GAI-BRI-based air pollutant concentration interpolation prediction method, device, and medium

Abstract

A graph attention network-bidirectional recurrent interpolation network (GAI-BRI)-based air pollutant concentration interpolation prediction method, a device, and a medium. The method comprises: on the basis of the extracted air pollutant concentrations, meteorological data and satellite data from multiple distributed sites within a given time range and region, constructing an air pollutant concentration map for describing pollutant conditions within the given time range and region; and inputting numerical values in the air pollutant concentration map into a GAT-BRI prediction model to obtain an air pollutant concentration interpolation prediction result for a future time period, wherein the GAT-BRI prediction model comprises a GAT that is arranged at the bottom layer and used for extracting spatial features from the air pollutant concentration map, a BRI that is arranged at the middle layer and used for extracting temporal features from the spatial features outputted by the GAT, and a feedforward network that is arranged at the top layer and used for outputting an interpolation prediction result. The method fully exploits the spatial and temporal dependencies among multiple cities, achieving high prediction accuracy.

Description

Methods, equipment, and media for predicting air pollutant concentrations based on GAI-BRI Technical Field

[0001] This invention relates to the field of air pollutant concentration prediction, and in particular to a method, device and medium for predicting air pollutant concentration by interpolation based on GAI-BRI. Background Technology

[0002] In recent years, air pollution has had multifaceted impacts on the environment and human health, drawing widespread attention from all sectors of society. Predicting air pollution concentrations in advance and promptly disseminating this information to the public is crucial. Extending the prediction window while maintaining accuracy allows for earlier warnings and preventative measures, effectively reducing the adverse effects of air pollution on people and the environment. Utilizing big data thinking combined with deep learning technology for effective data analysis of air pollution to make accurate predictions is a cutting-edge research topic in the interdisciplinary field of environmental science and computer science.

[0003] Air pollutant concentration interpolation is a method that predicts, estimates, or fills in missing air pollutant concentrations during periods of unmonitored data using existing air pollutant monitoring data. The most common method is linear interpolation. With further research, machine learning methods have been applied to interpolation. Compared to traditional statistical methods, machine learning methods can adaptively handle nonlinear and high-dimensional data, and can better capture correlations between data points. With the widespread application of deep learning methods in prediction tasks, the prediction performance of air pollutant concentrations has significantly improved. Therefore, using prediction methods to interpolate missing values ​​in real datasets has attracted widespread attention from researchers. Compared to traditional GAI-BRI-based air pollutant concentration interpolation prediction methods, deep learning methods have stronger nonlinear modeling capabilities.

[0004] While traditional interpolation methods for urban air pollution concentrations can achieve certain interpolation results, they all face several problems: they ignore the temporal changes in air pollutant concentration data; if there is a large area of ​​missing data, these interpolation methods will fill in a large number of identical values, leading to significant data bias and affecting prediction results; they rely on numerical models to learn the statistical characteristics of the data and lack the ability to jointly extract the spatiotemporal correlation of air pollutant concentrations; and they only focus on one aspect of the complex spatiotemporal correlation of air pollutant concentrations, lacking research on combining it with models suitable for extracting spatial correlations of data. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method, device, and medium for predicting air pollutant concentrations based on GAI-BRI with higher accuracy.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] According to a first aspect of the present invention, a method for interpolating and predicting air pollutant concentrations based on GAI-BRI is provided, comprising:

[0008] Based on the extracted air pollutant concentrations, meteorological data, and satellite data from multiple distributed stations within a given time range and area, an air pollutant concentration map describing the pollutant situation within the given time range and area is constructed.

[0009] The air pollutant concentration map values ​​are input into the GAT-BRI prediction model to obtain the interpolation prediction results of air pollutant concentrations for a future period of time. The GAT-BRI prediction model includes a graph attention network (GAT) at the bottom layer for extracting spatial features from the air pollutant concentration map, a bidirectional cyclic interpolation network (BRI) at the middle layer for extracting temporal features from the spatial features output by the graph attention network, and a feedforward network at the top layer for outputting the interpolation prediction results.

