A method for predicting air pollutant concentration and related equipment

Abstract

The present invention discloses a method for predicting air pollutant concentration and related equipment, which relates to the technical field of data processing. The method includes: obtaining the current pollutant concentration data, current meteorological data, and target site geographical location data corresponding to a target environmental monitoring site; using the current pollutant concentration data, current meteorological data, and geographical location data of the target site as target input data for a prediction model. Among them, the input module of the prediction model is used to perform modal decomposition processing on the input data to obtain multiple modal data; the time feature extraction module is used to extract time feature data; the spatial feature extraction module is used to extract spatial feature data; the convolutional network module for spatio-temporal fusion is used to extract spatio-temporal features; the output module is used to output a prediction result; and determining the prediction result of the air pollutant concentration corresponding to the target environmental monitoring site according to the output result. It realizes the effects of improving the comprehensiveness and accuracy of the prediction result and expanding the applicable scenarios.

Description

Technical Field

[0001] This application relates to the technical field of data processing, and particularly to a method for predicting air pollutant concentration and related devices. Background Art

[0002] Air quality is directly related to people's health, quality of life, and the sustainable development of society and economy. By predicting the concentration of air pollutants, it helps to take measures in advance to reduce pollution emissions and protect the environmental quality; and timely warn of high pollution weather to help relevant departments take necessary protective measures to reduce the harm to health.

[0003] The commonly used methods for predicting air pollutant concentration in related technologies mainly include the following: (1) Statistical models, such as establishing a linear relationship based on historical data for prediction; (2) Machine learning models, such as using Random Forest, which improves the prediction accuracy by integrating multiple decision trees. (3) Deep learning models, such as using Long Short-Term Memory (LSTM) networks to process time series data and capture long-term dependencies for prediction. Although the above methods perform well in some aspects, they also have some limitations: Most models only use single-type and single-site data, ignoring multi-type, multi-modal, and multi-site data, and failing to fully consider the correlations in time and space, resulting in less comprehensive prediction results and lower accuracy of the prediction results; and many models use fixed parameter settings, making it difficult to adapt to dynamically changing environmental conditions, restricting the applicable scenarios of the air pollutant concentration prediction scheme; at the same time, important information may be lost or a large amount of unimportant data may be retained during the data preprocessing stage, affecting the prediction effect and the final prediction result.

[0004] In view of the above problems in related technologies, no effective solution has been proposed yet. Summary of the Invention

[0005] The prediction method for air pollutant concentration and related devices provided by the embodiments of the present invention at least solves the problems of poor comprehensiveness and accuracy of the prediction results and narrow applicable scenarios in the prediction methods for air pollutant concentration in related technologies.

[0006] To solve the above problems, one aspect of the embodiments of the present invention provides a method for predicting air pollutant concentration, including:

[0007] Obtaining current pollutant concentration data, current meteorological data, and target site geographical location data corresponding to a target environmental monitoring site;

[0008] Take the current pollutant concentration data, current meteorological data, and target site geographical location data as the target input data of the prediction model; wherein, the prediction model includes an input module, a time feature extraction module, a spatial feature extraction module, a spatio-temporal fusion convolutional network module, and an output module; the input module is used to perform modal decomposition processing on the pollutant concentration data in the input data to obtain multiple modal data; the time feature extraction module is used to extract time feature data from the multiple modal data; the spatial feature extraction module is used to construct an adjacency matrix that fuses the pollutant concentration data, meteorological data, and site location data in the input data, and extract spatial feature data by a graph convolutional network; the spatio-temporal fusion convolutional network module is used to extract spatio-temporal features corresponding to each mode; the output module is used to aggregate the prediction outputs corresponding to multiple modes to output a prediction result;

[0009] Determine the prediction result of the air pollutant concentration corresponding to the target environmental monitoring site according to the output result of the prediction model.

