Air quality prediction method and device

Abstract

The invention discloses an air quality prediction method and device. The air quality prediction method comprises the following steps: constructing a directed graph structure associated with a plurality of monitoring objects; in the directed graph structure, node attribute information of each node is determined, and the node attribute information at least comprises pollutant emission information of the monitoring object corresponding to the corresponding node; determining a track point proportion of at least one entrance edge pointing to the same subsequent node in the directed graph structure, and determining edge attribute information of the at least one entrance edge according to the track point proportion; based on the node attribute information and the edge attribute information, constructing graph structure data corresponding to a digraph structure; inputting the graph structure data into a graph neural network model for message passing to obtain corresponding graph structure features; and performing air quality prediction related to air pollutants on a target object in the plurality of monitoring objects based on the graph structure features. According to the invention, the accuracy of air quality prediction is improved.

Description

Technical Field

[0001] This application relates to the technical field of air quality detection, and particularly to an air quality prediction method and device. Background Art

[0002] In the existing technical field of air quality prediction, related methods based on deep learning usually perform model training according to meteorological data and geographical information, without considering the factor of pollution emissions. In the process of processing backward trajectory data, due to the failure to deeply integrate pollution emission data with trajectory data, it is impossible to completely restore the traceability path and transmission evolution process of pollutants, making it difficult to quantitatively evaluate the impact degree of pollution sources in different regions on the air quality of the target region. Eventually, there are large deviations in the prediction results, which cannot meet the requirements of accurate air quality forecasting.

[0003] Aiming at the technical problem of inaccurate air quality prediction existing in the above-mentioned prior art, no effective solution has been proposed yet. Summary of the Invention

[0004] Embodiments of the present disclosure provide an air quality prediction method and device to at least solve the technical problem of inaccurate air quality prediction existing in the prior art.

[0005] According to one aspect of the embodiments of the present disclosure, an air quality prediction method is provided, including constructing a directed graph structure associated with a plurality of monitoring objects, where each of the plurality of monitoring objects is used to monitor air pollutants, nodes of the directed graph structure respectively correspond to different monitoring objects, and directed edges of the directed graph structure are established according to the flowability of air pollutants between the monitoring objects; in the directed graph structure, determining node attribute information of each node, where the node attribute information at least includes pollution emission information of the monitoring object corresponding to the corresponding node; determining the trajectory point ratio of at least one incoming edge pointing to the same successor node in the directed graph structure, and determining edge attribute information of the at least one incoming edge according to the trajectory point ratio, where the trajectory point ratio is used to indicate the allocation ratio of multiple backward trajectory points of air pollutants of a first monitoring object corresponding to the same successor node between the at least one incoming edge; based on the node attribute information and the edge attribute information, constructing graph structure data corresponding to the directed graph structure; inputting the graph structure data into a graph neural network model for message passing to obtain corresponding graph structure features; and based on the graph structure features, performing air quality prediction related to the air pollutants on a target object among the plurality of monitoring objects.

[0006] According to another aspect of the embodiments of the present disclosure, there is also provided a storage medium, which includes a stored program, wherein the method described in any one of the above is executed by a processor when the program runs.

[0007] According to another aspect of the embodiments of the present disclosure, there is also provided an air quality prediction device, including: a directed graph structure construction module, configured to construct a directed graph structure associated with a plurality of monitoring objects, wherein each of the plurality of monitoring objects is used to monitor air pollutants, nodes of the directed graph structure respectively correspond to different monitoring objects, and directed edges of the directed graph structure are established according to the flowability of air pollutants between the monitoring objects; a determination module, configured to determine node attribute information of each node in the directed graph structure, wherein the node attribute information at least includes pollutant emission information of the monitoring object corresponding to the corresponding node; determine a trajectory point ratio of at least one incoming edge pointing to the same successor node in the directed graph structure, and determine edge attribute information of the at least one incoming edge according to the trajectory point ratio, wherein the trajectory point ratio is used to indicate the distribution ratio of multiple backward trajectory points of air pollutants of a first monitoring object corresponding to the same successor node among the at least one incoming edge; a graph structure data construction module, configured to construct graph structure data corresponding to the directed graph structure based on the node attribute information and the edge attribute information; a processing module, configured to input the graph structure data into a graph neural network model for message passing to obtain corresponding graph structure features; and based on the graph structure features, perform air quality prediction related to the air pollutants on a target object among the plurality of monitoring objects.

[0008] According to another aspect of the embodiments of the present disclosure, an air quality prediction device is further provided, including: a processor; and a memory connected to the processor for providing instructions for the processor to process the following steps: constructing a directed graph structure associated with a plurality of monitoring objects, wherein each of the plurality of monitoring objects is used to monitor air pollutants, the nodes of the directed graph structure respectively correspond to different monitoring objects, and the directed edges of the directed graph structure are established according to the flowability of air pollutants between the monitoring objects; in the directed graph structure, determining the node attribute information of each node, wherein the node attribute information at least includes the pollutant emission information of the monitoring object corresponding to the corresponding node; in the directed graph structure, determining the trajectory point ratio of at least one incoming edge pointing to the same successor node, and determining the edge attribute information of the at least one incoming edge according to the trajectory point ratio, wherein the trajectory point ratio is used to indicate the distribution ratio of multiple backward trajectory points of air pollutants of the first monitoring object corresponding to the same successor node among the at least one incoming edge; based on the node attribute information and the edge attribute information, constructing graph structure data corresponding to the directed graph structure; inputting the graph structure data into a graph neural network model for message passing to obtain corresponding graph structure features; and based on the graph structure features, performing air quality prediction related to the air pollutants on a target object among the plurality of monitoring objects.

[0009] In the embodiments of the present disclosure, by associating a plurality of monitoring objects with a directed graph structure, each monitoring object corresponds to a node, and the directed edges are established according to the flowability of air pollutants between the monitoring objects, clearly representing the propagation relationship of pollutants between each monitoring object. Using the pollutant emission information of the monitoring object as the node attribute of the node enables the model to consider the pollution emission information and more comprehensively reflect the influencing factors of air quality. By calculating the trajectory point ratio of at least one incoming edge pointing to the same successor node to determine the edge attribute information, it helps to quantify the distribution of pollutants on different transmission paths, thereby more accurately describing the source and transmission process of pollutants, and being able to comprehensively reflect the relationship between monitoring objects and the relevant characteristics of pollutants. Performing air quality prediction on the target object based on the extracted graph structure features comprehensively considers the influence of multiple factors on air quality, thereby achieving accurate prediction of air quality. Furthermore, it solves the technical problem of inaccurate air quality prediction existing in the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] The drawings described herein are used to provide a further understanding of the present disclosure, form a part of this application, and the illustrative embodiments and descriptions of the present disclosure are used to explain the present disclosure and do not constitute an improper limitation to the present disclosure. In the drawings: Figure 1It is a hardware structure block diagram of a computing device for implementing the method described in Embodiment 1 of the present disclosure; Figure 2A It shows a schematic diagram of an air pollution monitoring system for implementing the air quality prediction method described in Embodiment 1 of the present embodiment; Figure 2B It shows Figure 2A The module architecture diagram of the monitoring platform in Figure 2C It shows a schematic diagram of a deep learning model for air quality prediction in Embodiment 1 of the present disclosure; Figure 3 It shows a flowchart of a method for implementing air quality prediction according to the first aspect of Embodiment 1 of the present disclosure; Figure 4A It shows a schematic diagram of the directed graph structure described in Embodiment 1 of the present disclosure; Figure 4B It shows a distribution schematic diagram of monitoring objects corresponding to some nodes in the directed graph structure; Figure 5A It shows a schematic diagram of a geographical grid corresponding to a monitoring object in the case where the monitoring object is an air monitoring station; Figure 5B It shows a schematic diagram of a geographical grid corresponding to a monitoring object in the case where the monitoring object is a city; Figure 6 It shows a schematic diagram of message passing for graph structure data using a graph neural network in a deep learning model; Figure 7 It shows a schematic diagram of air quality prediction using a GRU neural network in a deep learning model; Figure 8 It shows a distribution schematic diagram of a city as a monitoring object and surrounding air monitoring stations; Figure 9 It shows a schematic diagram of an air quality prediction device according to Embodiment 2 of the present disclosure; Figure 10 It shows a schematic diagram of an air quality prediction device according to Embodiment 3 of the present disclosure. Detailed implementation manners

[0011] In order to enable those skilled in the art to better understand the technical solutions of the present disclosure, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present disclosure. Obviously, the described embodiments are only part of the embodiments of the present disclosure, rather than all of the embodiments. Based on the embodiments in the present disclosure, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present disclosure.

[0012] It should be noted that the terms "first", "second", etc. in the specification, claims and the above-mentioned drawings of the present disclosure are used to distinguish similar objects, and do not necessarily describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present disclosure described here can be implemented in an order other than those illustrated or described here. In addition, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device comprising a series of steps or units does not necessarily limit to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0013] First, some nouns or terms that appear in the process of describing the embodiments of the present disclosure are applicable to the following explanations: Monitoring object: A facility for monitoring air pollutants. In some examples of the present disclosure, the monitoring object can be understood as an air monitoring station or a city capable of monitoring air pollutants.

[0014] Directed graph structure: A graph with directionality, which is composed of a set of vertices and a set of directed edges, and each directed edge connects a pair of ordered vertices. In some examples of the present disclosure, the vertices in the directed graph structure can be understood as nodes, and each monitoring object corresponds to its own node in the directed graph structure. Whether there is a directed edge between nodes is determined in the directed graph structure according to the flowability of air pollutants between monitoring objects.

[0015] Backward trajectory point: A trajectory point on the propagation path of air pollutants determined by backward trajectory analysis, which is determined according to the propagation path of air pollutants. In some examples of the present disclosure, the edge attribute information of each directed edge can be determined in the directed graph structure according to the backward trajectory point.

[0016] Successor node: In some examples of the present disclosure, the successor node can be understood as the node pointed to by a directed edge in the directed graph structure.

[0017] Predecessor node: In some examples of the present disclosure, the predecessor node can be understood as the source node where a directed edge starts in the directed graph structure.

[0018] Designated time: In some examples of the present disclosure, the designated time can be understood as a certain future time with respect to the current time, or can also be understood as a certain historical time with respect to the current time. When the designated time is a certain future time in the future, the air quality prediction method of the present application can be used in the scenario of air quality forecasting; when the designated time is a certain historical time, the air quality prediction method of the present application can be used in the scenario of supplementing or verifying historical pollutant information.

[0019] Embodiment 1 According to this embodiment, a method embodiment for air quality prediction is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than here.