[0010] Preferably, the air pollutant concentrations include PM2.5, PM10, SO2, NO2, O3, and CO; the meteorological data include temperature, humidity, wind speed, and air pressure; and the satellite data include boundary layer height, shortwave radiation, sunshine duration, specific humidity, and sensible heat flux.

[0011] Preferably, the construction of a spatiotemporal map describing the pollutant situation within a given time range and a given area, specifically including:

[0012] The adjacency matrix is ​​initialized using the distance between regions as the basis for the existence of edges between nodes;

[0013] We learn the correlation weights between cities using the attention mechanism in graph attention networks;

[0014] Based on the data collected at each moment from all distributed stations in the region, a spatiotemporal map describing the pollutant situation within a given area over a given time range is constructed.

[0015] Preferably, the graph attention network set at the bottom layer for extracting spatial features from the air pollutant concentration map updates the node's own features by assigning different attention coefficients to neighboring nodes, thereby extracting the nonlinear spatial dependence of pollutant propagation, specifically including:

[0016] The feature vectors from the air pollutant concentration map are input into a graph attention network.

[0017] The feature vectors of the target node and each of its neighboring nodes are concatenated and then input into a feedforward network to obtain the attention weights between nodes.

[0018] The aggregated feature vector is obtained by weighted summing of the feature vectors of all neighboring nodes of the target node with their corresponding attention weights.

[0019] The aggregated feature vectors are subjected to nonlinear transformation to obtain the spatial characteristics of air pollutant concentrations.

[0020] Preferably, the graph attention network uses attention scores between nodes to characterize the influence of air pollutant concentrations between regions. The specific network training process includes:

[0021] The similarity coefficient e between the target node i and its neighboring node j is calculated using a self-attention mechanism. ij Among them, the self-attention mechanism introduces the LeakyReLU activation function;

[0022] The similarity coefficient e between the target city node i and its neighboring city node j ij After normalization, the attention scores a of the target city node i and its neighboring city nodes j are obtained. ij , ;

[0023] Aggregate the information of all neighboring nodes of target node i to obtain the node features h of target node i. i :

[0024] The feature vector is obtained by summing the feature vectors of all neighboring nodes of the target node i with their corresponding attention weights.

[0025] Preferably, the graph attention network employs a multi-head attention mechanism, which captures different features through different graph attention layers, and then connects multiple captured features to obtain the multi-head attention calculation result.

[0026] Preferably, the spatial features of the air pollutant concentration map extracted by the graph attention network (GAT) are tensor transposed to obtain the first data. The spatial features of the air pollutant concentration map extracted by the GAT are then concatenated with the original data to obtain the second data. The first and second data are then jointly input into a bidirectional cyclic interpolation network. The time dependence of air pollutant propagation and diffusion is extracted through a gating mechanism to obtain the time features of air pollutant concentration.

[0027] Preferably, the extraction of the temporal characteristics of air pollutant concentration using a bidirectional cyclic interpolation network includes two stages of extraction, specifically:

[0028] The first stage of extraction involves determining the input data X at the current moment. t If there are no missing values, then input the current time value X. t The previous time step's estimate is used as the filler value input to the bidirectional cyclic interpolation network; otherwise, it is used as the filler value. The filler values ​​are input into the bidirectional cyclic interpolation network to obtain the complete first estimate based on historical observations. and fill value τ is the length of the time window;

[0029] The second stage of extraction uses the fill values ​​obtained in the first stage. As input, a second estimate based on other features is calculated by setting the diagonal of the parameter matrix to 0.

[0030] The first and second estimates are weighted and summed to obtain the combined vector.

[0031] Data X at the next moment t+1 If it does not exist, use a combined vector. Fill in the next moment's data X t+1 The value of is used to obtain the first vector.

[0032] The first vector The input is fed into a bidirectional cyclic interpolation network to extract time information in two directions, resulting in the hidden vector {h}. t h t+1 h t+2 , ..., h t+τ};

[0033] Based on the hidden vector {h t h t+1 ,h t+2 , ..., h t+τ This is used to subsequently obtain the predicted air pollutant concentration values ​​for the target city within a given region through a feedforward network.

