Atmospheric pollutant prediction and tracing method and device, electronic equipment and storage medium

Abstract

The application provides an atmospheric pollutant prediction and tracing method and device, an electronic device and a storage medium, and belongs to the field of environmental monitoring. The method comprises the following steps: acquiring multi-source heterogeneous data in each preset period in a target research area; generating a plurality of dynamic graphs based on the multi-source heterogeneous data in each preset period, one dynamic graph corresponding to one preset period; training a spatio-temporal adaptive graph neural network model based on the plurality of dynamic graphs to obtain a trained spatio-temporal adaptive graph neural network model, which is used for predicting and tracing the ozone concentration in the target research area. The application can improve the prediction accuracy of the ozone concentration and realize rapid tracing of pollutants.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of environmental monitoring, and more particularly relates to an atmospheric pollutant prediction and tracing method and device, an electronic device, and a storage medium. BACKGROUND

[0002] Near-surface ozone (O3) as a major secondary pollutant poses a serious threat to human health and ecosystems. High-concentration ozone pollution events occur frequently, making accurate monitoring, prediction, and source analysis (i.e., tracing) of ozone a top priority in atmospheric pollution prevention and control.

[0003] Current researchers begin to use machine learning and deep learning methods to integrate atmospheric environmental data obtained from multiple channels to estimate near-surface ozone concentration. However, traditional machine learning models (such as random forests and gradient boosting machines) are difficult to effectively capture complex spatio-temporal nonlinear relationships. In recent years, recurrent neural networks represented by long short-term memory (LSTM) and convolutional neural networks (CNN) have been used to process spatio-temporal data, but they have inherent defects in processing irregular non-Euclidean spatial structure data such as monitoring sites.

[0004] Graph neural networks (GNNs) provide a new approach to solving this problem. They can directly model graph-structured data and effectively capture spatial dependencies between nodes. Spatio-temporal graph neural networks (STGNNs) further combine time series models and can handle both spatial and temporal dependencies, showing great potential in air quality prediction. However, most STGNNs models are based on static adjacency matrices to analyze pollution sources, which cannot analyze the time-varying characteristics of pollutants, leading to inaccurate ozone concentration prediction. SUMMARY

[0005] The application aims to provide an atmospheric pollutant prediction and tracing method and device, an electronic device, and a storage medium to improve the accuracy of ozone concentration prediction and achieve rapid tracing of pollutants, meeting the needs of fine control of atmospheric pollution in the new era.

[0006] In a first aspect, the application provides an atmospheric pollutant prediction and tracing method, comprising:

[0007] Acquire multi-source heterogeneous data for the target study area during various preset time periods. The multi-source heterogeneous data includes ground-based monitoring data of pollutants associated with ozone, satellite remote sensing data of pollutants, meteorological parameter data, and pollutant emission inventory data.

[0008] Multiple dynamic graphs are generated based on multi-source heterogeneous data within each preset time period, with one dynamic graph corresponding to each preset time period.

[0009] The spatiotemporal adaptive graph neural network model is trained based on multiple dynamic graphs to obtain the trained spatiotemporal adaptive graph neural network model. The trained spatiotemporal adaptive graph neural network model is used to predict and trace the source of ozone concentration in the target study area.

[0010] Among them, the methods for generating a dynamic graph corresponding to the multi-source heterogeneous data within each preset time period include:

[0011] The target study area is divided into grids, with each grid serving as a node in the dynamic graph. The preset time period is divided into multiple time steps. Based on the multi-source heterogeneous data and nodes corresponding to each time step, an adjacency matrix is ​​generated for each time step. Based on the adjacency matrix for each time step, the dynamic graph corresponding to the preset time period is generated.

[0012] A second aspect of this application provides an air pollutant prediction and source tracing device, comprising:

[0013] The data acquisition unit is used to acquire multi-source heterogeneous data for each preset time period in the target study area. The multi-source heterogeneous data includes ground-based monitoring data of pollutants associated with ozone, satellite remote sensing data of pollutants, meteorological parameter data, and pollutant emission inventory data.

[0014] The graph generation unit is used to generate multiple dynamic graphs based on multi-source heterogeneous data within each preset time period, with one preset time period corresponding to one dynamic graph.

[0015] The training unit is used to train the spatiotemporal adaptive graph neural network model based on multiple dynamic graphs, resulting in a trained spatiotemporal adaptive graph neural network model. This trained model is then used to predict and trace the ozone concentration in the target study area.

[0016] Specifically, for multi-source heterogeneous data within each preset time period, the graph generation unit is used for:

[0017] The target study area is divided into grids, with each grid serving as a node in the dynamic graph. The preset time period is divided into multiple time steps. Based on the multi-source heterogeneous data and nodes corresponding to each time step, an adjacency matrix is ​​generated for each time step. Based on the adjacency matrix for each time step, the dynamic graph corresponding to the preset time period is generated.

[0018] A third aspect of this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of the above-described method for predicting and tracing atmospheric pollutants.

[0019] In a fourth aspect of this application, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the steps of the above-described method for predicting and tracing the sources of air pollutants.

[0020] The beneficial effects of the air pollutant prediction and source tracing method, device, electronic equipment, and storage medium provided in this application are as follows:

[0021] This application embodiment avoids the limitations of single-source data analysis by acquiring multi-source heterogeneous data and training the model based on this data. Furthermore, this application embodiment divides the target study area into grids, using each grid as a node in a dynamic graph. A preset time period is divided into multiple time steps. Adjacency matrices are generated for each time step based on the corresponding multi-source heterogeneous data and nodes. A dynamic graph corresponding to the preset time period is then generated based on these adjacency matrices. By training the spatiotemporal adaptive graph neural network model using the dynamically changing adjacency matrices, the limitations of the static graph structure of traditional GNNs models are overcome. This allows the trained model to capture the dynamic evolution of the spatiotemporal correlation of ozone concentrations in different grid units in the atmosphere in real time, significantly improving the prediction accuracy and precision of the model in complex pollution processes. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 A flowchart illustrating an embodiment of the atmospheric pollutant prediction and source tracing method provided in this application;

[0024] Figure 2 A flowchart illustrating an air pollutant prediction and source tracing method provided in another embodiment of this application;

[0025] Figure 3 This is a structural block diagram of an air pollutant prediction and source tracing device provided in an embodiment of this application;

[0026] Figure 4 This is a schematic block diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0027] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0028] To make the objectives, technical solutions, and advantages of this application clearer, the following description will be provided in conjunction with the accompanying drawings and specific embodiments.

