A method and system for assessing the importance of monitoring sites based on graph neural networks.

Abstract

This invention belongs to the fields of air pollution monitoring and interpretable machine learning, specifically relating to a method and system for assessing the importance of monitoring stations based on graph neural networks. The method includes: acquiring relevant data of monitoring stations within a target area and constructing a graph structure representation; training a graph neural network model using the graph structure representation; obtaining the first graph structure representation and its first prediction result for the latest time period; searching the first graph structure representation using a greedy best-first search algorithm to obtain a second graph structure representation and its second prediction result, with an improved Shapley Value as the heuristic function; comparing the first and second prediction results to obtain graph structure representations of important monitoring stations; ranking the importance of the graph structure representations of important monitoring stations by perturbation; and designing a verification algorithm to verify the accuracy of the importance ranking results. This invention provides a station assessment method that helps assist in station site selection and the demolition of existing stations.

Description

Technical Field

[0001] This invention belongs to the fields of air pollution monitoring and interpretable machine learning, specifically relating to a method and system for assessing the importance of monitoring stations based on graph neural networks. Background Technology

[0002] With the acceleration of national industrialization and urbanization, more and more environmental pollution problems have attracted public attention. Air quality environments, which are different from water, marine and soil environments, are not easily distinguished by the naked eye and are more difficult to manage and control. In order to monitor and prevent air pollution, many cities have established their own air quality monitoring stations.

[0003] Currently, the site selection for air quality monitoring stations requires a comprehensive consideration of factors such as population size, built-up area conditions, topography, and meteorological conditions in different monitoring areas to determine the layout of environmental quality grid points. Station locations should be distributed as evenly as possible, covering the entire monitoring area. The distribution of buildings and traffic within the monitoring area, as well as planned cleanup, must be considered to enhance operability. Because the coverage area of ​​each air quality monitoring station is limited and sparsely distributed, it is impossible to completely cover an entire city; therefore, algorithms are needed to determine the importance of each station.

[0004] Machine learning algorithms focus on fitting data and their non-intuitive internal mechanisms, resulting in high computational complexity in machine learning models. Furthermore, machine learning algorithms often simply use historical air pollutant data as input, failing to determine which data points are significant. Shapley Value offers some interpretability, but its time complexity is high. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a method and system for assessing the importance of monitoring sites based on graph neural networks.

[0006] In a first aspect, the present invention provides a method for assessing the importance of monitoring sites based on graph neural networks, comprising:

[0007] S1. Obtain relevant data for each monitoring station within the target area, including geographical information and pollutant concentrations at different historical time periods; construct a graphical representation of all monitoring stations within the target area at different historical time periods using the relevant data;

[0008] S2. The graph structure representations of all monitoring stations at different historical time periods are used as input to train the graph neural network model, and the output is the air quality index of the target area.

[0009] S3. Obtain the pollutant concentrations of all monitoring stations in real time within the latest time period, construct the first graph structure representation corresponding to the latest time period, and input the first graph structure into the trained graph neural network model to obtain the first prediction result;

[0010] S4. Search the first graph structure representation using the greedy best-first search algorithm to obtain the second graph structure representation; wherein, the improved Shapley Value is used as the heuristic function of the greedy best-first search algorithm;

[0011] S5. Input the second graph structure representation into the trained graph neural network model to obtain the second prediction result. Determine whether the second prediction result is less than the similarity threshold of the first prediction result. If so, the second graph structure representation is output as the graph structure representation of the important monitoring station. If not, return to S4 to continue the search.

[0012] S6. Construct an importance ranking module, which includes multiple ranking methods. First, the importance of the important monitoring station map structure representation is ranked by perturbation method.

[0013] S7. Design a verification algorithm to verify whether the importance ranking result is accurate. If so, output the importance ranking result; otherwise, change the ranking method through the importance ranking module.

