An ozone concentration prediction method and system based on multi-graph attention spatio-temporal fusion

Abstract

This invention discloses an ozone concentration prediction method and system based on multi-graph attention spatiotemporal fusion, belonging to the field of air pollution prediction technology. The method includes: acquiring the original ozone concentration time series from multiple monitoring stations; performing adaptive secondary decomposition on the series to obtain multi-scale temporal features; constructing a multi-view spatial graph structure including a distance graph, adjacency graph, and heuristic graph based on the station's geographical location; inputting the temporal features and spatial graph structure into a gated fusion-based multi-graph attention module to extract dynamically fused spatial features; encoding the spatial features in the temporal dimension using a bidirectional long short-term memory network to output a hidden state sequence; aggregating the weights of different time steps through a multi-head temporal attention mechanism to obtain a temporal context vector; and finally outputting the ozone concentration value through a fully connected layer. This invention solves the problems of simple splicing or fixed weights in multi-graph fusion, and the inability to adaptively adjust the contribution of different graph structures.

Description

Technical Field

[0001] This invention belongs to the field of atmospheric pollution prediction technology, specifically relating to an ozone concentration prediction method and system based on multi-graph attention spatiotemporal fusion. Background Technology

[0002] ozone( As a typical secondary pollutant, ozone concentration time series exhibits strong non-stationarity, multi-scale fluctuations, and complex spatial dependence. Existing ozone concentration prediction methods mainly suffer from the following problems: (1) Single signal decomposition methods, such as Empirical Mode Decomposition (EMD) and Fully Adaptive Noise Ensemble Empirical Mode Decomposition (CEEMDAN), are difficult to completely separate high-frequency noise from effective signals. Furthermore, the decomposition parameters (such as the number of modes K in Variational Mode Decomposition (VMD)) are usually manually set, lacking adaptability and easily leading to over-decomposition or under-decomposition. (2) Spatial map structures are usually constructed based solely on geographical distance, ignoring the dynamic evolution characteristics of pollutant transport and the similarity of statistical patterns between stations. Moreover, the map construction parameters (such as the number of bins) are empirically set, making it difficult to adapt to the data distribution characteristics of different regions. (3) When fusing multiple maps, simple splicing or fixed weight strategies are often adopted, which cannot adaptively adjust the contribution of different map structures. Furthermore, the fusion effect is limited when the feature dimensions of different map structures are inconsistent. Therefore, it is urgent to construct an ozone concentration prediction method that can adaptively process multi-scale time series features and dynamically fuse multi-perspective spatial dependence. Summary of the Invention

[0003] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0004] An ozone concentration prediction method based on multi-graph attention spatiotemporal fusion includes:

[0005] S1, Obtain the original historical ozone concentration time series of multiple air quality monitoring stations in the study area;

[0006] S2, perform adaptive secondary decomposition on the original historical ozone concentration time series of each air quality monitoring station to obtain the multi-scale time feature set of the air quality monitoring station;

[0007] S3. Based on the geographical location information of each air quality monitoring station, construct a multi-view spatial graph structure set; the multi-view spatial graph structure set includes at least a distance graph, an adjacency graph, and a heuristic graph;

[0008] S4 inputs the multi-scale temporal feature set obtained in S2 and the multi-view spatial graph structure set constructed in S3 into a multi-graph attention module based on a gated fusion mechanism to extract dynamically fused spatial feature representations. ;

[0009] S5, Bidirectional Long Short-Term Memory Network for Dynamically Fusion Spatial Feature Representation Encode the hidden state sequence at each time step in the time dimension; the multi-head temporal attention mechanism generates query, key and value by linear projection of the hidden state sequence at each time step, calculates the importance weight of different time steps in different feature subspaces, and aggregates them to obtain the final temporal context vector.

[0010] S6 inputs the temporal context vector into the fully connected layer and outputs the ozone concentration value for one or more predetermined prediction time steps in the future. The predetermined prediction time step is an integer hour within the range of 1 hour to the maximum prediction duration.

[0011] An ozone concentration prediction system based on multi-graph attention spatiotemporal fusion includes:

[0012] The data acquisition module acquires the original historical ozone concentration time series from multiple air quality monitoring stations within the study area;

[0013] The adaptive quadratic decomposition module performs adaptive quadratic decomposition on the original historical ozone concentration time series of each air quality monitoring station to obtain the multi-scale time feature set of the air quality monitoring station.

