Air quality prediction method based on multi-scale block sensing spatio-temporal dynamic graph network

Abstract

The invention relates to the technical field of air quality prediction, and particularly provides an air quality prediction method based on a multi-scale block sensing spatio-temporal dynamic graph network. The method comprises the following steps: by analyzing air quality data of a plurality of air quality monitoring stations at a plurality of time steps and combining a spatial dependency relationship among the stations, defining task indexes of air quality prediction; according to the method, the space-time dynamic graph network MSPA-STDGN for multi-scale block sensing is constructed according to the defined task indexes of air quality prediction so as to carry out air quality prediction, and through the constructed space-time dynamic graph network for multi-scale block sensing, the accurate prediction of the air quality is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of air quality prediction, and in particular to an air quality prediction method based on a multi-scale block-perceived spatiotemporal dynamic graph network. Background Art

[0002] Domestic and foreign scholars mainly use classical statistical methods, traditional machine learning methods, and deep learning methods in air quality prediction. Statistical methods are based on the construction of statistical models for regression analysis or time series prediction, such as the Autoregressive Integrated Moving Average model (ARIMA) and the XGBoost-GARCH hybrid model. They have the advantage of strong interpretability but rely on linear assumptions. Traditional machine learning methods use nonlinear modeling capabilities and comprehensive analysis of multiple factors to achieve predictions. Commonly used traditional machine learning methods include support vector machines, random forests, and artificial neural networks. For example, the Support Vector Regression (SVR) model was applied to study the air quality in Spain, combining differential evolution and random forests to predict pollutants. This type of method is computationally efficient but has difficulty handling long-term dependencies.

[0003] In recent years, deep learning has demonstrated significant advantages in the field of air quality prediction. Compared to traditional methods, deep learning, with its powerful nonlinear modeling capabilities, automatic feature extraction, and ability to process high-dimensional data, can more effectively capture the complex spatiotemporal characteristics of air quality data. Mainstream methods are mainly based on Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Transformer architectures. In terms of model innovation, scholars have proposed a variety of hybrid architectures. For example, the MvS CNN-BiLSTM model combines convolutional neural networks (CNN) and bidirectional long short-term memory (BiLSTM) networks, which has performed well in PM2.5 prediction in 17 Indian cities. The CDBGRU model integrates CNN with a deep learning model with a bidirectional and gated fusion mechanism (Bi-GRU), improving prediction accuracy. The improved vanilla LSTM model achieved performance improvements through parameter optimization. The GRU method based on multiple sensor nodes has verified the importance of spatial information for prediction. However, the current Recurrent Neural Network (RNN) structure still has the problems of gradient vanishing and gradient exploding; the Transformer model effectively captures global dependencies through the self-attention mechanism, effectively alleviating the gradient vanishing problem and promoting the research boom based on the attention mechanism model; TE-Informer combines periodic coding and performs outstandingly in multi-step prediction; the deep neural network ADNNet based on the attention mechanism significantly improves the prediction effect through the attention mechanism; the constructed MGSFformer effectively solves the time-space correlation problem; Airformer focuses on the special needs of nationwide prediction. However, the traditional self-attention-based model needs to perform global self-attention calculations on the entire sequence, so it faces O(N) when processing long sequences. 2 ) complexity; the Transformer-based model (Patch Time Series Transformer, PatchTST) significantly reduces the computational burden while maintaining prediction accuracy through a block-based processing mechanism. These innovative methods provide more efficient and accurate solutions for air quality forecasting.

[0004] Spatiotemporal convolution techniques, by simultaneously capturing both spatial and temporal features of data, have demonstrated significant advantages in areas such as air quality prediction. Research has shown that graph convolutional networks (GCNs) can effectively model the spatial dependencies of air pollution and handle non-Euclidean data. For example, a spatiotemporal graph convolutional recurrent neural network (GCRNN) combines GCNs and RNNs to process spatial and temporal features, respectively. A hybrid model using ResNet and ConvLSTM improves prediction accuracy. A multi-graph spatiotemporal attention network, designed for this purpose, optimizes data fusion through multi-scale spatiotemporal graphs and an attention mechanism. These methods significantly improve air quality prediction performance by effectively integrating spatiotemporal information. However, current air quality prediction methods still have shortcomings. For example, cyclical and trend changes in air quality data often overlap, forming complex nonlinear patterns. Spatiotemporal convolutional models struggle to accurately separate and model these overlapping patterns, leading to biased predictions. Spatiotemporal convolutions typically construct spatial relationships through graph structures, which often rely on a predefined spatial adjacency matrix. However, in real-world scenarios, the spatial propagation patterns of air quality often vary depending on factors such as time of day and meteorological conditions. Traditional spatiotemporal convolution cannot effectively capture this dynamically changing spatial dependency due to the lack of a dynamic graph generation mechanism. Summary of the Invention

[0005] In view of this, the present invention provides an air quality prediction method based on a multi-scale block-perceived spatiotemporal dynamic graph network to improve the accuracy of air quality prediction.

[0006] In a first aspect, the present invention provides an air quality prediction method based on a multi-scale block-perceived spatiotemporal dynamic graph network, the method comprising:

[0007] Step 1: By analyzing the air quality data of multiple air quality monitoring stations at multiple time steps and combining the spatial dependencies between the stations, define the task indicators for air quality prediction;

[0008] Step 2: Based on the defined task indicators of air quality prediction, a multi-scale block-aware spatiotemporal dynamic graph network MSPA-STDGN is constructed to perform air quality prediction.

[0009] Optionally, step 1 includes:

[0010] Define the input multivariate time series and write it as: That is, there are N air quality monitoring stations, each of which collects C characteristic variables in T consecutive time steps, and predicts the air quality index in the next τ time steps based on the input sequence X. The output is recorded as To model the spatial dependency between sites, an undirected graph G = (V, E, A) is introduced, where the site set V = {v1, v2, ..., v N}, represents N monitoring sites; E represents the spatial correlation between sites; A represents the initial adjacency matrix, It is constructed based on geographical distance, and the task index of air quality prediction is expressed as:

[0011] Y=F(X,G);

[0012] in, represents the feature input of N sites at T time steps, represents the value of the target variable of N sites in the next τ time steps; F is the prediction model, which is used to simultaneously capture the dynamic evolution law of the time dimension and the correlation pattern of the spatial dimension.