[0020] The method embodiment provided by this embodiment can be executed on a mobile terminal, a computer terminal, a server, or a similar computing device. Figure 1 A hardware structure block diagram of a computing device for implementing a method for air quality prediction is shown. As Figure 1 shown, the computing device may include one or more processors (the processor may include, but is not limited to, a processing device such as a microprocessor MCU or a programmable logic device FPGA), a memory for storing data, a transmission device for communication functions, and an input / output interface. Among them, the memory, the transmission device, and the input / output interface are connected to the processor through a bus. In addition, it may further include: a display, a keyboard, and a cursor control device connected to the input / output interface. Those of ordinary skill in the art can understand that Figure 1 the structure shown is only schematic and does not limit the structure of the above-mentioned electronic device. For example, the computing device may further include more or fewer components than those shown in Figure 1 or have a different configuration from that shown in Figure 1 shown.

[0021] It should be noted that the above one or more processors and / or other data processing circuits can generally be referred to as "data processing circuits" in this article. The data processing circuit can be embodied in whole or in part as software, hardware, firmware, or any combination thereof. In addition, the data processing circuit can be a single independent processing module, or can be incorporated in whole or in part into any one of other elements in the computing device. As involved in the embodiments of the present disclosure, the data processing circuit is a processor control (such as the selection of a variable resistance terminal path connected to an interface).

[0022] The memory can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to air quality prediction in the embodiments of the present disclosure. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory, that is, implements the air quality prediction method of the above application program. The memory may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memories. In some instances, the memory may further include a memory remotely provided with respect to the processor, and these remote memories can be connected to the computing device through a network. Examples of the above network include but are not limited to the Internet, intranet, local area network, mobile communication network, and combinations thereof.

[0023] The transmission device is used to receive or send data via a network. Specific examples of the above network may include a wireless network provided by a communication provider of the computing device. In one instance, the transmission device includes a network adapter (Network Interface Controller, NIC), which can be connected to other network devices through a base station and thus communicate with the Internet. In one instance, the transmission device may be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0024] The display can be, for example, a touch-screen liquid crystal display (LCD), and the liquid crystal display enables a user to interact with the user interface of the computing device.

[0025] It should be noted here that in some alternative embodiments, the above Figure 1 illustrated computing device may include hardware elements (including circuits), software elements (including computer code stored on a computer-readable medium), or a combination of both hardware elements and software elements. It should be pointed out that Figure 1 is only an example of a specific specific instance and is intended to show the types of components that may exist in the above computing device.

[0026] Figure 2A shows a schematic diagram of an air pollution monitoring system for implementing the air quality prediction method described in this embodiment. Referring to Figure 2A as shown, the air pollution monitoring system includes: a plurality of monitoring objects S 1~ S m ; and an air pollution monitoring platform 100 (hereinafter referred to as the "monitoring platform") communicatively connected to the monitoring objects S 1~ S m Although Figure 2AMultiple air monitoring stations (hereinafter also referred to as "stations") are used as the monitoring objects S 1~ S m , but the city can also be used as the monitoring object, so that each city can use air monitoring equipment to monitor the air pollution of the corresponding city, which will not be elaborated here.

[0027] Figure 2B The module architecture diagram of the monitoring platform 100 is shown. As Figure 2B shown, the monitoring platform 100 includes a data acquisition unit 110, a model training unit 120, and a daily forecast unit 130. The data acquisition unit 110 is used to collect air quality data, meteorological data, and transmission trajectory data required by the model training unit 120 and the daily forecast unit 130. The model training unit 120 is used to construct a training sample set for a deep learning model including a graph neural network based on the data collected by the data acquisition unit 110, and use the training sample set to train the deep learning module. The daily forecast unit 130 is used to predict the air quality (such as the concentration of pollutants) corresponding to the specified target monitoring object at the specified time (such as a future time of the current time or a time in the past history).

[0028] In Figure 2B , the data acquisition unit 110 includes an air quality data download module 111, a reanalysis data download module 112, a backward trajectory data download module 113, and a meteorological forecast data download module 114.

[0029] Among them, the air quality data download module 111 is used to obtain the hourly air quality data monitored by the monitoring object in units of hours. Among them, the air pollution factors in the air quality data include, for example, PM 2.5 , PM 10 , O3, NO2, SO2, and CO and other six types of air pollutants.

[0030] The reanalysis data download module 112 is used to obtain meteorological information and a gridded inventory related to air pollutant emissions in units of the running frequency of the training unit 120 (for example, in units of months or weeks). The gridded inventory is used to indicate the pollutant emission values in each geographical grid covering the geographical area as the grid values of each geographical grid. Among them, the running frequency of the reanalysis data download module 112 depends on the running frequency of the model training unit 120.

[0031] The backward trajectory data download module 113 is used to obtain the historical trajectory point data of air pollutants and the backward trajectory data point data of air pollutants at a specified moment. For example, the GDAS1 (Global Data Assimilation System, 1-degree) global data assimilation system product data can be obtained from the National Centers for Environmental Prediction (NCEP) of the United States respectively, and the GDAS1 data is used as the historical trajectory point data and the NCEP GFS (Global Forecast System) forecast data, and the NCEP GFS data is used as the backward trajectory point data of air pollutants at a future moment.

[0032] The meteorological forecast data download module 114 is used to obtain the meteorological forecast data for the future simulation period corresponding to the monitoring object. For example, the running frequency of the meteorological forecast data download module depends on the update frequency of the meteorological forecast data.

[0033] The model training unit 120 includes a first meteorological data processing module 121, a first air quality data processing module 122, a first backward trajectory data processing module 123, a first graph structure generation module 124, and a deep learning model training module 125. Among them, refer to Figure 2C As shown, the deep learning model used in this application includes a graph neural network GNN and a recurrent neural network GRU.

[0034] Among them, before running the model training unit 120, first obtain the basic information list of the site and the city (such as name, longitude and latitude, code, etc.) through the data acquisition unit 110; take the year as the unit to obtain the time range of the training data set for training the model. The specific types of meteorological elements and predicted air pollution factors used by the model training unit 120, the specific duration of the backward propagation of the transmission trajectory, the grid list, and the storage path of the digital elevation data are all modifiable information, and they are defined in the corresponding configuration file. Thus, according to the data in the configuration file and combined with the data collected in the data acquisition unit 110, the input of each module of the model training unit 120 can be controlled.

[0035] The first air quality data processing module 122 is used to divide the data obtained by the data acquisition unit 110 according to the specified time. Among them, the historical pollutant information corresponding to the historical time before the specified time (for example, the historical O3 concentration data of the site in the 24 hours or 48 hours before the specified time) can be used as part of the node attribute information of the nodes in the directed graph structure; while the data at the specified time can be used as the training set label data in the model training unit 120. That is, the first air quality data processing module 122 is responsible for processing and respectively generating the training set label data corresponding to the city or site required by the deep learning model training module 125 and the historical pollutant information as part of the node attribute information.

[0036] Among them, the operation process of the first air quality data processing module 122 can be understood as follows: First, obtain the codes of the cities or sites that need to be forecasted according to the basic information list of cities and sites in the configuration file, and then further determine the training time range and the air pollution factors to be simulated (such as O3) according to the configuration file. Then, screen the data obtained from the air quality data download module to obtain the air quality data (such as O3 concentration data) of the target city or site. After that, divide the data into label data and historical pollutant information.

[0037] The first meteorological data processing module 121 extracts the corresponding meteorological factors corresponding to the specified time according to the configuration file, and then forms the node attribute information together with the historical pollutant information of the air quality data processing module.

[0038] Among them, the first meteorological data processing module 121 is responsible for processing and generating part of the node attribute information required by the deep learning model training module 125. The operation process of the first meteorological data processing module 121 can be understood as follows: Based on the meteorological elements to be extracted, the time range, and the basic information list of cities and sites in the configuration file, extract the meteorological elements, time, and grid values from the historical meteorological reanalysis data downloaded by the atmospheric reanalysis data download module 112 (where the grid value corresponds to the pollutant emission information within the geographical grid, such as O3 emissions), obtain the historical meteorological data and pollutant emission information of the corresponding cities and sites, and use them as part of the node attribute information in the directed graph structure.

[0039] The first backward trajectory data processing module 123 simulates the backward propagation trajectory of the nodes, so as to determine the edge attribute information of the directed edges in the directed structure diagram according to the backward trajectory points.

[0040] The first graph structure generation module 124 constructs the connection features between nodes and edges considering the horizontal distance between nodes and the terrain height; combines the atmospheric emission inventory data and time feature data with the aforementioned historical pollutant information and meteorological factors to form the node attribute information; and forms the graph structure data based on the node attribute information and the edge attribute information.

[0041] The deep learning model training module 125 inputs the graph-structured data into the deep learning model, calculates the loss function based on the prediction results output by the deep learning model and the training set labels, and then updates the parameters through backpropagation to finally obtain the trained model parameters.

[0042] The daily forecast unit 130 includes a second meteorological data processing module 131, a second air quality data processing module 132, a second backward trajectory data processing module 133, a second graph structure generation module 134, and an air quality forecasting module 135.

[0043] Among them, the second air quality data processing module 132 obtains the pollutant historical information corresponding to the historical moments before the specified moment (such as the historical O3 concentration data of the station in the 24 hours or 48 hours before the specified moment) based on the data obtained by the data acquisition unit 110 as part of the node attribute information of the nodes of the directed graph structure. That is, the second air quality data processing module 132 is responsible for processing and generating the pollutant historical information required by the air quality forecasting module 135 as part of the node attribute information for subsequent air quality prediction. Specifically, reference can be made to the first air quality data processing module 122.

[0044] The second meteorological data processing module 131 extracts the corresponding meteorological factors at the specified moment according to the configuration file, and then forms part of the node attribute information together with the pollutant historical information of the air quality data processing module for subsequent air quality prediction. Specifically, reference can be made to the first meteorological data processing module 121.

[0045] The second backward trajectory data processing module 133 simulates the backward propagation trajectory of the nodes and determines the edge attribute information of the directed edges in the directed graph structure according to the trajectory for subsequent air quality prediction.

[0046] The second graph structure generation module 134 considers the horizontal distance between nodes and the terrain height to construct the connection features between nodes and edges; combines the atmospheric emission inventory data and time feature data with the aforementioned pollutant historical information and meteorological factors to form the node attribute information; and jointly forms the graph-structured data based on the node attribute information and the edge attribute information.

[0047] The air quality forecasting module 135 inputs the graph-structured data into the trained deep learning model, and after aggregating and updating the spatial information through the graph neural network, inputs it into the recurrent neural network GRU for time information update, so as to perform air quality prediction of the target object in the monitoring object with air pollutants (such as predicting the O3 concentration at the specified moment).