[0014] The multi-graph construction module constructs a set of multi-view spatial graph structures based on the geographical location information of each air quality monitoring station; the set of multi-view spatial graph structures includes at least a distance graph, an adjacency graph, and a heuristic graph;

[0015] The multi-graph attention module takes the multi-scale temporal feature set obtained in S2 and the multi-view spatial graph structure set constructed in S3 as inputs to the multi-graph attention module based on the gated fusion mechanism to extract the dynamically fused spatial feature representation. ;

[0016] The temporal modeling module uses a bidirectional long short-term memory network to represent dynamically fused spatial features. Encode the hidden state sequence at each time step in the time dimension; the multi-head temporal attention mechanism generates query, key and value by linear projection of the hidden state sequence at each time step, calculates the importance weight of different time steps in different feature subspaces, and aggregates them to obtain the final temporal context vector.

[0017] The prediction output module inputs the temporal context vector into the fully connected layer and outputs the ozone concentration value for one or more predetermined prediction time steps in the future. The predetermined prediction time step is an integer hour within the range of 1 hour to the maximum prediction duration.

[0018] The present invention has the following beneficial effects:

[0019] (1) This invention achieves a refined secondary decomposition of high-frequency components by using an adaptive secondary decomposition strategy based on the center frequency observation method. This involves first using CEEMDAN to decompose the original ozone sequence, then using normalized mutual information to adaptively screen high-frequency IMF components, and finally using the center frequency observation method (stopping to increase the number of modes when the center frequency difference between adjacent VMFs is less than the threshold) to adaptively determine the optimal number of modes K for VMD. This solves the problems of existing VMD mode number K relying on manual trial parameters, lacking adaptability, and easily leading to over-decomposition or under-decomposition.

[0020] (2) This invention constructs a multi-perspective spatial graph structure that includes a distance graph, an adjacency graph, and a heuristic graph. The heuristic graph uses Sturgess rules to adaptively determine the number of bins and combines Laplace smoothing and symmetric KL divergence to characterize the potential spatial dependence between sites from the perspective of statistical distribution similarity. This solves the problems of existing spatial graph structures that are based only on geographical distance, ignore the dynamic transport characteristics of pollutants and the similarity of statistical patterns, and have graph construction parameters that depend on empirical settings.

[0021] (3) This invention designs a node-level gating fusion mechanism with exponential decay regularization, that is, dynamically fuses spatial attention and graph attention outputs through learnable node-level gating factors, and adds a regularization term to the loss function in conjunction with an exponential decay strategy (forcing the two attention contributions to be balanced in the early stage of training, and freely learning the optimal weight in the later stage), which solves the problems of existing multi-graph fusion using simple splicing or fixed weights, unable to adaptively adjust the contribution of different graph structures, and easily leading to model mode collapse. Attached Figure Description

[0022] Figure 1 This is a flowchart of the ozone concentration prediction method based on multi-graph attention spatiotemporal fusion of the present invention;

[0023] Figure 2 This is a flowchart of the adaptive quadratic decomposition of the present invention;

[0024] Figure 3 A flowchart for constructing the heuristic graph of this invention;

[0025] Figure 4 This is a structural diagram of the gated fusion submodule of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0027] like Figure 1 As shown, this invention provides a method for predicting ozone concentration based on multi-graph attention spatiotemporal fusion, comprising the following steps:

[0028] S1, obtain the original historical ozone concentration time series of multiple air quality monitoring stations in the study area.

[0029] S2, an adaptive secondary decomposition is performed on the original historical ozone concentration time series of each air quality monitoring station to obtain a multi-scale time feature set for that air quality monitoring station. For example... Figure 2 As shown, the adaptive quadratic decomposition specifically includes:

[0030] S21, using Fully Adaptive Noise Ensemble Empirical Mode Decomposition (CEEMDAN) to analyze the original historical ozone concentration time series. The decomposition yields multiple intrinsic mode functions (IMFs) and a residual term.

[0031] CEEMDAN effectively suppresses mode aliasing by adaptively adding white noise and ensemble averaging. The stopping condition for CEEMDAN decomposition is: the decomposition process terminates when the number of extreme points in the residual signal obtained by decomposition is less than 2, and the residual signal at this point is the final residual term.

[0032] S22, Calculate each IMF component Compared with the original historical ozone concentration time series The normalized mutual information is used as the correlation degree, and its calculation formula is as follows:

[0033] ;

[0034] in, To normalize mutual information, its range is [0,1], and the larger the value, the stronger the correlation between the two variables; For mutual information, This refers to information entropy.

[0035] Set the first preset threshold Select the intrinsic mode function (IMF) with the lowest correlation as the component to be decomposed in the second step; if the lowest correlation is lower than the first preset threshold... The components to be decomposed are subjected to secondary decomposition in steps S23 and S24; if the minimum correlation degree is not lower than the first preset threshold... If so, skip steps S23 and S24.