[0029] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the atmospheric pollutant prediction and source tracing method provided in this application. The method can be executed by an electronic device and may include: S101-S103.

[0030] S101: Obtain multi-source heterogeneous data for each preset time period in the target study area.

[0031] Multi-source heterogeneous data includes ground-based monitoring data of pollutants associated with ozone, satellite remote sensing data of pollutants, meteorological parameter data, and pollutant emission inventory data.

[0032] In this embodiment, the pollutant ground-based monitoring data can be high temporal resolution ozone concentration data and ozone precursor concentration data collected by thousands of monitoring stations deployed by the environmental monitoring center, for example, hourly data, i.e., ozone concentration data is collected once per hour. Ozone precursors can include substances such as nitrogen dioxide (NO2), carbon monoxide (CO), and formaldehyde (HCHO). However, due to the vast area and uneven distribution of monitoring stations, monitoring blind spots exist in vast non-urban areas, making it impossible to fully characterize the continuous spatial distribution characteristics of ozone pollution. Therefore, this embodiment also utilizes pollutant satellite remote sensing data.

[0033] Satellite remote sensing data for pollutants is based on large-scale, continuous observational data such as ozone column concentration and ozone precursor column concentration collected by satellite sensors. The data collected by satellite sensors represents the concentration of various pollutants throughout the entire atmospheric column, which complements data collected by various ground-based monitoring stations near the ground. Satellite remote sensing data includes O3 column concentration, NO2 column concentration, formaldehyde column concentration, and aerosol optical depth (AOD), among others.

[0034] Meteorological parameter data can be hourly, gridded meteorological parameter data provided by the weather forecast center, such as 10-meter U / V wind speed, 2-meter temperature, and relative humidity.

[0035] Pollutant emission inventory data can be extracted from the database for nitrogen oxides (NOx). x Gridded emission intensity data for ozone precursors such as volatile organic compounds (VOCs) categorized by source type (e.g., industry, transportation, agriculture, power, regional background / transmission).

[0036] In this embodiment, the pollutant ground-based monitoring data, pollutant satellite remote sensing data, meteorological parameter data, and pollutant emission inventory data are historical data. These can be used to train the spatiotemporal adaptive graph neural network model described below, making the predictions and source tracing results of the trained model more accurate. The multi-source heterogeneous data in this embodiment includes multi-source heterogeneous data from various preset historical time periods, which can be daily, weekly, or monthly.

[0037] S102: Generate multiple dynamic graphs based on multi-source heterogeneous data within each preset time period, with one dynamic graph corresponding to each preset time period.

[0038] In one embodiment, for each preset time period, a dynamic graph is generated based on multi-source heterogeneous data within that preset time period, including:

[0039] Spatial alignment, temporal alignment, outlier handling, and normalization are performed on the multi-source heterogeneous data within the preset time period to obtain preprocessed multi-source heterogeneous data. A dynamic graph is then generated based on the preprocessed multi-source heterogeneous data.

[0040] In this embodiment, spatial alignment means: resampling the multi-source heterogeneous data in S101 to a uniform spatial resolution, such as 0.1°×0.1°, through bilinear interpolation or nearest neighbor interpolation, and matching the data collected by discrete monitoring stations to the grid cell where the monitoring station is located.

[0041] Time alignment means unifying the multi-source heterogeneous data in S101 to an hourly temporal resolution. Because pollutant satellite remote sensing data is collected on a daily basis, time interpolation can be used or it can be used as a static feature of all hours of the day.

[0042] Outlier handling refers to imputing missing values ​​or removing outliers. For example, for ground-based monitoring data of pollutants, temporal interpolation and spatial interpolation can be used for imputation. For raster data such as meteorological parameters, spatiotemporal kriging interpolation can be used. A preset threshold is set, and if the data missing rate of a single monitoring station or a single raster exceeds the preset threshold within a certain preset time period, it will not be included in subsequent calculations.

[0043] Normalization refers to the application of Z-score standardization to ensure that the spatially aligned, temporally aligned, and outlier-handled data follow a standard normal distribution with a mean of 0 and a variance of 1, thus eliminating the influence of dimensions. The spatial alignment, temporal alignment, and outlier handling processes are performed in no particular order.

[0044] In one embodiment, the method for generating a dynamic graph corresponding to multi-source heterogeneous data within each preset time period includes:

[0045] The target study area is divided into grids, with each grid serving as a node in the dynamic graph. The preset time period is divided into multiple time steps. Based on the multi-source heterogeneous data and nodes corresponding to each time step, an adjacency matrix is ​​generated for each time step. Based on the adjacency matrix for each time step, the dynamic graph corresponding to the preset time period is generated.

[0046] In this embodiment, the multi-source heterogeneous data within a preset time period can be preprocessed multi-source heterogeneous data within the preset time period. Based on the multi-source heterogeneous data within the preset time period and the rasterized target research area, a dynamic graph corresponding to the preset time period can be generated. The nodes of the dynamic graph are multiple grids obtained by rasterizing the target research area, and the edge weights of the dynamic graph are a dynamic adjacency matrix obtained based on the preprocessed multi-source heterogeneous data within the preset time period. This dynamic adjacency matrix can dynamically change based on time steps.

[0047] This embodiment constructs an adjacency matrix that dynamically changes over time. Compared to the traditional adjacency matrix generated solely based on static geographical distance, this matrix reflects the dynamic changes of atmospheric pollutants (ozone), making it easier to obtain the time-varying characteristics of ozone. This is beneficial for predicting ozone concentration in the target study area based on a spatiotemporal adaptive graph neural network model, and can improve the accuracy of the prediction.