[0048] Under the above operating environment, according to the first aspect of this embodiment, an air quality prediction method is provided. This method is performed byFigure 2A Implemented by the monitoring platform 100 shown in Figure 3 The figure shows a flowchart of a method for implementing air quality prediction. Refer to Figure 3 As shown, the method includes: S302: Construct a directed graph structure associated with multiple monitoring objects, where each of the multiple monitoring objects is used to monitor air pollutants, the nodes of the directed graph structure correspond to different monitoring objects respectively, and the directed edges of the directed graph structure are established according to the flowability of air pollutants between the monitoring objects; S304: In the directed graph structure, determine the node attribute information of each node, where the node attribute information at least includes the pollutant emission information of the monitoring object corresponding to the corresponding node; S306: In the directed graph structure, determine the trajectory point ratio of at least one incoming edge pointing to the same successor node, and determine the edge attribute information of at least one incoming edge according to the trajectory point ratio, where the trajectory point ratio is used to indicate the allocation ratio of multiple backward trajectory points of air pollutants of the first monitoring object corresponding to the same successor node among at least one incoming edge; S308: Based on the node attribute information and the edge attribute information, construct graph structure data corresponding to the directed graph structure; S310: Input the graph structure data into a graph neural network model for message passing to obtain corresponding graph structure features; and S312: Based on the graph structure features, perform air quality prediction related to air pollutants on the target object among the multiple monitoring objects.

[0049] Specifically, after the monitoring platform 100 obtains the list of basic information (name, longitude and latitude, code, etc.) of national stations and cities, it establishes a directed graph structure (S302) associated with the monitoring object S 1~ S m associated. Among them Figure 4A The figure shows a schematic diagram of the directed graph structure. The directed graph structure includes nodes corresponding to each monitoring object S 1~ S m corresponding N 1~ N m . And further refer to Figure 4A As shown, each edge in the directed graph structure E 1~ E L is a directed edge. And each edge E 1~ E L is based on each monitoring object S 1~ S mThe circulation of air pollutants (e.g., O3) is established. For example, in Figure 4A there is no edge constructed between node N 3 and node N 4, which means there is no circulation of air pollutants between node N 3 and node N 4. Among them, the connection relationship between nodes can be represented by an adjacency matrix. As for the specific method of establishing a directed graph structure, it will be described in detail below.

[0050] Then, the monitoring platform 100 determines the node attribute information corresponding to each node N 1 to N m through the second meteorological data processing module 131 and the second air quality data processing module 132 An 1 to An m , where An 1 to An m ∈R 39*1 (S304). The node attribute information An 1 to An m at least includes the pollutant emission information of the monitoring object corresponding to the corresponding node N 1 to N m , such as the O3 emission amount in the area where the monitoring object S 1 to S m is located. As for the specific method of determining the node attribute information S 1 to S m m An 1 to An m , it will be described in detail below.

[0051] Then, the monitoring platform 100 determines the trajectory point ratio of at least one incoming edge pointing to the same successor node in the directed graph structure. Specifically, referring to Figure 4A shown, the successor node described in this application refers to the node specified by a directed edge. For example, for the directed edge E 1, the successor node is N 1; for the directed edge E 2, the successor node is N 4; and so on. For the directed edge E L , the successor node is N m . And for these successor nodes, the directed edge pointing to this node is an incoming edge. For example, the directed edge E1 is a successor node N The incoming edge of 1; a directed edge E 2 is a successor node N The incoming edge of 4; and so on, a directed edge E L is a successor node N m The incoming edge of. Thus, continuing to refer to Figure 4A As shown, taking the node N 1 as an example, this node N 1 as a successor node has multiple incoming edges E 1, E 3 and E 5. That is, for the same successor node, there can be multiple incoming edges corresponding to it. It should be noted that even for the same node, it may be a successor node or a predecessor node relative to different directed edges. For example, the node N 1 is a successor node relative to the directed edge E 5, but it is a predecessor node relative to the directed edge E 6, and so on.

[0052] In addition, the second backward trajectory data processing module 133 of the monitoring platform 100 can simulate the backward propagation trajectory of O3 corresponding to the monitored object. Among them, Figure 4B It further shows some nodes in the directed graph N 1~ N 4 corresponding to the monitored object S 1~ S 4 of the distribution schematic diagram. Among them Figure 4B The black dots shown in are the multiple backward trajectory points of O3 corresponding to the monitored object S 1 (i.e., the first monitored object) relative to the specified time t For example, in this embodiment, it is the backward trajectory points of the first 24 or 48 moments before the specified time t . And, the second backward trajectory data processing module 133 further determines Figure 4B The multiple backward trajectory points shown in the nodes N 1 of the multiple incoming edges E 1, E 3 and E 5 of the allocation ratio. And according to the determined allocation ratio, determine the incoming edges E 1, E 3 and E 5 of the edge attribute information Ae 1, Ae 3 and Ae 5 (S306). And so on, the monitoring platform 100 can in the same way based on each successor node to determine each directed edge's E1 to E L edge attribute information Ae 1 to Ae L 。

[0053] Then, the second graph structure generation module 134 of the monitoring platform 100 constructs graph structure data corresponding to the directed graph structure described in An 1 to An m based on the determined node attribute information of each node Ae 1 to Ae L and the edge attribute information of each directed edge Figure 4A (S308).

[0054] Then, the air quality prediction module 135 of the monitoring platform 100 inputs the graph structure data into the graph neural network of the deep learning model to obtain graph structure features corresponding to the graph structure data (S310). The specific process will be described in detail below.

[0055] Finally, the air quality prediction module 135 of the monitoring platform 100 performs air quality prediction related to air pollutants on the target objects among multiple monitoring objects S 1 to S m . For example, it predicts the O3 concentration of the target object S 1 at a specified time t (S312). The specific prediction process will be described in detail below.

[0056] In the embodiment of the present disclosure, a directed graph structure is established based on the circulation of air pollutants between monitoring objects, which intuitively and accurately shows the degree of association between the propagation directions of air pollutants among different monitoring objects. And pollutant emission information is added to the node attribute information of the directed graph structure. Compared with the method of predicting air quality using meteorological data and geographical information, the accuracy of air quality prediction is improved. In addition, the proportion of in-edge trajectory points pointing to the same successor node is determined and the edge attribute information is determined accordingly. Since the backward trajectory point proportion reflects the transmission trajectory of air pollutants for monitoring objects, the influence weight of the transmission path of air pollutants on the air quality of monitoring objects is clarified. Graph structure data is constructed based on node attributes and edge attributes and input into the graph neural network model for message passing, enabling the model to fully learn the spatio-temporal features in the air quality data. Thus, the graph neural network model can better capture the internal connections of air quality data between different monitoring objects and at different time points, thereby improving the prediction ability of air quality change trends. Furthermore, the technical problem of inaccurate air quality prediction existing in the prior art is solved.

[0057] Optionally, the operation of constructing a directed graph structure associated with multiple monitoring objects includes: when the first distance between two monitoring objects is less than a first threshold and there is no terrain with an altitude greater than a second threshold between the two monitoring objects, constructing a directed edge between the nodes corresponding to the two monitoring objects.

[0058] Specifically, during the construction of the directed graph structure, the second graph structure generation module 134 obtains the digital elevation stored in the path according to the configuration file and obtains the list of station and urban basic information, so as to determine whether a directed edge can be constructed between the nodes of the directed graph structure.

[0059] Specifically, the second graph structure generation module 134 determines whether the distance (i.e., the first distance) between two monitoring objects is less than the first threshold according to the list of station and urban basic information. Refer to Figure 4B As shown, for example, it can be determined the distance (i.e., the first distance) between monitoring objects S 1 to S 4, and determine whether it is less than the first threshold.

[0060] Then, when the distance between monitoring objects S 1 to S 4 is less than the first threshold, it is determined whether there is terrain with an altitude greater than the second threshold between the two monitoring objects according to the digital elevation data. When there is no terrain with an altitude greater than the second threshold between the two monitoring objects, a directed edge is constructed between the nodes corresponding to the two monitoring objects. For example, refer to Figure 4A As shown, since the distance between monitoring objects S 1 to S 3 is less than the first threshold, and there is no terrain with an altitude greater than the second threshold between each monitoring object, so nodes N 1 to N 3 can construct directed edges with each other. For another example, the distance between monitoring objects S 3 and S 4 may be greater than the first threshold, or there may be terrain with an altitude greater than the second threshold between monitoring objects S 3 and S 4, so that nodes N 3 and N 4 cannot construct a directed edge.

[0061] According to the embodiments of the present disclosure, by using digital elevation data and comprehensively considering the distance and terrain factors between different monitoring objects to construct directed edges, the transmission path of air pollutants can be accurately simulated, and then the flow situation of pollutants between different monitoring objects can be accurately reflected, improving the accuracy of air quality prediction.

[0062] In the case of determining the existence of directed edges in a directed graph structure, further determine the edge attribute information of each directed edge.

[0063] Optionally, the operation of determining the trajectory point ratio of each incoming edge pointing to the same successor node in the directed graph structure includes: determining a plurality of backward trajectory points of air pollutants of the first monitoring object at a future specified time; determining the farthest backward trajectory point with the farthest distance from the first monitoring object among the plurality of backward trajectory points, and determining the second distance between the first monitoring object and the farthest backward trajectory point; determining the adjacent predecessor node among the predecessor nodes of the successor node, where the distance between the second monitoring object corresponding to the adjacent predecessor node and the first monitoring object is less than the second distance, and determining the first incoming edge corresponding to the adjacent predecessor node among each incoming edge; and setting the trajectory point ratio of the second incoming edge other than the first incoming edge among each incoming edge to zero, and determining the trajectory point ratio corresponding to the first incoming edge based on the plurality of backward trajectory points.

[0064] Specifically, taking Figure 4A the node N 1 in

[0065] as an example of the successor node, illustrate how to determine the trajectory point ratio of each incoming edge pointing to the same successor node. N First, the second graph structure generation module 134 determines, from the second transmission trajectory data processing module 133, a plurality of backward trajectory points of the monitoring object S 1 (i.e., the first monitoring object) corresponding to the node t with respect to the specified time Figure 4B as shown in the reference t . Among them, the plurality of backward trajectory points can be determined, for example, by the second transmission trajectory data processing module 133 using the hysplit model to simulate the backward propagation trajectory of the air pollutant (such as O3) corresponding to the monitoring object. For example, the plurality of backward trajectory points can be 24 hourly trajectory points of the backward propagation of O3 corresponding to the monitoring object S 1 for 24 hours starting from the specified time

[0066] Then, further referring to Figure 4B as shown, the second graph structure generation module 134 determines, among the plurality of backward trajectory points, the backward trajectory point with the farthest distance from the monitoring object S 1 (for example, it can be the trajectory point at the 24th hour of backward propagation). And the distance between the monitoring object S 1 and the backward trajectory point with the farthest distance (i.e., the second distance) can be determined.

[0067] Then, with the monitoring object S 1 as the center, and with the farthest backward trajectory point and the monitoring object SWith the distance between 1 as the radius, a circular area centered on the monitoring object S 1 is determined (see Figure 4B the circular area shown in N ). Then, among the monitoring objects corresponding to the predecessor nodes N 2 to N 4 of node S 2 to S 4, the monitoring objects located within the range of this circular area are determined S 2 and S 3 (i.e., the second monitoring object). And the nodes S 2 and S 3 corresponding to the monitoring objects N 2 and N 3 (i.e., the adjacent predecessor nodes) are further determined.

