An air quality prediction method based on a space-time graph convolution network

Abstract

The application provides an air quality prediction method based on a space-time graph convolution network, and specific steps are as follows: 1, obtaining air quality data, grid meteorological data and point of interest (POI: POI Data Feature Map) data; 2, performing consistency checking, abnormal value detection processing and missing value filling on the collected heterogeneous data; 3, using a multi-source data fusion method, a multi-dimensional grid feature map construction method is proposed, and different types of data are converted into grid layers; 4, an AE-STGCN space-time graph convolution network prediction model with a fusion attention mechanism is constructed; the space-time convolution layer in the space-time convolution block is used for extracting the space-time features in the air quality data; the data dimension is integrated by using a full connection layer; 5, air quality multi-channel image data is input into the trained prediction model, and the atmospheric concentration of pollutants at each detection station is obtained. The application comprehensively considers the space-time features of various air quality influencing factors, and the prediction is more comprehensive and accurate.

Description

Technical Field

[0001] This application relates to an air quality prediction method based on spatiotemporal graph convolutional networks, belonging to the fields of air quality prediction technology and deep learning technology. Background Technology

[0002] Air quality affects public health and constrains economic development. Most existing air quality forecasting methods rely solely on air quality data, neglecting factors beyond traditional pollutant indicators. In reality, the spatial distribution of pollutant concentrations is not only related to the characteristics of the pollutants themselves but also closely linked to external factors such as meteorology. Air quality varies regionally under different meteorological conditions. For example, rising temperatures and increased wind speeds can promote pollutant dispersion. Furthermore, the distribution of different areas within urban regions also impacts air quality; for instance, industrial parks with concentrated factories tend to have significantly poorer air quality. Therefore, air quality forecasting is not limited to the analysis of time-series data but also requires consideration of meteorological conditions, urban industrial layout, and other factors. The combined effect of these factors influences the accuracy of air quality forecasts. When studying air quality in a specific region, considering the region's meteorological conditions and POI distribution is essential. Existing air quality forecasting models struggle to capture the complex, non-linear, deep-seated relationships between data when processing massive amounts of data. Moreover, most models focus on time-dimensional forecasts, insufficiently considering the spatial dependence of data and the correlation between time periods and spatial distribution, thus affecting the accuracy of the forecast results. Summary of the Invention

[0003] This invention addresses the technical problems existing in the prior art by disclosing an air quality prediction method based on spatiotemporal graph convolutional networks. This method proposes a multi-source data fusion approach, which integrates air quality data, meteorological data, and POI data to establish a multi-dimensional network feature map dataset. This strengthens the correlation analysis between multi-source data and the description of local important features, resulting in better prediction performance.

[0004] To achieve the above objectives, the technical solution of the present invention is as follows:

[0005] An air quality prediction method based on spatiotemporal graph convolutional networks, the method comprising the following steps:

[0006] Step 1: Collect air quality data, grid meteorological data, and POI data for the area to be monitored.

[0007] Step 2: Perform data preprocessing, including checking the consistency of data measurement units, the consistency of detection site information, and the continuity of time series.

[0008] Step 3: Perform multi-source data fusion to generate a multi-dimensional grid feature map.

[0009] Step 4: Construct the spatiotemporal graph convolutional network prediction model AE-STGCN, which incorporates an attention mechanism, and train it using the multidimensional grid feature map of historical data as input.

[0010] Step 5: Use the trained model to predict future air quality and obtain the predicted results of changes in atmospheric pollutant concentrations.

[0011] Step 1 is as follows:

[0012] The collected air quality data covers six types of environmental pollutants monitored by the monitoring stations: PM2.5, PM10, SO2, NO2, O3, and CO; meteorological data includes temperature, humidity, precipitation, air pressure, wind speed, and wind direction; POI data includes POI categories for different monitoring locations such as factories, roads, and parks. All of the above data is referred to as multidimensional data.