[0034] According to a second aspect of the present invention, an electronic device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to implement any of the methods described above.

[0035] According to a third aspect of the invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements any of the methods described above.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] (1) This invention makes predictions based on air pollutant concentration data and meteorological data from multiple cities, and constructs a GAT-BRI prediction model. By using a graph neural network at the bottom layer to extract the spatial dependencies of the graph structure composed of multiple cities, the attention mechanism can be fully utilized to obtain the influence between multiple cities and mine the spatial dependencies between multiple cities. Then, a bidirectional cyclic interpolation network is used to fully extract the spatial features obtained by the graph neural network in both directions to mine the spatiotemporal information. Finally, the prediction results are output through a feedforward network, which can realize the fine assignment of real air pollutant concentration data.

[0038] (2) The bidirectional cyclic interpolation network uses a bidirectional long short-time memory network, which can describe the degree of change of the input latent vector. When the input data changes drastically, the memory unit can be reset in time, which can fully mine the spatiotemporal information in the air pollutant concentration data in both directions and improve the accuracy of pollutant concentration prediction. Attached Figure Description

[0039] Figure 1 is a flowchart of the steps of the present invention;

[0040] Figure 2 is a flowchart of an embodiment of the present invention;

[0041] Figure 3 shows the air pollutant concentration interpolation model proposed in this invention. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0043] Example

[0044] First, the interpolation of air pollutant concentrations is defined:

[0045] Air pollutant concentration interpolation: This method mainly uses historical pollutant and meteorological information to predict, estimate, or fill in the concentration values ​​of air pollutants such as PM2.5 and PM10 within a certain period of time in the future. It is one of the key research topics in environmental science, meteorological science, computer science, etc., and therefore has a certain degree of interdisciplinary nature.

[0046] Traditional GAI-BRI-based air pollutant concentration interpolation prediction methods: Non-deep learning-based GAI-BRI-based air pollutant concentration interpolation prediction methods mostly include traditional air pollutant concentration interpolation models, such as linear interpolation models, moving average data filling models, previous value interpolation models, and traditional machine learning models based on small sample data. These are all traditional prediction models.

[0047] As shown in Figures 1 to 3, this embodiment provides an air pollutant concentration interpolation prediction method based on GAI-BRI, including:

[0048] Based on the extracted air pollutant concentrations, meteorological data, and satellite data from multiple distributed stations within a given time range and area, an air pollutant concentration map describing the pollutant situation within the given time range and area is constructed.

[0049] The air pollutant concentration map values ​​are input into the GAT-BRI prediction model to obtain the interpolation prediction results of air pollutant concentration for a future period of time. The GAT-BRI prediction model includes a graph attention network (GAT) at the bottom layer to extract spatial features from the air pollutant concentration map, a bidirectional cyclic interpolation network (BRI) at the middle layer to extract temporal features from the spatial features output by the graph attention network, and a feedforward network at the top layer to output the prediction results.

[0050] The method of this embodiment will now be described in detail.

[0051] 1. Data Acquisition

[0052] In this embodiment, the following data were selected from all detection stations in the prediction area:

[0053] The air pollutant concentrations include PM2.5, PM10, SO2, NO2, O3 and CO. The data comes from the National Urban Air Quality Real-time Release Platform of the China National Environmental Monitoring Centre. The air pollutant data collected by the ground monitoring stations are at 1-hour intervals.

[0054] The meteorological data includes meteorological factors such as temperature, humidity, wind speed, and air pressure, and the data comes from the U.S. National Climate Data Center (NCDC).

[0055] The satellite data includes boundary layer height, shortwave radiation, sunshine duration, specific humidity, and sensible heat flux. The data was obtained through an influence experiment and comes from the European Centre for Medium-Range Weather Forecasts.

[0056] Specifically, taking PM2.5 as an example, the linear and nonlinear correlations between PM2.5 and satellite characteristics are extracted from the dataset. The correlation is first calculated using the Pearson correlation coefficient:

[0057] Where: X represents PM2.5 data, Y represents the mean of PM2.5 data, and Y represents different satellite data. The mean of the satellite data is represented by t, where t represents the observed values ​​of samples X and Y at time t, and n represents the number of samples.