[0013] Optionally, step 2 includes:

[0014] MSPA-STDGN consists of a PurePatchEmbedding module, a decomposable temporal convolutional network (DTCN) module, an ASTG-Net module, and a prediction layer. The PurePatchEmbedding module uses a block extraction strategy to segment the original sequence into local segments using a non-overlapping sliding window. The block data is decoupled from multi-scale temporal features by the DTCN module, and different components are respectively subjected to a stacked dilated causal convolutional network (TCN) to extract pattern features at different time scales. To simulate the dynamic relationships between various sites, including the spatial and temporal patterns related to air quality indicators and auxiliary factors at each site, a hybrid spatiotemporal graph is generated using the ASTG-Net module. After obtaining the short-term spatiotemporal features and long-term temporal patterns between each site, the gated fusion mechanism in the ASTG-Net module adaptively balances the contributions of the spatiotemporal graph features and the long-term temporal patterns, outputting a joint representation that fuses global spatiotemporal dependencies with fine-grained temporal evolution. Finally, the linear layer in the prediction layer is used for final regression prediction.

[0015] Optionally, the PurePatchEmbedding module includes two basic components: channel independence component and block partitioning component; the PurePatchEmbedding module first models each feature independently, and the original input sequence is a multivariate time series Contains C feature variables;

[0016] The channel-independent component first slices the original sequence along the feature dimension and represents each feature variable as an independent univariate time series, which is expressed as: where X crepresents the time series data of all air quality monitoring stations at the cth feature, so as to decouple the multi-scale time series into c independent univariate time series, each of which contains only the time evolution pattern of a single feature;

[0017] The block partitioning component first performs X c The time dimension is padded at the end. Let the block length be L and the padding length S satisfy: if T mod L≠0, then S=L-(T mod L); if T mod L=0, then S=0, and the padding value is X c The last time step data, that is, X c [:,T-1], get the filled time series data Its expression is:

[0018]

[0019] Among them, T′=T+S; T mod L means the original sequence T is modulo the block length L, which is used to calculate the number of time steps to be filled; the time series data after filling Ensure that T′ is an integer multiple of L;

[0020] After filling the data Perform non-overlapping sliding window segmentation, the window length is L, the step length S = L; the number of blocks P and the data of each block The expressions are:

[0021]

[0022] After block division, the output of each feature channel is a three-dimensional tensor, which is expressed as:

[0023]

[0024] Optionally, the DTCN module includes two basic components: a data decomposition component and a stacked TCN component; in the data decomposition component, local weighted regression is used to Perform smoothing with a window length of W T , get the trend component Its expression is:

[0025] Among them, w t (t,i) is the weighted value of the three-core function, and the closer the distance |ti| is, the higher the weight is;

[0026] For the detrended series Perform periodic fitting, w S To control the granularity of periodic smoothing, L S is the period length, and the periodic component is obtained Its expression is:

[0027]

[0028] According to the expression Get the residual component, for each block Perform STL decomposition to get the trend cycle residual Three components; among them, Represents the long-term macro-evolution law, Used to capture short-term cyclic fluctuations, Represents modeling random noise and nonlinear disturbances;

[0029] In stacked TCN components, according to the trend cycle residual Based on the characteristics of the three components, an independent dilated causal convolutional network (TCN) is designed. Through differentiated component modeling, multi-scale temporal features are extracted. The specific process is as follows:

[0030] The trend after decomposition cycle residual The three components are fed into three parallel TCN architectures, each of which contains L layers of dilated causal convolution blocks. The trend component TCN uses a wide kernel to directly cover the long-span time window and capture the macro trend between blocks. The lth layer of dilated causal convolution is:

[0031]

[0032] in, is the input trend component; is the kernel size K T , expansion factor d T causal convolution; 1×1 convolution weights for skip connections, used to adjust dimension matching;

[0033] The periodic component TCN uses the dilation factor to incrementally expand the temporal receptive field and cover the multi-scale dependency of the periodic pattern; the dilated causal convolution of the first layer is:

[0034]

[0035] in, is the input periodic component; d S =2 l-1 is the expansion factor of the lth layer, and the receptive field increases with the number of layers;

[0036] The residual component TCN suppresses sensor noise while retaining the instantaneous characteristics of sudden events through a shallow network; the l-th layer dilated causal convolution is:

[0037]

[0038] The lth layer of dilated causal convolution becomes the l+1th layer of causal convolution after causal padding. The output features of the three components are aligned by channel concatenation and linear projection to generate a fused block representation, which is expressed as:

[0039]

[0040] Where D = D T +D S +D R is the feature dimension after fusion; Linear is the linear projection layer, which is used to unify the dimensions of each component;

[0041] The fusion features of all blocks are spliced according to the time dimension to form a global temporal representation, which is expressed as:

[0042]

[0043] Concat1 and Concat2 represent two different combination operations, respectively, and the global timing representation The input is sent to the ASTG-Net module, and the spatial dependencies are fused through dynamic spatiotemporal graph convolution. Finally, the prediction results are output through the gated fusion mechanism and the linear layer in the prediction layer.

[0044] Optionally, the ASTG-Net module integrates static geographic constraints with data-driven spatiotemporal associations, and constructs a dynamic spatiotemporal graph through a bidirectional spatiotemporal attention mechanism and geographic prior knowledge. The static adjacency matrix is generated based on the geographic distance between sites to represent stable spatial constraints; the dynamic graph captures real-time changes, and the two are weightedly fused to form a hybrid spatiotemporal graph.