[0068] Then, the second graph structure generation module 134 further determines, from the incoming edges E 1, E 3, and E 5, the incoming edges N 2 and N 3 that point from nodes N 1 (i.e., the first incoming edges). E 5 and E 3 (i.e., the first incoming edges).

[0069] Then, since the distance between the monitoring object S 4 and the monitoring object S 1 is greater than the above-mentioned second distance, the proportion of the trajectory points corresponding to the incoming edge E 1 (i.e., the second incoming edge) is set to 0. And the second graph structure generation module 134 determines the proportion of the trajectory points of the incoming edges Figure 4B 5 and E 3 (i.e., the first incoming edges) according to the E backward trajectory points shown. And the method for specifically determining E 5 and E 3 (i.e., the first incoming edges) of the proportion of the trajectory points will be described in detail below.

[0070] Thus, according to the embodiments of the present disclosure, by determining the backward trajectory points of the air pollutants of the first monitoring object at a future specified time t and finding the farthest backward trajectory point, the possible farthest source location of the air pollutants can be clarified, and further the range of the air pollutants that affect the air quality of the first monitoring object can be accurately identified.

[0071] Optionally, the operation of determining the trajectory point ratio corresponding to the first incoming edge based on multiple backward trajectory points includes: determining the connection lines between each second monitoring object and the first monitoring object; from the first incoming edges, determining the incoming edge corresponding to the connection line with the smallest distance to the backward trajectory point as the incoming edge corresponding to the backward trajectory point; and determining the corresponding trajectory point ratio according to the number of backward trajectory points corresponding to the first incoming edge.

[0072] Specifically, further referring to Figure 4A and Figure 4B , when determining the incoming edges N 2 and N 3 pointing to the node N 1 (i.e., the first incoming edge), the second graph structure generation module 134 determines the connection lines E 5 and E 3 (i.e., the first incoming edge) between the monitoring objects S 2 and S 3 and the monitoring object S 1 L 12 and L 13 (as shown by the dotted lines in Figure 4B ). Among them, the connection line L 12 corresponds to the incoming edge E 5, and the connection line L 13 corresponds to the incoming edge E 3.

[0073] Then, further referring to Figure 4B shown, the second graph structure generation module 134 calculates the distance between each backward trajectory point and the connection lines L 12 and L 13 respectively. Specifically, it calculates the perpendicular distances from the backward trajectory point to the connection lines L 12 and L 13 respectively. Then, the incoming edge corresponding to the connection line with the smallest distance is used as the incoming edge corresponding to the backward trajectory point. For example, for a backward trajectory point, if the distance from it to the connection line L 12 is less than the distance to the connection line L 13 , then the incoming edge E 5 is determined as the incoming edge corresponding to the backward trajectory point. For another example, for another backward trajectory point, if the distance from it to the connection line L 13 is less than the distance to the connection line L 12 , then the incoming edgeE 3 is the incoming edge corresponding to the backward trajectory point. And so on, for each backward trajectory point shown in Figure 4B , the incoming edge corresponding to it is respectively confirmed.

[0074] Then, the second graph structure generation module 134 determines the number of backward trajectory points corresponding to the incoming edge E 5 and the incoming edge E 3, and further determines the trajectory point ratio corresponding to the incoming edge E 5 and the incoming edge E 3 according to the determined number of backward trajectory points.

[0075] Regarding the incoming edge E 1, since the distance between the monitored object S 4 and the monitored object S 1 is greater than the above-mentioned second distance, the trajectory point ratio corresponding to the incoming edge E 1 is set to 0.

[0076] Thus, according to the embodiments of the present disclosure, by determining the connection lines between each second monitored object and the first monitored object, and then finding the incoming edge corresponding to the connection line with the smallest distance from the backward trajectory point, the key transmission path affecting the air quality of the first monitored object is accurately located. Further, the trajectory point ratio is determined according to the number of backward trajectory points corresponding to each incoming edge, so as to quantify the weights of different transmission paths during the pollution transmission process.

[0077] Optionally, the operation of determining the edge attribute information of each incoming edge according to the trajectory point ratio includes: determining the distance information between the monitored objects corresponding to the predecessor node and the successor node of each incoming edge; and determining the edge attribute information of each incoming edge according to the distance information and the trajectory point ratio.

[0078] Specifically, still referring to Figure 4A and Figure 4B shown, after the second graph structure generation module 134 determines the incoming edges N 1 corresponding to the node E 1, E 3 and E 5, it further determines the distance between the monitored object S 1 and S 2 as the distance information corresponding to the incoming edge E 5, and determines the distance between the monitored object S 1 and S 3 as the distance information corresponding to the incoming edge E 3, and determines the distance between the monitored object S 1 and S 4 as the distance information corresponding to the incoming edge E 1.

[0079] Then, according to the distance information and the trajectory point ratio corresponding to the incoming edges E 1, E 3, and E 5, the second graph structure generation module 134 determines the edge attribute information of the incoming edges E 1, E 3, and E 5 Ae 1, Ae 3, and Ae 5 .

[0080] Therefore, according to the embodiments of the present disclosure, by determining the distance information between the monitoring objects corresponding to the predecessor node and the successor node of the incoming edge and combining the trajectory point ratio to determine the edge attribute information, the characteristics of the directed edges in the directed graph can be described more accurately. The distance information reflects the proximity of different monitoring objects in space, while the trajectory point ratio reflects the distribution of pollutants on different transmission paths. The combination of the two enables the edge attribute information to more comprehensively and meticulously reflect the difficulty and likelihood of pollutant transmission from one monitoring object to another.

[0081] Optionally, the operation of determining the node attribute information of each node includes: obtaining the pollutant historical information related to air pollutants at historical times before the specified time of the monitoring object and the meteorological information of the monitoring object at the specified time; determining the pollutant emission information corresponding to the monitoring object at the specified time; and determining the node attribute information of the node corresponding to the monitoring object according to the pollutant historical information, the pollutant emission information, and the meteorological information.

[0082] Specifically, taking the node Figure 4A shown in N 1 and Figure 4B the monitoring object in S 1 as an example for illustration. During the construction of the directed graph data, the second graph structure generation module 134 obtains the pollutant historical information of the monitoring object N 1 corresponding to the node S 1 from the second air quality data processing module 132. For example, in this embodiment, the O3 concentration information corresponding to the monitoring object t 1 at 24 historical times before the specified time can be used as the pollutant historical information, denoted as S ~ an 1,1 ~ an 1,24 . That is, the first 24 elements An 1 of the node attribute information an 1,1 ~ an 1,24 are the pollutant historical information.

[0083] Then, the second graph structure generation module 134 obtains from the second meteorological data processing module 131 the monitoring object N corresponding to node S 1 at a specified time t meteorological information. In this embodiment, 10 different meteorological factors can be taken as meteorological information, such as wind force, temperature, humidity, precipitation, and so on. Thus, the second graph structure generation module 134 records the meteorological information of the monitoring object S 1 at the specified time t as an 1,25 ~ an 1,34 , representing the 25th to 34th elements of the node attribute information An of node 1.

[0084] Then, the second graph structure generation module 134, according to the O3 emission amount (i.e., pollutant emission information) of the monitoring object N corresponding to node S 1 at the specified time t obtained from the second meteorological data processing module 131, records it as an 1,35 , representing the 35th element of the node attribute information An of node 1.

[0085] Finally, the second graph structure generation module 134 obtains from the second meteorological data processing module 131 the year, month, week, and hour corresponding to the specified time t , and records them as an 1,36 ~ an 1,39 , representing the 36th to 39th elements of the node attribute information An of node 1.

[0086] Thus, through the above operations, the second graph structure generation module 134 constructs the node attribute information N corresponding to node An 1 ∈ R 39*1 . For other nodes N 2 to N m , the corresponding node attribute information can be constructed by referring to the above operations.

[0087] In addition, this embodiment schematically shows that the node attribute information has 39 elements, but the specific number of elements can be adjusted according to the actual situation, as long as it conforms to the type of node attribute information defined above.

[0088] According to the embodiments of the present disclosure, by comprehensively obtaining the historical pollutant information, pollutant emission information, and meteorological information of the monitored object, the node attribute information is determined, which comprehensively covers the key factors affecting air quality. The historical pollution status of the monitored object is determined based on the historical pollutant information, thereby reflecting the changing trend of air quality at historical moments. In addition, since the air quality data corresponding to different monitored objects are different, and the factors affecting the air quality of different monitored objects are also different, the node attribute information of the node is determined through the historical pollutant information, pollutant emission information, and meteorological information, so that the model can subsequently accurately predict the air quality of different monitored objects according to the specific conditions of different monitored objects.

[0089] Optionally, when the monitored object is an air monitoring station, the operation of determining the pollutant emission information of the station at a specified moment includes: determining a plurality of geographical grids corresponding to the air monitoring station, where the geographical grid corresponding to the air monitoring station is located within the area covered by the plurality of geographical grids; averaging the pollutant emission information corresponding to the plurality of geographical grids at the specified moment as the pollutant emission information corresponding to the air monitoring station at the specified moment.

[0090] Specifically, when the second graph structure generation module 134 determines the pollutant emission information at a specified moment t , it obtains the data of the grid list from the storage path according to the configuration file, and determines the geographical grid corresponding to the monitored object according to the geographical location information of the monitored object.

[0091] Specifically, taking the monitored object S 1 as an example for illustration. Referring to Figure 5A as shown, when the monitored object S 1 is a station, the geographical grid corresponding to the monitored object S 1 determined by the second graph structure generation module 134 includes the geographical grid of the monitored object S 1 and the eight geographical grids surrounding this geographical grid.

[0092] Then, the second graph structure generation module 134 determines the grid values corresponding to each geographical grid in the grid list at the specified moment t , where the grid value is used to indicate the pollutant emission information (such as the O3 emission amount) within the corresponding geographical grid. And the second graph structure generation module 134 averages the grid values of each geographical grid as the pollutant emission information (such as the O3 emission amount) of the station at the specified moment t .

[0093] Optionally, when the monitoring object is a city, the operation of determining the pollutant emission information of the city at a specified moment includes: determining a plurality of geographical grids corresponding to the area of the city; and averaging the pollutant emission information corresponding to the plurality of geographical grids at the specified moment as the pollutant emission information corresponding to the city at the specified moment.

[0094] Specifically, taking the monitoring object S 1 as an example for illustration. Refer to Figure 5B As shown, when the monitoring object S1 is a city, the city area may include, for example, a plurality of geographical grids. Therefore, refer to Figure 5B As shown, the geographical grids determined by the second graph structure generation module 134 corresponding to the monitoring object S 1 include the geographical grids within the city shp area.

[0095] Then, the second graph structure generation module 134 determines the grid values corresponding to each geographical grid at the specified moment t in the grid list, where the grid values are used to indicate the pollutant emission information (such as O3 emission) within the corresponding geographical grid. And the second graph structure generation module 134 averages the grid values of each geographical grid as the pollutant emission information (such as O3 emission) of the city at the specified moment t .