[0048] S103: The spatiotemporal adaptive graph neural network model is trained based on multiple dynamic graphs to obtain the trained spatiotemporal adaptive graph neural network model. The trained spatiotemporal adaptive graph neural network model is used to predict and trace the ozone concentration in the target study area.

[0049] In one embodiment, training a spatio-temporal adaptive graph neural network (ST-AGNN) model based on multiple dynamic graphs includes:

[0050] For each dynamic graph, the dynamic graph is processed by spatiotemporal convolution through multiple cascaded spatiotemporal graph convolution modules to obtain intermediate feature representations;

[0051] The intermediate feature representation is input into the ozone concentration prediction branch to obtain the ozone concentration prediction value, and the intermediate feature representation is input into the ozone source tracing branch to obtain the ozone source prediction value.

[0052] Determine the concentration prediction loss between the predicted and actual ozone concentrations, and determine the source apportionment loss between the predicted and actual ozone sources.

[0053] The spatiotemporal adaptive graph neural network model is trained based on concentration prediction loss and source resolution loss.

[0054] In this embodiment, the ST-AGNN model adopts an encoder-decoder architecture, which includes a structure of multiple cascaded spatiotemporal graph convolutional modules (encoder) and two parallel task branches (decoder).

[0055] Each spatiotemporal graph convolutional module includes a spatial graph convolutional layer and a temporal convolutional layer, which are connected in series.

[0056] In one embodiment, the structure of multiple cascaded spatiotemporal graph convolutional modules includes: M cascaded spatiotemporal graph convolutional modules; inputting the dynamic graph into the structure of the multiple cascaded spatiotemporal graph convolutional modules to obtain intermediate feature representations includes:

[0057] The dynamic graph is sequentially processed through M cascaded spatiotemporal graph convolution modules to obtain intermediate feature representations. The dynamic graph serves as the input to the first-level spatiotemporal graph convolution module, while the inputs to other levels of spatiotemporal graph convolution modules are the outputs of the previous-level spatiotemporal graph convolution module.

[0058] That is, the dynamic graph is sequentially processed by spatiotemporal convolution through a structure of M cascaded spatiotemporal graph convolution modules to obtain intermediate feature representations;

[0059] Specifically, the dynamic image is processed by spatiotemporal convolution through the first-level spatiotemporal graph convolution module, including: inputting the dynamic image into the spatial graph convolution layer of the first-level spatiotemporal graph convolution module for spatial information aggregation, and inputting the aggregation result into the temporal convolution layer of the first-level spatiotemporal graph convolution module for temporal convolution processing to obtain spatiotemporal convolution features;

[0060] Among them, spatiotemporal convolution processing is performed through any other level of spatiotemporal graph convolution module, including:

[0061] Obtain the spatiotemporal convolutional features output by the previous level module of this level spatiotemporal convolutional module; aggregate the spatiotemporal convolutional features through the spatial graph convolutional layer of this level for spatial information, and input the aggregation result into the temporal convolutional layer of this level to obtain the spatiotemporal convolutional features; perform temporal convolution processing on the spatiotemporal convolutional features through the temporal convolutional layer of this level to obtain the spatiotemporal convolutional features corresponding to this layer;

[0062] The spatiotemporal convolutional features output by the last layer are used as intermediate feature representations.

[0063] In this embodiment, the spatial graph convolutional layer adopts the idea of ​​adaptive graph convolutional networks and combines it with an attention mechanism. This allows for more intelligent aggregation of effective information in the space through dynamically learnable attention weights. The core of the spatial graph convolutional layer is to update the features of the current node by aggregating information from neighboring nodes; this process is the spatial information aggregation process. For the current node i, its spatial information aggregation process is as follows:

[0064]

[0065] Where t represents the t-th time step, and j represents other nodes. This represents the input features of node j. This represents the learnable weight matrix. Indicates the attention coefficient. This represents the output feature of node i after spatial information aggregation. ( ) represents the activation function. Let j represent the set of node i and its neighboring nodes.

[0066] Attention coefficient Not only dependent on node features It is also modulated by a dynamic graph to achieve a "learnable edge weight mechanism". For each time step t, the input feature vector of each node i is: The preprocessed multi-source heterogeneous data is then stitched together at this node. Based on this attention coefficient, the spatiotemporal adaptive graph neural network model can focus on the neighboring nodes that have the greatest impact on the current node, which is beneficial for tracing the source of ozone.

[0067] Temporal convolutional layers employ gated recurrent units or temporal convolutional networks to capture the temporal evolution of each node. For example, a gated recurrent unit is used, whose input is the output of the spatial graph convolutional layer. H' The output of this unit is the output of the spatiotemporal graph convolution module. H'' .

[0068] The two parallel task branches are an ozone concentration prediction branch and an ozone source tracing branch. Inputting the aforementioned intermediate feature representations into the ozone concentration prediction branch and the ozone source tracing branch, respectively, yields the predicted ozone concentration and ozone source. The ozone concentration prediction branch is a fully connected layer that outputs predicted near-surface ozone concentrations for one or more future time steps. The ozone source tracing branch is also a fully connected layer, followed by a Softmax activation function, which outputs a K-dimensional vector p = {p1, p2, ..., p...}. K}, where K is the number of predefined ozone pollution source categories, for example, K = 5, representing industrial sources, transportation sources, agricultural sources, power sources, and industrial, transportation, agricultural, power, and regional background / transmission sources, etc. K This represents the proportion of the contribution of pollution source K to the increase in ozone concentration, satisfying Σp K =1.

[0069] This application innovatively achieves integrated ozone "prediction-source tracing," embedding quantitative pollution source analysis functionality into a spatiotemporal graph neural network architecture for the first time. By designing multi-task outputs and a hybrid loss function based on physical priors, the model can simultaneously and quantitatively analyze the contribution ratio of different pollution sources while predicting ozone concentration. This solves the problems of traditional prediction models being unable to trace sources, and source tracing models being computationally expensive and difficult to integrate with data-driven methods.