[0058] Then, the correlation was calculated using the Spearman rank correlation coefficient:

[0059] Where: d i For X i and Y i The difference in grade;

[0060] The above 15 features were selected for predicting PM2.5 concentration values. The data was then normalized using the Z-score standardization method, and the calculation expression is as follows:

[0061] in: The data represents the normalized data, μ is the mean of the original data, δ is the standard deviation of the original data, and x represents the normalized data. t This represents the normalized numbers, which follow a standard normal distribution with a mean of 0 and a variance of 1.

[0062] 2. GAT-BRI Prediction Model

[0063] A GAT-BRI prediction model is constructed based on the extracted data to interpolate missing values ​​in the real air pollutant concentration dataset.

[0064] The GAT-BRI prediction model includes a graph attention network (GAT) at the bottom layer to extract spatial features from the air pollutant concentration map, a bidirectional cyclic interpolation network (BRI) at the middle layer to extract temporal features from the spatial features output by the graph attention network, and a feedforward network at the top layer to output the prediction results.

[0065] 2.1 Graph Attention Network (GAT)

[0066] As shown in Figure 3, the input to the Graph Attention Network (GAT) is the feature vector obtained from modeling air pollutant concentration map data. The feature vectors of the target node and each of its neighboring nodes are concatenated and passed through a feedforward network to calculate the attention weights between nodes. Then, the feature vectors of all the target node's neighboring nodes are weighted and summed with their corresponding attention weights to obtain an aggregated feature vector. Finally, the aggregated feature vector undergoes a nonlinear transformation to obtain a new feature vector. In this way, GAT can extract spatial information from the structural data of air pollutant concentration maps and can also adaptively calculate the attention weights between each node and its neighboring nodes.

[0067] The output of the Graph Attention Network (GAT) contains an implicit graph representation with neighbor node information. This representation, along with the original input data, is fed into the BRI (Browser Injection) part through operations such as tensor transposition and splicing. A gating mechanism is then used to extract the time dependency of pollutant propagation and diffusion. Specifically, the spatial features of the air pollutant concentration map extracted by the GAT are tensor transposed to obtain the first data. The spatial features of the air pollutant concentration map extracted by the GAT are then spliced ​​with the original data to obtain the second data. The first and second data are then fed into the BRI part. The original input data consists of air pollutant concentrations, meteorological data, and satellite data from multiple distributed stations within a given time range and region.

[0068] The Graph Attention Network (GAT) is responsible for learning the representation information of nodes and extracting the spatial dependence of pollutant concentrations in the graph;

[0069] The training of GAT begins with calculating the attention coefficient. For the target city i, the similarity coefficient between i and each of the surrounding cities is calculated one by one.

[0070] Where: attention represents the calculation method of self-attention, vector h is the feature vector of the node, and its subscripts i and j represent different nodes respectively.

[0071] The softmax function is used to normalize node i and its neighboring node j. The attention score reflects the strength of the relationship between the two nodes, that is, the magnitude of the influence of air pollutant concentration between cities.

[0072] By introducing the non-linear activation function LeakyReLU into the attention mechanism, we arrive at the final calculation method for the attention mechanism. Specifically, in GAT, the feature vectors of two nodes are concatenated together, and then the inner product is performed with the weight vector.

[0073] Where: || represents a concatenation operation. This is the weight vector.

[0074] After obtaining the normalized attention coefficients between different nodes, the information of adjacent nodes is aggregated, and finally the node features of each node after GAT calculation are obtained:

[0075] The aggregated feature vector is obtained by summing the feature vectors of all neighboring nodes of node i with their corresponding attention weights.

[0076] The final result of multi-head attention is calculated using the following formula:

[0077] Where: K represents the number of attention calculations required, k represents the calculation of the k-th attention, and the k-th attention is a. k The feature linear transformation weights under the k-th attention mechanism are W. k .

[0078] 2.2 Bidirectional Cyclic Interpolation Network

[0079] Traditional recurrent neural networks: h t =σ(W h h t-1 +U h X t +b h ), where σ is the sigmoid function, W h U h and b h It's a network parameter, h t It is the hidden state of the current time step.