[0010] In some of these embodiments, the method further includes a training step of the prediction model:

[0011] Obtain the historical pollutant concentration data, historical meteorological data, and site geographical location data corresponding to multiple environmental monitoring sites respectively;

[0012] Through the input module of the prediction model, use the variational mode decomposition algorithm to perform variational mode decomposition on the historical pollutant concentration data of each site, use the particle swarm optimization algorithm to optimize the variational mode decomposition result, obtain multiple modal data corresponding to each site respectively, and perform data normalization processing on the multiple modal data;

[0013] Extract time feature data from the multiple modal data corresponding to each site respectively through the temporal convolutional network in the time feature extraction module;

[0014] Input the historical pollutant concentration data, historical meteorological data, and site geographical location data of each site into the spatial feature extraction module, construct the adjacency matrix in the graph convolutional network, and extract spatial feature data through the graph convolutional network;

[0015] Input the time feature data and the spatial feature data into the convolutional gated recurrent unit in the spatio-temporal fusion convolutional network module to fuse the time feature data and the spatial feature data, and introduce an attention mechanism to obtain spatio-temporal features corresponding to different modes; wherein, the fully connected layers of the spatio-temporal fusion convolutional network module are respectively connected to the spatio-temporal features of a single mode, and output the prediction outputs of the air pollutant concentration corresponding to each mode of each site;

[0016] The predicted outputs corresponding to all modalities are aggregated through an output module to obtain predicted concentration output feature data, and the predicted concentration output feature data is de-normalized to obtain the prediction result.

[0017] In some of these embodiments, after the step of obtaining the historical pollutant concentration data, historical meteorological data, and site geographical location data corresponding to multiple environmental monitoring stations respectively, the method further includes:

[0018] Preprocess the historical pollutant concentration data and historical meteorological data corresponding to multiple environmental monitoring stations respectively through an input module; wherein, the preprocessing includes missing value processing and outlier processing;

[0019] Calculate the mutual information value between the historical pollutant concentration data and the historical meteorological data, sort the mutual information value, and select the historical pollutant concentration types and meteorological data types that meet the target quantity.

[0020] In some of these embodiments, the prediction model further includes a Mogrifier gating operation module. Before the step of extracting time feature data from the multiple modality data corresponding to each site respectively through the temporal convolutional network in the time feature extraction module, the training step of the prediction model further includes:

[0021] Optimize the multiple modality data corresponding to each site respectively through the Mogrifier gating operation module.

[0022] In some of these embodiments, the method further includes an optimization step of the prediction model:

[0023] Adopt a particle swarm optimization algorithm to search for the optimal parameter configuration of the prediction model to achieve the optimization of the prediction model.

[0024] In some of these embodiments, the optimization step of the prediction model further includes:

[0025] Determine the loss function according to the predicted output result, the actual output result, and the predicted time step value;

[0026] Determine the initial parameters and particle swarm parameters of the prediction model; wherein, the particle swarm parameters include the particle swarm size, the maximum number of iterations, the inertia weight, the acceleration factor, the initial velocity, and the initial position;

[0027] Iteratively update the particle swarm to obtain the velocity and position of each particle, the individual extreme value of each particle, and the global extreme value of the particle swarm after each iterative update;

[0028] Update the output parameters of the prediction model with the position of the particle corresponding to the final global extreme value of the particle swarm obtained when the maximum number of iterations is satisfied to complete the optimization of the prediction model.

[0029] In some of these embodiments, when the method is applied to predict the air pollutant concentration in a target area, the target environmental monitoring site is multiple environmental monitoring sites within the target area;

[0030] When the method is applied to predict the air pollutant concentration of a single or multiple specified environmental monitoring sites, the target environmental monitoring site is the single or multiple specified environmental monitoring sites.

[0031] To solve the above problems, one aspect of the embodiments of the present invention provides a prediction device for air pollutant concentration, including:

[0032] A data acquisition module, configured to acquire the current pollutant concentration data, current meteorological data, and target site geographical location data corresponding to the target environmental monitoring site;

[0033] A model processing module, configured to use the current pollutant concentration data, current meteorological data, and target site geographical location data as the target input data of the prediction model; wherein, the prediction model includes an input module, a time feature extraction module, a spatial feature extraction module, a spatio-temporal fusion convolutional network module, and an output module; the input module is configured to perform modal decomposition processing on the pollutant concentration data in the input data to obtain multiple modal data; the time feature extraction module is configured to extract time feature data (time series features) from the multiple modal data; the spatial feature extraction module is configured to construct an adjacency matrix that fuses the pollutant concentration data, meteorological data, and site location data in the input data, and extract spatial feature data by a graph convolutional network; the spatio-temporal fusion convolutional network module is configured to extract spatio-temporal features corresponding to each mode; the output module is configured to aggregate the prediction outputs corresponding to multiple modes to output a prediction result;

[0034] A prediction result determination module, configured to determine the prediction result of the air pollutant concentration corresponding to the target environmental monitoring site according to the output result of the prediction model.

