Graph Transform-based sand and dust aerosol concentration prediction method

Abstract

The invention relates to the field of sand and dust prediction, in particular to a sand and dust aerosol concentration prediction method based on a Transform graph, which comprises the following steps: collecting and preprocessing multi-source sand and dust related data, and constructing a feature set; a graph structure is constructed based on spatial distribution of monitoring sites, and a self-adaptive gating feature enhancement module and a global representation learning module for spatial relationship perception are developed for sand and dust characteristic differences of different regions and seasons; the method comprises the following steps: designing a graph Transform encoder and a time sequence enhancement decoder, and constructing a graph Transform; and designing a comprehensive loss function to optimize the model to obtain an optimal model for predicting the dust aerosol concentration. According to the invention, through the organic fusion of the graph structure modeling and the Transform architecture, the spatial dependence characteristic and the time sequence evolution characteristic of the sand and dust transmission can be captured at the same time, the accuracy and the stability of the sand and dust aerosol concentration prediction are improved, the calculation efficiency is optimized, an effective basis is provided for environment monitoring and air quality management, and the method is suitable for popularization and application. The method is of great significance in improving the sand and dust disaster early warning capability and reducing the sand and dust weather influence.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of sand-dust prediction, and particularly relates to a sand-dust aerosol concentration prediction method based on a graph Transformer. BACKGROUND

[0002] Sand-dust aerosol concentration prediction is a core component of air quality management systems, and has important guiding significance for formulating sand-dust prevention and reduction measures and environmental protection policies. However, the spatial and temporal distribution of sand-dust aerosols shows high complexity, mainly influenced by multiple factors such as source area surface characteristics, meteorological elements, human activities, and physical and chemical transformations in the atmospheric transport process of aerosols. Among them, spatial correlation and long-term time series dependence are two key characteristics of sand-dust aerosol distribution. Spatial correlation includes the mutual influence relationship of sand-dust concentration caused by adjacent regions and long-distance transport, mainly driven by atmospheric circulation, diffusion process and long-distance transport; long-term time series dependence reflects the characteristics that the current sand-dust concentration is influenced by the atmospheric state and sand-dust emission history in the past several days to several weeks. In addition, the irregular distribution of sand-dust aerosol monitoring sites increases the difficulty of accurate modeling.

[0003] The existing sand dust prediction methods can be mainly divided into three categories: physical models, statistical models, and deep learning models. Physical models such as WRF-Chem, CMAQ, and GEOS-Chem optimize parameter settings and boundary conditions to improve simulation accuracy by simulating atmospheric dynamics and chemical processes (Wang et al., 2001; Yoon et al., 2024; Kerimray et al., 2020). However, these methods have high computational cost, poor real-time performance, and are highly sensitive to initial conditions, making them difficult to adapt to complex and changing real-world environments. Statistical models such as ARIMA, SVR, and random forest predict future trends by mining historical data patterns (Chen et al., 2013; Wang et al., 2022; Huang et al., 2018). However, they are difficult to capture the nonlinear changes in sand dust aerosol concentration at different time scales and spatial scales, and cannot model complex sand dust transport processes. Deep learning models such as CNN-LSTM, SpatioTemporal-CNN, and attention networks improve prediction accuracy by learning complex nonlinear mapping relationships (Xue et al., 2023; Wang et al., 2020; Zhou et al., 2024). However, they still have significant shortcomings in handling irregularly distributed monitoring data, capturing complex spatial dependencies, and modeling long-term temporal changes (Know et al., 2020; Pu et al., 2020; Zhu et al., 2021; Dao et al., 2024), which severely restrict the practicality and reliability of sand dust warning systems.