[0013] Step 2 is as follows:

[0014] The process involves checking all parameters in the multidimensional data, including temperature, humidity, and air pollutant concentrations, using consistent units of measurement. If inconsistencies arise, appropriate conversions must be performed to ensure data consistency. The check also ensures that information for each monitoring station, including station number, latitude and longitude, and monitoring period, remains consistent across the entire dataset, identifying and correcting any inconsistencies. Furthermore, the timestamps are examined to determine the data's temporal completeness and continuity; missing data is supplemented using linear interpolation, and duplicate data with the same timestamp is removed. Finally, the area to be inspected is divided into grid regions.

[0015] Step 3 is as follows:

[0016] The various types of data processed in step 2 are used as single features. Each feature is converted into a grid layer, and all grid layers are superimposed layer by layer to obtain a higher-dimensional multi-dimensional grid feature map.

[0017] Step 4 is as follows:

[0018] To address the temporal and spatial features inherent in multidimensional grid feature map sequences, an improved STGCN network model, AE-STGCN, is constructed. This model consists of an input layer, two stacked spatiotemporal convolutional blocks (ST-ConvBlock), a fully connected layer, and an output layer. Each spatiotemporal convolutional block contains three main components: a spatiotemporal attention layer, a temporally gated convolutional layer, and a spatial graph convolutional layer.

[0019] Step 5 is as follows: Input the newly collected and processed multidimensional network feature map to complete the air quality prediction.

[0020] Step 4 is as follows:

[0021] The input layer reads in the sequence of multidimensional grid feature maps generated in step 3;

[0022] Each spatiotemporal convolutional block contains three main components: a spatiotemporal attention layer, a temporally gated convolutional layer, and a spatial graph convolutional layer. The spatiotemporal attention layer is further divided into a temporal attention layer and a spatial attention layer. The spatial attention layer performs dynamic correlation analysis on nodes with different attributes, identifying the feature regions that have the greatest impact on the prediction results. This is described using an attention matrix S, which is calculated as follows:

[0023]

[0024]

[0025] in, It is the input of the r-th spatiotemporal convolutional block, C r-1 This is the number of input data channels in the r-th layer. When r = 1, we can obtain C0 = F,T r-1 Let T represent the time dimension in the r-th layer. When r = 1, we can obtain T0 = T h , These are the parameters set for model learning, and the Sigmoid function is used as the activation function. The attention matrix S is calculated using formula (2), and the elements S... i,j This represents the association strength between node i and node j. The Softmax function is used to ensure that the sum of the attention weights of the nodes is 1. During graph convolution operations, the adjacency matrix A and the spatial attention matrix are combined. To dynamically adjust the influence weights between nodes.

[0026] The spatial attention layer uses formula (3) to calculate the different weights of the data detected at each station in different time periods, describing their correlation, and uses formula (4) to calculate the temporal correlation matrix E, where element E i,j This represents the dependency between node i and node j, and matrix E is normalized using the Softmax function.

[0027]

[0028] These are the parameters set for model learning, element E. i,j The matrix E represents the degree of dependency between node i and node j. The Softmax function is used to normalize the matrix E. The normalized time attention matrix is ​​applied to the model input. Formula (5) is used to merge information and dynamically adjust the input.

[0029]

[0030] The first gated temporal convolutional layer further processes the output of the spatiotemporal attention layer. This layer is used to extract temporal features and contains a causal convolution and a gated linear unit (GLU). The GLU input sequence is divided into two parts and operated on using formula (6). Different operations are performed on the left part, which will pass through the Sigmoid function, and the right part will be connected by a residual (Fusion) connection. Then, the two parts are processed by a Hadamard product, and finally the output is obtained.

[0031]

[0032] P and Q are the two inputs of the GLU; ⊙ represents the Hadamard product. σ(Q) is the Sigmoid function, which is output after passing through the Hadamard product.

[0033] After the first gated temporal convolutional layer, the data enters the graph convolutional layer. This layer uses graph convolution operations in the spectral domain to process the data and uses the Chebyshev polynomial (7) to approximate the eigenvalues ​​of the graph Laplace operator, thereby reducing computational costs.