[0070] In one embodiment, training a spatiotemporal adaptive graph neural network model based on concentration prediction loss and source resolution loss includes:

[0071] Determine the coefficients corresponding to the concentration prediction loss and the source apportionment loss, respectively;

[0072] The mixing loss is determined based on the concentration prediction loss, source apportionment loss, and their respective coefficients;

[0073] Based on the hybrid loss, the parameters of the spatiotemporal adaptive graph neural network model are adjusted to obtain the trained spatiotemporal adaptive graph neural network model.

[0074] In this embodiment, the loss function of the spatiotemporal adaptive graphical neural network model is a hybrid loss function. :

[0075]

[0076] in, Indicates mixed loss, Indicates the concentration prediction loss, Indicates the source resolution loss. λ This represents a hyperparameter used to balance the accuracy of concentration prediction with the physical consistency of source apportionment results.

[0077] Concentration prediction loss It is used to measure the difference between predicted and actual ozone concentrations. For example, it can be expressed as root mean square error (RMSE) or mean absolute error (MAE). This represents the predicted ozone concentration. This represents the actual ozone concentration value, and |||2 represents the 2-norm.

[0078] Source resolution loss The source contribution vector p used to measure the difference between the model output and the physical prior vector e The differences in distribution between them can be represented, for example, by KL divergence or cross-entropy loss. Because the actual source contribution ratios are often unknown, this embodiment proposes a physical constraint loss based on prior knowledge of emission inventories, namely, source apportionment loss, to accurately locate ozone sources. Specifically, the precursor emission ratio vectors of various sources for each grid cell at each time step can be calculated based on pollutant emission inventory data. e= {e 1 , e 2 , ..., e K } That is, the physical prior vector, which is used as the "physical prior" or "weak label" of the source contribution ratio.

[0079] This embodiment minimizes... By adjusting the parameters of the spatiotemporal adaptive graph neural network model, while the model learns to fit real ozone concentration data, its internal source resolution logic is guided in a direction consistent with the physical reality reflected in the emission inventory. This results in more accurate ozone concentration predictions for future periods and quantitative proportions of various pollution sources, thereby identifying the most important pollution sources affecting ozone formation.

[0080] In this embodiment, the source apportionment loss term incorporates prior physical knowledge based on the emission inventory, imposing physical constraints on the model's learning process. This avoids the deep learning model becoming a complete "black box," making the source tracing results not only data-driven but also consistent with the basic principles of atmospheric chemistry, greatly enhancing the reliability and interpretability of the results.

[0081] As can be seen from the above, the embodiments of this application, by acquiring multi-source heterogeneous data and training the model based on this data, avoid the limitations of single-source data analysis. The embodiments of this application also divide the target study area into grids, using each grid as a node in a dynamic graph. A preset time period is divided into multiple time steps. Adjacency matrices are generated for each time step based on the multi-source heterogeneous data and nodes corresponding to each time step. A dynamic graph corresponding to the preset time period is then generated based on these adjacency matrices. By training the spatiotemporal adaptive graph neural network model using the dynamically changing adjacency matrices based on time steps, the limitations of the static graph structure of traditional GNNs models are overcome. This allows the trained model to capture the dynamic evolution of the spatiotemporal correlation of ozone concentrations in different grid units in the atmosphere in real time, significantly improving the prediction accuracy and precision of the model in complex pollution processes.

[0082] In one embodiment of this application, for each time step, an adjacency matrix corresponding to that time step is generated based on the node and the multi-source heterogeneous data corresponding to that time step, including:

[0083] For each node, the meteorological correlation between that node and other connected nodes is calculated based on the meteorological parameter data corresponding to that time step.

[0084] Based on the ozone precursor emission data in the pollutant emission inventory data corresponding to this time step, calculate the environmental similarity between this node and other connected nodes.

[0085] Calculate the node embedding vector of the multi-source heterogeneous data corresponding to the node at the time step, and calculate the similarity between the node embedding vectors of the node and the node embedding vectors of other connected nodes;

[0086] The first side weight matrix of the adjacency matrix is ​​obtained based on the meteorological correlation set corresponding to all nodes, the second side weight matrix is ​​obtained based on the environmental similarity set corresponding to all nodes, and the third side weight matrix is ​​obtained based on the similarity set corresponding to all nodes.

[0087] The target edge weight matrix is ​​obtained by weighted fusion of the first edge weight matrix, the second edge weight matrix, and the third edge weight matrix.

[0088] Generate the adjacency matrix corresponding to this time step based on the node and target edge weight matrix.

[0089] In one embodiment, meteorological parameter data includes wind speed or wind direction;

[0090] Based on the meteorological parameter data corresponding to this time step, calculate the meteorological correlation between this node and other connected nodes, including:

[0091] Based on the wind speed or direction corresponding to this time step, as well as the wind speed or direction corresponding to multiple historical time steps, the meteorological correlation between this node and other connected nodes is calculated using a lag correlation analysis algorithm.

[0092] In this embodiment, the wind speed or direction corresponding to this time step, as well as the wind speed or direction corresponding to multiple historical time steps, are all training data. Multiple historical time steps refer to multiple time steps prior to this time step. For example, using a hysteresis correlation analysis algorithm, the correlation coefficient between the wind speed (U / V component) of node i and the wind speed of node j within the past 6 hours is calculated.

[0093] In one embodiment, the meteorological correlation between the node and other connected nodes is calculated based on the meteorological parameter data corresponding to the time step, including: using the Hybrid Single-Particle Lagrangian Integrated Trajectory (HYSPLIT) model to calculate the backward trajectory overlap between node i and node j to obtain the meteorological correlation of the atmospheric transport relationship between the two nodes.

[0094] In one embodiment, based on ozone precursor emission data in the pollutant emission inventory data corresponding to the time step, the environmental similarity between the node and other connected nodes is calculated, including:

[0095] Based on the pollutant emission inventory data corresponding to this time step, the cosine similarity of the NOx to VOCs ratio (NV-Ratio) of this node and other connected nodes is calculated to obtain the environmental similarity. This is because regions with similar NV-Ratios often have similar ozone formation potential.