[0004] There are four main problems in the field of sand dust concentration prediction: (1) Insufficient data representation and spatial modeling. Existing prediction models are mostly designed based on regular grid data structures, which assume that monitoring points are uniformly distributed and cannot effectively handle the irregular distribution characteristics of sand dust monitoring stations. Traditional CNN models need to interpolate irregularly distributed monitoring data to regular grids, resulting in loss and distortion of original spatial information. (2) Feature extraction problem. Existing models use fixed weights for different features, which cannot distinguish the importance of different environmental factors in different regions and time periods, resulting in loss of key information. (3) Temporal modeling problem. Sand dust transport has complex long-term dependencies, and existing models mostly use linear temporal processing methods, which cannot capture nonlinear temporal dependencies. (4) Computational efficiency problem. When dealing with large-scale spatio-temporal data, the computational complexity is high, making it difficult to meet real-time prediction requirements. These problems severely restrict the application effect of sand dust prediction systems, and new prediction methods need to be developed to overcome these challenges.

[0005] Current mainstream dust aerosol concentration prediction methods still have significant limitations in handling complex spatiotemporal data and adapting to environmental variability: (1) Insufficient data representation and spatial modeling: Existing prediction models are mostly designed based on regular grid data structures, making it difficult to effectively handle the irregular distribution characteristics of dust monitoring sites. Traditional CNN models require interpolation of irregularly distributed monitoring data to regular grids, resulting in loss and distortion of original spatial information. While basic graph models can represent irregular data, they have limited ability to capture complex spatial dependencies and cannot accurately depict the transport rules and mutual influence relationships of dust aerosol between different regions.

[0006] (2) Insufficient environmental factor expression and adaptive ability: Existing models generally use a fixed weight processing mechanism for input meteorological and surface characteristics factors, lacking the ability to dynamically adjust the importance of factors according to different regions and time periods. Dust aerosol concentration is influenced by multiple meteorological and environmental factors, and the influence weights of these factors differ significantly in different regions and seasons. Fixed weight models are difficult to adapt to such differences, leading to the weakening of key environmental influence factors or the overemphasis of secondary factors, which seriously affects prediction accuracy.

[0007] (3) Limited temporal dependency modeling capability: Dust aerosol transport is a complex process that spans multiple time scales, exhibiting significant long-term temporal dependency. Existing models such as LSTM and GRU can capture some temporal dependencies, but still face the problem of gradient vanishing when dealing with long sequence data. While standard Transformer has advantages in long sequence modeling, it lacks optimization design for dust aerosol temporal characteristics, and cannot fully capture the influence of atmospheric state and dust emission history over the past few days to weeks on current concentration, resulting in limited long-term prediction capability.

[0008] (4) Insufficient integration of global information and balance of local features: Existing models have a clear contradiction between integrating global atmospheric circulation information and preserving local diffusion characteristics. CNN-based models focus on local feature extraction while ignoring global dependencies. Attention-based models focus on global relationships but often ignore local spatiotemporal structures. Basic graph models lack sufficient consideration of monitoring site spatial location relationships during global information propagation, making it difficult for the model to effectively balance global and local information, affecting the accuracy of dust aerosol concentration prediction at different spatial scales.

[0009] Overall, the existing prediction methods have obvious deficiencies in irregular monitoring network data representation, dynamic meteorological and surface factor weight adjustment, complex time series dependence modeling, and global-local information balance, which restrict the precision and applicability of dust aerosol concentration prediction. Therefore, it is urgent to develop a dust prediction model that can solve the above problems at the same time, and improve the precision of dust concentration prediction, which has important significance for environmental monitoring, public health protection and ecological system management. SUMMARY