[0096] Optionally, the operation of inputting the graph structure data into a graph neural network model for message passing to obtain the corresponding graph structure features includes: performing message passing on the directed edges of the graph structure data to determine the edge features corresponding to the directed edges; and performing message passing on the successor nodes according to the node attributes of the successor nodes in the graph structure data and the edge features of the incoming edges corresponding to the successor nodes to determine the node features of the successor nodes.

[0097] Specifically, Figure 6 shows Figure 2C a schematic diagram of the graph neural network GNN shown in Figure 2C . Refer to

[0098] Then, further refer to Figure 6 As shown, a first perceptron and a second perceptron are deployed in the graph neural network GNN. The first perceptron is used to perform message passing operations on the edge attribute information in the graph structure data. Preferably, the first perceptron can be a three-layer linear perceptron. The second perceptron is used to perform message passing operations on the node attribute information in the graph structure data. Preferably, the second perceptron can be a one-layer linear perceptron.

[0099] Thus, referring to Figure 6 As shown, the graph neural network GNN first performs a message passing operation on the edge attribute information of the graph structure data through a first perceptron to generate corresponding edge features; then the graph neural network GNN performs a message passing operation on the node attribute information of the graph structure data through a second perceptron to generate corresponding node features.

[0100] In this way, the graph neural network GNN generates node features corresponding to each node in the directed graph structure N 1~ N m respectively, and edge features corresponding to each directed edge E 1~ E L respectively. Thus, the graph structure features are constructed.

[0101] According to the embodiments of the present disclosure, message passing is performed on the directed edges of the graph structure data, so that the information contained in the directed edges can be deeply mined, and the edge features corresponding to the directed edges can be accurately determined. In the air quality prediction scenario, the directed edges represent the circulation relationship of air pollutants between different monitoring objects, and the edge features reflect characteristics such as the intensity and direction of this circulation. Through message passing, information such as the distance between monitoring objects and the pollutant transmission rate can be incorporated into the edge features, enabling the model to more accurately capture the transmission law of pollutants between different regions, thereby being able to more precisely depict the attributes of the directed edges and providing a richer information basis for subsequent analysis. Message passing is performed based on the node attributes of the successor nodes and the edge features of the corresponding incoming edges, and then the node features of the successor nodes are determined, fully considering the node's own attributes and the information of the directed edges connected to it, so that the node features can more comprehensively and accurately reflect the state of the monitoring object in the entire system. Compared with determining node features based on only single information, it can more realistically reflect its actual situation during the air quality change process, improving the accuracy and reliability of the node features.

[0102] Optionally, the operation of performing message passing on the directed edges of the graph structure data to determine the edge features corresponding to the directed edges includes: determining the associated successor nodes and associated predecessor nodes associated with the directed edges; determining the pollutant information corresponding to the previous moment of the associated successor nodes at a specified moment, and the pollutant information corresponding to the previous moment of the associated predecessor nodes; inputting the node attribute information of the associated successor nodes, the pollutant information corresponding to the previous moment of the associated successor nodes, the node attribute information of the associated predecessor nodes, the pollutant information corresponding to the previous moment of the associated predecessor nodes, and the edge attribute information of the directed edges into a pre-trained first linear perceptron to generate the edge features corresponding to the directed edges.

[0103] Specifically, for the directed edge in the graph structure data ETaking [1] as an example for illustration. When performing message passing operations on this directed edge, the graph neural network GNN first determines this directed edge E 1's associated successor nodes N 1 and associated predecessor nodes N 4.

[0104] Then, the graph neural network GNN determines the following information: the node attribute information of the associated successor nodes N 1 An 1 and at the specified time t the previous time t -1's O3 concentration; the node attribute information of the associated predecessor nodes N 4 An 4 and at the specified time t the previous time t -1's O3 concentration; and the edge attribute information of the directed edge E 1.

[0105] Then, after the graph neural network GNN fuses the above information (for example, by concatenation), it inputs it into the first perceptron, thereby generating an edge feature corresponding to the directed edge E 1, thus corresponding to the directed edge FE 1 in the graph structure feature.

[0106] Thus, for each directed edge E 1~ E L in the graph structure data, its corresponding edge feature can be determined respectively according to the above operations, corresponding to the directed edge FE 1~ FE L in the graph structure feature.

[0107] According to the embodiments of the present disclosure, accurately determining the predecessor and successor nodes associated with the directed edge and their pollutant information at the previous time, and combining the node attribute and edge attribute information realizes the deep fusion of the spatial dimension and the time dimension. In the air quality monitoring scenario, the transmission and diffusion of pollutants depend not only on the spatial distance between monitoring points but also on the concentration state at the previous time. By fusing this information to generate edge features, the transmission law of pollutants in the spatio-temporal dimension can be more realistically depicted, and then the model can more accurately capture the pollution propagation path and reduce the prediction error caused by feature loss or deviation.

[0108] Optionally, the operation of performing message passing on the successor nodes according to the node attributes of the successor nodes in the graph structure data and the edge features of the incoming edges corresponding to the successor nodes to determine the node features of the successor nodes includes: obtaining the average value of the edge features of the incoming edges corresponding to the successor nodes to obtain the aggregated edge features; and inputting the aggregated edge features and the node attributes of the successor nodes into a second perceptron to generate the node features corresponding to the successor nodes.

[0109] Specifically, taking node N 1 as an example, the operation of determining the node features of the successor node N 1 will be described. According to Figure 4A it is known that the incoming edges corresponding to node N 1 include incoming edges E 1, E 3, and E 5. According to the above, determine the edge features corresponding to the incoming edges E 1, E 3, and E 5, and then determine the average value of the edge features corresponding to the incoming edges E 1, E 3, and E 5, so as to obtain the aggregated edge features of the multiple incoming edges existing relative to node N 1.

[0110] Then, input the aggregated edge features corresponding to node N 1 and the node attribute information N 1 of node An 1 into the second perceptron to obtain the node features N 1 of node fn 1, corresponding to node FN 1 in the graph structure features.

[0111] Thus, for the nodes N 1 to N m in the graph structure data, adopt the same method as node N 1 to determine their respective corresponding node features fn 1 to fn m , corresponding to the nodes FN 1 to FN m in the graph structure features.

[0112] According to an embodiment of the present disclosure, the edge features of the incoming edges corresponding to the successor nodes are averaged to obtain aggregated edge features. This can effectively integrate the information of multiple incoming edges connected to the successor nodes and avoid the one-sidedness of a single incoming edge feature. In the air quality monitoring scenario, a monitoring object (successor node) may be affected by multiple different predecessor nodes, and the edge features of each incoming edge reflect the influence characteristics of different pollution sources on this monitoring point. By taking the average, the role of each pollutant information can be comprehensively considered, and the information such as the pollution transmission intensity and path characteristics carried by multiple incoming edges can be fused to form a more representative aggregated edge feature, which comprehensively reflects the comprehensive influence on the successor node. Inputting the aggregated edge feature and the node attributes of the successor node into the second perceptron to generate node features further improves the integrity and accuracy of the node features.

[0113] Optionally, based on the graph structure features, the operation of performing air quality prediction related to air pollutants on a target object among multiple monitoring objects includes: determining a first target node corresponding to the target object in the graph structure data and determining a second target node corresponding to the target object in the graph structure features; determining the pollutant information corresponding to the target object at the previous moment; and determining the pollutant information corresponding to the target object at the specified moment by using a pre-set GRU-based prediction model according to the pollutant information corresponding to the target object at the previous moment, the node attribute information of the first target node, and the node features of the second target node.

[0114] Specifically, Figure 7 Further shows a schematic diagram of the gated recurrent unit (GRU) of the recurrent neural network. Refer to Figure 7 As shown, the recurrent neural network GRU includes an input unit, a GRU model, and an MLP linear layer. The following takes the monitoring object S 1 as an example of the target object for illustration.

[0115] Thus, after the deep learning model obtains the graph structure features corresponding to the graph structure data by using the graph neural network according to the above operations, for example, it can determine the first target node S 1 corresponding to the target object N 1 in the graph structure data, and determine the second target node S 1 corresponding to the target object FN 1 in the graph structure features.

[0116] Then, the deep learning model can use the O3 concentration of the target object S 1 at the previous moment t -1 of the specified moment, the node attribute information t 1 of the first target node N 1, and the node features an 1 of the second target node FN 1.fn 1 Input to the input unit.

[0117] Thus, based on the target object, the GRU neural network S 1 at a specified time t at the previous time t -1 O3 concentration, the first target node N 1 node attribute information an 1 and the second target node FN 1 node features fn 1 to determine the target object S 1 at a specified time t O3 concentration (pollutant information).

[0118] According to the embodiments of the present disclosure, by determining the target node corresponding to the target object in the graph structure features, precise positioning of specific monitoring objects is achieved. In a complex air quality monitoring network, there are multiple monitoring objects. Precise locking of the target node can avoid interference from irrelevant information and focus subsequent analysis on key objects. Determining the pollutant information at the previous time of the target node and combining the node attributes and node features realizes the deep fusion of information in the time dimension (historical pollutant concentration) and the space dimension. The change of air quality has spatio-temporal dynamic characteristics. The historical pollutant concentration reflects the pollution evolution trend, the node attributes reflect the characteristics of the target object itself, and the node features integrate the influence of the surrounding environment. Compared with the prediction method that only considers single-dimensional information, it can more realistically simulate the pollution diffusion process and significantly improve the accuracy and reliability of the prediction.

[0119] Optionally, according to the pollutant information corresponding to the target object at the previous time, the node attribute information of the first target node, and the node features of the second target node, the operation of determining the pollutant information corresponding to the target object at the specified time by using a pre-set GRU-based prediction model includes: determining the first hidden state information corresponding to the target node at the previous time, where the first hidden state information is generated by the GRU model of the prediction model; inputting the pollutant information corresponding to the target object at the previous time, the node attribute information of the first target node, the node features of the second target node, and the first hidden state information into the GRU model to determine the second hidden state information corresponding to the target node at the specified time; and inputting the second hidden state information into a linear layer to determine the pollutant information corresponding to the target object at the specified time.

[0120] Specifically, as shown in Figure 7 , the input unit also receives the hidden state t -1 calculated by the GRU model at the previous time H t-1 (i.e., the first hidden state information), and the target objectS At a specified moment t The previous moment t The O3 concentration of -1, the first target node N The node attribute information of 1 an 1 and the second target node FN The node characteristics of 1 fn 1 and the hidden state H t-1 Input into the GRU model, so that the GRU model calculates the hidden state corresponding to the specified moment according to the input information t (i.e., the second hidden state information). H t (That is, the second hidden state information).

[0121] Hidden state H t Is input into the MLP linear layer of the recurrent neural network GRU, so as to predict the O3 concentration (i.e., pollutant information) of the target object 1 at the specified moment S (That is, the pollutant information).

