Method for predicting multiple air pollutants based on multi-scale spatio-temporal spectrum feature fusion framework

Abstract

The invention belongs to the field of air pollutant prediction, discloses a multiple air pollutant prediction method based on a multi-scale spatio-temporal spectrum feature fusion framework, and solves the technical problem of low prediction accuracy of multiple air pollutants caused by single air pollutant prediction type and neglect of a dynamic dependency relationship among different pollutants in the prior art. The method comprises the following steps of: constructing an air quality characteristic data set by acquiring collected data of a plurality of air quality monitoring stations in a city; preprocessing the air quality characteristic data set; constructing a multi-scale space-time spectrum feature fusion model; inputting the preprocessed air quality feature data set into a multi-scale spatio-temporal spectrum feature fusion model for training; and predicting the pollutant concentration of each station in a future time period by using the trained multi-scale spatio-temporal spectrum feature fusion model. According to the method, the dynamic dependency relationship among various pollutants is explicitly modeled when the model is constructed, and the prediction precision of various air pollutants is effectively improved.

Description

Technical Field

[0001] The present invention belongs to the field of air pollutant prediction, and specifically is a method for predicting multiple air pollutants based on a multi-scale spatiotemporal feature fusion framework. Background Art

[0002] With the rapid development of the global economy and the acceleration of urbanization, urban air quality issues have become increasingly prominent, and air pollution has caused serious harm to ecosystems and human activities. Air pollutants mainly include gaseous substances (NO2, CO, O3 and SO2) and particulate matter (PM2.5 and PM10). Short-term exposure to high concentrations of air pollution can cause serious health risks such as dyspnea and cardiopulmonary diseases, and long-term exposure may cause cancer. In recent years, the government and the public have increasingly demanded real-time, accurate and rapid predictions of air quality. It is crucial to scientifically predict the concentrations of multiple air pollutants within a given time and space range. Multi-pollutant predictions help reveal the interactions between pollutants and provide a scientific basis for the formulation of targeted pollution prevention and control measures.

[0003] Existing technologies mainly model air quality data by considering spatial and temporal characteristics. Existing air pollutant prediction methods include time series analysis methods (ARIMA models), machine learning methods (random forests, support vector regression), and deep learning models (RNN, LSTM). With the development of graph neural networks and graph attention mechanisms, dynamically changing spatial information is incorporated into the prediction modeling process, significantly improving the prediction accuracy. However, existing air pollutant prediction technologies are mostly limited to the prediction of a single pollutant (such as PM2.5). In order to predict the levels of other pollutants, it is usually necessary to train separate models for each pollutant. In addition, there are complex interactions between different pollutants, some pollutants promote or inhibit the formation of other pollutants, or multiple pollutants originate from the same emission source. Existing technologies fail to fully consider the interactions and intrinsic connections between multiple pollutants when predicting, resulting in limited prediction accuracy. Summary of the invention

[0004] The purpose of the present invention is to solve the problem of difficulty in accurately predicting multiple air pollutants at the same time and ignoring the interaction of multiple pollutants in the modeling method, and provide a method for predicting multiple air pollutants based on a multi-scale spatiotemporal spectral feature fusion framework, the method comprising the following steps:

[0005] S1: Obtain data collected from multiple air quality monitoring stations in the city and construct an air quality characteristic dataset. Each station contains time series data of multiple pollutant concentration indicators.

[0006] S2: preprocessing the air quality characteristic data set;

[0007] S3: Construct a multi-scale spatiotemporal feature fusion model;

[0008] S4: inputting the air quality feature data set into the multi-scale spatiotemporal feature fusion model for training;

[0009] S5: Input the pollutant time series data of multiple air quality monitoring stations in the historical period into the trained multi-scale spatiotemporal spectral feature fusion model to predict the pollutant concentration of each station in the future period.

[0010] A storage device stores instructions and data for implementing a method for predicting multiple air pollutants based on a multi-scale spatiotemporal spectral feature fusion framework.

[0011] A method and device for predicting multiple air pollutants based on a multi-scale spatiotemporal spectral feature fusion framework, comprising: a processor and a storage device; the processor loads and executes instructions and data in the storage device to implement a method for predicting multiple air pollutants based on a multi-scale spatiotemporal spectral feature fusion framework.