[0010] In order to solve the problems of the existing dust aerosol concentration prediction method, such as limited environmental factor expression ability, insufficient spatial correlation modeling, insufficient time series relationship capture and low long-term time series prediction accuracy, the present application provides a dust aerosol concentration prediction method based on graph Transformer, which significantly improves the accuracy, stability and adaptability of dust aerosol concentration prediction, and provides more reliable technical support for environmental monitoring, public health protection and ecological management. The method mainly includes: S1: Collecting multi-source dust related data and preprocessing, constructing a feature set, and dividing the feature set into a training set and a validation set; S2: Based on the spatial distribution of monitoring stations, a graph structure is constructed, and an adaptive gating feature enhancement module and a global representation learning module with spatial relationship perception are developed according to the differences in dust characteristics of different regions and seasons. The adaptive gating feature enhancement module is used to realize the dynamic adjustment mechanism of the importance of environmental factors, and the global representation learning module is used to enhance the capture ability of long-distance transmission mode of dust aerosol; S3: Designing a graph Transformer encoder based on S2; S4: Designing a time series enhancement decoder combined with the encoder output of S3 to obtain the final graph Transformer; S5: Training the final graph Transformer using the training set, designing a comprehensive loss function, and using the validation set to obtain the optimal graph Transformer for dust aerosol concentration prediction.

[0011] A computer readable storage medium stores a computer program, when the program is executed by a processor, the steps of the above method are realized.

[0012] The beneficial effects brought by the technical solutions provided by the present application are: the present application effectively processes the irregular distribution characteristics of sand monitoring stations through innovative graph structure representation and enhanced Transformer architecture, and realizes dynamic identification of the importance of different environmental factors through an adaptive gating mechanism, greatly improving the prediction accuracy and stability. Compared with other technologies, the present application comprehensively considers the multi-level challenges of sand prediction, improves the environmental factor expression ability through the adaptive gating mechanism, introduces spatial relationship perception to enhance spatial correlation modeling, and develops a time series enhanced attention mechanism to improve long-term prediction accuracy. Specifically as follows: (1) The spatial distribution characteristics of the sand monitoring network are represented by a graph structure, the monitoring stations are constructed as a node network in the graph model, the spatial connectivity is expressed through the edge relationship between nodes, the limitations of the traditional regular grid model are solved, and the spatial correlation of the sand concentration is effectively captured; (2) An adaptive gating mechanism is designed, the importance weights of different environmental factors are dynamically adjusted through learnable parameters, so that the model can automatically identify the key influencing factors in different regions and different periods, and the dominant factors affecting the change of the sand aerosol concentration, enhance the perception ability of the model to the key influencing factors, and improve the prediction accuracy; (3) A global vector updating mechanism with spatial relationship perception is introduced, the spatial relationship perception is introduced into global representation learning, the local diffusion and long-distance transmission characteristics are considered by using the graph topology, a spatial relationship matrix of the monitoring stations is constructed, so as to model the spatial dependence relationship, enhance the understanding ability of the model to the long-distance transmission mode of the sand, and the capture ability of the model to the spatial correlation of the sand aerosol and the atmospheric transport mode, and improve the prediction accuracy of the sand concentration change between regions; (4) A time series enhanced attention mechanism is developed, the history atmospheric state and sand emission information are accumulated in combination with the GRU recursive structure, the prediction ability of the model to the long-term time series dependence of the sand aerosol and the transmission process is improved. Combined with the state memory ability of GRU and the parallel processing advantage of Transformer, the long-term time series dependence relationship of the sand concentration is effectively captured, and the prediction ability of the model at different time scales is improved.

[0013] Through the above technical innovations, the present application significantly improves the accuracy and stability of the sand aerosol concentration prediction, and optimizes the calculation efficiency, providing an effective tool for environmental monitoring and air quality management, which has important significance for improving the sand disaster early warning ability and reducing the impact of sand weather. Accurate sand prediction can provide scientific decision-making basis for government departments, optimize emergency response measures, provide timely early warning information for the public to reduce health risks, and provide reliable data for researchers to deepen the understanding of the sand formation mechanism and environmental impact. BRIEF DESCRIPTION OF DRAWINGS

[0014] The application will be further described below in conjunction with the accompanying drawings and embodiments. In the drawings: Figure 1 is a flowchart of a sand dust aerosol concentration prediction method based on graph Transformer in an embodiment of the application. DETAILED DESCRIPTION

[0015] In order to have a clearer understanding of the technical features, objectives and effects of the application, the specific embodiments of the application will be described in detail below with reference to the accompanying drawings.