[0045] Constructing a static adjacency matrix: The static adjacency matrix is used to model the pollutant diffusion relationship between stations. The distance between stations is calculated using the Euclidean distance and mapped to similarity weights using a Gaussian function. To reduce noise and computational complexity, the adjacency matrix is sparsely processed, retaining only station connections within a certain station threshold. The weights of stations greater than the station threshold are set to 0. Its expression is:

[0046]

[0047] Among them, A static [i,j] is the predefined graph, d i,j Indicates site v i With v jThe Euclidean distance of each station is (lat i ,lon i );A static [i, j] is the spatial similarity weight converted by using the Gaussian function to reduce the geographic distance. σ represents the hyperparameter that controls the weight decay speed and is set to the average distance between stations. τ represents the threshold between stations. If the distance between two stations is less than τ, the two stations are connected.

[0048] The bidirectional spatiotemporal attention mechanism dynamically models the bidirectional influence between sites by separating the attention directions of time and space: first, the time series of each site output by the DTCN module is Data is converted into graph structure data The time series length T is decomposed into a grid dimension H×W to simulate the spatial distribution of the time series. The converted data is input into a graph convolutional network to extract spatial features, and then a bidirectional spatiotemporal attention mechanism is used to generate a dynamic graph. In the bidirectional spatiotemporal attention mechanism, high-dimensional features are mapped to a low-dimensional space through convolutional layers, allowing the model to focus more on key features and ignore minor fluctuations. For two sites x1 and x2, temporal attention and spatial attention are separated, and the influence of the temporal changes and spatial features of the other site are calculated respectively. In the temporal attention mechanism, query and key features are generated through convolutional layers, and their expressions are:

[0049] θ x1 =Conv(x1),φ x2 =Conv(x2);

[0050]

[0051] y1=Energy_Time_2·V1·Energy_Space_2;

[0052] y2=Energy_Time_1·V2·Energy_Space_1;

[0053] Among them, θ x1 is the query feature of x1; φ x2 is the key feature of x2; Energy_Time_1 is the time attention weight of x1 on x2, where Energy_Time_1[i,j] represents the influence strength of the i-th time step of x1 on the j-th time step of x2; Energy_Time_2=Energy_Time_1 Tis the temporal attention weight of x2 on x1; in the spatial attention mechanism, Energy_Space_1 is the spatial attention weight of x1 on x2, where Energy_Space_1[i,j] represents the influence strength of the i-th spatial position of x1 on the j-th spatial position of x2, and Energy_Space_2 = Energy_Space_1 T is the spatial attention weight of x2 on x1; the results of temporal and spatial attention are combined to generate the final feature representation y1, y2, and the fused features are added to the original input to obtain the output;

[0054] The dynamic graph generated by the bidirectional spatiotemporal attention mechanism is added to the static adjacency matrix according to the weights to generate a hybrid spatiotemporal graph. The weight ratio is dynamically adjusted through the gated fusion mechanism, and the hybrid adjacency matrix is symmetric normalized. Its expression is:

[0055] A=Sigmoid(W gate )·A static +(1-Sigmoid(W gate ))·A dynamic ;

[0056]

[0057] in, is a parameter matrix that can be learned and used to dynamically adjust weights; Sigmoid (W gate ) is to convert W gate Mapped to the interval [0,1], it represents the weight of the static graph. The sum of the weights of the dynamic graph and the static graph is 1. D is the degree matrix used to Perform normalization to enhance numerical stability.

[0058] Optionally, the gated fusion mechanism effectively combines the temporal features output by the DTCN module with the spatiotemporal graph features that fuse the static graph and the dynamic graph to generate the final spatiotemporal dependency representation; the input features include the temporal features output by the DTCN module and the dynamic spatiotemporal graph features; the gated fusion mechanism dynamically adjusts the weight ratio of the temporal features and the spatiotemporal graph features through the reset gate and the update gate to generate a joint representation that combines local details with global dependencies; the role of the reset gate is to select the spatiotemporal graph features that need to be ignored, and the temporal features and the reset spatiotemporal graph features effectively interact to generate candidate states; the role of the update gate is to dynamically adjust the weight ratio of the temporal features and the spatiotemporal graph features, and fuse the temporal features with the candidate states to generate the final spatiotemporal dependency representation, which is expressed as follows:

[0059] r t =σ(W r ·[x grid ,x graph ]+br );

[0060] z t =σ(W z ·[x grid ,x graph ]+b z );

[0061]

[0062] Among them, x grid is the temporal feature output by the DTCN module; x graph is the spatiotemporal graph feature; W r ,W z ,b r ,b z ,b h is the parameter that can be learned; r t is the output of the reset gate; z t is the output of the update gate, which is used to determine the weight ratio of time features and spatiotemporal graph features; σ, tanh is the activation function; is a candidate state, x fused is the final state of the fusion candidate state and time features.

[0063] Optionally, the prediction layer converts the spatiotemporal features X output by the ASTG-Net module into ST Mapped to the air quality prediction result of a specific time step in the future; First, the spatiotemporal feature X is mapped to the air quality prediction result of a specific time step in the future using a one-dimensional convolution. ST Converted into sequence data, normalized by combining layer normalization and activation function ReLU; finally, predicted by the fully connected layer; prediction of the future τ time steps of N sites Its expression is:

[0064] Y=Linear(LayerNorm(Relu(Conv1D(X ST ))));

[0065] Among them, Linear represents the linear layer, d m is the spatial and temporal characteristic dimension of N stations at the input time step T.

[0066] In the technical solution provided by the present invention, the method includes defining task indicators for air quality prediction by analyzing air quality data of multiple air quality monitoring sites at multiple time steps and combining the spatial dependencies between the sites; constructing a multi-scale block-aware spatiotemporal dynamic graph network MSPA-STDGN according to the defined task indicators for air quality prediction to perform air quality prediction. This method improves the accuracy of air quality prediction by constructing a multi-scale block-aware spatiotemporal dynamic graph network. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0068] Figure 1 A flow chart of an air quality prediction method provided by an embodiment of the present invention;

[0069] Figure 2 An architectural diagram of a multi-scale, block-aware spatiotemporal dynamic graph network provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0070] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0071] It should be understood that the embodiments described are only a portion of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without creative work are within the scope of protection of the present invention.