[0036] The component with the smallest NMI (Normalized Mutual Information) is chosen for secondary decomposition because it has the weakest correlation with the original historical ozone concentration time series and is most likely to represent high-frequency noise or complex fluctuations that are not strongly associated with the prediction target. Separating it can minimize the non-stationarity of the series.

[0037] S23, Variational Mode Decomposition (VMD) is used to decompose the components to be decomposed in the quadratic decomposition, wherein the number of modes in VMD is... The center frequency is determined adaptively using the center frequency observation method, which is specifically as follows:

[0038] Modal number of VMD The value starts from 2 and gradually increases, for each Value, obtained from decomposition Calculate the center frequency of each variational mode function (VMF) and sort the VMFs in ascending order of their center frequencies; calculate the absolute value of the difference between the center frequencies of all adjacent VMFs. And take the minimum value among them. ; , These are the numbers arranged in ascending order of center frequency. The and the first The center frequency of each VMF, The index number of VMF, It is the absolute value symbol.

[0039] When the minimum value Less than the second preset threshold Stop increasing The value of determines the current The value of is the optimal mode number; if The value is increased to the preset maximum value. If the condition is still not met, then the current condition will be used. The value of is the optimal number of modes.

[0040] The physical meaning of this criterion is that when the center frequencies of any two adjacent modes are too close, it indicates that mode aliasing has occurred, and there is no need to increase the number of decomposition layers, thus achieving an adaptive optimal selection of the number of modes.

[0041] S24. Combine the variational mode functions (VMFs) obtained from VMD decomposition in S23 with the remaining IMFs and residual terms obtained from CEEMDAN decomposition in S21, excluding the components to be decomposed a second time, to form a multi-scale time feature set of air quality monitoring stations.

[0042] In S23, the method for calculating the center frequency of VMF (Hilbert transform method) is as follows:

[0043] S231, for each VMF component obtained from VMD decomposition Perform Hilbert transform to obtain .

[0044] in, For the first VMF in time The value, Let τ be the value after its Hilbert transform, and τ be the integration variable. Indicates time offset.

[0045] S232, Constructing an analytic signal .

[0046] in, For the first The analytical signal of a VMF, It is the imaginary unit.

[0047] S233, Calculate the instantaneous phase sequence : And perform phase unwinding.

[0048] Phase unwinding specifically involves:

[0049] For length of Instantaneous phase sequence , The discrete time point index is used to initialize the unwound phase sequence. ;for The value ranges from 2 to Calculate the current instantaneous phase Phase after untangling in the previous step The difference , ;like For the current instantaneous phase Make a downward correction: ;like Then for the current instantaneous phase Make an upward correction: ;otherwise, Repeat this process until all time points have been processed, obtaining the unwound instantaneous phase sequence. .

[0050] S234, obtaining the instantaneous phase sequence after untangling. Then, it is first smoothed and filtered to suppress phase jitter caused by high-frequency noise.

[0051] In a preferred embodiment, a moving average with a window length of 5 sampling points is used for smoothing filtering.

[0052] Then, the instantaneous angular frequency is calculated using the central difference formula. The central difference formula is as follows:

[0053] ;

[0054] in, The sampling interval (e.g., 1 hour). Time index and The instantaneous phase value after unwinding at the location.

[0055] instantaneous frequency for: .

[0056] Finally, regarding instantaneous frequency Median filtering is performed to remove physically unreasonable frequency values. In a preferred embodiment, median filtering with a window length of 3 sampling points is used. For data with a sampling interval of 1 hour, the theoretical maximum instantaneous frequency should be less than 0.5 cycles / hour (i.e., 0.5 cycles / hour, corresponding to the Nyquist frequency). Data points exceeding this value are considered outliers and replaced with the mean of adjacent valid points. The location of the detected outliers is then determined. The filling method is performed using the mean of adjacent valid points: finding the nearest non-outlier locations both forward and backward. and Then the location of the wild point value Replaced with ,in, , Non-outlier locations , The value is used. If no valid point can be found in a certain direction (such as the beginning or end of the sequence), the only valid point that can be found on the other side is used as the replacement value. If multiple outliers appear consecutively, the entire interval of consecutive outliers is filled with global linear interpolation. Finally, the processed instantaneous frequency is... The center frequency of the VMF is obtained by taking the arithmetic mean over all time axes. ,in, The instantaneous phase sequence length defined for S233.

[0057] S3. Based on the geographical location information of each air quality monitoring station, construct a multi-view spatial graph structure set. The multi-view spatial graph structure set includes at least a distance graph, an adjacency graph, and a heuristic graph.