[0014] Furthermore, step S1, which generates a graph structure representation of all monitoring stations over the historical time period u, is as follows:

[0015] S11. Obtain the latitude and longitude information of each monitoring station within the target area, and calculate the distance between any two monitoring stations. The calculation formula is as follows:

[0016]

[0017]

[0018]

[0019] in, Indicates monitoring stations and monitoring stations The distance between, Indicates monitoring stations and monitoring stations The difference in longitude, Indicates monitoring stations and monitoring stations The difference in latitude, Represents the Earth's radius. Indicates monitoring stations latitude, Indicates monitoring stations latitude, Indicates monitoring stations longitude, Indicates monitoring stations longitude; , This indicates the total number of monitoring stations within the target area;

[0020] S12. Select any monitoring station within the target area as the first monitoring station;

[0021] S13. Sort the distances between the first monitoring station and the other monitoring stations in ascending order, and select the first... Each monitoring station adds an edge to the first monitoring station; if an edge already exists between monitoring stations, it is not necessary to add it again.

[0022] S14. Determine if there are any monitoring stations that have not been the first monitoring station. If so, return to step S13 as the first monitoring station. If not, obtain the graph structure representation of all monitoring stations.

[0023] S15. In the current graph structure representation, each node represents a monitoring station, and the pollutant concentration of the monitoring station in the historical time period u is used as the node feature. The feature vector of the edge is represented as [distance, longitude 1, latitude 1, longitude 2, latitude 2], which is used to retain the distance information and location information between monitoring stations. Among them, latitude 1 and longitude 1 are the location information of one monitoring station connected by the edge, and latitude 2 and longitude 2 are the location information of another monitoring station connected by the edge.

[0024] Furthermore, the improved Shapley Value is expressed as:

[0025]

[0026]

[0027]

[0028]

[0029] in, express exist The contribution of the middle to the output, Represents a random graph structure. Represents the set of nodes A randomly selected subset of nodes in the middle. This represents a graph neural network model. This represents the corrected coefficient. This represents the set of all monitoring stations within the target area. express The set of nodes in the middle, Indicates the number of samples. Represents a subset of nodes The weight, This represents the difference between the sum of the weights of 1 and a random subset of T nodes. Represents a subset of nodes The number of nodes in This indicates the number of all monitoring stations within the target area.

[0030] Furthermore, the improved Shapley Value is used as the heuristic function for the greedy best-first search algorithm. The process includes:

[0031] S21. Initialize the second graph structure representation within the latest time period. ;

[0032] S22. Based on the heuristic function, search for important monitoring stations in the first graph structure representation and add them to the second graph structure representation. middle;

[0033] S23. Represent the structure of the second diagram. The trained graph neural network model is input to obtain a second prediction result. It is then determined whether the second prediction result is less than a similarity threshold compared to the first prediction result. If so, the second graph structure represents... If the output is not a representation of the important monitoring station map structure, return to S22 to continue the search;

[0034] The formula for determining whether the second prediction result is close to the first prediction result is:

[0035]

[0036] in, This indicates the first prediction result. The first graph structure represents all monitoring stations within the target area in the latest time period. This indicates the second prediction result. Indicates control and exist The hyperparameter for the degree of similarity is the similarity threshold.

[0037] Furthermore, step S6 uses a perturbation method to represent the map structure of important monitoring stations. Ranking by importance includes:

[0038] S31. Set the iteration threshold Let the number of iterations be ;

[0039] S32. Randomly select from the structural representation of the first figure. Each monitoring station is masked, that is, in the structural representation of the first figure, this... The node feature corresponding to each monitoring station is represented as 0;

[0040] S33. A greedy best-first search algorithm is used to search the results obtained in S32 to obtain the structure of the important monitoring stations behind the mask. ;

[0041] S33. Determine if the condition is satisfied. If so, then Then return to step S32; otherwise, the iteration ends, and the perturbation set is obtained. Record the structure of important monitoring stations The importance ranking of important monitoring stations is obtained by ranking the number of times each station appears in the disturbance set from highest to lowest frequency.

[0042] Furthermore, the verification process in step S7 includes:

[0043] S41. Based on the importance ranking results, select the top... Each monitoring station and subsequent Each monitoring station performs a masking operation on the graph structure representation of important monitoring stations to obtain the first mask image and the second mask image;

[0044] S42. Input the first mask image and the second mask image into the trained graph neural network model to obtain the first predicted value and the second predicted value;

[0045] S43. If the first difference is greater than the second difference, output the importance ranking result; otherwise, change the ranking method.

[0046] In a second aspect, based on the method provided in the first aspect, this invention proposes a monitoring site importance assessment system based on a graph neural network, comprising a historical data storage module, a graph structure composition module, a graph neural network training module, a real-time data acquisition module, a greedy search module, a first visualization module, a first judgment module, a sorting module, a second judgment module, and a second visualization module, wherein:

[0047] The historical data storage module is used to store pollutant concentration data for all monitoring stations in the region for historical time periods, as well as the geographical information of each monitoring station.