[0016] Embodiment 1 Please refer to Figure 1 , Figure 1 is a flowchart of a sand dust aerosol concentration prediction method based on graph Transformer in an embodiment of the application, specifically including: S1: data acquisition and preprocessing S1.1: Collect multi-source sand dust related data, including sand dust concentration data (PM10, PM2.5) of 367 monitoring stations nationwide provided by China Environmental Monitoring Center, ground meteorological observation data (temperature, humidity, wind speed, wind direction, air pressure, etc.) provided by China Meteorological Administration, NASA MODIS satellite remote sensing data (aerosol optical depth, land cover type, etc.) and ECMWF reanalysis data (boundary layer height, vertical velocity profile, etc.).

[0017] S1.2: Quality control and preprocessing of original data, including outlier detection (using 3σ rule and expert knowledge combination method), missing value filling (using time and space interpolation and multiple linear regression combination method), data standardization (using Z-score method to normalize different dimensional features) and time alignment (unifying different frequency data to hourly scale).

[0018] S1.3: Construct the feature set for sand dust prediction, including basic features (sand dust concentration historical value, meteorological elements), time features (hour, date, season index), spatial features (geographical location, elevation, land surface type) and derived features (wind power factor, dryness index, visibility, etc.).

[0019] S2: graph structure construction and feature enhancement S2.1: Construct the graph structure based on the spatial distribution of monitoring stations, the monitoring stations as nodes, the relationship between stations as links, and the node features including the historical records of sand dust concentration and environmental characteristics of the station. This structure constructs the monitoring network into a network in the sense of graph theory, and uses adaptive K-nearest neighbor method to determine the connection relationship of the graph, and the connection weight comprehensively considers the geographical distance, sand dust concentration historical correlation and dominant wind direction transmission relationship.

[0020] S2.2: Develop an adaptive gating feature enhancement module for the differences in dust characteristics of different regions and seasons. This module implements a dynamic adjustment mechanism for the importance of environmental factors, automatically identifying key influencing factors according to spatiotemporal characteristics. The specific implementation includes: (1) First, calculate the gating weight ; (2) Then calculate the candidate feature ; (3) Finally, update the feature through the gating mechanism , where represents the Hadamard product, element-level multiplication. represents the sigmoid activation function, represents the gating weight matrix, represents the input feature, represents the gating bias vector, represents the hyperbolic tangent activation function, represents the candidate feature bias vector, g represents the gating weight, and z represents the candidate feature.

[0021] S2.3: Develop a spatial relationship-aware global representation learning module to enhance the ability to capture long-distance transport patterns of dust aerosols. The implementation process of this module is as follows: (A): First, construct a spatial relationship matrix that considers both local diffusion and long-distance transport characteristics .

[0022] The specific construction method of the spatial relationship matrix is as follows: 1) Calculate the basic adjacency relationship based on geographical distance, and use an adaptive threshold method to determine the local diffusion connection; 2) Identify potential long-distance transport paths based on dominant wind direction and meteorological data, and construct long-distance connections between source areas and potential impact areas; 3) Dynamically adjust the weight coefficients of long-distance transport connections in the adjacency matrix for different seasons and weather conditions. Wherein the element in the spatial relationship matrix A represents the weighted relationship strength between site and site , which can reflect the local diffusion influence caused by geographical proximity and the long-distance transport relationship identified by atmospheric circulation patterns.

[0023] (B): Then calculate the normalized Laplacian matrix , and realize spatial relationship awareness through L·G, where is the degree matrix and G is the graph node feature; (C): Finally, update the global representation through the attention mechanism wherein, represents the attention mechanism, represents the graph node feature representation after spatial relationship perception processing as a query vector, represents the spatial graph feature is spliced with the input feature X, and simultaneously serves as a key vector and a value vector to enhance the feature fusion effect.