[0072] The terms used in the embodiments of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The singular forms "a", "the" and "the" used in the embodiments of the present invention are also intended to include plural forms, unless the context clearly indicates other meanings.

[0073] It should be understood that the term "and / or" as used herein simply describes a relationship between associated objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A alone, A and B together, or B alone. Furthermore, the character " / " in this document generally indicates an "or" relationship between the associated objects.

[0074] The word "if," as used herein, may be interpreted as "at the time of" or "when" or "in response to determining" or "in response to detecting," depending on the context. Similarly, the phrases "if it is determined" or "if (stated condition or event) is detected" may be interpreted as "when it is determined" or "in response to the determination" or "when detecting (stated condition or event)" or "in response to detecting (stated condition or event)," depending on the context.

[0075] Global socioeconomic development and urbanization have exacerbated air pollution, with harmful substances such as PM2.5 threatening human health. Long-term exposure to polluted environments increases the risk of cardiovascular and respiratory diseases. Accurate pollution forecasts are crucial for governments to formulate prevention and control policies and for the public to take protective measures.

[0076] To address the challenges of spatiotemporal convolution in air quality prediction, this paper proposes a Multi-Scale Block-Aware Spatiotemporal Dynamic Graph Network (MSPA-STDGN). This model extracts local temporal features in blocks using the PurePatchEmbedding module. It then uses a Decomposed Temporal Convolutional Network (DTCN) to model trend, cycle, and residual components to capture multi-scale temporal patterns. It then dynamically learns spatiotemporal dependencies using an Adaptive Spatiotemporal Graph Network (ASTG-Net). Finally, it integrates features through a gated fusion mechanism to achieve accurate predictions.

[0077] Figure 1 The flow chart of the air quality prediction method provided by the embodiment of the present invention is as follows: Figure 1 As shown, the method includes:

[0078] Step 1: By analyzing the air quality data of multiple air quality monitoring stations at multiple time steps and combining the spatial dependencies between the stations, define the task indicators for air quality prediction.

[0079] In the embodiment of the present invention, step 1 includes:

[0080] Air quality forecasting is a spatiotemporal series prediction task based on historical data and environmental factors. It analyzes air quality data from multiple monitoring stations at multiple time steps and combines the spatial dependencies between stations to predict the air quality indicators of each station within a certain time period in the future. Define the input multivariate time series and record it as: That is, there are N air quality monitoring stations, each of which collects C characteristic variables in T consecutive time steps, and predicts the air quality index in the next τ time steps based on the input sequence X. The output is recorded as To model the spatial dependency between sites, an undirected graph G = (V, E, A) is introduced, where the site set V = {v1, v2, ..., v N}, represents N monitoring sites; E represents the spatial correlation between sites; A represents the initial adjacency matrix, It is constructed based on geographical distance, and the task index of air quality prediction is expressed as:

[0081] Y=F(X,G);

[0082] in, represents the feature input of N sites at T time steps, represents the value of the target variable of N sites in the next τ time steps; F is the prediction model, which is used to simultaneously capture the dynamic evolution law of the time dimension and the correlation pattern of the spatial dimension.

[0083] Step 2: Based on the defined task indicators of air quality prediction, a multi-scale block-aware spatiotemporal dynamic graph network MSPA-STDGN is constructed to perform air quality prediction.

[0084] In the embodiment of the present invention, Figure 2 As shown, step 2 includes:

[0085] MSPA-STDGN achieves high-precision air quality prediction through multi-scale block perception and dynamic graph collaborative modeling. MSPA-STDGN employs a hierarchical, progressive structure for prediction. It extracts local feature primitives through block embedding, decouples them into multi-scale temporal patterns through factorized temporal convolution, and correlates spatiotemporal dependencies based on dynamic graph networks, forming a progressive modeling process of "blocking-decoupling-correlation." MSPA-STDGN consists of a PurePatchEmbedding module, a Decomposable Temporal Convolutional Network (DTCN) module, an Adaptive Spatiotemporal Graph (ASTG-Net) module, and a prediction layer. To address the information redundancy and complexity of long time series data, the PurePatchEmbedding module adopts a block extraction strategy, dividing the original sequence into local segments using a non-overlapping sliding window. The block data is decoupled from multi-scale time series features by the DTCN module, and different components are respectively subjected to a stacked dilated causal convolutional network (TCN) to extract pattern features at different time scales. To simulate the dynamic relationships between various stations, including the spatial and temporal patterns related to air quality indicators and auxiliary factors at each station, a hybrid spatiotemporal graph is generated using the ASTG-Net module. After obtaining the short-term spatiotemporal features and long-term temporal patterns between each station, the gated fusion mechanism in the ASTG-Net module adaptively balances the contributions of the spatiotemporal graph features and the long-term temporal patterns, outputting a joint representation that fuses global spatiotemporal dependencies with fine-grained temporal evolution. Finally, the linear layer in the prediction layer is used for final regression prediction.

[0086] In order to refine the extraction of local dependency features of time series, the present invention proposes a PurePatchEmbedding module. First, the multivariate time series of multiple monitoring stations are processed channel-independently, and then the original series is divided into local fragments (Patch) through a non-overlapping sliding window. Each fragment captures the fine-grained time series fluctuation pattern through high-dimensional embedding. The block design decouples complex time series into multiple local units, enabling the model to focus on the microscopic evolution law of each fragment, and generate physically meaningful feature primitives through linear projection. After the block features are spliced and integrated, they retain local details and construct a global time series representation through cross-fragment collaborative optimization, providing an input basis for subsequent modules that is both fine-grained and robust.