[0012] The beneficial effects brought by the technical solution provided by the present invention are:

[0013] By acquiring the collected data from multiple air quality monitoring stations in the city, an air quality characteristic data set is constructed; the air quality characteristic data set is preprocessed; a multi-scale spatiotemporal spectral feature fusion model is constructed; the air quality characteristic data set is input into the multi-scale spatiotemporal spectral feature fusion model for training; the time series data of multiple pollutants from multiple air quality monitoring stations in the historical period are input into the trained model together, and the concentration values ​​of the corresponding pollutants at multiple stations in the future period are output. The steps and techniques explicitly extract the dependency between multiple pollutants, and realize the full integration of the time, space and variable dimension characteristics, and complete the accurate prediction of the concentration of multiple pollutants at multiple stations in the future period, providing a reference for government departments and decision makers to formulate relevant air pollution control measures. It solves the problem of ignoring the dependency between multiple pollutants and the difficulty in accurately predicting the changing trends of multiple pollutants at the same time in previous technologies. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

[0015] Figure 1 is a step diagram of a method for predicting multiple air pollutants based on a multi-scale spatiotemporal spectral feature fusion framework in an embodiment of the present invention;

[0016] Figure 2 is a schematic diagram of a method for predicting multiple air pollutants based on a multi-scale spatiotemporal spectral feature fusion framework in an embodiment of the present invention;

[0017] Figure 3 Schematic diagram of a temporal feature extraction module in a multi-scale spatiotemporal feature fusion model in an embodiment of the present invention;

[0018] Figure 4 Schematic diagram of a spatial feature extraction module in a multi-scale spatiotemporal feature fusion model in an embodiment of the present invention;

[0019] Figure 5 Schematic diagram of a variable feature extraction module in a multi-scale spatiotemporal feature fusion model in an embodiment of the present invention;

[0020] Figure 6 is a schematic diagram of the results of an ablation experiment in an embodiment of the present invention;

[0021] Figure 7 It is a schematic diagram of the working of the hardware device of an embodiment of the present invention. DETAILED DESCRIPTION

[0022] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, specific embodiments of the present invention are now described in detail with reference to the accompanying drawings.

[0023] The embodiments of the present invention provide a method and device for predicting multiple air pollutants based on a multi-scale spatiotemporal feature fusion framework.

[0024] Please refer to Figure 1 , Figure 1 : is a flowchart of a method for predicting multiple air pollutants based on a multi-scale spatiotemporal spectral feature fusion framework in an embodiment of the present invention, which specifically includes the following steps:

[0025] S1: Acquire data collected from multiple air quality monitoring stations in the city and construct an air quality characteristic data set, which includes hourly concentration data of pollutants PM2.5, PM10, O3, and NO2;

[0026] S2: preprocessing the air quality characteristic data set;

[0027] S3: Construct a multi-scale spatiotemporal feature fusion model;

[0028] S4: inputting the air quality feature data set into the spatiotemporal spectrum multi-scale feature fusion model for training;

[0029] S5: Input the time series data of multiple pollutants from multiple air quality monitoring stations in the historical period into the trained model, and output the concentration values ​​of corresponding pollutants at multiple stations in the future period;

[0030] Step S2 preprocesses the air quality data set including: missing value processing, outlier processing, and training set and test set division.

[0031] The multi-scale spatiotemporal feature fusion model constructed in step S3 is as follows: Figure 2 As shown in the figure, it includes input layer, feature extraction and fusion layer, and output layer. The construction process is:

[0032] S31: The input layer organizes the time series data of the four air pollutant concentrations at each station into a multivariate feature matrix X∈R containing time series information T×N×m ; Where T is the time window length, N is the number of air quality monitoring stations, and m is the number of pollutants;

[0033] S32: Feature extraction and fusion layer includes decoupled autoencoder and decoder. Decoupled autoencoder includes Figure 3 The temporal feature extraction module shown in Figure 4 The spatial feature extraction module shown in Figure 5 The variable feature extraction module shown. The decoder includes multiple cross-attention feature fusion modules.