[0035] In some of these embodiments, the device further includes a prediction model training module, which is configured to: obtain the historical pollutant concentration data, historical meteorological data, and site geographical location data respectively corresponding to a plurality of environmental monitoring stations; perform variational mode decomposition on the historical pollutant concentration data of each station by using the variational mode decomposition algorithm through the input module of the prediction model, perform an optimization process on the variational mode decomposition result by using the particle swarm optimization algorithm to obtain a plurality of modal data respectively corresponding to each station, and perform data normalization processing on the plurality of modal data; extract time feature data from the plurality of modal data respectively corresponding to each station through the temporal convolutional network in the time feature extraction module; input the historical pollutant concentration data, historical meteorological data, and site geographical location data of each station into the spatial feature extraction module to construct an adjacency matrix in the graph convolutional network, and extract spatial feature data through the graph convolutional network; input the time feature data and the spatial feature data into the convolutional gated recurrent unit in the spatio-temporal fusion convolutional network module to fuse the time feature data and the spatial feature data, and introduce an attention mechanism to obtain spatio-temporal features corresponding to different modalities; wherein, the fully connected layer of the spatio-temporal fusion convolutional network module is respectively connected to the spatio-temporal features of a single modality, and outputs the prediction output of the air pollutant concentration corresponding to each modality of each station; aggregate the prediction outputs corresponding to all modalities through the output module to obtain prediction concentration output feature data, and perform anti-normalization processing on the prediction concentration output feature data to obtain the prediction result.

[0036] To solve the above problems, in one aspect of the embodiments of the present invention, there is provided an electronic device, including: a processor, and a memory storing a program, the program including instructions that, when executed by the processor, cause the processor to execute any one of the above air pollutant concentration prediction methods.

[0037] To solve the above problems, in one aspect of the embodiments of the present invention, there is provided a non-transitory machine-readable medium storing computer instructions, the computer instructions being used to cause a computer to execute any one of the above air pollutant concentration prediction methods.

[0038] Advantages of the embodiments of the present invention: By obtaining the current pollutant concentration data, current meteorological data, and target site geographical location data corresponding to the target environmental monitoring site; using the current pollutant concentration data, current meteorological data, and target site geographical location data as the target input data of the prediction model; wherein, the prediction model includes an input module, a time feature extraction module, a spatial feature extraction module, a spatio-temporal fusion convolutional network module, and an output module; the input module is used to perform modal decomposition processing on the pollutant concentration data in the input data to obtain multiple modal data; the time feature extraction module is used to extract time feature data from the multiple modal data; the spatial feature extraction module is used to construct an adjacency matrix that fuses the pollutant concentration data, meteorological data, and site location data in the input data, and extract spatial feature data by means of a graph convolutional network; the spatio-temporal fusion convolutional network module is used to extract spatio-temporal features corresponding to each mode; the output module is used to aggregate the prediction outputs corresponding to multiple modes to output a prediction result; according to the output result of the prediction model, the prediction result of the air pollutant concentration corresponding to the target environmental monitoring site is determined. The technical means overcomes the problems in the related art that the prediction method of air pollutant concentration has poor comprehensiveness and accuracy of prediction results and a narrow applicable scenario. Through modal decomposition and feature extraction in the prediction model, key information is ensured to be retained during the prediction process, enhancing the robustness of the model. The spatio-temporal feature fusion module based on the prediction model enables the prediction method to adapt to more diverse application scenarios, improving its versatility and flexibility. Finally, by aggregating the prediction outputs of multiple modes, the deviation that may be brought by single-mode prediction is reduced, improving the stability and reliability of the final prediction result. Through multi-modal data fusion, time feature extraction, spatial feature extraction, and spatio-temporal feature fusion by the prediction model, the technical effect of significantly improving the comprehensiveness and accuracy of the prediction result and expanding the applicable scenario of the prediction method is achieved.

[0039] Details of one or more embodiments of the present invention are set forth in the following drawings and description to make other features, objects, and advantages of the present invention more concise and understandable. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention, and those of ordinary skill in the art can obtain other embodiments based on these drawings without creative efforts.