[0087] In the embodiment of the present invention, the PurePatchEmbedding module includes two basic components: a channel-independent component and a block-partitioning component. To avoid information interference across feature variables, the PurePatchEmbedding module first independently models each feature, and the original input sequence is a multivariate time series. Contains C feature variables;

[0088] The channel-independent component first slices the original sequence along the feature dimension and represents each feature variable as an independent univariate time series, which is expressed as: where X c represents the time series data of all air quality monitoring stations at the cth feature, so as to decouple the multi-scale time series into c independent univariate time series, each of which contains only the time evolution pattern of a single feature;

[0089] To ensure that the block operation covers the full time span, the block partitioning component first c The time dimension is padded at the end. Let the block length be L and the padding length S satisfy: if T mod L≠0, then S=L-(T mod L); if T mod L=0, then S=0, and the padding value is X c The last time step data, that is, X c [:,T-1], get the filled time series data Its expression is:

[0090]

[0091] Among them, T′=T+S; T mod L means the original sequence T is modulo the block length L, which is used to calculate the number of time steps to be filled; the time series data after filling Ensure that T′ is an integer multiple of L;

[0092] After filling the data Perform non-overlapping sliding window segmentation, the window length is L, the step length S = L; the number of blocks P and the data of each block The expressions are:

[0093]

[0094] After block division, the output of each feature channel is a three-dimensional tensor, which is expressed as:

[0095]

[0096] The Decomposition-based Temporal Convolutional Network (DTCN) module achieves efficient modeling of complex time dependencies through multi-scale temporal decoupling and feature fusion. The module first uses the STL method to decompose the time series into three components: trend, cycle, and residual. Subsequently, an expanded causal convolution layer is designed for each component to expand the temporal receptive field by increasing the dilation coefficient layer by layer, and to adaptively extract features at different time scales. The trend component uses wide kernel convolution to strengthen long-term dependencies, the cycle component combines with void convolution to focus on short- and medium-term fluctuations, and the residual component suppresses noise through a lightweight network. After independent feature extraction, the three are combined through a cross-component fusion mechanism to achieve multi-scale information complementarity, outputting a temporal representation that combines local details with global correlation, providing robust input for the downstream spatiotemporal graph network.

[0097] In the embodiment of the present invention, the DTCN module includes two basic components: a data decomposition component and a stacked TCN component. In the data decomposition component, local weighted regression is used to Perform smoothing with a window length of W T , get the trend component Its expression is:

[0098] Among them, w t (t,i) is the weighted value of the three-core function, and the closer the distance |ti| is, the higher the weight is;

[0099] For the detrended series Perform periodic fitting, w S To control the granularity of periodic smoothing, L S is the period length, and the periodic component is obtained Its expression is:

[0100]

[0101] According to the expression Get the residual component, for each block Perform STL decomposition to get the trend cycle residual Three components; among them, Represents the long-term macro-evolution law, Used to capture short-term cyclic fluctuations, Represents modeling random noise and nonlinear disturbances;

[0102] According to the characteristics of each component, an independent and parallel stacked TCN architecture is designed; in the stacked TCN components, according to the trend cycle residual Based on the characteristics of the three components, an independent dilated causal convolutional network (TCN) is designed. Through differentiated component modeling, multi-scale temporal features are extracted. The specific process is as follows:

[0103] The trend after decomposition cycle residual The three components are fed into three parallel TCN architectures, each of which contains L layers of dilated causal convolution blocks. The trend component TCN uses a wide kernel to directly cover the long-span time window and capture the macro trend between blocks. The lth layer of dilated causal convolution is:

[0104]

[0105] in, is the input trend component; is the kernel size K T (Kernel), expansion factor d T Causal convolution of (Dilation); 1×1 convolution weights for skip connections, used to adjust dimension matching;

[0106] The periodic component TCN uses the dilation factor to incrementally expand the temporal receptive field and cover the multi-scale dependency of the periodic pattern; the dilated causal convolution of the first layer is:

[0107]

[0108] in, is the input periodic component; d S =2 l-1 is the expansion factor of the lth layer, and the receptive field increases with the number of layers;

[0109] The residual component TCN suppresses sensor noise while retaining the instantaneous characteristics of sudden events through a shallow network; the l-th layer dilated causal convolution is:

[0110]

[0111] The lth layer of dilated causal convolution becomes the l+1th layer of causal convolution after causal padding. The output features of the three components are aligned by channel concatenation and linear projection to generate a fused block representation, which is expressed as:

[0112]

[0113] Where D = D T +D S +D R is the feature dimension after fusion; Linear is the linear projection layer, which is used to unify the dimensions of each component;

[0114] The fusion features of all blocks are spliced according to the time dimension to form a global temporal representation, which is expressed as:

[0115]

[0116] Concat1 and Concat2 represent two different combination operations, respectively, and the global timing representation The input is sent to the ASTG-Net module, and the spatial dependencies are fused through dynamic spatiotemporal graph convolution. Finally, the prediction results are output through the gated fusion mechanism and the linear layer in the prediction layer.

[0117] In an embodiment of the present invention, the temporal pattern output by the DTCN module is converted into an initial graph. The ASTG-Net module integrates static geographic constraints with data-driven spatiotemporal associations. The air quality stations are irregularly distributed and can be abstracted into an undirected graph, with nodes representing stations. A dynamic spatiotemporal graph is constructed through a bidirectional spatiotemporal attention mechanism and geographic prior knowledge. The static adjacency matrix is generated based on the geographic distance between stations to represent stable spatial constraints; the dynamic graph captures real-time changes. The two are weightedly fused to form a mixed space-time graph, which not only retains the stability of geographic relationships but also adapts to the associated changes in dynamic scenarios. The mixed space-time graph and the fine-grained temporal features output by the DTCN are adaptively fused through a gated fusion mechanism to output a joint representation of global spatiotemporal dependencies and local temporal evolution.