[0034] The specific steps of the time feature extraction module are:

[0035] Input the multivariate feature matrix X to position-code the time series of each site;

[0036] The position-encoded feature matrix is ​​input into the multi-site Transformer to calculate the multi-head attention. The calculation method is:

[0037]

[0038] in, are the query vector, key vector, and value vector. are learnable parameters. represents the i-th attention head, Attention represents the multi-head attention mechanism, and softmax represents the activation function;

[0039] Concatenate the output of each attention head according to the feature dimension:

[0040]

[0041] The output of the multi-site transformer Input into the bilinear time attention to calculate the interaction weights of each time point in multiple time series, and establish a global context representation to capture the global time correlation of all sites. The calculation method is:

[0042]

[0043] Among them, U1, U2, U3, V e ,be is a learnable parameter, σ is an activation function. E is the temporal attention matrix obtained by bilinear temporal attention calculation, E i,j is the element in the temporal attention matrix, is the normalized temporal attention matrix, whose elements are composition, Represents the correlation between time points i and j. The normalized time attention matrix With the output of the multi-site transformer Multiply the output to get the global time feature

[0044] The global time feature Input into the local time convolution, specifically:

[0045]

[0046] Among them, Conv2D represents a two-dimensional convolution operation and ReLU is a nonlinear activation function.

[0047] The convolution kernel size of the two-dimensional convolution is 3 and the padding is 1. The output of the local time convolution is the time feature matrix

[0048] The specific steps of the spatial feature extraction module are:

[0049] The multivariate feature matrix X is input into the bilinear spatial attention to capture the global spatial dependencies between all sites in the region. The calculation method of bilinear spatial attention is:

[0050] S=V s ·σ((XW1)W2(W3X) T +b s )

[0051]

[0052] Among them, W1, W2, W3, b s ,V s is a learnable parameter, and σ is an activation function. The input matrix undergoes two layers of linear transformation to obtain the global spatial attention matrix S, which is then normalized to obtain the spatial dependency graph is the global spatial attention weight composition, Represents the correlation between sites i and j.

[0053] Construct the initial adjacency matrix, specifically:

[0054]

[0055] Among them, A ij represents the initial adjacency matrix between sites i and j, d ij Represents the geographical distance between sites i and j. R represents the distance threshold. If the geographical distance between two sites is less than R, it is considered that there is an edge between the two sites.

[0056] Calculate the Laplacian matrix: L = DA, where D is the degree matrix; A represents the initial adjacency matrix;

[0057] Normalize the Laplacian matrix: Among them I N is the identity matrix, λ max represents the maximum eigenvalue of L; represents the normalized Laplacian matrix;

[0058] Compute Chebyshev polynomials: in and denote the k-1th and k-2th order Chebyshev polynomials respectively.

[0059] Perform the Hadamard product of the Chebyshev polynomials with the spatial dependency graph and compute the Chebyshev graph convolution:

[0060]

[0061] Among them, θ k is a learnable parameter, ⊙ represents the Hadamard product. The spatial feature matrix is ​​obtained g θ *Gx is the graph convolution operation, X is the multivariate feature matrix mentioned above, is the spatial dependence diagram mentioned above.

[0062] The specific steps of the variable feature extraction module are:

[0063] Flatten the time and space dimensions of the multivariate feature matrix X into vectors to represent the spatiotemporal information of each pollutant;

[0064] The vector of each pollutant is regarded as a node, and the multi-head attention mechanism is used to calculate the dependencies between different pollutants;

[0065] The obtained feature matrix reshapes the time dimension and space dimension

[0066] The specific steps of the multi-cross attention feature fusion module are:

[0067] The temporal feature matrix and the spatial feature matrix are input into the first layer of cross attention, specifically:

[0068]

[0069] in, is a learnable parameter, Embedding dimension for spatiotemporal features, with spatial feature matrix After linear mapping as the query subvector, the time feature matrix After linear mapping as key sub-vector and value sub-vector, a spatiotemporal fusion feature matrix is ​​obtained through dot product attention calculation.

[0070] The variable feature matrix and the spatiotemporal fusion feature matrix are input into the second layer of cross attention, specifically:

[0071]

[0072] in, is a learnable parameter, is the spatiotemporal feature embedding dimension, with variable feature matrix Through linear mapping As the query sub-vector, the spatiotemporal fusion feature matrix After linear mapping as key sub-vector and value sub-vector, a spatiotemporal feature matrix is ​​obtained through dot product attention calculation.

[0073] S33: The output layer uses two-dimensional convolution to predict the pollutant concentration of each station in the future period. Specifically:

[0074]

[0075] Among them, t represents the time dimension, v represents the variable dimension, K is the kernel function, and the convolution kernel is used to input data Slide up and calculate the weighted sum of the input area covered by the convolution kernel. τ represents the convolution window offset in the time dimension, is the embedding dimension of the above-mentioned spatiotemporal spectral features, q is the convolution window offset on the variable dimension, and T represents the time window length.