[0096] In one embodiment, the node embedding vector of the node in the multi-source heterogeneous data corresponding to the node at the given time step is calculated. First, the four types of data obtained in step S101 are preprocessed. The preprocessed and normalized multi-source heterogeneous data are then concatenated to obtain the node feature vector of the node. The dimensionality of this node feature vector is then reduced to obtain a low-dimensional feature vector, which is the node embedding vector of the node. In another embodiment, the similarity between the node embedding vectors of the node and those of other connected nodes is calculated, including: calculating the inner product of the node embedding vectors of the node and those of other connected nodes, and performing a normalization operation to obtain the similarity.

[0097] In this embodiment, it is assumed that the node is the source node. The node connected to it is the target node. Then the inner product of the node embedding vectors corresponding to the source node and the target node is ( The similarity between the source node and the target node is: ,in ( ) indicates a normalization operation. ( ) denotes a non-linear activation function. The third edge weight matrix of the adjacency matrix is ​​obtained by weighting the similarity sets corresponding to all nodes. This refers to an adaptive learning view, which can capture hidden, non-explicit spatiotemporal relationships in data.

[0098] The target edge weight matrix is ​​obtained by weighted fusion of the first edge weight matrix, the second edge weight matrix, and the third edge weight matrix; the adjacency matrix corresponding to the time step is generated based on the node and the target edge weight matrix.

[0099] The adjacency matrix can be represented as follows:

[0100]

[0101] in, , , This represents the learnable weight parameters. This represents the first side weight matrix, i.e., the meteorological similarity view. This represents the weight matrix on the second side, i.e., the precursor emission similarity view. Let represent the third side weight matrix, i.e., the adaptive learning view, and t represent the t-th time step.

[0102] In one embodiment of this application, after weighting and fusing the first side weight matrix, the second side weight matrix, and the third side weight matrix to obtain the target side weight matrix, the atmospheric pollutant prediction and source tracing method further includes:

[0103] The pollutant diffusion coefficient between each node is determined based on pollutant ground-based monitoring data and pollutant satellite remote sensing data, and the pollutant half-life between each node is determined based on meteorological parameter data.

[0104] The spatial decay factor is determined based on the distance between each node and the pollutant diffusion coefficient, and the temporal decay factor is determined based on the pollutant half-life between each node and the duration of each meteorological parameter in the meteorological parameter data.

[0105] Generate the adjacency matrix corresponding to this time step based on the node and target edge weight matrices, including:

[0106] The target edge weight matrix is ​​modified by adjusting the spatial decay factor and the time decay factor to obtain the modified target edge weight matrix.

[0107] Generate the adjacency matrix corresponding to this time step based on the nodes and the corrected target edge weight matrix.

[0108] In this embodiment, the timestamp and latitude / longitude information corresponding to each monitoring station at the current time step are extracted, and the above information is matched to the corresponding grid (node). The satellite remote sensing data of pollutants is downscaled in terms of spatial resolution (consistent with the grid division scale), and matched to each time step based on the temporal interpolation method to obtain the remote sensing inversion concentration data of each grid (node) corresponding to the current time step; the ground-based monitoring data and remote sensing inversion concentration data of the same node are checked for consistency and then weighted and fused to obtain the comprehensive pollutant concentration data of each node at the current time step.

[0109] The pollutant diffusion coefficient between nodes is calculated based on the following formula:

[0110]

[0111] in, Let be the pollutant diffusion coefficient between node i and node j. Let be the diffusion flux of pollutants from node i to node j. The straight-line distance between node i and node j and These represent the combined pollutant concentrations at nodes i and j at time step t, respectively. Ensure that the diffusion direction is from high concentration to low concentration.

[0112] In this embodiment, after obtaining the pollutant diffusion coefficient, the spatial attenuation factor can be determined based on the distance between each node and the pollutant diffusion coefficient:

[0113]

[0114] in, Indicates the spatial decay factor. , This represents the wind direction correction factor, determined based on meteorological parameter data at time step t. If node j is located downwind of node i (i.e., the wind direction is from i to j), then... , This is the wind direction enhancement factor, such as 0.2. If node j is located upwind or perpendicular to the wind direction of node i, then... .

[0115] In one embodiment, the half-life of a pollutant (such as ozone precursors NOx and VOCs) is essentially the time it takes for it to degrade to half its initial concentration in the atmosphere due to reactions such as photolysis and oxidation, while meteorological parameters (such as light, temperature, and oxidants) are key drivers of the reaction rate. Therefore, a baseline half-life of the pollutant can be obtained based on laboratory data, and then the corrected pollutant half-life can be obtained by determining the baseline half-life based on meteorological parameter data at each node.

[0116] In this embodiment, the pollutant half-life can dynamically respond to changes in meteorological conditions, and more accurately reflect the attenuation characteristics of pollutants in spatiotemporal propagation.

[0117] In one embodiment, the time decay factor is determined based on the pollutant half-life between nodes and the duration of each meteorological parameter in the meteorological parameter data, and can be obtained by the following formula:

[0118]

[0119] in, This represents the time decay factor of node i at a historical time step to node j at the current time step. It is used to quantify the degree of decay of the influence of pollutants over time, and its value ranges from (0,1]. It is a natural exponential function. This represents the time interval between two adjacent time steps. This represents the half-life of a pollutant as it propagates from node i to node j. This represents the function that takes the minimum value. This indicates the duration during which meteorological parameters (such as wind speed, wind direction, and sunlight) remain stable within a time interval Δt. This represents the stability gain coefficient, which is a constant used to avoid the attenuation factor being too small when meteorological conditions are unstable, thereby enhancing model stability.

[0120] As can be seen from the above, this embodiment determines the diffusion coefficient by combining ground-based pollutant monitoring data and satellite remote sensing data, determines the pollutant half-life based on meteorological parameter data, and then calculates the spatial decay factor and temporal decay factor. These factors are then used to correct the target edge weight matrix to generate an adjacency matrix. The method provided in this embodiment considers multiple factors, making the generated adjacency matrix more realistic and able to more accurately reflect the propagation relationship of pollutants, thus improving the accuracy and reliability of atmospheric pollutant prediction and source tracing.

[0121] In one embodiment of this application, reference is made to Figure 2 Using a specific region and its surrounding areas from June 1 to August 31, 2024 as the target research area, this paper presents the complete process of the atmospheric pollutant prediction and source tracing method proposed in this application.

