A Cross-Regional Air Pollution Prediction Method and System Based on Graph Neural Networks

Abstract

The present invention belongs to the field of air pollution prediction, and relates to a cross-regional air pollution prediction method and system based on a graph neural network. The method includes: acquiring sensor collection data of different stations to construct an air quality feature data set; based on the constructed air quality feature data set, using a spatio-temporal graph neural network to build a cross-regional air pollution spatial relationship network graph; based on the air pollution spatial relationship network graph and the air quality feature data set, using GRU and Transformer layers to extract multi-dimensional features, and inputting the extracted multi-dimensional features into a fully connected neural network to obtain multi-step predictions of air pollution concentration. The present invention uses a graph neural network to process historical pollutant concentrations across regions, predict and analyze pollutants. The present invention provides a new solution for air pollutant prediction, and is thus widely applied in the field of air pollution prediction, facilitating the construction of a scientific system for air pollution prevention and control and promoting the development of carbon neutrality.

Description

Technical Field

[0001] The present invention belongs to the field of air pollution prediction, and particularly relates to a cross-regional air pollution prediction method and system based on a graph neural network. Background Art

[0002] Studying the spatio-temporal evolution of air pollution and predicting air pollution concentration have important practical significance in aspects such as environmental protection, urban life, and intelligent transportation. The air quality problem has evolved into a real people's livelihood issue. For health and safety considerations, accurate prediction can help citizens select appropriate travel plans and take protective measures to reduce the harm of air pollution to the body. In addition, air quality prediction can be used for pollution warning, such as closing schools or reducing outdoor sports, so as to reduce the damage caused by pollution and provide effective decision-making support for urban managers.

[0003] Atmospheric pollution is a cross-regional environmental problem. Since the atmosphere is circulating and air pollutants are flowing, air pollution between cities or regions will affect each other. The pollutants in a city come from local emissions and external transportation. Wind is the main driving force for transmission, and other meteorological factors such as boundary layer height, rainfall, humidity, etc. will all affect the accumulation and dissipation of pollutants locally. The Community Multiscale Air Quality (CMAQ) model is an important research tool in the current field of atmospheric environmental management and scientific research, and has been widely applied at home and abroad. However, due to the complexity, high professionalism, and excessive time consumption required for calculation of this system during actual use, the advantages of CMAQ are greatly reduced. In addition, currently, methods such as neural networks, regression analysis, support vector machines, and deep learning are commonly used for prediction. However, the above methods only use air pollution data in one region for prediction, and have good prediction effects on air pollution in a single region, but cannot meet the needs of air pollution modeling between regions. Summary of the Invention

[0004] To solve the problem in the prior art that cross-regional air pollution prediction cannot be achieved with high precision, the present invention provides a cross-regional air pollution prediction method and system based on a graph neural network, which can better reflect the spatio-temporal characteristics of the research object by using the graph neural network and solve the problem of cross-regional air pollution prediction.

[0005] The object of the present invention is achieved by the following technical solutions:

[0006] In the first aspect, the present invention provides a cross-regional air pollution prediction method based on a graph neural network, and the method includes the following steps:

[0007] Step 1: Obtain the sensor collection data of different stations and construct an air quality feature data set;

[0008] Step 2: Based on the constructed air quality feature dataset, use a spatio-temporal graph neural network to build a cross-regional air pollution spatial relationship network diagram;

[0009] Step 3: Based on the air pollution spatial relationship network diagram and the air quality feature dataset, use GRU and Transformer layers to extract multi-dimensional features;

[0010] Step 4: Input the extracted multi-dimensional features into a fully connected neural network to obtain multi-step predictions of air pollution concentration.

[0011] In some alternative embodiments, the data collected by the sensors is air pollution data, meteorological data, GIS data, etc. of the observation station sites collected by various sensors, which are used to build an air pollution relationship network diagram. Among them, data preprocessing is also included for the collected data. The data preprocessing is to fill in missing values, process outliers and standardize the historical time series data of indicators to obtain the final historical time series dataset.

[0012] In some alternative embodiments, before using a spatio-temporal graph neural network to build a cross-regional air pollution spatial relationship network diagram, it further includes:

[0013] Calculate the spatial distance between different sites through the Euclidean distance;

[0014] Determine the geographical environment between different sites according to the geographical environment data;

[0015] Construct a spatial network topology diagram between cross-regional sites according to the spatial distance and geographical environment relationship between different sites.