[0122] According to the embodiments of the present disclosure, obtain the first hidden state information generated by the GRU model, which aggregates multi-dimensional information such as the pollutant change trend and the influence of the surrounding environment of the target node in the past moments. And input it together with the pollutant information, node attributes and node characteristics of the current node into the GRU model, so that the GRU model can deeply remember the historical pollution evolution process and dynamically update the hidden state (i.e., the second hidden state information) according to the current input information. Avoid prediction deviations caused by ignoring historical trends. Air quality data has significant non-linear dynamic change characteristics, and the GRU model shows unique advantages in processing such data. The GRU model can fully mine the complex laws of the pollutant information of the target node in the time series.

[0123] Optionally, when the target object is a city, the method further includes: obtaining the pollutant information corresponding to the air monitoring stations around the city as the monitoring objects at the specified moment; averaging the pollutant information of the air monitoring stations to obtain the average pollutant information; and performing a weighted sum of the predicted pollutant information and the average pollutant information to determine the pollutant information of the city at the specified moment.

[0124] Specifically, as shown in Figure 8 When the air quality forecasting module 135 predicts the predicted O3 concentration (predicted pollutant information) of the city at the specified moment according to the above method, obtain the O3 concentration of the stations around the city at the specified moment, where the O3 concentration of the stations at the specified moment can also be predicted by the above method, for example.

[0125] Then, the air quality forecasting module 135 calculates the average of the O3 concentrations at each station at the specified time to obtain the average O3 concentration (i.e., the average pollutant information).

[0126] Then, the final pollutant information of the city at the specified time can be determined by linear regression: y = a * x1 + b * x2 x1 and x2 are the average O3 concentration calculated from the station mean and the O3 concentration directly forecasted for the city respectively, and y is the city forecast concentration value obtained from the final regression calculation.

[0127] Thus, through the above method, this application forecasts the air quality of stations and cities simultaneously, and performs a secondary forecast for the city by linear regression to improve the accuracy of the city forecast.

[0128] Therefore, according to this embodiment, both pollutant emission factors are considered, and backward trajectory data is fully utilized through methods such as the proportion of backward trajectory points, comprehensively covering multiple influencing factors of air quality, avoiding the limitations of only focusing on single factors such as meteorological data and geographical information, and enabling the model to more realistically reflect the actual situation of air quality changes. The directed graph structure and its edge attribute information can accurately describe the fluidity and propagation path of air pollutants between monitoring objects, quantify the distribution ratio of pollutants on different paths, and help accurately judge the actual impact of different regions on the air quality of the target region, thereby improving the accuracy of prediction.

[0129] Taking the hourly O3 concentration of national stations from 00:00 to 23:00 on April 1, 2025 as an example: (1) Data acquisition unit 110 The air quality data download module 111 downloads the O3 hourly data of national stations from 00:00 on December 31, 2019 to 23:00 on March 31, 2025, denoted as data A. The atmospheric reanalysis data download module 112 downloads the era5 data from 00:00 on January 1, 2020 to 23:00 on March 31, 2025, denoted as data B. The backward trajectory data download module 113 downloads the gdas1 data from 01:00 on December 31, 2019 to 23:00 on March 31, 2025, denoted as data C; downloads the hysplit - gfs meteorological forecast data from 01:00 on March 31, 2025 to 23:00 on April 1, 2025, denoted as data D. Assuming the backward propagation duration is 24 hours, the data required for the trajectory to be propagated at 00:00 on April 1 is 24 data from 01:00 on March 31 to 00:00 on April 1. The meteorological forecast data download module 114 downloads the gfs meteorological data from 00:00 to 23:00 on April 1, 2025, denoted as data E.

[0130] (2) Model training unit 120 Simulation period: 00:00 on January 1, 2020 - 23:00 on March 31, 2025.

[0131] Both the labeled data and the pollutant historical information extracted by the first air quality data processing module 122 are O3 concentration data. The labeled data is data A; the pollutant historical information is the data of the previous 24 hours, that is, the O3 hourly data from 00:00 on December 31, 2019 to 23:00 on March 30, 2025. The reason for the previous 24 hours: considering the influence of the same period of the previous day on the node. The pollutant historical information is the first node attribute information of the node.

[0132] The first meteorological data processing module 121 extracts the meteorological reanalysis data at the site. Assuming there are 10 factors, the 2nd to 11th node attribute information of the node is these 10 factors.

[0133] The first backward trajectory data processing module 123 inputs data C to simulate the 24-hour backward trajectory of each hour at the site, calculates the proportion of the trajectory points obtained within the trajectory circle, and obtains the weight characteristics of the associated edges of each site, as the first edge attribute information of the directed edge.

[0134] The first graph structure generation module 124 first connects the nodes to form edges according to the distance between sites (less than 200 km) and the altitude difference between sites (less than 1000 km) to form an adjacency matrix, and at the same time takes the distance between the sites corresponding to the nodes as the second edge attribute information of the edge; according to the gridded emission inventory, extracts the emissions corresponding to the sites as the 12th feature of the node. Here, the MEIC inventory is used for the emission inventory. There are direct emission statistics for all five factors except O3. O3 is replaced by VOCs data (considering that the national control sites are mainly distributed in urban areas, which are mainly VOCs control areas); at the same time, add year, month, week, and hour as the 13th to 16th node attribute information of the node. Therefore, the above-mentioned graph structure is composed of an adjacency matrix, nodes (16 features), and edges (2 features).

[0135] The deep learning model training module 125 divides the aforementioned graph-structured data into a training set (from January 1, 2020 to December 31, 2023), a validation set (from January 1, 2024 to December 31, 2024), and a test set (from January 1, 2025 to March 31, 2025). The training set is used to learn model parameters, the validation set is used for hyperparameter tuning, and the test set is used to evaluate the model's generalization ability. Among them, the hyperparameters in the validation set are externally defined (configuration file), do not require the model to learn, and can be set according to human experience or determined after several experimental simulations. Therefore, the simulation of the training set is mainly introduced. For example, the prediction of a certain node A at 01: The O3 concentration of node A at 00 is used as an input information, and the O3 concentration of node B connected to node A at 00 is used as another input information. Then, it is concatenated with the node attribute information of nodes A and B at 01 and the edge attribute information of the edge between A and B at 01, and then input into a three-layer perceptron for the edge to obtain the edge feature of the incoming edge between node A and node B at 01 with node A as the target. According to the above method, the edge features of the incoming edges connected to node A by other nodes at 01 are obtained, and then the edge features of the incoming edges at 01 are aggregated (averaged) and input into a one-layer perceptron of the node to obtain the updated node feature of node A at 01. The O3 concentration of node A at 00, the node attribute information of node A itself at 01, the updated node feature of node A at 01, and the hidden state at 00 are input into the GRU to obtain the hidden state at 01, and then output through a linear layer to obtain the prediction result. The measured O3 concentration of node A at 01 is used as the loss function with the prediction result, and then backpropagation is performed to update the weight parameters, and finally the trained model parameters are obtained.

[0136] (3) Forecast unit 130 Forecast period: from 00:00 on April 1, 2025 to 23:59 on April 1, 2025. The node features are the same as those of the training unit.

[0137] The second air quality data processing module 132 uses the data from 00:00 on March 31, 2025 to 23:59 on March 31, 2025, which has been extracted in the training unit, as the historical O3 concentration information of the station during forecasting, and it is also the first node attribute information of the node. The second meteorological data processing module 131 extracts the meteorological forecast data gfs of the station, and for the same 10 factors as mentioned above, the 2nd to 11th node attribute information of the node is these 10 factors. The second backward trajectory data processing module 133 inputs the data D to simulate the 24-hour backward trajectory of each station per hour, and calculates the proportion of the trajectory points within the trajectory circle to obtain the weight information of the incoming edges associated with each station, which is used as the first edge attribute information of the incoming edges.

[0138] The second graph structure generation module 134 first connects nodes to form edges based on the distance between sites (less than 200 km) and the altitude difference between sites (less than 1000 km) to form an adjacency matrix. At the same time, the distance between sites is used as the second feature of the edge. According to the gridded emission inventory, the emissions corresponding to the sites are extracted as the 12th feature of the nodes. At the same time, year, month, week, and hour are added as the 13th to 16th features of the nodes (these training units have been completed and do not need to be repeated). Therefore, the above graph structure is composed of an adjacency matrix, nodes (16 features), and edges (2 features).

[0139] Air quality prediction module 135 (still taking the prediction of node A at 01:00 as an example): The O3 concentration of node A at 00:00 is used as one input information, and the O3 concentration of node B connected to node A at 00:00 is used as another input information. Then, it is concatenated with the node attribute information of nodes A and B at 01:00 and the edge attribute information of the edge between A and B at 01:00, and then input into a three-layer perceptron for the edge. The parameters are provided by the training unit to obtain the edge feature of the edge between node A and node B at 01:00 with node A as the target. According to the above method, the edge features of the incoming edges connected to node A by other nodes are obtained, and then the edge features of the edges at 01:00 are aggregated (averaged) and input into a one-layer perceptron for the node. The parameters are provided by the training unit to obtain the updated node feature of node A at 01:00. The O3 concentration of node A at 00:00, the node attribute information of node A itself at 01:00, the updated node feature of node A at 01:00, and the hidden state at 00:00 are input into the GRU model. The parameters are provided by the training unit to obtain the hidden state at 01:00, and then output through a linear layer. The parameters are provided by the training unit to obtain the prediction result.

[0140] Taking the hourly O3 concentration prediction of Wuxi City, a national city, from 00:00 to 23:00 on April 1, 2025 as an example: In this system, whether it is site or city prediction, it is regarded as node prediction. A set of systems is adopted to construct different graph structures for separate training and prediction. The following case takes national cities as nodes to construct a graph structure to obtain the prediction result of Wuxi City. It is also possible to select city regions containing Wuxi City, such as Jiangsu Province and the Yangtze River Delta region, for construction, taking into account the impacts of transmission, computing costs, etc.

[0141] (1) Data acquisition unit 110 The time period is the same as that of the previous sites. Air quality data needs to be downloaded at the national city level, and the rest of the data is the same and can be used interchangeably.

[0142] (2) Model training unit 120 Simulation time period: The same as that of the previous sites.

[0143] The first air quality data processing module 122, which is the same as the foregoing, except that the data of national stations is extracted.

[0144] The first meteorological data processing module 121, where the foregoing extracts meteorological data within the grid where the station longitude and latitude are located, and for cities, the average value within the shp area is extracted.

[0145] The first backward trajectory data processing module 123, which is the same as the foregoing, except that the city is used as the receptor point of the backward trajectory.