[0058] The method for constructing the distance graph includes: calculating edge weights using a thresholded Gaussian kernel function based on the Euclidean distance between air quality monitoring stations. :

[0059] ;

[0060] like Then calculate the edge weights according to the above formula. Otherwise edge weight ;in, For the site With the site Euclidean distance, For bandwidth parameters, This is the distance threshold.

[0061] In a preferred embodiment, the bandwidth parameter Take the median of the Euclidean distances between all stations, and the distance threshold. Take the 90th percentile of the Euclidean distance between all stations.

[0062] The method for constructing an adjacency graph includes: if the Euclidean distance between two air quality monitoring stations does not exceed 50km, then the edge weights are... The edge weight is 1, otherwise it is 1. .

[0063] like Figure 3 As shown, the method for constructing the heuristic graph includes: First, aligning the original historical ozone concentration time series of all air quality monitoring stations on the time axis and cropping them to the maximum continuous time length common to all air quality monitoring stations, denoted as . The maximum common continuous time length is defined as: the existence of a continuous time window across all air quality monitoring stations' time series. , These represent the start and end times of the continuous time window. Within the continuous time window, the ozone concentration time series for each air quality monitoring station has no missing values, and the length of the window is... The maximum value is found within all such continuous windows. The global minimum and maximum values ​​for all air quality monitoring stations are determined, forming a global value range. This global value range is then divided into equally spaced intervals. There are several intervals, among which... Determined using Sturgess rules: , This is a rounding function. For each air quality monitoring station, the observed values ​​of its original historical ozone concentration time series fall within the range of... The frequency within each interval is calculated, and the zero-probability interval is smoothed using Laplace smoothing to obtain the probability distribution for each air quality monitoring station. The formula for Laplace smoothing is: ,in, For the site The observed ozone concentration fell on the first Smooth probability estimates within each interval For range index, , The total number of intervals to be divided is determined by Sturgess's rule. For the first Frequency within each interval For sequence length, The Laplace smoothing parameter and .

[0064] In a preferred embodiment Calculate the symmetric KL divergence (symmetric relative entropy) between the probability distributions of any two air quality monitoring stations, and obtain the edge weights through exponential mapping. Construct a heuristic graph, where the stations... With the site Symmetric KL divergence between probability distributions for:

[0065] ;

[0066] in, and Sites and sites The probability distribution of the th The probability quality of each interval, Given the total number of intervals, the edge weights of the heuristic graph constructed based on KL divergence are: . For the site The probability distribution, For the site The probability distribution, It is an exponential function. It is the logarithm with the natural constant e as the base.

[0067] S4 inputs the multi-scale temporal feature set obtained in S2 and the multi-view spatial graph structure set constructed in S3 into the multi-graph attention module based on the gating fusion mechanism to extract the dynamically fused spatial feature representation.

[0068] The intrinsic mode functions, variational mode functions, and residual terms in the multi-scale time feature set of each air quality monitoring station obtained from S2 are analyzed at the same time. The eigenvalues ​​of all intrinsic mode functions, variational mode functions, and residual terms of the air quality monitoring station at that time are vectorized and concatenated (i.e., the eigenvalues ​​of each component of the station at the same time are concatenated). The eigenvalues ​​are concatenated into a 3D eigenvectors , for This allows us to transform the original historical ozone concentration time series into a feature vector sequence using a real-valued vector space. ), to obtain that moment The original feature matrices of all air quality monitoring stations are then processed through a fully connected layer to map the feature vectors of each air quality monitoring station to the same preset feature dimension. (e.g., 128), to obtain the spatial feature input matrix. ( For the number of stations, For N rows The real matrix space of the column (i.e., the space to which the feature matrices of all stations belong under the preset feature dimensions) is input into the multi-graph attention module.

[0069] The multi-graph attention module utilizes the distance graph, adjacency graph, and heuristic graph constructed in S3 as priors for spatial dependencies. The multi-graph attention module includes a spatial attention submodule, a graph attention submodule, and a gated fusion submodule. The spatial attention submodule computes spatial attention outputs for the distance graph, adjacency graph, and heuristic graph respectively. The graph attention submodule performs weighted fusion of the spatial attention outputs from the distance graph, adjacency graph, and heuristic graph to obtain a graph attention fused output. The gated fusion submodule dynamically fuses the spatial attention output and the graph attention fused output using node-level gating factors, outputting a dynamically fused spatial feature representation.