[0016] In some alternative embodiments, the method for building a cross-regional air pollution spatial relationship network diagram includes:

[0017] Obtain the spatial relationship between observation stations, between observation station sites, and the attribute information of the observation station sites. The attribute information is detected air pollution data, meteorological data, GIS data, etc.;

[0018] Based on the obtained observation stations, the spatial relationship between observation station sites, and the attribute information of the observation station sites, construct a weighted directed graph to represent the air pollution network topology relationship between different regions;

[0019] Extract the spatial features on the site time series to form a spatial feature time series, predict the time series based on the time series dependence relationship of the spatial features, and obtain the graph signal prediction data on the time series.

[0020] In some alternative embodiments, when using GRU and Transformer layers to extract time features, it further includes:

[0021] Based on the air pollution spatial relationship network diagram, use the Gaussian diffusion model to simulate the diffusion of atmospheric pollutants, and obtain the air pollutant concentration through simulation;

[0022] Take the air pollutant concentration and the air quality characteristic data set as input variables and input them into the spatio-temporal graph neural network to extract multi-dimensional features of air pollution;

[0023] Input the extracted multi-dimensional features into a fully connected neural network to obtain multi-step predictions of air pollution concentration.

[0024] In some alternative embodiments, based on the air pollution spatial relationship network diagram, use the Gaussian diffusion model to simulate the diffusion of atmospheric pollutants, and obtain the air pollutant concentration through simulation.

[0025] In some alternative embodiments, when extracting multi-dimensional features of air pollution, use GRU and Transformer layers to extract temporal features, including the following steps:

[0026] Use the graph convolutional neural network GCN to extract spatial features of the spatial relationship of air pollution at different stations;

[0027] The extracted spatial features form a time series matrix, which is used as input to traverse GRU to capture short-term temporal relationships of air pollution;

[0028] Take the output features of GRU as the input of Transformer, capture the global temporal dependence features of air pollution through the Transformer layer, construct an air pollution spatio-temporal evolution model, and extract multi-dimensional features of air pollution.

[0029] In some alternative embodiments, the graph convolutional neural network GCN updates nodes with the neighbor information of directed graph nodes, which is used to capture the longitudinal propagation of air pollution, and the GRU and Transformer layers are used to update the aggregation and diffusion of pollution laterally.

[0030] In a second aspect, the present invention provides a cross-regional air pollution prediction system based on a graph neural network, which realizes cross-regional air pollution prediction by using the aforementioned cross-regional air pollution prediction method based on a graph neural network; the cross-regional air pollution prediction system based on a graph neural network includes a data set construction module, a network construction module, and a prediction module.

[0031] The data set construction module is used to construct an air quality characteristic data set;

[0032] The network construction module is used to build a topological graph of the spatio-temporal relationship of air pollution between regions, and use GRU and Transformer to extract temporal features;

[0033] The prediction module is used to input the output features of the network construction module into the fully connected layer for concentration prediction.

[0034] The technical solution provided by the present invention has the following beneficial effects:

[0035] The cross-regional air pollution prediction method and system based on graph neural network provided by the present invention uses graph neural network to fuse historical data to construct a cross-regional air pollution spatial relationship network diagram, then uses GRU and Transformer to extract features, and then inputs multi-dimensional features into a fully connected neural network to obtain multi-step predictions of air pollution concentration. This method overcomes the problem that traditional air pollution quality prediction cannot achieve cross-regional multi-step prediction, and at the same time solves the deficiency of low prediction accuracy, contributes to the development of carbon neutrality, and assists relevant departments in formulating air pollution prevention and control policies.

[0036] These aspects or other aspects of the present invention will be more clearly understood in the following description of the embodiments. It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention, and do not constitute a limitation to the present invention. In the drawings:

[0038] Figure 1 It is a flowchart of a cross-regional air pollution prediction method based on graph neural network according to an embodiment of the present invention.

[0039] Figure 2 It is a technical roadmap of a cross-regional air pollution prediction model based on graph neural network in an embodiment of the present invention.

[0040] Figure 3 It is a schematic diagram of the GCN model structure in a cross-regional air pollution prediction method based on graph neural network according to an embodiment of the present invention.