[0146] The first graph structure generation module 124, also based on the distance between stations (less than 200 km) and the altitude difference between cities (less than 1000 km) to connect nodes to form edges and form an adjacency matrix. At the same time, the distance between nodes is used as the second feature of the edge; according to the gridded emission inventory, the average emissions within the shp area of the city are extracted as the 12th node attribute information of the node. Here, the MEIC inventory is used for the emission inventory, and there are direct emission statistics for all five factors except O3. For O3, VOCs data is used instead (the urban area is mainly a VOCs control area); at the same time, year, month, week, and hour are added as the 13th to 16th node attribute information of the node. Therefore, the above graph structure is composed of an adjacency matrix, nodes (16 node attribute information), and edges (2 node attribute information).

[0147] The deep learning model training module 125, the forecast for Wuxi at 01:00: The O3 concentration of Wuxi at 00:00 is used as one input information, and the O3 concentration of Suzhou at 00:00 connected to Wuxi is used as another input information. Then, it is concatenated with the 01:00 node attribute information of Wuxi and Suzhou themselves and the 01:00 edge attribute information of the edge between Wuxi and Suzhou, and then input into the three-layer perceptron of the edge to obtain the 01:00 edge feature of the edge between Wuxi and Suzhou with Wuxi as the target. According to the above method, the 01:00 edge features of the edges connecting other cities to Wuxi are obtained, and then the 01:00 edge features of the edges are aggregated (averaged) and input into the one-layer perceptron of the Wuxi node to obtain the updated 01:00 node feature of Wuxi. The O3 concentration of Wuxi at 00:00, the 01:00 node attribute information of Wuxi itself, the updated 01:00 node feature of Wuxi, and the 00:00 hidden state are input into the GRU to obtain the 01:00 hidden state, and then output through a linear layer to obtain the prediction result. The measured O3 concentration of Wuxi at 01:00 is used as the loss function with the prediction result, and then backpropagation is used to update the weight parameters, and finally the trained model parameters are obtained.

[0148] (3)Daily Forecast Unit 130 Forecast period: The same as the foregoing stations.

[0149] The second air quality data processing module 132, which is the same as the foregoing, except that it extracts data from national stations.

[0150] The second meteorological data processing module 131, where the foregoing extracts the gfs meteorological data within the grid where the station longitude and latitude are located, and for cities, it extracts the average value within the shp area.

[0151] The second backward trajectory data processing module 133, which is the same as the foregoing, except that the city is used as the receptor point of the backward trajectory.

[0152] The second graph structure generation module consists of the adjacency matrix, city node features (16 features), and edges (2 features) in the previous training unit to form the structure of the graph.

[0153] Taking the air quality forecast module 135 as an example of the forecast for Wuxi at 01:00: The O3 concentration of Wuxi at 00:00 is used as one input information, and the O3 concentration of Suzhou at 00:00 connected to Wuxi is used as another input information. Then, it is concatenated with the 01:00 node attribute information of Wuxi and Suzhou themselves and the 01:00 edge attribute information of the edge between Wuxi and Suzhou, and then input into a three-layer perceptron for the edge. The parameters are provided by the training unit 110 to obtain the 01:00 edge features of the edge between Wuxi and Suzhou with Wuxi as the target. According to the above method, the 01:00 edge features of the edges connected between other cities and Wuxi are obtained, and then the 01:00 edge features of the edges are aggregated (averaged) and input into a one-layer perceptron of the Wuxi node. The parameters are provided by the training unit to obtain the updated 01:00 node features of Wuxi. The O3 concentration of Wuxi at 00:00, the 01:00 node attribute information of Wuxi itself, the updated 01:00 node features of Wuxi, and the hidden state at 00:00 are input into the GRU model. The parameters are provided by the training unit to obtain the hidden state at 01:00, and then output through a linear layer. The parameters are provided by the training unit to obtain the direct prediction result of Wuxi at 01:00.

[0154] (4) Regression calculation Further calculate the forecast for Wuxi. Obtain the O3 forecast results of the 9 national control stations in Wuxi from 00:00 on April 1, 2025 to 23:00 on April 1, 2025 in Case 1, and calculate the hourly average value to obtain the hourly concentration value of Wuxi (the concentration value of Wuxi at 00:00 is equal to the average concentration of the 9 stations at 00:00), and perform a regression with the directly forecasted concentration value of Wuxi in Case 2, that is: y = a * x1 + b * x2, where a and b are the weights (obtained by fitting the data, and the values will change dynamically due to the continuous increase of data), x1 and x2 are the city concentration calculated from the station average value and the concentration directly forecasted for the city respectively, and y is the final city forecast concentration value obtained from the regression calculation.

[0155] In addition, with reference to Figure 1 as shown, according to the third aspect of this embodiment, a storage medium is provided. The storage medium includes a stored program, wherein the method described in any one of the above is executed by a processor when the program runs.

[0156] Thus, in this embodiment, by associating multiple monitoring objects with a directed graph structure, each monitoring object corresponds to a node, and the directed edges are established based on the air pollutant circulation between the monitoring objects, clearly representing the pollutant transmission relationship between each monitoring object. Using the pollutant emission information of the monitoring object as the node attribute of the node enables the model to consider the pollution emission information and more comprehensively reflect the influencing factors of air quality. By calculating the proportion of trajectory points of at least one incoming edge pointing to the same successor node to determine the edge attribute information, it helps to quantify the distribution of pollutants on different transmission paths, thus more accurately describing the source and transmission process of pollutants, and being able to comprehensively reflect the relationship between monitoring objects and the relevant characteristics of pollutants. Based on the extracted graph structure features, air quality prediction is performed on the target object, comprehensively considering the influence of multiple factors on air quality, thereby achieving accurate prediction of air quality. Furthermore, the technical problem of inaccurate air quality prediction existing in the prior art is solved.

[0157] It should be noted that, for the foregoing method embodiments, for the sake of simple description, they are all expressed as a series of action combinations. However, those skilled in the art should know that the present invention is not limited by the described action sequence, because according to the present invention, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0158] Through the description of the above embodiments, those skilled in the art can clearly understand that the method according to the above embodiments can be implemented by means of software plus a necessary general hardware platform. Of course, it can also be implemented by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art can be embodied in the form of a software product. The computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disc), and includes several instructions for causing a terminal device (which can be a mobile phone, computer, server, or network device, etc.) to execute the methods described in various embodiments of the present invention.

[0159] Embodiment 2 Figure 9An air quality prediction device 900 according to the present embodiment is shown. The device 900 corresponds to the method according to the first aspect of Embodiment 1. Refer to Figure 9 As shown, the device 900 includes: a directed graph structure construction module 910 for constructing a directed graph structure associated with a plurality of monitoring objects, where each of the plurality of monitoring objects is used to monitor air pollutants, the nodes of the directed graph structure correspond to different monitoring objects respectively, and the directed edges of the directed graph structure are established according to the circulation of air pollutants between the monitoring objects; a node attribute determination module 920 for determining the node attribute information of each node in the directed graph structure, where the node attribute information at least includes the pollutant emission information of the monitoring object corresponding to the corresponding node; a trajectory point ratio determination module 930 for determining the trajectory point ratio of at least one incoming edge pointing to the same successor node in the directed graph structure, and determining the edge attribute information of at least one incoming edge according to the trajectory point ratio, where the trajectory point ratio is used to indicate the distribution ratio of multiple backward trajectory points of the air pollutants of the first monitoring object corresponding to the same successor node among at least one incoming edge; a graph structure data construction module 940 for constructing graph structure data corresponding to the directed graph structure based on the node attribute information and the edge attribute information; a message passing module 950 for inputting the graph structure data into a graph neural network model for message passing to obtain the corresponding graph structure features; and an air quality prediction module 960 for performing air quality prediction related to air pollutants on a target object among the plurality of monitoring objects based on the graph structure features.

[0160] Thus, in the present embodiment, by associating a plurality of monitoring objects with a directed graph structure, each monitoring object corresponds to a node, and the directed edges are established according to the circulation of air pollutants between the monitoring objects, clearly representing the pollutant propagation relationship between each monitoring object. Using the pollutant emission information of the monitoring object as the node attribute of the node enables the model to consider the pollution emission information situation and more comprehensively reflect the influencing factors of air quality. By calculating the trajectory point ratio of at least one incoming edge pointing to the same successor node to determine the edge attribute information, it helps to quantify the distribution of pollutants on different transmission paths, thus more accurately describing the source and transmission process of pollutants, and being able to comprehensively reflect the relationship between monitoring objects and the relevant characteristics of pollutants. Performing air quality prediction on the target object based on the extracted graph structure features comprehensively considers the influence of multiple factors on air quality, thereby achieving accurate prediction of air quality. Furthermore, it solves the technical problem of inaccurate air quality prediction existing in the prior art.

[0161] Embodiment 3 Figure 10Fig. 0 shows an air quality prediction device 1000 according to the first aspect of the present embodiment, and the device 1000 corresponds to the method according to the first aspect of Embodiment 1. Refer to Figure 10 As shown, the device 1000 includes: a processor 1010; and a memory 1020 connected to the processor 1010 for providing instructions for the processor 1010 to process the following steps: constructing a directed graph structure associated with a plurality of monitoring objects, where each of the plurality of monitoring objects is used to monitor air pollutants, nodes of the directed graph structure respectively correspond to different monitoring objects, and directed edges of the directed graph structure are established according to the circulation of air pollutants between the monitoring objects; in the directed graph structure, determining node attribute information of each node, where the node attribute information at least includes pollutant emission information of the monitoring object corresponding to the corresponding node; in the directed graph structure, determining a trajectory point ratio of at least one incoming edge pointing to the same successor node, and determining edge attribute information of the at least one incoming edge according to the trajectory point ratio, where the trajectory point ratio is used to indicate the distribution ratio of multiple backward trajectory points of air pollutants of a first monitoring object corresponding to the same successor node among the at least one incoming edge; based on the node attribute information and the edge attribute information, constructing graph structure data corresponding to the directed graph structure; inputting the graph structure data into a graph neural network model for message passing to obtain corresponding graph structure features; and based on the graph structure features, performing air quality prediction related to the air pollutants on a target object among the plurality of monitoring objects.

[0162] In the embodiments of the present disclosure, by associating a plurality of monitoring objects with a directed graph structure, each monitoring object corresponds to a node, and directed edges are established according to the circulation of air pollutants between the monitoring objects, clearly representing the pollutant propagation relationship between each monitoring object. Using the pollutant emission information of the monitoring object as the node attribute of the node enables the model to consider the pollution emission information and more comprehensively reflect the influencing factors of air quality. By calculating the trajectory point ratio of at least one incoming edge pointing to the same successor node to determine the edge attribute information, it helps to quantify the distribution of pollutants on different transmission paths, thereby more accurately describing the source and transmission process of pollutants, and being able to comprehensively reflect the relationship between monitoring objects and the relevant characteristics of pollutants. Performing air quality prediction on the target object based on the extracted graph structure features comprehensively considers the influence of multiple factors on air quality, thereby achieving accurate prediction of air quality. Furthermore, it solves the technical problem of inaccurate air quality prediction existing in the prior art.