[0118] Constructing a static adjacency matrix: The static adjacency matrix is used to model the pollutant diffusion relationship between stations. The closer the station distance, the stronger the correlation and the greater the weight; the farther the distance, the smaller the weight. The distance between stations is calculated using Euclidean distance and mapped to similarity weights using a Gaussian function. To reduce noise and computational complexity, the adjacency matrix is sparsely processed, retaining only station connections within a certain station threshold. The weights of stations greater than the station threshold are set to 0. The expression is:

[0119]

[0120] Among them, A static [i,j] is a pre-defined graph, d i,j Indicates site v i With v j The Euclidean distance of each station is (lat i ,lon i );A static [i, j] is the spatial similarity weight converted by using the Gaussian function to reduce the geographic distance. σ represents the hyperparameter that controls the weight decay speed and is set to the average distance between stations. τ represents the threshold between stations. If the distance between two stations is less than τ, the two stations are connected.

[0121] The bidirectional spatiotemporal attention mechanism dynamically models the bidirectional influence between sites by separating the attention directions of time and space: first, the time series of each site output by the DTCN module is Data is converted into graph structure data The time series length T is decomposed into grid dimensions H×W to simulate the spatial distribution of the time series. The transformed data is input into the graph convolutional network to extract spatial features, and then a bidirectional spatiotemporal attention mechanism is used to generate a dynamic graph. In the bidirectional spatiotemporal attention mechanism, the input of the x1 and x2 sites is The operations include normalization (Norm), convolution (Conv), normalized exponential function (Softmax), and matrix multiplication function (Matmul); the convolution layer maps high-dimensional features to low-dimensional space, allowing the model to focus more on key features and ignore minor fluctuations; for the two sites x1 and x2, separate the temporal attention and spatial attention, and calculate the influence of the temporal changes and spatial features of the other site respectively, and obtain the output In the temporal attention mechanism, query and key features are generated through the convolutional layer, and its expression is:

[0122] θ x1 =Conv(x1),φ x2 =Conv(x2);

[0123]

[0124] y1=Energy_Time_2·V1·Energy_Space_2;

[0125] y2=Energy_Time_1·V2·Energy_Space_1;

[0126] Among them, θ x1 is the query feature of x1; φ x2is the key feature of x2; Energy_Time_1 is the time attention weight of x1 on x2, where Energy_Time_1[i,j] represents the influence strength of the i-th time step of x1 on the j-th time step of x2; Energy_Time_2=Energy_Time_1 T is the temporal attention weight of x2 on x1; in the spatial attention mechanism, Energy_Space_1 is the spatial attention weight of x1 on x2, where Energy_Space_1[i,j] represents the influence strength of the i-th spatial position of x1 on the j-th spatial position of x2, and Energy_Space_2 = Energy_Space_1 T is the spatial attention weight of x2 on x1; the results of temporal and spatial attention are combined to generate the final feature representation y1, y2, and the fused features are added to the original input to obtain the output;

[0127] The dynamic graph generated by the bidirectional spatiotemporal attention mechanism is added to the static adjacency matrix according to the weights to generate a hybrid spatiotemporal graph, which can reflect stable spatial relationships and capture real-time changes. The weight ratio is dynamically adjusted through the gated fusion mechanism, and the hybrid adjacency matrix is symmetric normalized to prevent gradient explosion or vanishing. Its expression is:

[0128] A=Sigmoid(W gate )·A static +(1-Sigmoid(W gate ))·A dynamic ;

[0129]

[0130] in, is a parameter matrix that can be learned and used to dynamically adjust weights; Sigmoid (W gate ) is to convert W gate Mapped to the interval [0,1], it represents the weight of the static graph. The sum of the weights of the dynamic graph and the static graph is 1. D is the degree matrix used to Perform normalization to enhance numerical stability.

[0131] In an embodiment of the present invention, the gated fusion mechanism effectively combines the temporal features output by the DTCN module with the spatiotemporal graph features that fuse the static graph and the dynamic graph to generate the final spatiotemporal dependency representation; the input features include the temporal features output by the DTCN module and the dynamic spatiotemporal graph features; the gated fusion mechanism dynamically adjusts the weight ratio of the temporal features and the spatiotemporal graph features through the reset gate and the update gate to generate a joint representation that combines local details with global dependencies; the role of the reset gate is to select the spatiotemporal graph features that need to be ignored, and the temporal features and the reset spatiotemporal graph features effectively interact to generate candidate states; the role of the update gate is to dynamically adjust the weight ratio of the temporal features and the spatiotemporal graph features, and fuse the temporal features with the candidate states to generate the final spatiotemporal dependency representation, which is expressed as follows:

[0132] r t =σ(W r ·[x grid ,x graph ]+b r );

[0133] z t =σ(W z ·[x grid ,x graph ]+b z );

[0134]

[0135] Among them, x grid is the temporal feature output by the DTCN module; x graph is the spatiotemporal graph feature; W r ,W z ,b r ,b z ,b h is the parameter that can be learned; r t is the output of the reset gate; z t is the output of the update gate, which is used to determine the weight ratio of time features and spatiotemporal graph features; σ, tanh is the activation function; is a candidate state, x fused is the final state of the fusion candidate state and time features.

[0136] In the embodiment of the present invention, the prediction layer converts the spatiotemporal features X output by the ASTG-Net module into ST Mapped to the air quality prediction result of a specific time step in the future; First, the spatiotemporal feature X is mapped to the air quality prediction result of a specific time step in the future using the one-dimensional convolution Conv1D. ST Convert it into sequence data, combine layer normalization LayerNorm and activation function Relu for standardization; finally, make predictions through the fully connected layer; predict the future τ time steps of N sites Its expression is:

[0137] Y=Linear(LayerNorm(Relu(Conv1D(X ST ))));

[0138] Among them, Linear represents the linear layer, d m is the spatial and temporal characteristic dimension of N stations at the input time step T.

[0139] In the technical solution provided by the present invention, the method includes defining task indicators for air quality prediction by analyzing air quality data of multiple air quality monitoring sites at multiple time steps and combining the spatial dependencies between the sites; constructing a multi-scale block-aware spatiotemporal dynamic graph network MSPA-STDGN according to the defined task indicators for air quality prediction to perform air quality prediction. This method improves the accuracy of air quality prediction by constructing a multi-scale block-aware spatiotemporal dynamic graph network.