[0034]

[0035] in This represents the graph convolution operation. It is a k-th order Chebyshev polynomial.

[0036] Spatial graph convolutional layers extract spatial features between grid nodes, using graph convolution to process non-Euclidean data, enabling the model to learn complex functional mappings between node features and their topological structures. This identifies and utilizes spatial relationships between grids, providing deeper spatial information for subsequent time series analysis.

[0037] The second temporally gated convolutional layer has the same structure as the first temporally gated convolutional layer. It is responsible for capturing dynamic changes in the time dimension, processing features of continuous time steps through convolution operations, and using a gating mechanism to enhance the model's selective learning ability for time series information. This enables the network to effectively process time series data and identify the importance of different time nodes.

[0038] A fully connected layer is used to integrate the features extracted and processed from the above layers, with ReLU used as the activation function to ensure that the output of each component has the same dimension and shape as the prediction target.

[0039] When training the AE-STGCN model, the number of channels in the gated temporal convolutional layer is 64, the number of channels in the graph convolutional layer is 16, the number of kernels in the graph convolution and the one-dimensional convolution in the temporal dimension is 3, the learning rate is set to 0.001, the number of training epochs is 100, and the batch size is 64. The specific parameter settings are shown in Table 1.

[0040] Table 1 Explanation of Hyperparameter Settings

[0041]

[0042] An air quality prediction model based on spatiotemporal graph convolutional networks is provided to implement the aforementioned air quality prediction method based on spatiotemporal graph convolutional networks.

[0043] The prediction model includes an input layer, two stacked spatiotemporal convolutional blocks (ST-Conv Blocks), a fully connected layer, and an output layer.

[0044] The input layer data is a sequence of multi-dimensional grid feature maps obtained through data preprocessing, including air quality data feature maps, meteorological data feature maps, and POI data feature maps.

[0045] Each spatiotemporal convolutional block contains three main components: a spatiotemporal attention layer, a temporally gated convolutional layer, and a spatial graph convolutional layer.

[0046] The spatiotemporal attention layer is divided into a temporal attention layer and a spatial attention layer. The spatial attention layer emphasizes the most important spatial information in the data for air quality prediction, identifying the regions with the greatest impact on the prediction results. It learns the relative importance between each grid node to assign attention weights, allowing the model to adaptively adjust the representation of spatial features and prioritize spatial information more critical to the prediction target. The temporal attention layer identifies the importance of different time periods in the multidimensional network feature map sequence, highlighting the time points with the greatest influence on the prediction results. It calculates attention weights for different time periods, enabling the network to capture long-term dependencies and emphasize the time periods with the greatest impact on the prediction results.

[0047] Temporally gated convolutional layers are responsible for capturing dynamic changes in the time dimension. They process features from consecutive time steps through convolution operations and utilize gating mechanisms to enhance the model's selective learning ability for time-series information. This enables the network to effectively process time-series data and identify the importance of different time points.

[0048] Spatial graph convolutional layers extract spatial features between grid nodes. By using graph convolution to process non-Euclidean data, the model can learn complex functional mappings between node features and their topological structure.

[0049] The fully connected layer integrates the processed features extracted from each layer and uses ReLU as the activation function, ensuring that the output of each component has the same dimension and shape as the prediction target, thus achieving the transformation from feature representation to final output.

[0050] The output layer contains the model's prediction results, constructing an air quality data feature map for a future time series, including parameters of various pollutant indicators, including the Air Quality Index (AQI).

[0051] A terminal includes a processor, a memory, and a program for an air quality prediction method based on a spatiotemporal graph convolutional network stored in the memory. When the processor runs the air quality prediction program based on the spatiotemporal graph convolutional network, it implements the steps of the air quality prediction method based on the spatiotemporal graph convolutional network.

[0052] A computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the steps of the air quality prediction method based on a spatiotemporal graph convolutional network.