[0163] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0164] In the above embodiments of the present invention, the descriptions of the various embodiments have their respective emphases. For the parts not detailed in a certain embodiment, reference may be made to the relevant descriptions of other embodiments.

[0165] In several embodiments provided by the present application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only illustrative. For example, the division of the units is only a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed coupling or direct coupling or communication connection between each other can be through some interfaces. The indirect coupling or communication connection of the units or modules can be in an electrical or other form.

[0166] The units described as separate components may or may not be physically separated. The components displayed as units may or may not be physical units, that is, they can be located in one place or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0167] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above-mentioned integrated units can be implemented in the form of hardware or in the form of software functional units.

[0168] If the above-mentioned integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to enable a computer device (which can be a personal computer, a server or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The foregoing storage medium includes: USB flash drives, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), mobile hard disks, magnetic disks or optical disks and other various media that can store program codes.

[0169] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. An air quality prediction method, characterized in that, Including: Construct a directed graph structure associated with multiple monitoring objects, where each of the multiple monitoring objects is used to monitor air pollutants, the nodes of the directed graph structure correspond to different monitoring objects respectively, and the directed edges of the directed graph structure are established according to the flowability of air pollutants between the monitoring objects; In the directed graph structure, determine the node attribute information of each node, where the node attribute information at least includes the pollutant emission information of the monitoring object corresponding to the corresponding node; Determine the trajectory point ratio of each incoming edge pointing to the same successor node in the directed graph structure, and determine the edge attribute information of each incoming edge according to the trajectory point ratio, where the trajectory point ratio is used to indicate the distribution ratio of multiple backward trajectory points of air pollutants of the first monitoring object corresponding to the same successor node among the incoming edges; Based on the node attribute information and the edge attribute information, construct graph structure data corresponding to the directed graph structure; Input the graph structure data into a graph neural network model for message passing to obtain corresponding graph structure features; and Based on the graph structure features, perform air quality prediction related to the air pollutants on a target object among the multiple monitoring objects.

2. The method according to claim 1, wherein The operation of constructing a directed graph structure associated with multiple monitoring objects includes: When the first distance between two monitoring objects is less than a first threshold and there is no terrain with an altitude greater than a second threshold between the two monitoring objects, construct a directed edge between the nodes corresponding to the two monitoring objects.

3. The method according to claim 1, characterized in that, The operation of determining the node attribute information of each node includes: obtaining the pollutant historical information related to the air pollutants of the monitoring object at historical moments before a specified moment and the meteorological information of the monitoring object at the specified moment; determining the pollutant emission information of the monitoring object corresponding to the specified moment; and determining the node attribute information of the node corresponding to the monitoring object according to the pollutant historical information, the pollutant emission information, and the meteorological information, and where When the monitoring object is an air monitoring station, the operation of determining the pollutant emission information of the monitoring object corresponding to the specified moment includes: determining a plurality of geographical grids corresponding to the air monitoring station, where the geographical grids corresponding to the air monitoring station are located within the area covered by the plurality of geographical grids; and averaging the pollutant emission information corresponding to the plurality of geographical grids and the specified moment as the pollutant emission information corresponding to the air monitoring station and the specified moment, or when the monitoring object is a city, the operation of determining the pollutant emission information of the monitoring object corresponding to the specified moment includes: Determining a plurality of geographical grids corresponding to the area of the city; and averaging the pollutant emission information corresponding to the plurality of geographical grids and the specified moment as the pollutant emission information corresponding to the city and the specified moment.

4. The method according to claim 3, wherein The operation of determining the node attribute information of the node corresponding to the monitoring object according to the pollutant historical information, the pollutant emission information, and the meteorological information includes: Concatenating the pollutant historical information, the pollutant emission information, and the meteorological information with the time information corresponding to the specified moment as the node attribute information of the node corresponding to the monitoring object.

5. The method according to claim 1, characterized in that, The operation of determining the trajectory point ratio of each incoming edge pointing to the same successor node in the directed graph structure includes: Based on the specified moment, determining the plurality of backward trajectory points corresponding to the air pollutants of the first monitoring object; Determining the farthest backward trajectory point with the farthest distance from the first monitoring object among the plurality of backward trajectory points, and determining the second distance between the first monitoring object and the farthest backward trajectory point; Determining the neighboring predecessor node among the predecessor nodes of the successor node, where the distance between the second monitoring object corresponding to the neighboring predecessor node and the first monitoring object is less than the second distance, and determining the first incoming edge corresponding to the neighboring predecessor node among the respective incoming edges; and Setting the trajectory point ratio of the second incoming edge other than the first incoming edge among the respective incoming edges to zero, and determining the trajectory point ratio corresponding to the first incoming edge based on the plurality of backward trajectory points, and wherein The operation of determining the trajectory point ratio corresponding to the first incoming edge based on the plurality of backward trajectory points includes: determining the connection line between each second monitoring object and the first monitoring object; determining, from the first incoming edge, the incoming edge corresponding to the connection line with the smallest distance from the backward trajectory point as the incoming edge corresponding to the backward trajectory point; and determining the corresponding trajectory point ratio according to the number of backward trajectory points corresponding to the first incoming edge, and wherein The operation of determining the edge attribute information of each incoming edge according to the trajectory point ratio includes: determining the distance information between the monitoring objects corresponding to the predecessor node and the successor node of each incoming edge; and determining the edge attribute information of each incoming edge according to the distance information and the trajectory point ratio.

6. The method according to claim 1, wherein The operation of inputting the graph structure data into a graph neural network model for message passing to obtain the corresponding graph structure features includes: Performing message passing on the directed edges of the graph structure data to determine the edge features corresponding to the directed edges; and Performing message passing on the successor nodes according to the node attributes of the successor nodes in the graph structure data and the edge features of the incoming edges corresponding to the successor nodes to determine the node features of the successor nodes, and wherein The operation of performing message passing on the directed edges of the graph structure data to determine the edge features corresponding to the directed edges includes: determining the associated successor nodes and associated predecessor nodes associated with the directed edges; determining the pollutant information corresponding to the previous moment of the associated successor nodes at a specified moment, and the pollutant information corresponding to the previous moment of the associated predecessor nodes; and inputting the node attribute information of the associated successor nodes, the pollutant information corresponding to the previous moment of the associated successor nodes, the node attribute information of the associated predecessor nodes, the pollutant information corresponding to the previous moment of the associated predecessor nodes, and the edge attribute information of the directed edges into a pre-trained first linear perceptron to generate the edge features corresponding to the directed edges, and wherein The operation of performing message passing on the successor nodes according to the node attributes of the successor nodes in the graph structure data and the edge features of the incoming edges corresponding to the successor nodes to determine the node features of the successor nodes includes: averaging the edge features of the incoming edges corresponding to the successor nodes to obtain aggregated edge features; and inputting the aggregated edge features and the node attributes of the successor nodes into a second perceptron to generate the node features corresponding to the successor nodes, and wherein The operation of performing air quality prediction related to the air pollutants on the target object among the multiple monitoring objects based on the graph structure features includes: determining a first target node corresponding to the target object in the graph structure data and a second target node corresponding to the target object in the graph structure features; determining the pollutant information corresponding to the previous moment of the target object; and determining the pollutant information corresponding to the target object at the specified moment by using a pre-set GRU-based prediction model according to the pollutant information corresponding to the previous moment of the target object, the node attribute information of the first target node, and the node features of the second target node.

7. The method according to claim 6, wherein The operation of determining the pollutant information corresponding to the target object at the specified moment by using a pre-set GRU-based prediction model according to the pollutant information corresponding to the previous moment of the target object, the node attribute information of the first target node, and the node features of the second target node includes: Determining the first hidden state information corresponding to the target node at the previous moment, wherein the first hidden state information is generated by the GRU model of the prediction model; Inputting the pollutant information corresponding to the previous moment of the target object, the node attribute information of the first target node, the node features of the second target node, and the first hidden state information into the GRU model to determine the second hidden state information corresponding to the target node at the specified moment; and Inputting the second hidden state information into the linear layer of the prediction model to determine the predicted pollutant information corresponding to the target object at the specified moment.

8. The method according to claim 7, characterized in that, In the case where the target object is a city, it further includes: Obtain the pollutant information corresponding to the air monitoring stations around the city as the monitoring objects at the specified time; Calculate the average value of the pollutant information of the air monitoring stations to obtain the average pollutant information; and Perform a weighted sum of the predicted pollutant information and the average pollutant information to determine the pollutant information of the city at the specified time.

9. An air quality prediction device, characterized in that, It includes: A directed graph structure construction module for constructing a directed graph structure associated with multiple monitoring objects, where each of the multiple monitoring objects is used to monitor air pollutants, the nodes of the directed graph structure correspond to different monitoring objects respectively, and the directed edges of the directed graph structure are established according to the fluidity of air pollutants between the monitoring objects; A node attribute determination module for determining the node attribute information of each node in the directed graph structure, where the node attribute information at least includes the pollutant emission information of the monitoring object corresponding to the corresponding node; A trajectory point ratio determination module for determining the trajectory point ratio of at least one incoming edge pointing to the same successor node in the directed graph structure, and determining the edge attribute information of the at least one incoming edge according to the trajectory point ratio, where the trajectory point ratio is used to indicate the distribution ratio of multiple backward trajectory points of the air pollutants of the first monitoring object corresponding to the same successor node among the at least one incoming edge; A graph structure data construction module for constructing graph structure data corresponding to the directed graph structure based on the node attribute information and the edge attribute information; A message passing module for inputting the graph structure data into a graph neural network model for message passing to obtain corresponding graph structure features; and An air quality prediction module for predicting the air quality related to the air pollutants of the target object among the multiple monitoring objects based on the graph structure features.

10. An air quality prediction device, characterized in that, It includes: A processor; And A memory connected to the processor for providing instructions for the processor to perform the following processing steps: Construct a directed graph structure associated with multiple monitoring objects, where each of the multiple monitoring objects is used to monitor air pollutants, the nodes of the directed graph structure correspond to different monitoring objects respectively, and the directed edges of the directed graph structure are established according to the fluidity of air pollutants between the monitoring objects; In the directed graph structure, determine the node attribute information of each node, where the node attribute information at least includes the pollutant emission information of the monitoring object corresponding to the corresponding node; Determine the trajectory point ratio of at least one incoming edge pointing to the same successor node in the directed graph structure, and determine the edge attribute information of the at least one incoming edge according to the trajectory point ratio, where the trajectory point ratio is used to indicate the distribution ratio of multiple backward trajectory points of the air pollutants of the first monitoring object corresponding to the same successor node among the at least one incoming edge; Construct graph structure data corresponding to the directed graph structure based on the node attribute information and the edge attribute information; Input the graph structure data into a graph neural network model for message passing to obtain corresponding graph structure features; and Based on the graph structure features, perform air quality prediction related to the air pollutants on the target object among the multiple monitored objects.

Images