[0070] The spatial attention submodule is for each graph The spatial attention output is computed using a graph attention network based on scaled dot product attention. Specifically, the node feature matrix First, the query matrix is ​​obtained through three independent linear transformations. Key matrix Sum matrix Attention weights are calculated by computing the query matrix. AND key matrix The dot product is obtained by scaling, and the attention score of non-neighbor nodes is set to negative infinity through a mask operation, wherein the mask is based on the current graph structure. Adjacency matrix: If nodes With nodes If there are no edges between them (i.e., edge weights are 0), then the corresponding position in the attention score matrix is ​​set to negative infinity. Attention output. for:

[0071] ;

[0072] in, Let be the dimension of the key vector. Let be the normalized exponential function. Let the outputs of the spatial attention submodule corresponding to the distance graph, adjacency graph, and heuristic graph be respectively... .

[0073] The graph attention submodule uses learnable graph attention weights. The spatial attention outputs of different graphs are weighted and fused to obtain the graph attention fusion output. ,in, Normalization is achieved using the softmax (normalization exponential function) function.

[0074] like Figure 4 As shown, the gating fusion submodule performs the following operations: By concatenating along the feature dimensions, we obtain the concatenated spatial feature matrix. ;Will The input to the first fully connected layer is mapped to the first fused feature. ;Will The input to the second fully connected layer is mapped to the second fused feature. The first fully connected layer and the second fully connected layer are independent of each other; then, a node-level gating factor with the same dimension as the first fused feature is calculated through learnable parameters. ; The activation function is Sigmoid. The final dynamically fused spatial features are represented as follows: ,in, This represents element-wise multiplication. In a preferred embodiment, the output dimensions of both the first fully connected layer and the second fully connected layer are 1. For example, 128, the gating factor The dimension is In the early stages of training, Initialize to 0.5 to balance the initial contributions of spatial attention and graph attention.

[0075] S5, Bidirectional Long Short-Term Memory Network for Dynamically Fusion Spatial Feature Representation Encode the hidden state sequence at each time step by time.

[0076] The multi-head temporal attention mechanism generates Query, Key, and Value by linear projection of the hidden state sequence at each time step. It calculates the importance weights of different time steps in different feature subspaces and aggregates them in a weighted manner to obtain the final temporal context vector.

[0077] In the training process of the ozone concentration prediction system based on multi-graph attention spatiotemporal fusion, the mean squared error (MSE) between the predicted output and the expected output is used as the basic loss function. To prevent gating factors... Premature saturation or degradation can be addressed by adding a regularization term to the loss function. The total loss function is ,in, Mean square error, The regularization coefficient is . This is the Frobenius norm. Regularization coefficient. With training rounds As the value increases, an exponential decay strategy is employed to reduce the value from the initial value to 0. The decay formula for the exponential decay strategy is: ,in, The attenuation constant is The value range is from 10 to 50. , Regularization coefficient The initial value. This allows the ozone concentration prediction system based on multi-graph attention spatiotemporal fusion to maintain a balanced exploration of the two attention methods in the early stages of training, and in the later stages of training, the ozone concentration prediction system based on multi-graph attention spatiotemporal fusion can freely learn the optimal value based on the data. value.

[0078] S6, input the time context vector into the fully connected layer, and output the ozone concentration value for one or more predetermined prediction time steps in the future. The predetermined prediction time step is an integer hour within the range of 1 hour to the maximum prediction duration (e.g., 72 hours).

[0079] This invention further provides an ozone concentration prediction system based on multi-graph attention spatiotemporal fusion, comprising:

[0080] The data acquisition module acquires the original historical ozone concentration time series from multiple air quality monitoring stations within the study area;

[0081] The adaptive quadratic decomposition module performs adaptive quadratic decomposition on the original historical ozone concentration time series of each air quality monitoring station to obtain the multi-scale time feature set of the air quality monitoring station.

[0082] The multi-graph construction module constructs a set of multi-view spatial graph structures based on the geographical location information of each air quality monitoring station; the set of multi-view spatial graph structures includes at least a distance graph, an adjacency graph, and a heuristic graph;

[0083] The multi-graph attention module takes the multi-scale temporal feature set obtained in S2 and the multi-view spatial graph structure set constructed in S3 as inputs to the multi-graph attention module based on the gated fusion mechanism to extract the dynamically fused spatial feature representation. ;

[0084] The temporal modeling module uses a bidirectional long short-term memory network to represent dynamically fused spatial features. Encode the hidden state sequence at each time step in the time dimension; the multi-head temporal attention mechanism generates query, key and value by linear projection of the hidden state sequence at each time step, calculates the importance weight of different time steps in different feature subspaces, and aggregates them to obtain the final temporal context vector.