[0076] Step S4 includes:

[0077] S41: inputting the multivariate feature matrix containing time series information into the spatiotemporal spectrum multi-scale feature fusion model for training;

[0078] S42: Use the MAE loss function to calculate the error and back propagate until the loss error is minimized, and save the model parameters. The MAE loss function calculation method is as follows:

[0079]

[0080] Among them, Y predis the predicted value, Y true is the true value.

[0081] Specifically, by introducing the time feature extraction module, the characteristics of individual, global and local time scales are fully modeled; by introducing the spatial feature extraction module, the global and local spatial characteristics are fully modeled; by introducing the variable feature extraction module, the dynamic dependency relationship between different pollutants is explicitly modeled. This solves the problem of low accuracy in predicting multiple pollutants due to neglect of the correlation between different pollutants in previous studies.

[0082] For example: construct an air quality feature dataset. Taking a city’s air quality dataset as an example, it contains four air pollutants: PM2.5, PM10, O3, and NO2.

[0083] The original air quality data set was preprocessed, and the missing values ​​of the time series in the data set were filled by linear interpolation. According to the "Ambient Air Quality Standards" issued by the Ministry of Ecology and Environment of the People's Republic of China, the maximum and minimum boundary values ​​were set for the four pollutants, and outliers were removed by screening. All data sets were divided into batches and samples according to the time series, and the order of all samples was randomly disrupted to ensure that the characteristics of the training data were evenly distributed. The training set, test set, and validation set were divided according to 6:2:2.

[0084] Inputting a multivariate feature matrix containing time series information into the spatiotemporal multiscale feature fusion model for training;

[0085] The MAE loss function is used to calculate the error and back propagate until the loss error is minimized and the model parameters are saved. The MAE loss function calculation method is as follows:

[0086]

[0087] Among them, Y pred is the predicted value, Y true is the true value, and n is the number of samples collected.

[0088] The time series data of multiple pollutants from multiple air quality monitoring stations in the historical period are input into the trained model together, and the concentration values ​​of the corresponding pollutants at multiple stations in the future period are output; RMSE (root mean square error), MAE (mean absolute error), R 2 (coefficient of determination) as an accuracy evaluation indicator:

[0089]

[0090] In the formula, m represents the total number of samples, y i represents the actual observed value, represents the predicted value of the i-th sample, Represents the average of all sample actual values. 2 The closer the value is to 1, the better the model fit is.

[0091] Taking 12h and 24h as the prediction time length, the accuracy change comparison of each model is shown in Table 1, including the STSMFM proposed in the present invention, that is, the multi-scale spatiotemporal feature fusion model, and other comparison models: ASTGCN, STTN, DSTAGNN, STIDGCN, Airformer, GDGCN, and STGDN.

[0092] Table 1

[0093]

[0094] As shown in Table 1, STSMFM achieves significant performance improvement in both 12-hour and 24-hour forecasting tasks compared with the best baseline model. STSMFM has a significant improvement in R of PM2.5, PM10, O3 and NO2 compared with the suboptimal model STGDN. 2 In the 24h forecasting task, STSMFM has a better R performance than STGDN for the four pollutants. 2 The accuracy of STSMFM is improved by 2.3%, 2.5%, 3.6% and 11.8% respectively. When predicting O3 and NO2, STSMFM has a more significant improvement than other advanced models.

[0095] In addition, this case conducted ablation experiments on each module described. Figure 6 ,The temporal feature extraction module (TFEM), spatial feature extraction module (SFEM), variable feature extraction module (VFEM) and multi-cross attention feature fusion module (MFFM) all play a positive role in the prediction accuracy, and the variable feature extraction module plays a vital role in this case.

[0096] The present invention uses an improved attention mechanism to explicitly model the dependencies between multiple pollutants, and extracts and fuses the multi-scale time series and spatial distribution characteristics of air pollutants from multiple dimensions of the spatiotemporal spectrum, solving the problem of ignoring the dependencies between multiple pollutants and the difficulty in accurately predicting the changing trends of multiple pollutants at the same time in the prior art. The method for predicting multiple air pollutants based on a multi-scale spatiotemporal spectrum feature fusion framework proposed by the present invention helps to improve the prediction efficiency and accuracy of changes in multiple air pollutants, and promotes relevant departments to promptly carry out and take corresponding environmental governance measures.

[0097] See also Figure 7 , Figure 7It is a schematic diagram of the working of the hardware device of an embodiment of the present invention. The hardware device specifically includes: a multiple air pollutant prediction device 401 based on a spatiotemporal spectrum multi-scale feature fusion framework, a processor 402 and a storage device 403.