[0024] The adaptive gating feature enhancement module and the global representation learning module introducing spatial relationship perception in this step are both feature processing modules designed based on a graph structure and act on the node features of the graph. The input data features are first enhanced by the adaptive gating feature enhancement module, and then spatial perception is obtained by the global representation learning module. The outputs of the two modules will be used as inputs of the graph Transformer encoder, and the encoder output will be transmitted to the temporal enhancement decoder for decoding.

[0025] S3: Design a graph Transformer encoder S3.1: Design a multi-layer graph attention network to capture the spatial dependency between nodes through a message passing mechanism. Each layer of graph attention is calculated as follows: first, calculate the attention coefficient wherein a is a learnable attention vector, || represents feature splicing, T represents transposition, represents a linear activation function, represents a weight transformation matrix, represents the feature representation of node i, represents the feature representation of node j; then obtain the attention weight through softmax normalization; finally, update the node representation .

[0026] S3.2: Introduce a multi-head attention mechanism to enhance the environmental factor expression capability and model stability. The single-head attention is expanded into K independent attention heads, each head has its own parameter set, and finally the outputs of the K heads are spliced and linearly transformed to obtain the final representation wherein k represents the kth attention head, represents the attention weight of the kth attention head, represents the weight transformation matrix of the kth attention head, represents the neighbor node set of node i; in this embodiment, K is set to 8.

[0027] S3.3: Implement a skip connection and layer normalization to alleviate the gradient problem of deep network training. Add a residual connection after each layer of graph attention, and then apply layer normalization to ensure the stability of the feature distribution and accelerate the training convergence.

[0028] S4: Designing a timing enhancement decoder S4.1: Develop a timing-enhanced attention module that combines the memory capabilities of GRU and the parallel processing capabilities of Transformer. First, encode the historical sequence through GRU, Then calculate the timing gate weight Finally, update the timing features through the gating mechanism . Among them, represents the hidden state output of the GRU unit at time t, represents the gated recurrent unit, represents the input features at time t, represents the hidden state at time t-1, represents the timing gate weight matrix, represents the feature concatenation operation, represents the timing gate bias vector.

[0029] S4.2: Design a multi-scale timing attention mechanism to capture dependencies at different time scales. Implement two parallel attention heads, focusing on short-term and medium-term dependencies, then integrate the results of the two scales through adaptive weights.

[0030] S4.3: Adopt the axial attention mechanism, apply attention calculation along the time dimension T, spatial dimension H, and W in turn, to reduce computational complexity. The specific operation is: first apply one-dimensional attention along the time dimension cuboid_size(T,1,1), then apply one-dimensional attention along the spatial dimension H(1,H,1), and finally apply one-dimensional attention along the spatial dimension W(1,1,W).

[0031] S5: Model training and optimization scheme S5.1: Design a comprehensive loss function that takes into account both prediction accuracy and model generalization ability. The loss function consists of a prediction loss and a regularization loss: Where the prediction loss Combines the mean squared error, mean absolute error, and Loss, parameters , , Adjust dynamically according to the performance of the validation set; the regularization loss Includes L2 regularization and graph structure regularization.

[0032] S5.2: Adopt two-stage training strategy to improve model performance and training efficiency. In the first stage, pre-trained graph neural networks are used for node representation learning; in the second stage, the complete graph Transformer model is trained on this basis. Adam optimizer is used in each stage, with an initial learning rate of 0.001, and a cosine annealing strategy is applied to adjust the learning rate.

[0033] S5.3: Implement model ensemble technology to improve prediction stability and generalization performance. Train 5 models with different initialization parameters and random seeds, and obtain the final prediction result by weighted averaging (weights determined by validation set performance). The training batch size is set to 64, the training round is 200, and the early stopping strategy is applied (stop training if the validation set performance does not improve for 15 consecutive rounds).

[0034] Embodiment 2 A computer readable storage medium storing a computer program, when the program is executed by a processor, the steps of the above method are realized.