[0048] The graph structure module is used to construct a graph structure representation using the geographical information of monitoring stations and pollutant concentration data;

[0049] The graph neural network training module is used to train a graph neural network based on a graph structure representation constructed from historical data, and outputs the air quality index of the region.

[0050] The real-time data acquisition module is used to acquire the pollutant concentrations of all monitoring stations within the latest time period, and construct the first graph structure representation corresponding to the latest time period through the graph structure composition module. The first graph structure is then input into the trained graph neural network model to obtain the first prediction result.

[0051] The greedy search module is used to search the first graph structure representation according to the greedy best-first search algorithm to obtain the second graph structure representation, and input the second graph structure representation into the trained graph neural network model to obtain the second prediction result;

[0052] The first visualization module is used to output the search results of the greedy search module and display them on the screen;

[0053] The first judgment module is used to determine whether the second prediction result is close to the first prediction result. If so, the second graph structure representation is output as the important monitoring station graph structure representation. If not, it returns to the greedy search module to continue searching.

[0054] The sorting module is used to sort the importance of the map structure representation of important monitoring stations;

[0055] The second judgment module is used to verify whether the importance ranking result is accurate through a verification algorithm. If it is, the importance ranking result is output; otherwise, the ranking module changes the ranking method and performs the ranking.

[0056] The second visualization module is used to output the final importance ranking results on the display screen.

[0057] The beneficial effects of this invention are:

[0058] This invention provides a method and system for assessing the importance of monitoring stations based on graph neural networks. First, discrete monitoring stations are connected by their latitude, longitude, and pollutant concentrations to construct a graph structure representation of the target area. A greedy best-first search algorithm is then used to search for important monitoring stations within the graph structure representation. Simultaneously, the Shapley Value is improved and used as a heuristic function for the search algorithm, ensuring the selection process for important monitoring stations is interpretable and understandable to humans. A verification method is designed to ensure the accuracy of the ranking of important monitoring stations by importance. Based on the ranking of station importance, the importance of a station can be distinguished, primarily providing a basis for the site selection of monitoring stations and the dismantling of existing monitoring stations. Attached Figure Description

[0059] Figure 1This is the overall flowchart of the graph neural network-based system for evaluating the importance of monitoring sites according to the present invention;

[0060] Figure 2 The pseudocode for searching important monitoring sites in this invention;

[0061] Figure 3 This is the pseudocode for sorting important monitoring stations by importance according to the present invention. Detailed Implementation

[0062] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0063] This invention proposes a method and system for assessing the importance of monitoring sites based on graph neural networks, aiming to provide a certain degree of interpretability for black-box machine learning models, such as... Figure 1 As shown, the main steps include: constructing a complete graph representation based on the geographical distribution information of monitoring stations; training a graph neural network model using monitoring data from the monitoring stations, and using this model to predict the air quality index level of all monitoring stations in the graph representation; using an improved Shapley Value as the heuristic function for a greedy best-first search algorithm to obtain a subgraph from the complete graph representation and output the subgraph information, i.e., the important monitoring stations found; using a perturbation method to obtain and rank the importance of all important monitoring stations in the subgraph; and finally, using a verification algorithm to verify the results and visually outputting them. The results can then be used to assist in the selection and removal of monitoring stations.

[0064] In one embodiment, such as Figure 1 As shown, the specific content of a monitoring site importance assessment method based on graph neural networks includes:

[0065] S1. Obtain relevant data for each monitoring station within the target area, including geographical information and pollutant concentrations at different historical time periods; construct a graphical representation of all monitoring stations within the target area at different historical time periods using the relevant data;

[0066] Specifically, the pollutant concentration monitored by each monitoring station will change over time. This invention can use one day as a time period to calculate the pollutant concentration, or it can use one hour as a time period to calculate the pollutant concentration, using the same time period for calculation.

[0067] S2. Use the graph structure of all monitoring stations at different historical time periods as input to train a graph neural network model, and output the air quality index of the target area.

[0068] Specifically, the Air Quality Index (AQI) has 6 levels, so the output dimension of the graph neural network model is 6-dimensional, with each dimension corresponding to the probability of each AQI level, and the one with the highest probability is the output result.