[0098] A device 401 for predicting multiple air pollutants based on a spatiotemporal spectrum multi-scale feature fusion framework: implements a method for predicting multiple air pollutants based on a spatiotemporal spectrum multi-scale feature fusion framework.

[0099] Processor 402: Loads and executes instructions and data in storage device 402 to implement a method for predicting multiple air pollutants based on a spatiotemporal spectrum multi-scale feature fusion framework.

[0100] Storage device 403: stores instructions and data; used to implement a method for predicting multiple air pollutants based on a spatiotemporal spectrum multi-scale feature fusion framework.

[0101] The above are only preferred embodiments of the present invention and are 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 protection scope of the present invention.

Claims

1. A method for predicting multiple air pollutants based on a multi-scale spatiotemporal feature fusion framework, characterized in that: The method comprises the following steps: S1: Obtain data collected from multiple air quality monitoring stations in the city and construct an air quality characteristic dataset. Each station contains time series data of multiple pollutant concentration indicators. S2: preprocessing the air quality characteristic data set; S3: Construct a multi-scale spatiotemporal feature fusion model; S4: inputting the air quality feature data set into the multi-scale spatiotemporal feature fusion model for training; S5: Input the pollutant time series data of multiple air quality monitoring stations in the historical period into the trained multi-scale spatiotemporal spectral feature fusion model to predict the pollutant concentration of each station in the future period.

2. The method for predicting multiple air pollutants based on a multi-scale spatiotemporal spectral feature fusion framework as claimed in claim 1, characterized in that: In step S2, the preprocessing includes missing value processing, outlier processing, and training set and test set division.

3. The method for predicting multiple air pollutants based on a multi-scale spatiotemporal spectral feature fusion framework as claimed in claim 1, characterized in that: In step S3, the multi-scale spatiotemporal spectral feature fusion model includes: an input layer, a feature extraction and fusion layer, and an output layer.

4. The method for predicting multiple air pollutants based on a multi-scale spatiotemporal spectral feature fusion framework as claimed in claim 3, characterized in that: The input layer organizes the time series data of various air pollutant concentrations at each station into a multivariate feature matrix X∈R containing time series information T×N×m ; Where T is the time window length, N is the number of air quality monitoring stations, and m is the number of pollutants.