[0053] Compared with the prior art, the advantages of the present invention are as follows:

[0054] This invention proposes a multi-source data fusion method that integrates air quality data, meteorological data, and POI data to establish a multi-dimensional network feature map dataset, thereby enhancing the correlation analysis between multi-source data and the description of important local features.

[0055] This invention selects STGCN as the base model, analyzes its shortcomings in existing prediction tasks, and proposes to incorporate an attention mechanism to better focus on locally important features in the data. The overall structure of the model is also modified to construct a spatiotemporal graph convolutional network prediction model incorporating the attention mechanism. Compared with benchmark models, the proposed model fully exploits the spatiotemporal characteristics of air quality data and the impact of meteorological and POI data on air quality, resulting in better prediction performance. Attached Figure Description

[0056] Figure 1 This is a schematic flowchart of the method of the present invention;

[0057] Figure 2 The process of generating a multidimensional grid feature dataset;

[0058] Figure 3 This is a schematic diagram of the structure of an air quality prediction model. Detailed Implementation

[0059] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0060] Example: Figure 1 As shown, this invention provides an air quality prediction method for spatiotemporal graph convolutional networks, comprising the following steps:

[0061] Step 1: Collect air quality data, grid meteorological data, and POI data for the area to be monitored.

[0062] The example retrieves historical air quality data, gridded meteorological data, and POI data collected daily for 24 hours from May 2020 to December 2022 from 13 air quality monitoring stations in Nanjing. The collected air quality data covers the six types of environmental pollutants monitored by the stations: PM2.5, PM10, SO2, NO2, O3, and CO; the meteorological data includes temperature, humidity, precipitation, air pressure, wind speed, and wind direction; and the POI data includes POI categories for different monitoring locations such as factories, roads, and parks.

[0063] Step 2: Perform data preprocessing. Check that all parameters in the data from Step 1 (such as temperature, humidity, air pollutant concentration, etc.) use a consistent unit of measurement. Convert units for data with inconsistent units. Check that the information of each monitoring station (such as station number, latitude and longitude, monitoring period, etc.) is consistent throughout the entire dataset, and identify and correct inconsistent station information. Check the timestamps of the data to determine whether the data is complete and continuous in time, use linear interpolation to supplement missing data, and delete duplicate data with the same timestamp. Divide the area to be inspected into grid regions.

[0064] Step 3: As Figure 2 As shown, multi-source data fusion is performed to generate a multi-dimensional grid feature map.

[0065] For example: The Nanjing city area is divided into several grid regions with a resolution of 0.25° × 0.25°. The entire region is represented by G, where each grid point is... This refers to the grid point located in the i-th row and j-th column of G. Each parameter in the preprocessed air quality data, gridded meteorological data, and city POI data from step 2 is extracted as a single feature. Each feature corresponds to a grid layer. All grid layers are stacked layer by layer to generate a multidimensional grid feature map. The multidimensional grid feature maps corresponding to each time point are arranged in chronological order to form a sequence of 533,784 multidimensional grid feature maps used for prediction model input and training.

[0066] Step 4: As Figure 3 As shown, a spatiotemporal graph convolutional network prediction model AE-STGCN with an attention mechanism is constructed. The model consists of an input layer (InputBlock), two stacked spatiotemporal convolutional blocks (ST-Conv Blocks), a fully connected layer, and an output layer (OutputBlock).

[0067] The input layer (InputBlock) reads in the multidimensional grid feature map sequence data generated in step 3.

[0068] Each spatiotemporal convolutional block (ST-ConvBlock) contains three main components: a spatiotemporal attention layer, a temporally gated convolutional layer, and a spatial graph convolutional layer. The spatiotemporal attention layer is further divided into temporal attention and spatial attention layers. The spatial attention layer performs dynamic correlation analysis on nodes with different attributes, identifying the feature regions that have the greatest impact on the prediction results, described using an attention matrix S. The calculation method is as follows:

[0069]

[0070] in, It is the input of the r-th spatiotemporal convolutional block, C r-1 This is the number of input data channels in the r-th layer. When r = 1, we can obtain C0 = F. r-1 This represents the time dimension in the r-th layer. When r = 1, we can obtain T0 = T h . These are the parameters set for model learning, and the Sigmoid function is used as the activation function. The attention matrix S is calculated using formula (2), and the elements S... i,j This represents the association strength between node i and node j. The Softmax function is used to ensure that the sum of the attention weights of the nodes is 1. During graph convolution operations, the adjacency matrix A and the spatial attention matrix are combined. To dynamically adjust the influence weights between nodes.