[0035] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for predicting dust aerosol concentration based on graph Transformer, characterized in that, include: S1: Collect multi-source dust-related data, preprocess them, construct a feature set, and divide the feature set into a training set and a validation set; S2: Based on the spatial distribution of monitoring stations, a graph structure is constructed. To address the differences in dust characteristics in different regions and seasons, an adaptive gating feature enhancement module and a global representation learning module that introduces spatial relationship awareness are developed. The adaptive gating feature enhancement module is used to realize the dynamic adjustment mechanism of the importance of environmental factors, and the global representation learning module is used to enhance the ability to capture the long-distance transmission mode of dust aerosols. S3: Design a Transformer encoder based on S2; S4: Combining the encoder output of S3, design a timing enhancement decoder to obtain the final graph Transformer; S5: Train the final graph Transformer using the training set, design a comprehensive loss function, and use the validation set to obtain the optimal graph Transformer for predicting dust aerosol concentration.

2. The method for predicting dust aerosol concentration based on graph Transformer as described in claim 1, characterized in that, In S1, the multi-source dust-related data includes: (1) dust concentration data: PM10, PM2.5; (2) ground meteorological observation data: temperature, humidity, wind speed, wind direction, air pressure; (3) NASA MODIS satellite remote sensing data: aerosol optical thickness, land cover type; (4) ECMWF reanalysis data: boundary layer height, vertical velocity profile; The preprocessing includes: outlier detection using a combination of the 3σ rule and expert knowledge; missing value imputation using a combination of spatiotemporal interpolation and multiple linear regression; and data standardization and time alignment using the Z-score method. The feature set includes basic features, temporal features, spatial features, and derived features. The basic features include historical dust concentration values ​​and meteorological elements. The temporal features include hour, date, and seasonal indicators. The spatial features include geographical location, elevation, and surface type. The derived features include wind force factor, dryness index, and visibility.

3. The method for predicting dust aerosol concentration based on graph Transformer as described in claim 1, characterized in that, In S2, the graph structure uses monitoring stations as nodes and the relationships between stations as connections. Node features include the station's historical dust concentration records and environmental characteristics. The adaptive K-nearest neighbor method is used to determine the graph's connection relationships, and the connection weights comprehensively consider geographical distance, historical correlation of dust concentration, and prevailing wind direction transmission.

4. The method for predicting dust aerosol concentration based on graph Transformer as described in claim 1, characterized in that, In S2, the implementation process of the adaptive gating feature enhancement module is as follows: (1) Calculate the gating weight ; (2) Calculate candidate features ; (3) Update features through gating mechanism ; in, Represents the Hadamard product. This represents the sigmoid activation function. Represents the gate weight matrix. Indicates input features, This represents the gating bias vector. This represents the hyperbolic tangent activation function. Let g represent the candidate feature bias vector, g represent the gating weight, and z represent the candidate feature.

5. The method for predicting dust aerosol concentration based on graph Transformer as described in claim 1, characterized in that, In S2, the implementation process of the global representation learning module is as follows: (A): First, construct a spatial relationship matrix that comprehensively considers both local diffusion and long-distance transmission characteristics. ; (B): Then calculate the normalized Laplace matrix. Spatial relationship perception is achieved through matrix multiplication L·G, where Let G be the degree matrix, and G be the graph node features; (C): Finally, the global representation is updated through an attention mechanism. ,in, This represents the attention mechanism. The graph node feature representation after spatial relationship perception processing is used as the query vector. Indicates spatial map features It performs feature concatenation with the input feature X, and serves as both a key vector and a value vector to enhance the feature fusion effect.

6. The method for predicting dust aerosol concentration based on graph Transformer as described in claim 1, characterized in that, In S2, the spatial relation matrix The specific construction method is as follows: 1) Based on geographic distance, the basic adjacency relationship is calculated, and the local diffusion connection is determined by an adaptive threshold method; 2) Identify potential long-distance transmission paths based on prevailing wind direction and meteorological data, and build long-distance connections between the source area and the potential affected area; 3) Dynamically adjust the weight coefficients of long-distance transmission connections in the adjacency matrix according to different seasons and weather conditions. ,in, The elements in the spatial relationship matrix A represent the stations. With the site The weighted relationship strength between them is used to reflect the local diffusion effects caused by geographical proximity and the long-distance transport relationships of atmospheric circulation pattern recognition.