5. The method for predicting multiple air pollutants based on a multi-scale spatiotemporal spectral feature fusion framework as claimed in claim 3, characterized in that: The feature extraction and fusion layer includes: a decoupled autoencoder and a decoder; the decoupled autoencoder includes a time feature extraction module, a spatial feature extraction module and a variable feature extraction module; the decoder includes a multi-cross attention feature fusion module; The time feature extraction module comprises: Position encoding the multivariate feature matrix X to capture individual temporal patterns of multiple sites; The position-encoded feature matrix is ​​input into the multi-site Transformer, and the multi-head self-attention mechanism is used to calculate the multi-head attention: in, are query vector, key vector and value vector, is a learnable parameter; represents the i-th attention head, Attention represents the multi-head attention mechanism, and softmax represents the activation function; Concatenate the output of each attention head according to the feature dimension: The output of the multi-site transformer Input into the bilinear time attention, use bilinear mapping to calculate the interaction weights of each time point in multiple time series, and establish a global context representation to capture the global time correlation of all sites. The calculation method is: Among them, U1, U2, U3, V e ,b e is the first learnable parameter, σ is the activation function, E is the temporal attention matrix obtained by the bilinear temporal attention calculation, E i,j is the element in the temporal attention matrix, is the normalized temporal attention matrix, whose elements are composition, Represents the correlation of time points i and j, the normalized time attention matrix With the output of the multi-site transformer Multiply the output to get the global time feature The global time feature Input into the local time two-dimensional convolution to extract the local time feature matrix, specifically: Among them, Conv2D represents a two-dimensional convolution operation, and ReLU is a nonlinear activation function; The spatial feature extraction module includes: inputting the multivariate feature matrix into bilinear spatial attention, using bilinear mapping to calculate the dependency of time series of different sites, and adaptively capturing the global spatial correlation between sites; inputting the output of the bilinear spatial attention into Chebyshev graph convolution to extract local spatial dependency; The calculation formula of bilinear spatial attention is: S=V s ·σ((XW1)W2(W3X) T +b s ) Among them, W1, W2, W3, b s ,V s is the second learnable parameter, σ is the activation function; the input matrix undergoes two layers of linear transformation to obtain the global spatial attention matrix S, which is then normalized to obtain the spatial dependency graph is the global spatial attention weight composition, represents the correlation of site i, j, and the superscript T indicates transposition; Construct the initial adjacency matrix, specifically: Among them, A ij represents the initial adjacency matrix between sites i and j, d ij represents the geographical distance between sites i and j, R represents the distance threshold, if the geographical distance between two sites is less than R, it is considered that there is an edge between the two sites; Calculate the Laplacian matrix: L = DA, where D is the degree matrix; A represents the initial adjacency matrix; Normalize the Laplacian matrix: Among them I N is the identity matrix, λ max represents the maximum eigenvalue of L, represents the normalized Laplacian matrix; Compute Chebyshev polynomials: in and denote the k-1th and k-2th order Chebyshev polynomials respectively; Perform the Hadamard product of the Chebyshev polynomials with the spatial dependency graph and compute the Chebyshev graph convolution: Among them, θ k is the third learnable parameter, ⊙ represents the Hadamard product, represents the spatial feature matrix; g θ *Gx is the graph convolution operation, X is the multivariate feature matrix, is the spatial dependency graph; The variable feature extraction module includes: flattening the time dimension and space dimension of the multivariate feature matrix X into vectors to represent the spatiotemporal information of each pollutant; treating the vector of each pollutant as a node, and calculating the dependencies between different pollutants using a multi-head attention mechanism; and reshaping the time dimension and space dimension of the obtained feature matrix. The specific steps of the variable feature extraction module are: Flatten the time and space dimensions of the multivariate feature matrix X into vectors to represent the spatiotemporal information of each pollutant; The vector of each pollutant is regarded as a node, and the multi-head attention mechanism is used to calculate the dependencies between different pollutants; The resulting feature matrix reshapes the time and space dimensions; The specific steps of the multi-cross attention feature fusion module are: The temporal feature matrix and the spatial feature matrix are input into the first layer of cross attention, specifically: in, is the fourth learnable parameter, Embedding dimension for spatiotemporal features, with spatial feature matrix After linear mapping as the query subvector, the time feature matrix After linear mapping as key sub-vector and value sub-vector, a spatiotemporal fusion feature matrix is ​​obtained through dot product attention calculation. The variable feature matrix and the spatiotemporal fusion feature matrix are input into the second layer of cross attention, specifically: in, is the fifth learnable parameter, is the spatiotemporal feature embedding dimension, with variable feature matrix Through linear mapping As the query sub-vector, the spatiotemporal fusion feature matrix After linear mapping as key sub-vector and value sub-vector, a spatiotemporal feature matrix is ​​obtained through dot product attention calculation. The output layer uses two-dimensional convolution to predict the pollutant concentration of each station in the future period, specifically: Among them, t represents the time dimension, v represents the variable dimension, K is the kernel function, and the convolution kernel is used to input data Slide up and calculate the weighted sum of the input area covered by the convolution kernel; τ represents the convolution window offset in the time dimension, is the embedding dimension of spatiotemporal spectral features, q is the convolution window offset on the variable dimension, and T represents the time window length.

6. The method for predicting multiple air pollutants based on a multi-scale spatiotemporal spectral feature fusion framework as claimed in claim 3, characterized in that: The output layer includes: using two-dimensional convolution to predict the pollutant concentration of each station in the future period.

7. The method for predicting multiple air pollutants based on a multi-scale spatiotemporal spectral feature fusion framework according to claim 1, characterized in that: Step S4 includes: S41: inputting the multivariate feature matrix containing time series information into the multi-scale spatiotemporal feature fusion model for training; S42: Using the MAE loss function to perform error calculation and back propagate until the loss error is minimized, and saving the multi-scale spatiotemporal feature fusion model parameters. The calculation formula of the MAE loss function is as follows: Among them, Y pred is the predicted value, Y true is the true value, and n is the number of samples collected.

8. A storage device, characterized in that: The storage device stores instructions and data for implementing any one of claims 1 to 7 of the method for predicting multiple air pollutants based on a multi-scale spatiotemporal spectral feature fusion framework.

9. A device for predicting multiple air pollutants based on a multi-scale spatiotemporal spectral feature fusion framework, characterized in that: It includes a processor and a storage device; the processor loads and executes instructions and data in the storage device to implement any one of claims 1 to 7 of a method for predicting multiple air pollutants based on a multi-scale spatiotemporal spectral feature fusion framework.

Images