[0069] S3. Obtain the pollutant concentrations of all monitoring stations in real time within the latest time period, construct the first graph structure representation corresponding to the latest time period, and input the first graph structure into the trained graph neural network model to obtain the first prediction result;

[0070] S4. Search the first graph structure representation corresponding to the latest time period using the greedy best-first search algorithm to obtain the second graph structure representation; wherein, the improved Shapley Value is used as the heuristic function of the greedy best-first search algorithm;

[0071] S5. Input the second graph structure representation into the trained graph neural network model to obtain the second prediction result. Determine whether the second prediction result is less than the similarity threshold of the first prediction result. If so, the second graph structure representation is output as the graph structure representation of the important monitoring station. If not, return to S4 to continue the search.

[0072] S6. Construct an importance ranking module, which includes multiple ranking methods. First, the importance of the important monitoring station map structure representation is ranked by perturbation method.

[0073] S7. Design a verification algorithm to verify whether the importance ranking result is accurate. If so, output the importance ranking result; otherwise, change the ranking method through the importance ranking module.

[0074] Specifically, when using the perturbation method for sorting, this method is unsuitable due to the inherent limitations of the data (for example, the number of nodes in the first graph structure is too small, resulting in a limited amount of data after perturbation). Therefore, an importance ranking module is provided that includes multiple methods for ranking the importance of network nodes, including node importance ranking based on local attributes, global attributes, network location attribute indicators, and random walks.

[0075] Specifically, in step S1, a graph structure representation for different historical time periods is constructed using the location information of the monitoring stations and the pollutant concentration information for different historical time periods. The generation process of the graph structure representation for any historical time period is as follows:

[0076] S11. Obtain the latitude and longitude information of each monitoring station within the target area, and calculate the distance between any two monitoring stations. The calculation formula is as follows:

[0077]

[0078]

[0079]

[0080] in, Indicates monitoring stations and monitoring stations The distance between, Indicates monitoring stations and monitoring stations The difference in longitude, Indicates monitoring stations and monitoring stations The difference in latitude, Represents the Earth's radius. Indicates monitoring stations latitude, Indicates monitoring stations latitude, Indicates monitoring stations longitude, Indicates monitoring stations longitude; , This indicates the total number of monitoring stations within the target area;

[0081] S12. Select any monitoring station within the target area as the first monitoring station;

[0082] S13. Sort the distances between the first monitoring station and the other monitoring stations in ascending order, and select the first... Each monitoring station adds an edge to the first monitoring station; if an edge already exists between monitoring stations, it is not necessary to add it again.

[0083] S14. Determine if there are any monitoring stations that have not been the first monitoring station. If so, return to step S13 as the first monitoring station. If not, obtain the graph structure representation of all monitoring stations.

[0084] S15. In the graph structure representation of historical time period u, each node represents a monitoring station, and the pollutant concentration of the monitoring station in historical time period u is used as the node feature. The feature vector of the edge is represented as [distance, longitude 1, latitude 1, longitude 2, latitude 2], which can preserve the distance and location information between stations.

[0085] Specifically, such as Figure 2 As shown, the improved Shapley Value is used as the heuristic function for the greedy best-first search algorithm to search for important monitoring sites. The process includes:

[0086] S21. Initialize the second graph structure representation within the latest time period. ;

[0087] S22. Based on the heuristic function, search for important monitoring sites in the first graph structure representation within the latest time period and add them to the second graph structure representation. middle;

[0088] Specifically, it is expressed as:

[0089]

[0090]

[0091] This represents the node with the largest heuristic function value. express The set of neighboring nodes;

[0092] S23. Represent the structure of the second diagram. The trained graph neural network model is input to obtain a second prediction result. It is then determined whether the second prediction result is close to the first prediction result obtained by the trained graph neural network model with the first graph structure representation. If so, the second graph structure representation is... If the output is not a representation of the important monitoring station map structure, return to S22 to continue the search;

[0093] The formula for determining whether the second prediction result is close to the first prediction result is:

[0094]

[0095] in, This indicates the first prediction result. The first graph structure represents all monitoring stations within the target area in the latest time period. This indicates the second prediction result. Indicates control and exist The degree of similarity in the data is a hyperparameter.