[0122] Data acquisition phase:

[0123] Pollutant ground-based monitoring data: Obtain hourly O3, NO2, and CO concentration data from six national monitoring stations within the historical target study area.

[0124] Pollutant satellite remote sensing data: NO2 and HCHO column concentration data were downloaded from a website, with a spatial resolution of 0.05°×0.05°.

[0125] Meteorological parameter data: Hourly reanalysis data were acquired, including 10 m U / V wind speed, 2 m temperature, boundary layer height, total cloud cover, and surface solar radiation, with a spatial resolution of 0.1°.

[0126] Pollutant emission inventory data: Historical monthly emission inventory data from the database were used, with source categories including power, industry, civil, transportation and agriculture, and a spatial resolution of 0.1°×0.1°.

[0127] Preprocessing stage: All the above data are spatially and temporally aligned, uniformly resampled to a 0.1°×0.1° grid, and the temporal resolution is uniformly 1 hour. Then, the spatiotemporally aligned data are imputed and the features are normalized. For example, spatiotemporal kriging is used to impute missing values ​​and Z-score standardization is performed.

[0128] Dynamic graph construction phase:

[0129] The target study area was divided into 80×80=6400 nodes.

[0130] The preset time period is one month, with a time step of 12 hours.

[0131] First side weight matrix For any two nodes i and j, calculate the Pearson correlation coefficient between the wind speed vectors at point i and j over the past 6 hours, and use it as the edge weight.

[0132] Second side weight matrix Calculate the cosine similarity of the (NOx emission / VOCs emission) ratio between nodes i and j, and use it as the edge weight to reflect the similarity of ozone generation areas.

[0133] Third side weight matrix : Set the node embedding dimension d=64, which is generated by two learnable node embedding matrices (6400×64) and (6400×64).

[0134] The learnable weights ω1, ω2, and ω3 are set to an initial value of 1 / 3 and are automatically adjusted during training.

[0135] Spatiotemporal adaptive graphical neural network model construction and training phases:

[0136] Spatiotemporal adaptive graph neural network model: Employs a 3-layer stacked spatiotemporal graph convolutional module. In each spatiotemporal graph convolutional module, the spatial graph convolutional layer uses a graph convolutional network with attention mechanism, the temporal convolutional layer uses a gated recurrent unit, and the hidden layer dimension is 128.

[0137] Source category definition: Define K=5 source categories: power, industrial, transportation, civil and agricultural (combined), regional background / transmission.

[0138] Mixed loss: Concentration prediction loss Lconc uses MAE. Source apportionment loss Lsrc uses KL divergence, and its "weak label" is composed of the proportion of NOx emissions from the five types of sources at each node in the MEIC emission inventory. The hyperparameter λ is set to 0.2.

[0139] Model training phase: Data from June-July 2024 was used as the training set, and data from August 2024 was used as the test set. The Adam optimizer was used with a learning rate of 0.001, a batch size of 32, and training for 100 epochs.

[0140] Results evaluation and analysis:

[0141] Ozone concentration prediction evaluation: On the test set, the prediction results of the method in this application were compared with the actual observations from ground monitoring stations. Evaluation metrics included root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). Compared with traditional benchmark models such as LSTM, static GCN, and Graph WaveNet, the RMSE of the method in this application was reduced by 15%–20%, and the R² was significantly lower. 2 It improved by 0.1-0.15, especially during a typical heavy pollution event, it can more accurately capture the timing and intensity of pollution peaks.

[0142] Source tracing results verification: In a typical case of cross-regional pollution transport caused by southeasterly winds, the model's source tracing results showed that the contribution of "regional background / transport" sources to the increase in ozone concentration in Beijing rose from 20% before pollution to 55% at the pollution peak, which is highly consistent with the qualitative analysis results using the HYSPLIT backward trajectory model. Meanwhile, during morning and evening rush hours, the contribution of "traffic" sources in the urban center increased significantly, consistent with objective laws.

[0143] Corresponding to the air pollutant prediction and source tracing method in the above embodiments, Figure 3 This is a structural block diagram of an air pollutant prediction and source tracing device provided according to an embodiment of this application. For ease of explanation, only the parts relevant to the embodiment of this application are shown. References Figure 3 The air pollutant prediction and tracing device 20 includes: a data acquisition unit 21, a graph generation unit 22, and a training unit 23.

[0144] Among them, the data acquisition unit 21 is used to acquire multi-source heterogeneous data in the target study area for each preset time period. The multi-source heterogeneous data includes ground-based monitoring data of pollutants associated with ozone, satellite remote sensing data of pollutants, meteorological parameter data and pollutant emission inventory data.

[0145] The graph generation unit 22 is used to generate multiple dynamic graphs based on multi-source heterogeneous data within each preset time period, with one preset time period corresponding to one dynamic graph.

[0146] Training unit 23 is used to train the spatiotemporal adaptive graph neural network model based on multiple dynamic graphs to obtain the trained spatiotemporal adaptive graph neural network model. The trained spatiotemporal adaptive graph neural network model is used to predict and trace the ozone concentration in the target study area.

[0147] Specifically, for multi-source heterogeneous data within each preset time period, the graph generation unit 22 is used for:

[0148] The target study area is divided into grids, with each grid serving as a node in the dynamic graph. The preset time period is divided into multiple time steps. Based on the multi-source heterogeneous data and nodes corresponding to each time step, an adjacency matrix is ​​generated for each time step. Based on the adjacency matrix for each time step, the dynamic graph corresponding to the preset time period is generated.

[0149] In one embodiment of this application, for each time step, the graph generation unit 22 is specifically used for:

[0150] For each node, the meteorological correlation between that node and other connected nodes is calculated based on the meteorological parameter data corresponding to that time step.

[0151] Based on the ozone precursor emission data in the pollutant emission inventory data corresponding to this time step, calculate the environmental similarity between this node and other connected nodes.