[0041] Figure 4 It is an air pollution prediction system based on graph neural network according to an embodiment of the present invention DETAILED DESCRIPTION OF THE EMBODIMENTS

[0042] In order to make the purpose, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0043] The following disclosure provides many different embodiments or examples for implementing different structures of the present invention. To simplify the disclosure of the present invention, components and settings of specific examples are described below. Of course, they are only examples and are not intended to limit the present invention. In addition, the present invention may repeat reference numerals and / or reference letters in different examples. Such repetition is for the purpose of simplification and clarity, and does not itself indicate the relationship between the various embodiments and / or settings discussed. In addition, the present invention provides examples of various specific processes and materials, but those of ordinary skill in the art can be aware of the application of other processes and / or the use of other materials.

[0044] Reference Figure 1 is a flowchart of a cross-regional air pollution prediction system based on a graph neural network.

[0045] An embodiment of the present invention provides a cross-regional air pollution prediction method based on a graph neural network. First, a cross-regional air pollution relationship network is constructed, and then a GRU and Transformer layer are used to construct an air pollution spatio-temporal evolution model to achieve multi-step prediction of air pollution. The cross-regional air pollution prediction method includes the following steps S1-S4;

[0046] S1. Obtain sensor collection data of different stations and construct an air quality feature data set.

[0047] In this embodiment, the sensor collection data is air pollution data, meteorological data, traffic data, etc. of observation station sites collected by various sensors, and is used to construct an air pollution relationship network diagram.

[0048] Among them, when collecting data, multi-source big data such as air pollution data, meteorological data, and traffic data (GIS data) collected by various sensors includes the following aspects:

[0049] Meteorological data includes: wind speed (Wind_speed), wind direction (Wind_direction), temperature (Temperature), relative humidity (Relative humidity), pressure (Pressure), rainfall (Rainfall), etc.;

[0050] Air pollution data includes: PM2.5 concentration, SO2 concentration, NO2 concentration, O3 concentration, etc.;

[0051] Traffic data: traffic road network, elevation of 1km digital elevation model in Chinese regions, x and y coordinates.

[0052] In this embodiment, the collected data also includes data preprocessing, and the data preprocessing is to fill in missing values of the historical time series data of indicators by the mean method, process outliers, and perform standardization to obtain the final historical time series data set.

[0053] S2. Based on the constructed air quality feature data set, use a spatio-temporal graph neural network to build a cross-regional air pollution spatial relationship network graph.

[0054] In this embodiment, calculate the distances between different stations and whether there are special geographical environments such as mountains between different stations, and use a graph neural network to build a cross-regional air pollution spatial relationship network graph.

[0055] Among them, the method for constructing a cross-regional air pollution spatial relationship network graph includes:

[0056] Obtain the spatial relationship between observation stations, the spatial relationship between observation station sites, and the attribute information of observation station sites, where the attribute information is detected air pollution data, meteorological data, GIS data, etc.;

[0057] Based on the obtained observation stations, the spatial relationship between observation station sites, and the attribute information of observation station sites, construct a weighted directed graph to represent the air pollution network topology relationship between different regions;

[0058] Extract the spatial features on the site time series to form a spatial feature time series, and predict the time series based on the time series dependence relationship of the spatial features to obtain the graph signal prediction data on the time series.

[0059] In this embodiment, step 2 specifically includes the following steps:

[0060] Step 2.1: The air pollution network topology relationship between different regions can be represented by a weighted directed graph G=(V, E, A), where V is the node representing the observation station, E is the edge representing the spatial relationship between stations, and A is the adjacency weight matrix, which is the detected geographical distance.

[0061] In this embodiment, p t ∈H N*h represents the attribute matrix of the node at time t. Among them, N is the number of observation stations, and h is the number of attributes of the observation station. The attributes include meteorological elements such as temperature, pressure, relative humidity, rainfall, and air pollution elements such as PM2.5.

[0062] In this embodiment, Q t ∈H M*q is the attribute matrix of the edge at time t, where M is the number of edges and q is the corresponding number of attributes. The edge attributes include wind speed, wind direction, etc.

[0063] In this embodiment, Ai,j Denote the node v i and the node v j The spatial distance between them is shown in Formula (1).

[0064]

[0065] Step 2.1: In Step S1, each monitoring site has its own time series. The spatial distribution of PM2.5 concentration data of all monitoring sites at a certain moment can be abstracted into a topological graph. Refer to Figure 2 and Figure 3 , first extract the spatial features at each time step of the site to form a spatial feature time series, and then predict the time series based on the temporal dependence relationship of the spatial features.

[0066] S3. Based on the air pollution spatial relationship network graph and the air quality feature data set, use GRU and Transformer layers to extract multi-dimensional features.