7. The method for predicting dust aerosol concentration based on graph Transformer as described in claim 1, characterized in that, The design process of the Transformer encoder in S3 is as follows: S3.1: Design a multi-layer graph attention network to capture spatial dependencies between nodes through a message passing mechanism; the attention calculation for each layer is as follows: First, calculate the attention coefficient. ,in a Let T be a learnable attention vector, where T represents transpose and || represents feature concatenation. Represents a linear activation function. Represents the weight transformation matrix. The feature representation of node i. This represents the feature representation of node j; Then, the attention weights are obtained through softmax normalization. ; Finally, update the node representation of the graph structure. ; S3.2: Introducing a multi-head attention mechanism, the single-head attention is extended into K independent attention heads, each with its own parameter set. Finally, the outputs of the K attention heads are concatenated and subjected to a linear transformation to obtain the final representation. , k represents the k-th attention head, This represents the attention weight of the k-th attention head. This represents the weight transformation matrix of the k-th attention head. Represents the set of neighboring nodes of node i; S3.3: Add residual connections after each layer's graph attention. Then, the application layer normalizes. This is used to ensure the stability of the feature distribution and accelerate training convergence.

8. The method for predicting dust aerosol concentration based on graph Transformer as described in claim 1, characterized in that, The design process of the timing enhancement decoder in S4 is as follows: S4.1: Develop a temporally enhanced attention module that combines the memory capabilities of GRU with the parallel processing capabilities of Transformer; First, the historical sequences are encoded using GRU. ; Then calculate the timing gating weights. ; Finally, the timing features are updated using a gating mechanism. ; in, This represents the hidden state output of the GRU cell at time t. Indicates a gated loop unit. This represents the input features at time t. This represents the hidden state at time t-1. Represents the time-gated weight matrix. This indicates a feature concatenation operation. Represents the timing gate bias vector; S4.2: Design a multi-scale temporal attention mechanism to capture dependencies at different time scales simultaneously, implement two parallel attention heads to focus on short-term and medium-term dependencies respectively, and then integrate the results of the two scales through adaptive weights; S4.3: An axial attention mechanism is adopted, and attention calculation is applied sequentially along the time dimension T, the spatial dimension H and W. Specifically, one-dimensional attention cuboid_size(T,1,1) is applied along the time dimension first, then one-dimensional attention (1,H,1) is applied along the spatial dimension H, and finally one-dimensional attention (1,1,W) is applied along the spatial dimension W.

9. The method for predicting dust aerosol concentration based on graph Transformer as described in claim 1, characterized in that, The training process for S5 is as follows: S5.1: The designed loss function consists of prediction loss and regularization loss: Among them, predicted loss , Mean square error, mean absolute error and Loss, parameters , , Dynamically adjust based on validation set performance; regularization loss. Including L2 regularization and graph structure regularization; S5.2: A two-stage training strategy is adopted to improve the performance and training efficiency of the graph Transformer. In the first stage, a pre-trained graph neural network is used to learn the representation of the nodes. In the second stage, the complete graph Transformer is trained on this basis. The Adam optimizer is used in each stage, and the cosine annealing strategy is applied to adjust the learning rate. S5.3: Use ensemble techniques to improve the prediction stability and generalization performance of graph Transformers. Train multiple models with different initialization parameters and random number seeds, and obtain the final prediction result by weighted averaging.

10. A computer-readable storage medium, characterized in that, The device contains a computer program that, when executed by a processor, implements the steps of the graph Transformer-based method for predicting dust aerosol concentration as described in any one of claims 1 to 9.

Images