[0152] Calculate the node embedding vector of the multi-source heterogeneous data corresponding to the node at the time step, and calculate the similarity between the node embedding vectors of the node and the node embedding vectors of other connected nodes;

[0153] The first side weight matrix of the adjacency matrix is ​​obtained based on the meteorological correlation set corresponding to all nodes, the second side weight matrix is ​​obtained based on the environmental similarity set corresponding to all nodes, and the third side weight matrix is ​​obtained based on the similarity set corresponding to all nodes.

[0154] The target edge weight matrix is ​​obtained by weighted fusion of the first edge weight matrix, the second edge weight matrix, and the third edge weight matrix.

[0155] Generate the adjacency matrix corresponding to this time step based on the node and target edge weight matrix.

[0156] In one embodiment of this application, the meteorological parameter data includes wind speed or wind direction;

[0157] The graph generation unit 22 is specifically used for:

[0158] Based on the wind speed or direction corresponding to this time step, as well as the wind speed or direction corresponding to multiple historical time steps, the meteorological correlation between this node and other connected nodes is calculated using a lag correlation analysis algorithm.

[0159] In one embodiment of this application, the graph generation unit 22 is specifically used for:

[0160] Calculate the inner product of the node embedding vectors of the node and the other connected nodes, and then perform a normalization operation to obtain the similarity.

[0161] In one embodiment of this application, the training unit 23 is specifically used for:

[0162] For each dynamic graph, the dynamic graph is processed by spatiotemporal convolution through multiple cascaded spatiotemporal graph convolution modules to obtain intermediate feature representations;

[0163] The intermediate feature representation is input into the ozone concentration prediction branch to obtain the ozone concentration prediction value, and the intermediate feature representation is input into the ozone source tracing branch to obtain the ozone source prediction value.

[0164] Determine the concentration prediction loss between the predicted and actual ozone concentrations, and determine the source apportionment loss between the predicted and actual ozone sources.

[0165] The spatiotemporal adaptive graph neural network model is trained based on concentration prediction loss and source resolution loss.

[0166] In one embodiment of this application, the training unit 23 is specifically used for:

[0167] Determine the coefficients corresponding to the concentration prediction loss and the source apportionment loss, respectively;

[0168] The mixing loss is determined based on the concentration prediction loss, source apportionment loss, and their respective coefficients;

[0169] Based on the hybrid loss, the parameters of the spatiotemporal adaptive graph neural network model are adjusted to obtain the trained spatiotemporal adaptive graph neural network model.

[0170] The structure of multiple cascaded spatiotemporal graph convolutional modules includes: M cascaded spatiotemporal graph convolutional modules;

[0171] Training unit 23 is specifically used for:

[0172] The dynamic graph is sequentially processed through M cascaded spatiotemporal graph convolution modules to obtain intermediate feature representations;

[0173] The dynamic graph serves as the input to the first-level spatiotemporal graph convolution module, while the inputs to other levels of spatiotemporal graph convolution modules are the outputs of the previous-level spatiotemporal graph convolution module.

[0174] See Figure 4 , Figure 4 This is a schematic block diagram of an electronic device provided according to an embodiment of this application. Figure 4The electronic device 300 in this embodiment may include one or more processors 301, one or more input devices 302, one or more output devices 303, and one or more memories 304. The processors 301, input devices 302, output devices 303, and memories 304 communicate with each other via a communication bus 305. The memories 304 store computer programs, including program instructions. The processors 301 execute the program instructions stored in the memories 304. Specifically, the processors 301 are configured to invoke the program instructions to perform the functions of the modules in the aforementioned device embodiments, for example... Figure 3 The functions of the data acquisition unit 21, graph generation unit 22, and training unit 23 are shown.

[0175] It should be understood that, in the embodiments of this application, the processor 301 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0176] Input device 302 may include a touchpad, a fingerprint sensor (for collecting the user's fingerprint information and fingerprint orientation information), a microphone, etc., and output device 303 may include a display (LCD, etc.), a speaker, etc.

[0177] The memory 304 may include read-only memory and random access memory, and provides instructions and data to the processor 301. A portion of the memory 304 may also include non-volatile random access memory.

[0178] In specific implementations, the processor 301, input device 302, and output device 303 described in the embodiments of this application can execute the implementation methods described in the atmospheric pollutant prediction and source tracing methods provided in the embodiments of this application, or they can execute the implementation methods of the electronic devices described in the embodiments of this application, which will not be repeated here.

[0179] In another embodiment of this application, a computer-readable storage medium is provided. This computer-readable storage medium stores a computer program, which includes program instructions. When executed by a processor, the program instructions implement all or part of the processes in the methods described above. Alternatively, the computer program can instruct related hardware to complete the process. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include any entity or device capable of carrying computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0180] The computer-readable storage medium can be an internal storage unit of the electronic device in any of the foregoing embodiments, such as a hard disk or memory of the electronic device. The computer-readable storage medium can also be an external storage device of the electronic device, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the electronic device. Furthermore, the computer-readable storage medium can include both internal and external storage units of the electronic device. The computer-readable storage medium is used to store computer programs and other programs and data required by the electronic device. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0181] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0182] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the electronic devices and units described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0183] In the several embodiments provided in this application, it should be understood that the disclosed electronic devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the couplings or direct couplings or communication connections shown or discussed may be indirect couplings or communication connections through some interfaces or units, or they may be electrical, mechanical, or other forms of connection.

[0184] The modules described as separate components may or may not be physically separate. Similarly, the components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of the embodiments of this application, depending on actual needs.