[0096] Specifically, the improved Shapley Value is expressed as:

[0097]

[0098]

[0099]

[0100]

[0101] in, express exist The contribution of the middle to the output, Represents a random graph structure. Represents the set of nodes A randomly selected subset of nodes in the middle. This represents a graph neural network model. This represents the corrected coefficient. This represents the set of all monitoring stations within the target area. express The set of nodes in the middle, Indicates the number of random samples. Represents a subset of nodes The weight, This represents the difference between the sum of the weights of 1 and a random subset of T nodes. Represents a subset of nodes The number of nodes in This indicates the number of all monitoring stations within the target area.

[0102] The original Shapley Value has a time complexity of O(n log n). The time complexity is exponential; however, the improved Shapley Value in this embodiment is related to the number of samples T. This T is set manually, so that when the number of nodes N in the graph is very small, T can be equal to... When the number of nodes is very large, the manually set T is much smaller than the value of T. This allows the accuracy to remain within a reasonable range. In short, it represents a trade-off between time complexity and accuracy.

[0103] In one embodiment, such as Figure 3 As shown, step S6 uses a perturbation method to represent the map structure of important monitoring stations. Ranking by importance includes:

[0104] S31. Set the iteration threshold Let the number of iterations be ;

[0105] S32. Randomly select from the structural representation of the first figure. Each monitoring station is masked, that is, in the structural representation of the first figure, this... The node feature corresponding to each monitoring station is represented as 0;

[0106] S33. A greedy best-first search algorithm is used to search the results obtained in S32 to obtain the structure of the important monitoring stations of the mask. ;

[0107] S33. Determine if the condition is satisfied. If so, then Then return to step S32; otherwise, the iteration ends, and the perturbation set is obtained. Record the structure of important monitoring stations The importance ranking of important monitoring stations is obtained by ranking the number of times each station appears in the disturbance set from highest to lowest frequency.

[0108] In one embodiment, the verification process in step S7 includes:

[0109] S41. Based on the importance ranking results, select the top... Each monitoring station and subsequent Each monitoring station performs a masking operation on the graph structure representation of important monitoring stations to obtain the first mask image and the second mask image;

[0110] S42. Input the first mask image and the second mask image into the trained graph neural network model to obtain the first predicted value and the second predicted value;

[0111] S43. If the first difference is greater than the second difference, output the importance ranking result; otherwise, change the ranking method.

[0112] In one embodiment, a monitoring site importance assessment system based on graph neural networks includes a historical data storage module, a graph structure composition module, a graph neural network training module, a real-time data acquisition module, a greedy search module, a first visualization module, a first judgment module, a sorting module, a second judgment module, and a second visualization module, wherein:

[0113] The historical data storage module is used to store pollutant concentration data for all monitoring stations in the region for historical time periods, as well as the geographical information of each monitoring station.

[0114] The graph structure module is used to construct a graph structure representation using the geographical information of monitoring stations and pollutant concentration data;

[0115] The graph neural network training module is used to train a graph neural network based on a graph structure representation constructed from historical data, and outputs the air quality index of the region.

[0116] The real-time data acquisition module is used to acquire the pollutant concentrations of all monitoring stations within the latest time period, and construct the first graph structure representation corresponding to the latest time period through the graph structure composition module. The first graph structure is then input into the trained graph neural network model to obtain the first prediction result.

[0117] The greedy search module is used to search the first graph structure representation according to the greedy best-first search algorithm to obtain the second graph structure representation, and input the second graph structure representation into the trained graph neural network model to obtain the second prediction result;

[0118] The first visualization module is used to output the search results of the greedy search module and display them on the screen;

[0119] The first judgment module is used to determine whether the second prediction result is close to the first prediction result. If so, the second graph structure representation is output as the important monitoring station graph structure representation. If not, it returns to the greedy search module to continue searching.

[0120] The sorting module is used to sort the importance of the map structure representation of important monitoring stations;

[0121] The second judgment module is used to verify whether the importance ranking result is accurate through a verification algorithm. If it is, the importance ranking result is output; otherwise, the ranking module changes the ranking method and performs the ranking.

[0122] The second visualization module is used to output the final importance ranking results on the display screen.