[0080] In this embodiment, a bidirectional cyclic interpolation network, specifically Bi-LSTM, is used to extract the temporal features of air pollutant concentrations. The extraction process involves two stages:

[0081] The first stage of extraction involves determining the input data X at the current moment. t If there are no missing values, then input the current time value X. t The previous time step's estimate is used as the filler value input to the bidirectional cyclic interpolation network; otherwise, it is used as the filler value. The filler values ​​are input into the bidirectional cyclic interpolation network to obtain the complete first estimate based on historical observations. and fill value τ is the length of the time window;

[0082] The second stage of extraction uses the fill values ​​obtained in the first stage. As input, a second estimate based on other features is calculated by setting the diagonal of the parameter matrix to 0. in, W z Let b be the parameter matrix. z This is the bias value;

[0083] The first and second estimates are weighted and summed to obtain the combined vector. The specific calculation expression is as follows:

[0084] Where: β t ∈[0,1] D As a combination and The weight, It can be either an observed value or a predicted value;

[0085] Data X at the next moment t+1 If it does not exist, use a combined vector. Fill in the next moment's data X t+1 The value of is used to obtain the first vector.

[0086] The first vector The input is fed into a bidirectional cyclic interpolation network to extract time information in two directions, resulting in the hidden vector {h}. t h t+1 h t+2 , ..., h t+τ The calculation expression is:

[0087] in: As the final fill value, it can be used to fill X. t+1 The value of h t It is the hidden state at time t;

[0088] Based on the hidden vector {h t h t+1, h t+2 , ..., h t+τ This is used to subsequently obtain the predicted air pollutant concentration values ​​for the target city within a given region through a feedforward network.

[0089] 2.3 Model Training

[0090] The model training process specifically includes:

[0091] 1) Input the air pollutant concentration data after feature extraction via the graph neural network GAT:

[0092] X = {X1, X2, ..., X} T};

[0093] 2) The data dimension is transformed into a time series distributed along the time dimension and merged with the original data as input to a bidirectional recurrent interpolation network. A bidirectional long short-time neural network (Bi-LSTM) is then used to perform stage prediction on the input data. Specifically, the input data is... Treating the input sequence as a variable, applying the above formula in reverse using Bi-LSTM yields a forward estimated sequence. and the reverse estimated sequence And the positive loss sequence {l l l2, ..., l t} and the reverse loss sequence

[0094] 3) Introduce consistency loss to force the predictions at each stage to be consistent in both directions:

[0095] 4) The training objective is to minimize the reconstruction error across all stages in both directions. The loss function is as follows:

[0096] in, This represents the reconstruction error in two directions.

[0097] The Graph Attention Network (GAT) is set to 2 layers, and the feature vectors of the nodes after processing in the first and second layers are both set to 10; the Bidirectional Long Short-Term Memory Network (Bi-LSTM) is set to 2 layers, and the number of hidden neurons is set to 64; the fully connected layer is set to 2 layers, the learning rate is set to between 0.001 and 0.1, and the number of iterations is 200.

[0098] The loss function during the model training phase is defined as the root mean square error function. The backpropagation algorithm is used to minimize the value of the model loss function and reduce the error between the predicted value and the true value.

[0099] The root mean square error (RMSE) is calculated as follows:

[0100] After training, the model's hyperparameters are adjusted using data from the validation set until the model is optimal. The test set is then input into the model to obtain prediction results.

[0101] Using mean squared error (SE), mean absolute error (MAE), and coefficient of determination (R²) 2 To measure model error and prediction performance, the calculation expressions are as follows:

[0102] Where: Y t Represents the true value, y t Y represents the predicted value. tThis represents the average of the true values, and n is the test sample size.

[0103] If the above four indicators are better than traditional methods, it means that the model meets expectations and the performance of the proposed model can be evaluated from multiple perspectives.

[0104] The validated spatiotemporal joint interpolation model is used to refine the values ​​of real air pollutant concentration data.