[0185] Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0186] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for predicting and tracing the sources of air pollutants, characterized in that, include: Acquire multi-source heterogeneous data for each preset time period in the target study area. The multi-source heterogeneous data includes ground-based monitoring data of pollutants associated with ozone, satellite remote sensing data of pollutants, meteorological parameter data, and pollutant emission inventory data. Multiple dynamic graphs are generated based on multi-source heterogeneous data within each preset time period, with one dynamic graph corresponding to each preset time period. The spatiotemporal adaptive graph neural network model is trained based on the multiple dynamic graphs to obtain the trained spatiotemporal adaptive graph neural network model. The trained spatiotemporal adaptive graph neural network model is used to predict and trace the source of ozone concentration in the target study area. Among them, the methods for generating a dynamic graph corresponding to the multi-source heterogeneous data within each preset time period include: The target research area is divided into grids, and each grid is used as a node of the dynamic graph. The preset time period is divided into multiple time steps. Based on the multi-source heterogeneous data corresponding to each time step and the node, an adjacency matrix corresponding to each time step is generated. Based on the adjacency matrix corresponding to each time step, the dynamic graph corresponding to the preset time period is generated. Specifically, for each time step, an adjacency matrix corresponding to that time step is generated based on the node and the corresponding multi-source heterogeneous data, including: For each node, the meteorological correlation between that node and other connected nodes is calculated based on the meteorological parameter data corresponding to that time step. Based on the ozone precursor emission data in the pollutant emission inventory data corresponding to this time step, calculate the environmental similarity between this node and other connected nodes. Calculate the node embedding vector of the multi-source heterogeneous data corresponding to the node at the time step, and calculate the similarity between the node embedding vectors of the node and the node embedding vectors of other connected nodes; The first side weight matrix of the adjacency matrix is ​​obtained based on the meteorological correlation set corresponding to all nodes, the second side weight matrix is ​​obtained based on the environmental similarity set corresponding to all nodes, and the third side weight matrix is ​​obtained based on the similarity set corresponding to all nodes. The target edge weight matrix is ​​obtained by weighted fusion of the first edge weight matrix, the second edge weight matrix, and the third edge weight matrix. Generate the adjacency matrix corresponding to the time step based on the node and the target edge weight matrix.

2. The method as described in claim 1, characterized in that, The meteorological parameter data includes wind speed or wind direction; The calculation of the meteorological correlation between the node and other connected nodes based on the meteorological parameter data corresponding to the time step includes: Based on the wind speed or direction corresponding to this time step, as well as the wind speed or direction corresponding to multiple historical time steps, the meteorological correlation between this node and other connected nodes is calculated using a lag correlation analysis algorithm.

3. The method as described in claim 1, characterized in that, The calculation of the similarity between the node's embedding vectors and those of other connected nodes includes: The similarity is obtained by calculating the inner product of the node embedding vectors of the node and the other connected nodes, and then performing a normalization operation.

4. The method as described in claim 1, characterized in that, The training of the spatiotemporal adaptive graph neural network model based on the multiple dynamic graphs includes: For each dynamic graph, the dynamic graph is processed by spatiotemporal convolution through multiple cascaded spatiotemporal graph convolution modules to obtain intermediate feature representations; The intermediate feature representation is input into the ozone concentration prediction branch to obtain the ozone concentration prediction value, and the intermediate feature representation is input into the ozone source tracing branch to obtain the ozone source prediction value. Determine the concentration prediction loss between the predicted ozone concentration and the actual ozone concentration, and determine the source apportionment loss between the predicted ozone source and the actual ozone source. The spatiotemporal adaptive graph neural network model is trained based on the concentration prediction loss and the source resolution loss.

5. The method as described in claim 4, characterized in that, The training of the spatiotemporal adaptive graph neural network model based on the concentration prediction loss and the source resolution loss includes: Determine the coefficients corresponding to the concentration prediction loss and the source apportionment loss, respectively; The mixing loss is determined based on the concentration prediction loss, the source apportionment loss, and their respective coefficients; Based on the hybrid loss, the parameters of the spatiotemporal adaptive graph neural network model are adjusted to obtain the trained spatiotemporal adaptive graph neural network model.

6. The method as described in claim 4, characterized in that, The structure of the multiple cascaded spatiotemporal graph convolutional modules includes: M cascaded spatiotemporal graph convolutional modules; The process of performing spatiotemporal convolution on the dynamic graph using multiple cascaded spatiotemporal graph convolution modules to obtain intermediate feature representations includes: The dynamic graph is sequentially processed through the M cascaded spatiotemporal graph convolution modules to obtain intermediate feature representations; The dynamic graph is the input to the first-level spatiotemporal graph convolution module, and the input to other-level spatiotemporal convolution modules is the output of the previous-level spatiotemporal graph convolution module.

7. An air pollutant prediction and source tracing device, characterized in that, include: The data acquisition unit is used to acquire multi-source heterogeneous data of the target study area in various preset time periods. The multi-source heterogeneous data includes ground-based monitoring data of pollutants associated with ozone, satellite remote sensing data of pollutants, meteorological parameter data, and pollutant emission inventory data. The graph generation unit is used to generate multiple dynamic graphs based on multi-source heterogeneous data within each preset time period, with one preset time period corresponding to one dynamic graph. The training unit is used to train the spatiotemporal adaptive graph neural network model based on the multiple dynamic graphs to obtain the trained spatiotemporal adaptive graph neural network model. The trained spatiotemporal adaptive graph neural network model is used to predict and trace the ozone concentration in the target study area. Specifically, for multi-source heterogeneous data within each preset time period, the graph generation unit is used for: The target research area is divided into grids, and each grid is used as a node of the dynamic graph. The preset time period is divided into multiple time steps. Based on the multi-source heterogeneous data corresponding to each time step and the node, an adjacency matrix corresponding to each time step is generated. Based on the adjacency matrix corresponding to each time step, the dynamic graph corresponding to the preset time period is generated. The graph generation unit is specifically used for: For each node, the meteorological correlation between that node and other connected nodes is calculated based on the meteorological parameter data corresponding to that time step. Based on the ozone precursor emission data in the pollutant emission inventory data corresponding to this time step, calculate the environmental similarity between this node and other connected nodes. Calculate the node embedding vector of the multi-source heterogeneous data corresponding to the node at the time step, and calculate the similarity between the node embedding vectors of the node and the node embedding vectors of other connected nodes; The first side weight matrix of the adjacency matrix is ​​obtained based on the meteorological correlation set corresponding to all nodes, the second side weight matrix is ​​obtained based on the environmental similarity set corresponding to all nodes, and the third side weight matrix is ​​obtained based on the similarity set corresponding to all nodes. The target edge weight matrix is ​​obtained by weighted fusion of the first edge weight matrix, the second edge weight matrix, and the third edge weight matrix. Generate the adjacency matrix corresponding to the time step based on the node and the target edge weight matrix.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 6.

Images