[0071] The spatial attention layer uses formula (3) to calculate the different weights of the data detected at each site in different time periods, describing their correlation. Formula (4) is used to calculate the temporal correlation matrix E, where element E... i,j This represents the degree of dependency between node i and node j. The matrix E is normalized using the Softmax function.

[0072]

[0073] These are the parameters set for model learning. Element Ei,j This represents the degree of dependency between node i and node j. The matrix E is normalized using the Softmax function, and the normalized time attention matrix is ​​applied to the model input. Information is merged using formula (5) to dynamically adjust the input.

[0074]

[0075] The first gated temporal convolutional layer further processes the output of the spatiotemporal attention layer. This layer is used to extract temporal features and contains a causal convolution and a gated linear unit (GLU). The GLU input sequence is divided into two parts and operated on using Equation (6). Different operations are performed on it. The left part in the figure will pass through the Sigmoid function, while the right part will be connected by a residual (Fusion) connection. Then, the two parts will pass through a Hadamard product, and finally output.

[0076]

[0077] P and Q are the two inputs of the GLU; ⊙ represents the Hadamard product. σ(Q) is the sigmoid function, which is output after passing through the Hadamard product.

[0078] After processing by the first gated temporal convolutional layer, the data enters the graph convolutional layer. This layer uses graph convolution operations in the spectral domain to process the data, and uses the Chebyshev polynomial of formula (7) to approximate the eigenvalues ​​of the graph Laplacian operator, thereby reducing computational cost.

[0079]

[0080] in This represents the graph convolution operation. It is a k-th order Chebyshev polynomial.

[0081] Spatial graph convolutional layers extract spatial features between grid nodes, using graph convolution to process non-Euclidean data, enabling the model to learn complex functional mappings between node features and their topological structures. This identifies and utilizes spatial relationships between grids, providing deeper spatial information for subsequent time series analysis.

[0082] The second temporally gated convolutional layer has the same structure as the first temporally gated convolutional layer. It is responsible for capturing dynamic changes in the time dimension, processing features of continuous time steps through convolution operations, and using a gating mechanism to enhance the model's selective learning ability for time series information. This enables the network to effectively process time series data and identify the importance of different time nodes.

[0083] A fully connected layer is used to integrate the features extracted and processed from the above layers, with ReLU used as the activation function to ensure that the output of each component has the same dimension and shape as the prediction target.

[0084] When training the AE-STGCN model, the number of channels in the gated temporal convolutional layer is 64, the number of channels in the graph convolutional layer is 16, the number of kernels in the graph convolution and the one-dimensional convolution in the temporal dimension is 3, the learning rate is set to 0.001, the number of training epochs is 100, and the batch size is 64. The specific parameter settings are shown in Table 1.

[0085] Table 1 Explanation of Hyperparameter Settings

[0086]

[0087]

[0088] Step 5: Use the trained model to predict future air quality and obtain the prediction results.

[0089] The newly collected and preprocessed data are used to generate a multidimensional network feature map sequence, which is then input into the AE-STGCN model. The prediction results show that the prediction accuracy of this invention is higher than that of other models, and it can better capture changes in air quality in the short term. It is suitable for application scenarios that require rapid response and decision-making.