[0123] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "setting," "connection," "fixing," "rotation," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Unless otherwise explicitly limited, those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0124] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for assessing the importance of monitoring stations based on graph neural networks, characterized in that, include: S1. Obtain relevant data from each monitoring station within the target area, including geographical information and pollutant concentrations for different historical time periods; A graphical representation of all monitoring stations within the target area for different historical time periods is constructed using relevant data; S2. The graph structure representations of all monitoring stations at different historical time periods are used as input to train the graph neural network model, and the output is the air quality index of the target area. S3. Obtain the pollutant concentrations of all monitoring stations in real time within the latest time period, construct the first graph structure representation corresponding to the latest time period, and input the first graph structure into the trained graph neural network model to obtain the first prediction result; S4. The first graph structure representation is searched using a greedy best-first search algorithm to obtain the second graph structure representation; wherein, the improved Shapley Value is used as the heuristic function of the greedy best-first search algorithm; the improved Shapley Value is represented as: in, express exist The contribution of the middle to the output, Represents a random graph structure. Represents the set of nodes A randomly selected subset of nodes in the middle. This represents a graph neural network model. This represents the corrected coefficient. This represents the set of all monitoring stations within the target area. express The set of nodes in the middle, Indicates the number of random samples. Represents a subset of nodes The weight, This represents the difference between the sum of the weights of 1 and a random subset of T nodes. Represents a subset of nodes The number of nodes in This indicates the number of all monitoring stations within the target area; S5. Input the second graph structure representation into the trained graph neural network model to obtain the second prediction result. Determine whether the second prediction result is less than the similarity threshold of the first prediction result. If so, the second graph structure representation is output as the graph structure representation of the important monitoring station. If not, return to S4 to continue the search. The improved Shapley Value is used as the heuristic function for the greedy best-first search algorithm. The process includes: S21. Initialize the second graph structure representation within the latest time period. ; S22. Based on the heuristic function, search for important monitoring stations in the first graph structure representation and add them to the second graph structure representation. middle; S23. Represent the structure of the second diagram. The trained graph neural network model is input to obtain a second prediction result. It is then determined whether the second prediction result is less than a similarity threshold compared to the first prediction result. If so, the second graph structure represents... If the output is not a representation of the important monitoring station map structure, return to S22 to continue the search; The formula for determining whether the second prediction result is close to the first prediction result is: in, This indicates the first prediction result. The first graph structure represents all monitoring stations within the target area in the latest time period. This indicates the second prediction result. Indicates control and exist The hyperparameter for similarity, i.e., the similarity threshold; S6. Construct an importance ranking module, which includes multiple ranking methods. First, the importance of the important monitoring station map structure representation is ranked by perturbation method. S7. Design a verification algorithm to verify whether the importance ranking result is accurate. If so, output the importance ranking result; otherwise, change the ranking method through the importance ranking module.

2. The method for assessing the importance of monitoring stations based on graph neural networks according to claim 1, characterized in that, Step S1, generating a graph structure representation of all monitoring stations over the historical time period u, is as follows: S11. Obtain the latitude and longitude information of each monitoring station within the target area, and calculate the distance between any two monitoring stations. The calculation formula is as follows: in, Indicates monitoring stations and monitoring stations The distance between, Indicates monitoring stations and monitoring stations The difference in longitude, Indicates monitoring stations and monitoring stations The difference in latitude, Represents the Earth's radius. Indicates monitoring stations latitude, Indicates monitoring stations latitude, Indicates monitoring stations longitude, Indicates monitoring stations longitude; , This indicates the total number of monitoring stations within the target area; S12. Select any monitoring station within the target area as the first monitoring station; S13. Sort the distances between the first monitoring station and the other monitoring stations in ascending order, and select the first... Each monitoring station adds an edge to the first monitoring station; if an edge already exists between monitoring stations, it is not necessary to add it again. S14. Determine if there are any monitoring stations that have not been the first monitoring station. If so, return to step S13 as the first monitoring station. If not, obtain the graph structure representation of all monitoring stations. S15. In the current graph structure representation, each node represents a monitoring station, and the pollutant concentration of the monitoring station in the historical time period u is used as the node feature. The feature vector of the edge is represented as [distance, longitude 1, latitude 1, longitude 2, latitude 2].