[0140] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0141] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An air quality prediction method based on a multi-scale block-perceived spatiotemporal dynamic graph network, characterized in that: The method comprises: Step 1: By analyzing the air quality data of multiple air quality monitoring stations at multiple time steps and combining the spatial dependencies between the stations, define the task indicators for air quality prediction; Step 2: Based on the defined task indicators of air quality prediction, a multi-scale block-aware spatiotemporal dynamic graph network MSPA-STDGN is constructed to perform air quality prediction.

2. The method according to claim 1, characterized in that The step 1 comprises: Define the input multivariate time series and write it as: That is, there are N air quality monitoring stations, each of which collects C characteristic variables in T consecutive time steps, and predicts the air quality index in the next τ time steps based on the input sequence X. The output is recorded as To model the spatial dependency between sites, an undirected graph G = (V, E, A) is introduced, where the site set V = {v1, v2, ..., v N }, represents N monitoring sites; E represents the spatial correlation between sites; A represents the initial adjacency matrix, It is constructed based on geographical distance, and the task index of air quality prediction is expressed as: Y=F(X,G); in, represents the feature input of N sites at T time steps, represents the value of the target variable of N sites in the next τ time steps; F is the prediction model, which is used to simultaneously capture the dynamic evolution law of the time dimension and the correlation pattern of the spatial dimension.

3. The method according to claim 1, characterized in that The step 2 includes: MSPA-STDGN consists of a PurePatchEmbedding module, a decomposable temporal convolutional network (DTCN) module, an adaptive spatiotemporal graph network (ASTG-Net) module, and a prediction layer. The PurePatchEmbedding module uses a block extraction strategy to segment the original sequence into local segments using a non-overlapping sliding window. The block data is decoupled from multi-scale temporal features by the DTCN module, and different components are respectively subjected to a stacked dilated causal convolutional network (TCN) to extract pattern features at different time scales. To simulate the dynamic relationships between various sites, including the spatial and temporal patterns related to air quality indicators and auxiliary factors at each site, a hybrid spatiotemporal graph is generated using the ASTG-Net module. After obtaining the short-term spatiotemporal features and long-term temporal patterns between each site, the gated fusion mechanism in the ASTG-Net module adaptively balances the contributions of the spatiotemporal graph features and the long-term temporal patterns, outputting a joint representation that fuses global spatiotemporal dependencies with fine-grained temporal evolution. Finally, the linear layer in the prediction layer is used for final regression prediction.

4. The method according to claim 3, characterized in that The PurePatchEmbedding module includes two basic components: channel independence component and block partitioning component. The PurePatchEmbedding module first models each feature independently, and the original input sequence is a multivariate time series. Contains C feature variables; The channel-independent component first slices the original sequence along the feature dimension and represents each feature variable as an independent univariate time series, which is expressed as: where X c represents the time series data of all air quality monitoring stations at the cth feature, so as to decouple the multi-scale time series into c independent univariate time series, each of which contains only the time evolution pattern of a single feature; The block partitioning component first performs X c The time dimension is padded at the end. Let the block length be L and the padding length S satisfy: if T mod L ≠ 0, then S = L - (T mod L); if T mod L = 0, then S = 0; the padding value is X c The last time step data, that is, X c [:,T-1], get the filled time series data Its expression is: in, T mod L means the original sequence T is modulo the block length L, which is used to calculate the number of time steps that need to be filled; the time series data after filling make sure is an integer multiple of L; After filling the data Perform non-overlapping sliding window segmentation with a window length of L and a step size of S = L; The number of blocks P and the data of each block The expressions are: After block division, the output of each feature channel is a three-dimensional tensor, which is expressed as:

5. The method according to claim 3, characterized in that The DTCN module includes two basic components: data decomposition component and stacked TCN component. In the data decomposition component, local weighted regression is used to Perform smoothing with a window length of W T , get the trend component Its expression is: Among them, w t (t,i) is the weighted value of the three-core function, and the closer the distance |ti| is, the higher the weight is; For the detrended series Perform periodic fitting, w S To control the granularity of periodic smoothing, L S is the period length, and the periodic component is obtained Its expression is: According to the expression Get the residual component, for each block Perform STL decomposition to get the trend cycle residual Three components; among them, Represents the long-term macro-evolution law, Used to capture short-term cyclic fluctuations, Represents modeling random noise and nonlinear disturbances; In stacked TCN components, according to the trend cycle residual Based on the characteristics of the three components, an independent dilated causal convolutional network (TCN) is designed. Through differentiated component modeling, multi-scale temporal features are extracted. The specific process is as follows: The trend after decomposition cycle residual The three components are fed into three parallel TCN architectures, each of which contains L layers of dilated causal convolution blocks. The trend component TCN uses a wide kernel to directly cover the long-span time window and capture the macro trend between blocks. The lth layer of dilated causal convolution is: in, is the input trend component; is the kernel size K T , expansion factor d T Causal convolution; 1×1 convolution weights for skip connections, used to adjust dimension matching; The periodic component TCN uses the dilation factor to incrementally expand the temporal receptive field and cover the multi-scale dependency of the periodic pattern; the dilated causal convolution of the first layer is: in, is the input periodic component; d S =2 l-1 is the expansion factor of the lth layer, and the receptive field increases with the number of layers; The residual component TCN suppresses sensor noise while retaining the instantaneous characteristics of sudden events through a shallow network; the l-th layer dilated causal convolution is: The lth layer of dilated causal convolution becomes the l+1th layer of causal convolution after causal padding. The output features of the three components are aligned by channel concatenation and linear projection to generate a fused block representation, which is expressed as: Where D = D T +D S +D R is the feature dimension after fusion; Linear is the linear projection layer, which is used to unify the dimensions of each component; The fusion features of all blocks are spliced according to the time dimension to form a global temporal representation, which is expressed as: Concat1 and Concat2 represent two different combination operations, respectively, and the global timing representation The input is sent to the ASTG-Net module, and the spatial dependencies are fused through dynamic spatiotemporal graph convolution. Finally, the prediction results are output through the gated fusion mechanism and the linear layer in the prediction layer.