Claims

1. An air quality prediction method based on spatiotemporal graph convolutional networks, characterized in that, The method includes the following steps: Step 1: Collect air quality data, grid meteorological data, and POI data for the area to be monitored; Step 2: Perform data preprocessing, including checking the consistency of data measurement units, the consistency of detection site information, and the continuity of time series; Step 3: Perform multi-source data fusion to generate a multi-dimensional grid feature map; Step 4: Construct the spatiotemporal graph convolutional network prediction model AE-STGCN with an attention mechanism, and train it by using the multidimensional grid feature map of historical data as the model input; Step 5: Use the trained model to predict future air quality and obtain the predicted results of changes in atmospheric pollutant concentrations; Step 4 is as follows: To leverage the temporal and spatial features inherent in multidimensional grid feature map sequences, an improved STGCN network model, AE-STGCN, is constructed. This model consists of an input layer, two stacked spatiotemporal convolutional blocks, a fully connected layer, and an output layer. Each spatiotemporal convolutional block contains three components: a spatiotemporal attention layer, a temporally gated convolutional layer, and a spatial graph convolutional layer. The input layer reads the multidimensional grid feature map sequence data generated in step 3. Each spatiotemporal convolutional block contains three components: a spatiotemporal attention layer, a temporally gated convolutional layer, and a spatial graph convolutional layer. The spatiotemporal attention layer is further divided into a temporal attention layer and a spatial attention layer. The spatial attention layer performs dynamic correlation analysis on nodes with different attributes, identifying the feature regions that have the greatest impact on the prediction results. This is described using an attention matrix S, which is calculated as follows: in, It is the first The input of a spatiotemporal convolutional block, It is the first The number of input data channels in the layer, when At that time, it can be obtained , Representing the The time dimension in the layer, when At that time, it can be obtained , , , , These are the parameters set for model learning, and the Sigmoid function is used as the activation function. The attention matrix S is calculated using formula (2), with elements... This represents the association strength between node i and node j. The Softmax function is used to ensure that the sum of the attention weights of the nodes is 1. During graph convolution operations, the adjacency matrix is ​​considered. Spatial attention matrix To dynamically adjust the influence weights between nodes. The spatial attention layer uses formula (3) to calculate the different weights of the detection data of each station at different time periods to describe their correlation, and uses formula (4) to calculate the temporal correlation matrix. Among the elements The degree of dependency between node i and node j is represented by the Softmax function on the matrix. Normalize, , , , These are the parameters, elements, set for model learning. To represent the dependency between node i and node j, the Softmax function is used to optimize the matrix. Normalization is performed, and the normalized time attention matrix is ​​applied to the model input. Information is merged using formula (5) to dynamically adjust the input. （5） The first gated temporal convolutional layer further processes the output of the spatiotemporal attention layer. This layer is used to extract temporal features and contains a causal convolution and a gated linear unit (GLU). The GLU input sequence is divided into two parts and operated on using formula (6). Different operations are performed on the left part, which will pass through the Sigmoid function, and the right part will be connected by a residual fusion. Then, the two parts will pass through a Hadamard product, and finally output the result. （6） P and Q are the two inputs of the GLU; It represents the Hadamah accumulation. It is the Sigmoid function, output after passing through the Hadamard product; After the first gated temporal convolutional layer, the data enters the graph convolutional layer. This layer uses graph convolution operations in the spectral domain to process the data, and uses the Chebyshev polynomial of formula (7) to approximate the eigenvalues ​​of the graph Laplacian operator, thereby reducing computational cost. ( 7 ) in* This represents the graph convolution operation. It is a k-th order Chebyshev polynomial. The second temporally gated convolutional layer has the same structure as the first one. It is responsible for capturing dynamic changes in the time dimension, processing features of continuous time steps through convolution operations, and using a gating mechanism to enhance the model's selective learning ability for time series information. This enables the network to effectively process time series data and identify the importance of different time nodes.

2. The air quality prediction method based on spatiotemporal graph convolutional networks according to claim 1, characterized in that, Step 1 is as follows: The collected air quality data covers six types of environmental pollutants monitored by the monitoring stations: PM2.5, PM10, SO2, NO2, O3, and CO; meteorological data includes temperature, humidity, precipitation, air pressure, wind speed, and wind direction; POI data includes POI categories for different monitoring locations such as factories, roads, and parks. All of the above data is referred to as multidimensional data.