[0085] The prediction output module inputs the temporal context vector into the fully connected layer and outputs the ozone concentration value for one or more predetermined prediction time steps in the future. The predetermined prediction time step is an integer hour within the range of 1 hour to the maximum prediction duration.

[0086] The technical solution of the present invention will be further described below with reference to the embodiments.

[0087] This example uses hourly ozone concentration data from 24 air quality monitoring stations in Beijing from 2021 to 2024. Data preprocessing employs... The criteria include outlier removal, time series interpolation to fill missing values, Kriging interpolation to match meteorological data, and min-max normalization.

[0088] The decomposition algorithm parameters are set as follows:

[0089] In this embodiment, the key parameters used in the CEEMDAN algorithm implementation are: a standard deviation of 0.2 for the added white noise, a set average of 100, and a maximum number of screening iterations of 500. The key parameters used in the VMD algorithm implementation are: a penalty factor. (Internal parameters of the VMD algorithm), noise tolerance Convergence discrimination accuracy (Internal parameters of the VMD algorithm). The above parameters are set based on conventional analysis of the characteristics of the original signal, and those skilled in the art can also fine-tune them according to the actual data characteristics.

[0090] In this embodiment, to ensure the effectiveness and reproducibility of the ozone concentration prediction system based on multi-graph attention spatiotemporal fusion, the hyperparameters of each key module are set as follows:

[0091] Adaptive quadratic decomposition parameters: First preset threshold Through comparative experiments, when When the value is 0.1, some low-frequency IMFs are misclassified as high-frequency, leading to excessive secondary decomposition; when When the value is 0.3, some high-frequency IMFs are not identified, resulting in residual non-stationarity. The time-based decomposition yields the best results. In this embodiment, the component with the smallest NMI among the IMFs with an NMI below 0.2 is selected as the component to be decomposed in the secondary decomposition. If no IMF has an NMI below 0.2, no secondary decomposition is performed, and all decomposition results from CEEMDAN are directly used as the multi-scale time feature set for this air quality monitoring station.

[0092] Second preset threshold Hz, this threshold is determined based on data sampled at 1-hour intervals. Preset maximum value. Center frequency observation method: Initially K=2, gradually increase K. For each K value, calculate the center frequency of all VMFs after VMD decomposition and arrange the VMFs in ascending order of center frequency. Then calculate the absolute value of the difference between the center frequencies of all adjacent VMFs and take the minimum value. .when Less than Stop increasing at Hz. The value of K is the current optimal number of modes. This is the second preset threshold. Set the preset maximum value. When K reaches the preset maximum value Stop iterating when the time comes.

[0093] Heuristic graph construction parameters: Laplacian smoothing parameters The Sturgess rule is used to determine: ,in, This represents the maximum continuous time length shared by all participants. In this embodiment, it is obtained after trimming. =8760 (full hours in a year), then .

[0094] Multi-Graph Attention Module Parameters: The intrinsic mode functions, variational mode functions, and residual terms from the multi-scale temporal feature sets of each air quality monitoring station are concatenated at the same time, and then mapped to a preset feature dimension through a fully connected layer. The spatial attention submodule has an output dimension of 128. The graph attention submodule also has an output dimension of 128. In gated fusion, two independent fully connected layers map the input to the same output dimension. Gating factor Initialized to 0.5.

[0095] Exponential decay strategy: A regularization term is added to the loss function during training, and the regularization coefficient... The initial value is set to An exponential decay strategy is adopted: In the early stages of training, Larger, regularization term forced Approaching 0.5 forces a balance between spatial attention and graph attention contributions; as training progresses, As the value approaches zero, the regularization constraint disappears, and the ozone concentration prediction system based on multi-graph attention spatiotemporal fusion can freely learn the optimal value. value.

[0096] Temporal modeling module parameters: BiLSTM hidden layer dimension 128, multi-head temporal attention mechanism head number 8.

[0097] Training parameters: learning rate 0.001, batch size 32, number of training epochs 200, using Adam (adaptive moment estimator) optimizer and dynamic learning rate scheduling.

[0098] On the test set, the root mean square error (RMSE) for single-step prediction was 5.84, and the mean absolute error (MAE) was 4.25. , The coefficient of determination. The RMSE for 72-hour multi-step prediction is 14.02. Ablation experiments showed that the RMSE increased by 77.4% after the adaptive quadratic decomposition module was implemented, and increased by 23.5% after removing the gated fusion submodule, thus verifying the effectiveness of the adaptive quadratic decomposition module and the gated fusion submodule.