[0041] Figure 1 It is a main process schematic diagram of a prediction method for air pollutant concentration in an embodiment of the embodiments of the present invention.

[0042] Figure 2It is a schematic diagram of the main process for training a prediction model in an embodiment of the present invention.

[0043] Figure 3 It is a schematic diagram of the main framework of a prediction device for air pollutant concentration in an embodiment of the present invention.

[0044] Figure 4 It is a schematic diagram of the structure of an electronic device of the present invention. Detailed implementation manners

[0045] Embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although some embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. On the contrary, these embodiments are provided to more thoroughly and completely understand the present invention. It should be understood that the drawings and embodiments of the present invention are only for exemplary purposes and are not used to limit the protection scope of the present invention.

[0046] Common air pollutant concentration prediction methods in related technologies mainly include the following several types:

[0047] (1) Statistical models. For example, CN109657842A discloses a method and device for predicting air pollutant concentration and an electronic device. It first obtains the location information, air pollutant concentration, and collection time of historical air pollutant concentration of several air detection stations adjacent to the target prediction station, and then predicts the air pollutant concentration of the target prediction station in a specified prediction time period based on the data obtained above. However, CN109657842A mainly relies on the historical air pollutant concentration data of adjacent stations, only considers partial spatial correlation, and fails to fully utilize other relevant data (such as meteorological data, geographical location, data of all stations in the region, etc.), and is easily affected by a single data source, resulting in unstable prediction results.

[0048] (2) Machine learning model. For example, CN115327041A discloses a method for predicting air pollutant concentration based on correlation analysis. First, it ranks the correlation strengths between the air pollutant concentrations at adjacent stations and the pollutant concentration at the target station through the Pearson coefficient. Then, the data obtained by combining the pollutant concentrations of several items with the highest correlation and the historical pollutant concentrations at the target station are put into a neural network model for training, so as to explore the impact of the changes in the air pollutant concentrations at adjacent stations on the changes in the pollutant concentrations at the target station, and thus be able to more accurately predict the changes in the air pollutant concentrations at the target station. However, CN109657842A relies on the pollutant concentration data of several items with the highest correlation and the historical pollutant concentration data at the target station, mainly focusing on optimizing the input data based on correlation analysis, and does not fully utilize other relevant data (such as meteorological data, station geographical location, etc.), failing to achieve deep integration in the time and space dimensions. The prediction results are easily affected by a single data source, resulting in unstable prediction results. At the same time, the model parameter settings are relatively fixed and it is difficult to adapt to complex environmental changes.

[0049] (3) Deep learning models. For example, CN114741972A discloses a method for constructing a seasonal prediction model of air pollutant concentration. It divides the historical data of air pollutant concentration into four seasonal historical data sets of spring, summer, autumn and winter according to seasons, then determines the time series of the data sets, and uses a long short-term memory network to learn the variation law of pollutant concentration contained in the time series of air pollutant concentration. At the same time, the Bayesian optimization algorithm is used to select the optimal hyperparameter combination for the long short-term memory network, and then a prediction model with the best prediction performance is obtained. CN113188968A discloses an air pollutant concentration prediction method and system based on a combined deep learning model. It obtains the air pollutant PM2.5 concentration of multiple air quality stations and gets the time series of air pollutant PM2.5; uses this time series to train the constructed combined deep learning network to obtain a prediction model; and then uses this prediction model to predict the predicted value of the air pollutant PM2.5 concentration at the target station. However, neither CN114741972A nor CN113188968A considers other relevant data such as meteorological data and geographical location. CN114741972 only considers feature extraction in the time dimension. Although CN113188968 extracts time and space features through a combined deep learning network, it fails to achieve deep spatio-temporal feature fusion and also has problems of poor comprehensiveness and accuracy of prediction results. Another example is that CN118035777A discloses an air pollutant concentration prediction method and device based on spatio-temporal graph data. This method constructs a data set by obtaining historical data of air pollutant concentration and corresponding meteorological data, and obtains monitoring site distribution data; initializes the MGATT-LSTM prediction model by using a graph convolutional network, a long short-term memory network and a fully connected network based on a multi-graph attention mechanism; trains and iteratively optimizes the MGATT-LSTM prediction model, and when the prediction result meets the preset prediction conditions, the final prediction model is obtained. CN118035777A considers the fusion of time features and space features, but does not consider the modal features of the data, and still has problems of insufficient comprehensiveness and poor accuracy of prediction results.