3. The air quality prediction method based on spatiotemporal graph convolutional networks according to claim 1, characterized in that, Step 2 is as follows: The process involves checking all parameters in the multidimensional data, including temperature, humidity, and air pollutant concentration, using uniform units of measurement. If inconsistencies in units are found, appropriate conversions are required to ensure data consistency. The check also includes ensuring that information for each monitoring station, including station number, latitude and longitude, and monitoring period, remains consistent throughout the dataset, identifying and correcting any inconsistencies. Furthermore, the timestamps are checked to determine the data's temporal completeness and continuity. Linear interpolation is used to supplement missing data, duplicate data with the same timestamp are removed, and the area to be checked is divided into grid regions.

4. The air quality prediction method based on spatiotemporal graph convolutional networks according to claim 1, characterized in that, Step 3 is as follows: The various types of data processed in step 2 are used as single features. Each feature is converted into a grid layer, and all grid layers are superimposed layer by layer to obtain a higher-dimensional multi-dimensional grid feature map.

5. The air quality prediction method based on spatiotemporal graph convolutional networks according to claim 1, characterized in that, Step 5 is as follows: Input the newly collected and processed multidimensional network feature map to complete the air quality prediction.

6. An air quality prediction model based on spatiotemporal graph convolutional networks, characterized in that, This method is used to implement the air quality prediction method based on spatiotemporal graph convolutional networks as described in any one of claims 1-5. The prediction model includes an input layer, two stacked spatiotemporal convolutional blocks (ST-Conv Blocks), a fully connected layer, and an output layer. The input layer data is a sequence of multi-dimensional grid feature maps obtained from data preprocessing, which includes air quality data feature maps, meteorological data feature maps, and POI data feature maps; Each spatiotemporal convolutional block contains three components: a spatiotemporal attention layer, a temporally gated convolutional layer, and a spatial graph convolutional layer; The spatiotemporal attention layer is divided into a temporal attention layer and a spatial attention layer. The spatial attention layer emphasizes the most important spatial information in the data when making air quality predictions. It identifies the areas that have the greatest impact on the prediction results and learns the relative importance between each grid node to allocate attention weights. The model can adaptively adjust the representation of spatial features and prioritize spatial information that is more critical to the prediction target. The temporal attention layer identifies the importance of different time periods in the multidimensional network feature map sequence, highlights the time points that have the greatest impact on the prediction results, and calculates the attention weights for different time periods. This enables the network to capture long-term dependencies and highlight the time periods that have the greatest impact on the prediction results. Temporally gated convolutional layers are responsible for capturing dynamic changes in the time dimension. They process features of consecutive time steps through convolution operations and use gating mechanisms to enhance the model’s selective learning ability for time series information, enabling the network to effectively process time series data and thus identify the importance of different time nodes. Spatial graph convolutional layers extract spatial features between grid nodes and use graph convolution to process non-Euclidean data, enabling the model to learn complex function mappings between node features and their topological structures. The fully connected layer integrates the processed features extracted from each layer and uses ReLU as the activation function, so that the output of each component has the same dimension and shape as the prediction target, realizing the transformation from feature representation to final output; The output layer contains the model's prediction results, constructing an air quality data feature map for a future time series, including parameters of various pollutant indicators, including the Air Quality Index (AQI).

7. A terminal, characterized in that: The system includes a processor, a memory, and a program for an air quality prediction method based on a spatiotemporal graph convolutional network stored in the memory. When the processor runs the air quality prediction program based on the spatiotemporal graph convolutional network, it implements the steps of the air quality prediction method based on the spatiotemporal graph convolutional network as described in any one of claims 1-5.

8. A computer-readable storage medium storing computer instructions thereon, characterized in that: When executed by a processor, the computer instructions implement the steps of the air quality prediction method based on spatiotemporal graph convolutional networks as described in any one of claims 1-5.

Images