[0099] The above description is merely an embodiment of the present invention and does not limit the scope of the invention. Any equivalent structural or procedural transformations made based on the description and drawings of this invention, or direct or indirect applications in other related system fields, are similarly included within the protection scope of this invention. Contents not described in detail in this specification are prior art known to those skilled in the art.

[0100] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The embodiments of the present invention can be implemented using various computer languages.

[0101] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0102] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0103] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

Claims

1. A method for predicting ozone concentration based on multi-graph attention spatiotemporal fusion, characterized in that, include: S1, Obtain the original historical ozone concentration time series of multiple air quality monitoring stations in the study area; S2, perform adaptive secondary decomposition on the original historical ozone concentration time series of each air quality monitoring station to obtain the multi-scale time feature set of the air quality monitoring station; S3. Based on the geographical location information of each air quality monitoring station, construct a multi-view spatial graph structure set; the multi-view spatial graph structure set includes at least a distance graph, an adjacency graph, and a heuristic graph; S4 inputs the multi-scale temporal feature set obtained in S2 and the multi-view spatial graph structure set constructed in S3 into a multi-graph attention module based on a gated fusion mechanism to extract dynamically fused spatial feature representations. ; S5, Bidirectional Long Short-Term Memory Network for Dynamically Fusion Spatial Feature Representation Encode the hidden state sequence at each time step in the time dimension; the multi-head temporal attention mechanism generates query, key and value by linear projection of the hidden state sequence at each time step, calculates the importance weight of different time steps in different feature subspaces, and aggregates them to obtain the final temporal context vector. S6 inputs the temporal context vector into the fully connected layer and outputs the ozone concentration value for one or more predetermined prediction time steps in the future. The predetermined prediction time step is an integer hour within the range of 1 hour to the maximum prediction duration.

2. The ozone concentration prediction method based on multi-graph attention spatiotemporal fusion according to claim 1, characterized in that, S2 include: S21, employing fully adaptive noise ensemble empirical mode decomposition to analyze the original historical ozone concentration time series. The decomposition yields multiple intrinsic mode functions and a residual term. S22, Calculate each intrinsic mode function component Compared with the original historical ozone concentration time series The normalized mutual information is used as the correlation degree; the intrinsic mode function with the smallest correlation degree is selected as the component to be decomposed in the second step; if the smallest correlation degree is lower than the first preset threshold... The components to be decomposed are subjected to secondary decomposition in steps S23 and S24; if the minimum correlation degree is not lower than the first preset threshold... If so, skip steps S23 and S24; S23, variational mode decomposition is used to decompose the components to be decomposed in the quadratic decomposition, wherein the number of modes in the variational mode decomposition is... The center frequency is adaptively determined using the center frequency observation method, which specifically involves determining the number of modes in the variational mode decomposition. The value starts from 2 and gradually increases, for each For each of the M variational mode functions obtained from the decomposition, calculate its center frequency and arrange the variational mode functions in ascending order of center frequency; calculate the absolute value of the difference between the center frequencies of all adjacent variational mode functions and take the minimum value; stop increasing when the minimum value is less than a second preset threshold. The value of determines the current The value of is the optimal mode number; if If the condition is still not met when the value is increased to the preset maximum value, then the current value will be used. The value of is the optimal number of modes; S24 combines the variational mode functions obtained from variational mode decomposition in S23 with the eigenmode functions and residual terms obtained from CEEMDAN decomposition in S21, excluding the components to be decomposed a second time, to form a multi-scale time feature set of the air quality monitoring station.

3. The ozone concentration prediction method based on multi-graph attention spatiotemporal fusion according to claim 2, characterized in that, In S22, each eigenmode function component Compared with the original historical ozone concentration time series The formula for calculating normalized mutual information is: ； in, To normalize mutual information, For mutual information, This refers to information entropy.

4. The ozone concentration prediction method based on multi-graph attention spatiotemporal fusion according to claim 3, characterized in that, In S23, the method for calculating the center frequency of the variational mode function is as follows: S231, for each variational mode function component obtained from variational mode decomposition Performing a Hilbert transform yields: ； in, For the k-th variational mode function in time The value, Let τ be the value after its Hilbert transform, and τ be the integration variable. Indicates time offset; S232, Constructing an analytical signal: ： in, For the first analytic signal of a variational mode function, The imaginary unit; S233, Calculate the instantaneous phase sequence : Then, phase unwrapping is performed to obtain the unwrapped instantaneous phase sequence. ; S234, for the instantaneous phase sequence after unwinding. Perform a smoothing filter and calculate the instantaneous angular frequency using the center difference formula. ; instantaneous frequency for: ; For instantaneous frequency Perform median filtering and remove physically unreasonable frequency values.