[0105] In summary, the prediction model based on graph attention network-bidirectional cyclic interpolation network designed in this invention can accurately and effectively predict the PM2.5 concentration in the air within a certain period of time in the future.

[0106] This invention addresses the shortcomings of existing air pollutant concentration interpolation prediction methods based on GAI-BRI by fully utilizing various spatiotemporal big data and existing research results. It constructs a spatiotemporal prediction model of GAT-BRI using graph attention networks (GAT) and an improved long short-term memory network (Bi-LSTM).

[0107] The spatiotemporal graph sequence is constructed by considering the spatial distance and feature similarity between graph nodes, which can both eliminate some interfering neighbor nodes and retain effective neighbors.

[0108] Using a graph attention network (GAT) as the bottom layer, different attention coefficients can be assigned to neighboring nodes to update the node's own features, thereby extracting the nonlinear spatial dependence of pollutant propagation; the output of the GAT network is used as the input of the high-level BiLSTM network to extract the time series features of air pollutant data.

[0109] The improved BiLSTM cell can describe the degree of change in the input latent vector and can reset the memory cell in time when the input data changes drastically.

[0110] In terms of the time series prediction model, a bidirectional recurrent interpolation network was used as a feature extractor, which fully leveraged the time series feature acquisition capability of the recurrent neural network to fully extract the information obtained by the graph attention network GAT in both directions and mine spatiotemporal information.

[0111] Experimental results show that, compared with existing methods, the present invention obtains more accurate prediction results and can assign refined values ​​to real air pollutant concentration data, thus having practical application value and prospects.

[0112] The electronic device of this invention includes a central processing unit (CPU), which can perform various appropriate actions and processes according to computer program instructions stored in read-only memory (ROM) or loaded from a storage unit into random access memory (RAM). The RAM may also store various programs and data required for device operation. The CPU, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0113] Multiple components in the device are connected to the I / O interface, including: input units such as keyboards and mice; output units such as various types of displays and speakers; storage units such as disks and optical discs; and communication units such as network interface cards (NICs), modems, and wireless transceivers. The communication unit allows the device to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0114] The processing unit performs the various methods and processes described above. For example, in some embodiments, the methods may be implemented as computer software programs tangibly contained in a machine-readable medium, such as a storage unit. In some embodiments, part or all of the computer program may be loaded and / or installed on the device via ROM and / or a communication unit. When the computer program is loaded into RAM and executed by the CPU, one or more steps of the methods described above may be performed. Alternatively, in other embodiments, the CPU may be configured to execute the methods by any other suitable means (e.g., by means of firmware).

[0115] The functions described above in this document can be performed at least in part by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: field programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload programmable logic devices (CPLDs), and so on.

[0116] The program code used to implement the methods of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0117] In the context of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0118] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention 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 the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for predicting air pollutant concentrations based on GAI-BRI, characterized in that, include: Based on the extracted air pollutant concentrations, meteorological data, and satellite data from multiple distributed stations within a given time range and area, an air pollutant concentration map describing the pollutant situation within the given time range and area is constructed. The air pollutant concentration map values ​​are input into the GAT-BRI prediction model to obtain the interpolation prediction results of air pollutant concentrations for a future period of time. The GAT-BRI prediction model includes a graph attention network (GAT) at the bottom layer for extracting spatial features from the air pollutant concentration map, a bidirectional cyclic interpolation network (BRI) at the middle layer for extracting temporal features from the spatial features output by the graph attention network, and a feedforward network at the top layer for outputting the interpolation prediction results.

2. The air pollutant concentration interpolation prediction method based on GAI-BRI according to claim 1, characterized in that, The air pollutant concentrations mentioned include PM2.5, PM10, SO2, NO2, O3, and CO; the meteorological data includes meteorological factors such as temperature, humidity, wind speed, and air pressure; and the satellite data includes boundary layer height, shortwave radiation, sunshine duration, specific humidity, and sensible heat flux.