[0067] In this embodiment, when using GRU and Transformer layers to extract time features, it further includes:[[]]

[0068] Use the Gaussian diffusion model to simulate the diffusion of air pollutants based on the air pollution spatial relationship network graph, and simulate the air pollutant concentration obtained;

[0069] Take the air pollutant concentration obtained by the Gaussian diffusion model and the air quality feature data set as input variables and input them into the spatio-temporal graph neural network to extract multi-dimensional features of air pollution.

[0070] In this embodiment, when extracting multi-dimensional features of air pollution, using GRU and Transformer layers to extract time features includes the following steps:[[]]

[0071] Use the graph convolutional neural network GCN to extract the spatial features of the spatial relationship of air pollution at different sites;

[0072] The extracted spatial features form a time series matrix, which is used as input to traverse GRU to capture the short-term time relationship of air pollution;

[0073] Take the output features of GRU as the input of Transformer, capture the global time-dependent features of air pollution through the Transformer layer, construct an air pollution spatio-temporal evolution model, and extract multi-dimensional features of air pollution.

[0074] The graph convolutional neural network GCN updates the nodes with the neighbor information of the directed graph nodes, which is used to capture the longitudinal propagation of air pollution, and the GRU and Transformer layers are used to update the aggregation and diffusion of pollution horizontally.

[0075] Reference Figure 2 , in order to improve the prediction accuracy, the predicted concentration of air pollutants in the Gaussian diffusion model is newly added to the input of the GCN.

[0076] The Gaussian diffusion model is the main model used to simulate the diffusion process of atmospheric pollutants. The model is mainly applicable to large-scale light clouds and neutral clouds, with stable and uniform pollution sources. According to different diffusion models, it can be divided into two types: the puff model and the plume model. For short-term diffusion situations such as emergency leaks, the puff model can be used to simulate the diffusion of atmospheric pollutants. For the situation where the pollution source diffuses continuously and stably for a long time, the Gaussian plume model can be used to simulate the diffusion of atmospheric pollutants.

[0077] First, establish a coordinate system. Take the position of the pollution source as the coordinate origin, the azimuth of the positive wind direction as the x-axis, the direction perpendicular to the x-axis on the horizontal plane as the y-axis, and the direction perpendicular to the horizontal plane as the z-axis.

[0078] The Gaussian puff model is shown in formula (2):

[0079]

[0080] In the formula, C(x, y, z, t) represents the pollutant concentration value at a certain position (x, y, z) at time t; σ x , σ y , σ z represent the diffusion coefficients in the x, y, and z directions. The diffusion coefficients are determined by factors such as distance and atmospheric stability.

[0081] The Gaussian plume model is shown in formula (2):

[0082]

[0083] For the convenience of representation, C t is used to represent C(x, y, z, t) later.

[0084] Reference Figure 2 , at time t, the air pollution concentration X t , as well as the attribute matrices of the next T steps [p t+1 , …, p t+T and [Q t+1 , …, Q t+T , the predicted value C of the pollutant concentration in the Gaussian diffusion model t and the graph structure are input into the model to obtain the predicted values at the next T times as shown in formula (4).

[0085]

[0086] Among them, F(·) = g(…g(g(·))), for a total of T times.

[0087]

[0088] In this embodiment, step S3 includes the following steps:

[0089] The air pollution concentration between regions depends not only on the sequence pattern in the time dimension but also on other regions in the space dimension. A spatio-temporal evolution model of inter-regional air pollution based on the graph convolutional neural network GCN and Transformer is proposed to achieve multi-step prediction of air pollutant concentrations.

[0090] Specifically, step 3.1: Use the graph convolutional neural network GCN to extract the spatial relationship of air pollution.

[0091] Reference Figure 3 , the graph neural network updates the nodes with the neighbor information of the nodes to capture the longitudinal propagation of air pollution. The GRU and Transformer layers are used to update the aggregation and diffusion of pollution horizontally.

[0092] The graph neural network iteratively aggregates neighbor information. From the predicted value at the previous moment Contemporary attributes And the pollutant concentration obtained by using the Gaussian diffusion model at time t Constitute, edge Consists of the vertices adjacent to the edge and the edge attributes. For a vertex i, the pollution has an input And output Two aspects of influence. Therefore, for the spatial correlation Of node i is the sum of the influences of all neighbor nodes on itself.