6. The method according to claim 3, characterized in that The ASTG-Net module integrates static geographic constraints with data-driven spatiotemporal associations. It constructs a dynamic spatiotemporal graph through a bidirectional spatiotemporal attention mechanism and geographic prior knowledge. The static adjacency matrix is generated based on the geographic distance between sites to represent stable spatial constraints; the dynamic graph captures real-time changes. The two are weightedly fused to form a hybrid spatiotemporal graph. Constructing a static adjacency matrix: The static adjacency matrix is used to model the pollutant diffusion relationship between stations. The distance between stations is calculated using the Euclidean distance and mapped to similarity weights using a Gaussian function. To reduce noise and computational complexity, the adjacency matrix is sparsely processed, retaining only station connections within a certain station threshold. The weights of stations greater than the station threshold are set to 0. Its expression is: Among them, A static [i,j] is a predefined graph, d i,j Indicates site v i With v j The Euclidean distance of each station is (lat i ,lon i );A static [i, j] is the spatial similarity weight converted by using the Gaussian function to reduce the geographic distance. σ represents the hyperparameter that controls the weight decay speed and is set to the average distance between stations. τ represents the threshold between stations. If the distance between two stations is less than τ, the two stations are connected. The bidirectional spatiotemporal attention mechanism dynamically models the bidirectional influence between sites by separating the attention directions of time and space: first, the time series of each site output by the DTCN module is Data is converted into graph structure data The time series length T is decomposed into a grid dimension H×W to simulate the spatial distribution of the time series. The converted data is input into a graph convolutional network to extract spatial features, and then a bidirectional spatiotemporal attention mechanism is used to generate a dynamic graph. In the bidirectional spatiotemporal attention mechanism, high-dimensional features are mapped to a low-dimensional space through convolutional layers, allowing the model to focus more on key features and ignore minor fluctuations. For two sites x1 and x2, temporal attention and spatial attention are separated, and the influence of the temporal changes and spatial features of the other site are calculated respectively. In the temporal attention mechanism, query and key features are generated through convolutional layers, and their expressions are: θ x1 =Conv(x1),φ x2 =Conv(x2); y1=Energy_Time_2·V1·Energy_Space_2; y2=Energy_Time_1·V2·Energy_Space_1; Among them, θ x1 is the query feature of x1; φ x2 is the key feature of x2; Energy_Time_1 is the time attention weight of x1 on x2, where Energy_Time_1[i,j] represents the influence strength of the i-th time step of x1 on the j-th time step of x2; Energy_Time_2=Energy_Time_1 T is the temporal attention weight of x2 on x1; in the spatial attention mechanism, Energy_Space_1 is the spatial attention weight of x1 on x2, where Energy_Space_1[i,j] represents the influence strength of the i-th spatial position of x1 on the j-th spatial position of x2, and Energy_Space_2 = Energy_Space_1 T is the spatial attention weight of x2 on x1; the results of temporal and spatial attention are combined to generate the final feature representation y1, y2, and the fused features are added to the original input to obtain the output; The dynamic graph generated by the bidirectional spatiotemporal attention mechanism is added to the static adjacency matrix according to the weights to generate a hybrid spatiotemporal graph. The weight ratio is dynamically adjusted through the gated fusion mechanism, and the hybrid adjacency matrix is symmetric normalized. Its expression is: A=Sigmoid(W gate )·A static +(1-Sigmoid(W gate ))·A dynamic ; in, is a parameter matrix that can be learned and used to dynamically adjust weights; Sigmoid (W gate ) is to convert W gate Mapped to the interval [0,1], it represents the weight of the static graph. The sum of the weights of the dynamic graph and the static graph is 1. D is the degree matrix used to Perform normalization to enhance numerical stability.

7. The method according to claim 6, characterized in that The gated fusion mechanism effectively combines the temporal features output by the DTCN module with the spatiotemporal graph features that fuse static and dynamic graphs to generate the final spatiotemporal dependency representation; the input features include the temporal features output by the DTCN module and the dynamic spatiotemporal graph features; The gated fusion mechanism dynamically adjusts the weight ratio of temporal features and spatiotemporal graph features through the reset gate and update gate to generate a joint representation that combines local details with global dependencies. The reset gate selects spatiotemporal graph features that need to be ignored, and the temporal features and the reset spatiotemporal graph features effectively interact to generate candidate states. The update gate dynamically adjusts the weight ratio of temporal features and spatiotemporal graph features, fusing the temporal features with the candidate states to generate the final spatiotemporal dependency representation, which is expressed as follows: r t =σ(W r ·[x grid ,x graph ]+b r ); z t =σ(W z ·[x grid ,x graph ]+b z ); Among them, x grid is the temporal feature output by the DTCN module; x graph is the spatiotemporal graph feature; W r ,W z ,b r ,b z ,b h is the parameter that can be learned; r t is the output of the reset gate; z t is the output of the update gate, which is used to determine the weight ratio of time features and spatiotemporal graph features; σ, tanh is the activation function; is a candidate state, x fused is the final state of the fusion candidate state and time features.

8. The method according to claim 3, characterized in that The prediction layer converts the spatiotemporal features X output by the ASTG-Net module into ST Mapped to the air quality prediction result of a specific time step in the future; First, the spatiotemporal feature X is mapped to the air quality prediction result of a specific time step in the future using the one-dimensional convolution Conv1D. ST Convert it into sequence data, combine layer normalization LayerNorm and activation function Relu for standardization; finally, make predictions through the fully connected layer; predict the future τ time steps of N sites Its expression is: Y=Linear(LayerNorm(Relu(Conv1D(X ST )))); Among them, Linear represents the linear layer, d m is the spatial and temporal characteristic dimension of N stations at the input time step T.

Images