, and finally obtain the site PM through the fully connected layer 2.5 Prediction output of concentration; S33, use the Pytorch framework to build the model, set the neural network parameters of the spatiotemporal convolutional graph model, and set the site PM 2.5 The corresponding wind field data set is input into the spatiotemporal convolutional graph model and trained with the loss function to convert PM 2.5 The parameters with the highest prediction accuracy are retained as the ideal parameters of the spatiotemporal convolutional graph model, and the spatiotemporal PM is obtained. 2.5 Predictive models.

10. A regional multi-site PM based on a semi-dynamic graph neural network according to claim 9. 2.5 The prediction method is characterized in that In step S32, the number of layers of the long short-term memory neural network is set to 2, the batch_first parameter is set to true, and the input size is N is the batch size, L is the length of the hidden layer H sequence, H in is the hidden layer embedding size, and the output size is H out is the size of the output layer; In the fully connected layer, the end of the sequence output by the long short-term memory neural network is extracted for input, and a one-dimensional output Y is obtained through the linear layer, whose size is the product of the number of sites and the number of prediction steps, and then the one-dimensional output is reshaped into an output Y of the same shape (N×pred) , N is the number of sites, and pred is the number of prediction steps.

Images