3. The method for assessing the importance of monitoring stations based on graph neural networks according to claim 1, characterized in that, Step S6 represents the map structure of important monitoring stations using a perturbation method. Ranking by importance includes: S31. Set the iteration threshold Let the number of iterations be ; S32. Randomly select from the structural representation of the first figure. Each monitoring station is masked, that is, in the structural representation of the first figure, this... The node feature corresponding to each monitoring station is represented as 0; S33. A greedy best-first search algorithm is used to search the results obtained in S32 to obtain the structure of the important monitoring stations behind the mask. ; S33. Determine if the condition is satisfied. If so, then Then return to step S32; otherwise, the iteration ends, and the perturbation set is obtained. Record the structure of important monitoring stations. The importance ranking of important monitoring stations is obtained by ranking the number of times each station appears in the disturbance set from highest to lowest frequency.

4. The method for assessing the importance of monitoring stations based on graph neural networks according to claim 1, characterized in that, The verification process in step S7 includes: S41. Based on the importance ranking results, select the top... Each monitoring station and subsequent Each monitoring station performs a masking operation on the graph structure representation of important monitoring stations to obtain the first mask image and the second mask image; S42. Input the first mask image and the second mask image into the trained graph neural network model to obtain the first predicted value and the second predicted value; S43. If the first predicted value is greater than the second predicted value, output the importance ranking result; otherwise, change the ranking method.

5. A monitoring site importance assessment system based on graph neural networks, characterized in that, It includes a historical data storage module, a graph structure composition module, a graph neural network training module, a real-time data acquisition module, a greedy search module, a first visualization module, a first judgment module, a sorting module, a second judgment module, and a second visualization module, wherein: The historical data storage module is used to store pollutant concentration data for all monitoring stations in the region for historical time periods, as well as the geographical information of each monitoring station. The graph structure module is used to construct a graph structure representation using the geographical information of monitoring stations and pollutant concentration data; The graph neural network training module is used to train a graph neural network based on a graph structure representation constructed from historical data, and outputs the air quality index of the region. The real-time data acquisition module is used to acquire the pollutant concentrations of all monitoring stations within the latest time period, and construct the first graph structure representation corresponding to the latest time period through the graph structure composition module. The first graph structure is then input into the trained graph neural network model to obtain the first prediction result. The greedy search module is used to search the first graph structure representation according to the greedy best-first search algorithm to obtain the second graph structure representation, and input the second graph structure representation into the trained graph neural network model to obtain the second prediction result; The first visualization module is used to output the search results of the greedy search module and display them on the screen; The first judgment module is used to determine whether the second prediction result is close to the first prediction result. If so, the second graph structure representation is output as the important monitoring station graph structure representation. If not, it returns to the greedy search module to continue searching. The sorting module is used to sort the importance of the map structure representation of important monitoring stations; The second judgment module is used to verify whether the importance ranking result is accurate through a verification algorithm. If it is, the importance ranking result is output; otherwise, the ranking module changes the ranking method and performs the ranking. The second visualization module is used to output the final importance ranking results on the display screen; The greedy search module uses the improved Shapley Value as the heuristic function for the greedy best-first search algorithm; the improved Shapley Value is expressed as: in, express exist The contribution of the middle to the output, Represents a random graph structure. Represents the set of nodes A randomly selected subset of nodes in the middle. This represents a graph neural network model. This represents the corrected coefficient. This represents the set of all monitoring stations within the target area. express The set of nodes in the middle, Indicates the number of random samples. Represents a subset of nodes The weight, This represents the difference between the sum of the weights of 1 and a random subset of T nodes. Represents a subset of nodes The number of nodes in This indicates the number of all monitoring stations within the target area; The improved Shapley Value is used as the heuristic function for the greedy best-first search algorithm. The process includes: S21. Initialize the second graph structure representation within the latest time period. ; S22. Based on the heuristic function, search for important monitoring stations in the first graph structure representation and add them to the second graph structure representation. middle; S23. Represent the structure of the second diagram. The trained graph neural network model is input to obtain a second prediction result. It is then determined whether the second prediction result is less than a similarity threshold compared to the first prediction result. If so, the second graph structure represents... If the output is not a representation of the important monitoring station map structure, return to S22 to continue the search; The formula for determining whether the second prediction result is close to the first prediction result is: in, This indicates the first prediction result. The first graph structure represents all monitoring stations within the target area in the latest time period. This indicates the second prediction result. Indicates control and exist The hyperparameter for the degree of similarity is the similarity threshold.

Images