3. The air pollutant concentration interpolation prediction method based on GAI-BRI according to claim 1, characterized in that, The construction of a spatiotemporal map describing the pollutant situation within a given time range and a given area, and an air pollutant concentration map describing the pollutant situation within a given time range and a given area, specifically includes: The adjacency matrix is ​​initialized using the distance between regions as the basis for the existence of edges between nodes; We learn the correlation weights between cities using the attention mechanism in graph attention networks; Based on the data collected at each moment from all distributed stations in the region, a spatiotemporal map describing the pollutant situation within a given area over a given time range is constructed.

4. The air pollutant concentration interpolation prediction method based on GAI-BRI according to claim 1, characterized in that, The graph attention network, set at the bottom layer for extracting spatial features from air pollutant concentration maps, updates node features by assigning different attention coefficients to neighboring nodes, thereby extracting the nonlinear spatial dependence of pollutant propagation. Specifically, it includes: The feature vectors from the air pollutant concentration map are input into a graph attention network. The feature vectors of the target node and each of its neighboring nodes are concatenated and then input into a feedforward network to obtain the attention weights between nodes. The aggregated feature vector is obtained by weighted summing of the feature vectors of all neighboring nodes of the target node with their corresponding attention weights. The aggregated feature vectors are subjected to nonlinear transformation to obtain the spatial characteristics of air pollutant concentrations.

5. The air pollutant concentration interpolation prediction method based on GAI-BRI according to claim 4, characterized in that, The graph attention network uses attention scores between nodes to represent the magnitude of the influence of air pollutant concentrations between regions. The specific network training process includes: The similarity coefficient e between the target node i and its neighboring node j is calculated using a self-attention mechanism. ij Among them, the self-attention mechanism introduces the LeakyReLU activation function; The similarity coefficient e between the target city node i and its neighboring city node j ij After normalization, the attention scores a of the target city node i and its neighboring city nodes j are obtained. ij , ; Aggregate the information of all neighboring nodes of target node i to obtain the node features h of target node i. i : The feature vector is obtained by summing the feature vectors of all neighboring nodes of the target node i with their corresponding attention weights.

6. The air pollutant concentration interpolation prediction method based on GAI-BRI according to claim 5, characterized in that, The graph attention network employs a multi-head attention mechanism, capturing different features through different graph attention layers, and then connecting multiple captured features to obtain the multi-head attention calculation result.

7. The air pollutant concentration interpolation prediction method based on GAI-BRI according to claim 1, characterized in that, The bidirectional cyclic interpolation network, set in the middle layer for extracting temporal features from the spatial features output by the graph attention network, specifically includes: The spatial features of the air pollutant concentration map extracted by the graph attention network (GAT) are tensor transposed to obtain the first data. The spatial features of the air pollutant concentration map extracted by the graph attention network (GAT) are concatenated with the original data to obtain the second data. The first and second data are jointly input into a bidirectional cyclic interpolation network. The time dependence of air pollutant propagation and diffusion is extracted through a gating mechanism to obtain the time features of air pollutant concentration.

8. The air pollutant concentration interpolation prediction method based on GAI-BRI according to claim 7, characterized in that, The extraction of temporal features of air pollutant concentrations using a bidirectional cyclic interpolation network involves two stages: The first stage of extraction involves determining the input data X at the current moment. t If there are no missing values, then input the current time value X. t The previous time step's estimate is used as the filler value input to the bidirectional cyclic interpolation network; otherwise, it is used as the filler value. The filler values ​​are input into the bidirectional cyclic interpolation network to obtain the complete first estimate based on historical observations. and fill value τ is the length of the time window; The second stage of extraction uses the fill values ​​obtained in the first stage. As input, a second estimate based on other features is calculated by setting the diagonal of the parameter matrix to 0. The first and second estimates are weighted and summed to obtain the combined vector. Data X at the next moment t+1 If it does not exist, use a combined vector. Fill in the next moment's data X t+1 The value of is used to obtain the first vector. The first vector The input is fed into a bidirectional cyclic interpolation network to extract time information in two directions, resulting in the hidden vector {h}. t h t+1 h t+2 , ..., h t+τ }; Based on the hidden vector {h t h t+1 ,h t+2 , ..., h t+τ This is used to subsequently obtain the predicted air pollutant concentration values ​​for the target city within a given region through a feedforward network.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 8.

Images