[0050] In summary, although the air pollutant concentration prediction methods in the related technologies perform well in some aspects, they also have some limitations: Most models only use a single type of data, ignoring multi-type and multi-modal data, and failing to fully consider the relevance in time and space, resulting in insufficient comprehensiveness of prediction results; and many models adopt fixed parameter settings, making it difficult to adapt to dynamic environmental conditions and restricting the applicable scenarios of air pollutant concentration prediction schemes; important information may be lost or a large amount of unimportant data may be retained in the data preprocessing stage, affecting the prediction effect and the final prediction result.

[0051] To solve the above problems, an embodiment of the present invention provides a method for predicting air pollutant concentration, as Figure 1 shown. The method for predicting air pollutant concentration mainly includes:

[0052] Step S101, obtaining current pollutant concentration data, current meteorological data, and target site geographical location data corresponding to a target environmental monitoring site.

[0053] By simultaneously collecting pollutant concentration, meteorological conditions (such as temperature, humidity, wind speed, etc.), and geographical information (such as site location), the integrity and diversity of the input data are ensured, providing a richer information basis for subsequent feature extraction and prediction.

[0054] According to the embodiment of the present invention, by obtaining various types of data, subsequent analysis of the correlation between different factors via a prediction model can be better performed, such as the impact of meteorological conditions on pollutant concentration, thereby improving the accuracy of prediction. At the same time, combining geographical location data helps to capture the spatial correlation between different sites, further enhancing the spatial perception ability of the prediction model.

[0055] Among them, the pollutants in the embodiment of the present invention include but are not limited to sulfur dioxide, nitrogen dioxide, ozone, carbon monoxide, inhalable particulate matter (PM10, referring to particulate matter with an aerodynamic equivalent diameter ≤ 10 microns), and fine particulate matter (PM2.5, that is, particulate matter in the atmosphere with an aerodynamic equivalent diameter less than or equal to 2.5 microns, also known as respirable particulate matter).

[0056] Among them, the data obtained in step S101 of the embodiment of the present invention should be the current latest data, which can reflect the current environmental changes and contribute to improving the timeliness and reliability of the prediction results.

[0057] In some embodiments, when the method for predicting air pollutant concentration provided by the embodiment of the present invention is applied to predict the air pollutant concentration in a target area, the target environmental monitoring site is multiple environmental monitoring sites within the target area; when the prediction method is applied to predict the air pollutant concentration of a single or multiple specified environmental monitoring sites, the target environmental monitoring site is a single or multiple specified environmental monitoring sites.

[0058] When applied to predict the air pollutant concentration in a target area, by simultaneously considering the data of multiple sites, the target area can be comprehensively covered, providing more extensive air quality information. And using the data of multiple sites can better capture the spatial correlation between different positions, improving the accuracy of regional overall prediction. The data of multiple sites also provide more redundant information, enhancing the robustness and reliability of the model and reducing the impact of data anomalies at a single site.

[0059] When applied to predicting the air pollutant concentration at a single or multiple specified environmental monitoring stations, it can provide more accurate and personalized prediction results according to the needs of specific stations, meeting the refined management requirements of local areas. Key stations can be selected for focused monitoring according to actual needs to optimize resource allocation and improve efficiency.

[0060] The prediction model adopted by the air pollutant concentration prediction method provided in the embodiments of the present invention has wide applicability and high prediction accuracy, can meet diverse prediction needs, and ensures the accuracy and comprehensiveness of prediction results by comprehensively utilizing data of different scales and types.

[0061] Step S102: Use the current pollutant concentration data, current meteorological data, and target station geographical location data as the target input data of the prediction model; wherein, the prediction model includes an input module, a time feature extraction module, a spatial feature extraction module, a spatio-temporal fusion convolutional network module, and an output module; the input module is used to perform modal decomposition processing on the pollutant concentration data in the input data to obtain multiple modal data; the time feature extraction module is used to extract time feature data from the multiple modal data; the spatial feature extraction module is used to construct an adjacency matrix that combines the pollutant concentration data, meteorological data, and station location data in the input data, and extract spatial feature data by means of a graph convolutional network; the spatio-temporal fusion convolutional network module is used to extract spatio-temporal features corresponding to each modality; the output module is used to aggregate the prediction outputs corresponding to multiple modalities to output a prediction result.