5. The ozone concentration prediction method based on multi-graph attention spatiotemporal fusion according to claim 4, characterized in that, In S233, phase unwinding includes: for a length of Instantaneous phase sequence , The discrete time point index is used to initialize the unwound phase sequence. ;for The value ranges from 2 to Calculate the current instantaneous phase Phase after untangling in the previous step The difference , ;like For the current instantaneous phase Make a downward correction: ;like Then for the current instantaneous phase Make an upward correction: ;otherwise, Repeat this process until all time points have been processed, obtaining the unwound instantaneous phase sequence. .

6. The ozone concentration prediction method based on multi-graph attention spatiotemporal fusion according to claim 5, characterized in that, In S234, the central difference formula is as follows: ； in, The sampling interval is... Time index and The instantaneous phase value after unwinding at the location.

7. The ozone concentration prediction method based on multi-graph attention spatiotemporal fusion according to claim 6, characterized in that, In S3, the method for constructing the distance graph includes: calculating edge weights using a thresholded Gaussian kernel function based on the Euclidean distance between air quality monitoring stations. : ； like Calculate the edge weights using the formula above. Otherwise, edge weight ;in, For the site With the site Euclidean distance, For bandwidth parameters, Distance threshold; The method for constructing an adjacency graph includes: if the Euclidean distance between two air quality monitoring stations does not exceed 50km, then the edge weights are... The edge weight is 1, otherwise it is 1. ; The heuristic graph construction method includes: aligning the original historical ozone concentration time series of all air quality monitoring stations on the time axis and cropping them to the maximum continuous time length common to all air quality monitoring stations; determining the global minimum and maximum values ​​of all air quality monitoring stations to form a global value range, and dividing the global value range into multiple intervals at equal intervals; for each air quality monitoring station, counting the frequency of its original historical ozone concentration time series observations falling within each of the multiple intervals, and using Laplace smoothing to process the zero probability intervals to obtain the probability distribution of each air quality monitoring station; Calculate the symmetric KL divergence between the probability distributions of any two air quality monitoring stations, and obtain the edge weights through exponential mapping to construct a heuristic graph.

8. The ozone concentration prediction method based on multi-graph attention spatiotemporal fusion according to claim 7, characterized in that, In S3, the global value range is divided into equal intervals. There are several intervals, among which... Determined using Sturgess rules: , It is the floor function. This is the maximum continuous time length shared by all air quality monitoring stations.

9. The ozone concentration prediction method based on multi-graph attention spatiotemporal fusion according to claim 8, characterized in that, In S4, the multi-graph attention module uses the distance graph, adjacency graph, and heuristic graph constructed in S3 as priors for spatial dependencies; The multi-graph attention module includes a spatial attention submodule, a graph attention submodule, and a gated fusion submodule. The spatial attention submodule is used to calculate the spatial attention output for the distance graph, adjacency graph, and heuristic graph, respectively. The graph attention submodule is used to perform weighted fusion of the spatial attention outputs of the distance graph, adjacency graph, and heuristic graph to obtain the graph attention fusion output. The gated fusion submodule is used to dynamically fuse the spatial attention output and the graph attention fusion output through node-level gating factors to output a dynamically fused spatial feature representation.

10. An ozone concentration prediction system based on multi-graph attention spatiotemporal fusion, characterized in that, include: The data acquisition module acquires the original historical ozone concentration time series from multiple air quality monitoring stations within the study area; The adaptive quadratic decomposition module performs adaptive quadratic decomposition on the original historical ozone concentration time series of each air quality monitoring station to obtain a multi-scale time feature set of the air quality monitoring station. The multi-graph construction module constructs a set of multi-view spatial graph structures based on the geographical location information of each air quality monitoring station; the set of multi-view spatial graph structures includes at least a distance graph, an adjacency graph, and a heuristic graph; The multi-graph attention module takes the multi-scale temporal feature set obtained in S2 and the multi-view spatial graph structure set constructed in S3 as inputs to the multi-graph attention module based on the gated fusion mechanism to extract the dynamically fused spatial feature representation. ; The temporal modeling module uses a bidirectional long short-term memory network to represent dynamically fused spatial features. Encode the hidden state sequence at each time step in the time dimension; the multi-head temporal attention mechanism generates query, key and value by linear projection of the hidden state sequence at each time step, calculates the importance weight of different time steps in different feature subspaces, and aggregates them to obtain the final temporal context vector. The prediction output module inputs the temporal context vector into the fully connected layer and outputs the ozone concentration value for one or more predetermined prediction time steps in the future. The predetermined prediction time step is an integer hour within the range of 1 hour to the maximum prediction duration.

Images