[0093] In this embodiment, the graph neural network iteratively aggregates neighbor information through the following formulas (5)-(7):

[0094]

[0095]

[0096]

[0097] Step 3.2: The spatial features extracted by the graph convolutional neural network form a time series matrix, which is used as the input to traverse the input GRU to capture the short-term time relationship (or local time correlation) of air pollution; where each cell is represented by a node And spatial correlation As the input, it enables the GRU to consider both the spatial transmission and the temporal diffusion, as described by formulas (8)-(12):

[0098]

[0099]

[0100]

[0101]

[0102]

[0103] Step 3.3: To better predict the PM2.5 concentration, after the GRU layer, a Transformer layer is added to capture the global features of PM2.5. The feature output of the GRU is used as the input of the Transformer, and after these features are fused, they are output. The relationships between different observation stations and the prediction site are different. This inter-site correlation will change relatively with different datasets. The attention mechanism can obtain weights during the training process and adjust the weights according to the characteristics of the data itself. By means of weighting, the accuracy of the model is improved, and its formula is (13):

[0104]

[0105] α i is the weight.

[0106] The attention function is composed of Query, Key, and Value. Q i = h i W Q , K i = h i W K , V i = h i W V , where W Q , W K , W V are mapping matrices, and the weight is given by formula (14):

[0107]

[0108] Compared with single-head attention, multi-head attention jointly aggregates information from different representation subspaces, thereby improving the model's representation ability. Therefore, multi-head attention is adopted, as shown in formulas (15)-(16):

[0109] Multihead(h i) = Concat(head1, …, head s )W o ; (15)

[0110] where

[0111] Among them, is the s-th attention head, and W o is the linear output mapping.

[0112] Figure 3 includes Transformer layers. GRU can capture local time information. Then in the air pollution system, the time information is not only sequentially related. Therefore, after GRU, Transformer layers are used to capture global information. The Transformer layer includes a multi-head attention layer, a shared feed-forward neural network layer, and a batch normalization layer.

[0113] S4. Input the extracted multi-dimensional features into a fully connected neural network to obtain multi-step predictions of air pollution concentration.

[0114] In this embodiment, step S4 is after the Transformer layer, and its output is used as the input of the prediction layer to make multi-step predictions of future air pollution concentration using past historical data.

[0115] In this embodiment, MAE (Mean Absolute Error) and RMSE (Root Mean Square Error) are used to verify the accuracy of the model, and common meteorological indicators are used to measure the performance near the pollution threshold, including: CSI (critical success index), POD (probability of detection), FAR (false alarm rate). The higher the CSI and POD, the better the model. The evaluation results of the final model are shown in Table 1. It can be seen that the algorithm in this patent has good practical prospects and generalization ability.

[0116] Table 1 Prediction result table

[0117]

[0118] The cross-regional air pollution prediction method based on graph neural network in this embodiment can be designed as an application software, such as a cross-regional air pollution prediction system based on graph neural network, and then the cross-regional air pollution prediction system based on graph neural network is loaded into a terminal device to achieve corresponding cross-regional air pollution prediction.

[0119] See Figure 4As shown in the figure, an embodiment of the present invention provides an air pollution prediction system based on a graph neural network. The system includes a dataset construction module 100, a network construction module 200, and a prediction module 300.

[0120] The dataset construction module 100 is used to construct an air quality feature dataset, including air pollution data, meteorological data, GIS data, etc.

[0121] The network construction module 200 is used to construct a spatio-temporal relationship network topology map of air pollution between regions, and uses GRU and Transformer to extract time features.

[0122] The network construction module 200 uses the collected data to construct a temporal graph convolutional network to obtain a spatial network topology map between stations; the dataset and the air pollutant concentration obtained through the Gaussian plume model are used as input variables and input into the graph convolutional neural network, and GRU and Transformer are used to obtain multi-dimensional feature values.

[0123] The prediction module 300 is used to input the output features of the network construction module into a fully connected layer for concentration prediction. When in use, the multi-dimensional feature prediction values are input into a fully connected neural network to output the final air pollutant concentration prediction value of the target station in the future time period.

[0124] The present invention realizes multi-step prediction of air pollution across regions by fusing the dynamic spatio-temporal correlation between stations, with the help of graph neural networks, Transformer, etc., which helps relevant departments to carry out and take corresponding control measures.

[0125] In an embodiment of the present invention, a computer device is provided, including a memory and a processor. A computer program is stored in the memory, and when the processor executes the computer program, the steps in the above method embodiments are implemented.

[0126] In an embodiment of the present invention, a computer-readable storage medium is also provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0127] Those of ordinary skill in the art can understand that all or part of the processes of implementing the above method embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the above method embodiments. Among them, any reference to a memory, storage, database, or other medium used in the various embodiments provided by the present invention can include at least one of non-volatile and volatile memories.

[0128] The advantages of the present invention are as follows: The cross-regional air pollution prediction method and system based on graph neural network provided by the present invention use graph neural network to fuse historical data and Gaussian diffusion model to construct a cross-regional air pollution spatial relationship network diagram, then use GRU and Transformer to extract features, and then input multi-dimensional features into a fully connected neural network to obtain multi-step prediction of air pollution concentration. This method overcomes the problem that traditional air pollution quality prediction cannot achieve cross-regional multi-step prediction, and at the same time solves the deficiency of low prediction accuracy, helps the development of carbon neutrality, and assists relevant departments in formulating air pollution prevention and control policies.

[0129] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A cross-regional air pollution prediction method based on graph neural network, characterized in that, The method includes: Obtaining sensor collection data from different sites and constructing an air quality feature dataset; Based on the constructed air quality feature dataset, using a spatio-temporal graph neural network to build a cross-regional air pollution spatial relationship network graph; Based on the air pollution spatial relationship network graph and the air quality feature dataset, using GRU and Transformer layers to extract multi-dimensional features; Inputting the extracted multi-dimensional features into a fully connected neural network to obtain multi-step predictions of air pollution concentration; Among them, the cross-regional air pollution spatial relationship network graph is constructed through the following steps: Calculating the spatial distance between different sites through the Euclidean distance; Determining the geographical environment between different sites according to geographical environment data; Constructing a spatial network topology graph between cross-regional sites according to the spatial distance and geographical environment relationship between different sites; Among them, when using GRU and Transformer layers to extract time features, it also includes: Based on the air pollution spatial relationship network graph, using a Gaussian diffusion model to simulate the diffusion of air pollutants and obtaining the concentration of air pollutants; Taking the concentration of air pollutants obtained through the Gaussian diffusion model and the air quality feature dataset as input variables and inputting them into the spatio-temporal graph neural network to extract multi-dimensional features of air pollution; Inputting the extracted multi-dimensional features into a fully connected neural network to obtain multi-step predictions of air pollution concentration; Among them, when extracting multi-dimensional features of air pollution, using GRU and Transformer layers to extract time features includes the following steps: Using a graph convolutional neural network GCN to extract spatial features of the spatial relationship of air pollution at different sites; The extracted spatial features form a time series matrix and are used as input to traverse GRU to capture the short-term time relationship of air pollution; Taking the output features of GRU as the input of Transformer, and capturing the global time-dependent features of air pollution through the Transformer layer to construct an air pollution spatio-temporal evolution model and extract multi-dimensional features of air pollution.

2. The cross-regional air pollution prediction method based on a graph neural network according to claim 1, wherein The sensor collection data also includes data preprocessing, and the data preprocessing is to fill in missing values, process outliers and standardize the historical time series data of indicators to obtain the final historical time series dataset.

3. The cross-regional air pollution prediction method based on graph neural network according to claim 1, wherein The method for constructing a cross-regional air pollution spatial relationship network graph includes: Obtaining observation stations, the spatial relationship between observation station sites, and the attribute information of observation station sites, where the attribute information is detected air pollution data, meteorological data, and GIS data; Based on the obtained observation stations, the spatial relationship between observation station sites, and the attribute information of observation station sites, constructing a weighted directed graph to represent the air pollution network topology relationship between different regions.

4. The cross-regional air pollution prediction method based on a graph neural network according to claim 3, wherein Based on the air pollution spatial relationship network graph, using a Gaussian diffusion model to simulate the diffusion of atmospheric pollutants and obtaining the concentration of air pollutants.

5. The cross-regional air pollution prediction method based on a graph neural network according to claim 4, wherein The graph convolutional neural network GCN updates nodes with the neighbor information of directed graph nodes and is used to capture the longitudinal propagation of air pollution, and the GRU and Transformer layers are used to update the lateral aggregation and diffusion of pollution.

6. A cross-regional air pollution prediction system based on graph neural network, characterized in that, Implement cross - regional air pollution prediction using the cross - regional air pollution prediction method based on graph neural network described in any one of claims 1 - 5; the system includes: A data set construction module for constructing an air quality feature data set; A network construction module for constructing a spatio - temporal relationship network topology of air pollution between regions, using GRU and Transformer to extract time features; a prediction module for inputting the output features of the network construction module into a fully - connected layer for concentration prediction.

Images