[0062] Based on the above prediction model, through multi-modal data processing and fusion, it is ensured that the data of each modality can be effectively utilized, avoiding the limitations brought by single-modal data and enhancing the robustness of the model. The time feature extraction module can accurately capture the variation law of pollutant concentration over time through methods such as TCN (Temporal Convolutional Network), enhancing the prediction ability for short-term and long-term trends and improving the prediction accuracy. The spatial feature extraction module constructs an adjacency matrix (the adjacency matrix is a parameter matrix in the graph convolutional network) and applies GCN (Graph Convolutional Network), fully considering the spatial correlation between different stations, improving the understanding of the overall pollution situation in the region and enhancing the prediction accuracy. The spatio-temporal fusion convolutional network module realizes a more comprehensive and in-depth feature representation by fusing features in the time and space dimensions, further improving the performance of the prediction model and adapting to more diverse application scenarios. The output module reduces the deviation that may be brought by single-modal prediction by aggregating the prediction outputs of multiple modalities, improving the robustness and stability of the model and ensuring the high accuracy and reliability of the final prediction result.

[0063] Among them, the above-mentioned TCN is a neural network architecture for processing time series data. Different from traditional recurrent neural networks (such as LSTM and GRU), TCN uses one-dimensional convolutional layers to capture long-term dependencies in time series data while maintaining causality (i.e., future information does not affect current predictions). TCN has the advantage of parallel computing and can process long sequence data more efficiently. The above-mentioned GCN (Graph Convolutional Network) is a neural network architecture specifically for processing graph-structured data. It extracts node features and the relationships between their neighbor nodes by applying convolutional operations on the nodes and edges of the graph. GCN can effectively capture complex relationships and structural information in graph data.

[0064] Based on the above settings, through means such as multi-modal data processing and fusion, time feature extraction, spatial feature extraction, and spatio-temporal feature fusion, the comprehensiveness and accuracy of air pollutant concentration prediction have been significantly improved. Additionally, through a flexible output aggregation mechanism, the robustness and reliability of the prediction results have been enhanced.

[0065] In some of these embodiments, the above method further includes the training step of the prediction model: obtaining historical pollutant concentration data, historical meteorological data, and site geographical location data corresponding to multiple environmental monitoring stations respectively; using the variational mode decomposition algorithm to perform variational mode decomposition on the historical pollutant concentration data of each station through the input module of the prediction model, using the particle swarm optimization algorithm to optimize the variational mode decomposition results, obtaining multiple modal data corresponding to each station respectively, and performing data normalization processing on the multiple modal data; extracting time feature data from the multiple modal data corresponding to each station respectively through the temporal convolutional network in the time feature extraction module; inputting the historical pollutant concentration data, historical meteorological data, and site geographical location data of each station into the spatial feature extraction module, constructing the adjacency matrix in the graph convolutional network, and extracting spatial feature data through the graph convolutional network in the spatial feature extraction module; inputting the time feature data and spatial feature data into the convolutional gated recurrent unit in the spatio-temporal fusion convolutional network module to fuse the time feature data and spatial feature data, and introducing an attention mechanism to obtain spatio-temporal features corresponding to different modalities; where the fully connected layer of the spatio-temporal fusion convolutional network module is respectively connected to the spatio-temporal features of a single modality and outputs the prediction outputs of the air pollutant concentrations corresponding to each modality of each station; aggregating the prediction outputs corresponding to all modalities through the output module to obtain the predicted concentration output feature data, and performing anti-normalization processing on the predicted concentration output feature data to obtain the prediction results.

[0066] According to an embodiment of the present invention, data of multiple environmental monitoring stations are collected. Each station corresponds to different pollutant concentration data, meteorological data, and geographical location data. The prediction model trained based on this has a distributed monitoring and forecasting function. Further, through the input module, multi-modal data processing and optimization are performed. Specifically, through VMD (Variational Mode Decomposition), variational mode decomposition is performed on the historical pollutant concentration data of each station to obtain multiple modes with different frequency characteristics, ensuring that the data of each mode can be effectively utilized. At the same time, through PSO (Particle Swarm Optimization), optimization processing is performed on the VMD result, further optimizing the decomposition result and enhancing the robustness of the model. By performing data normalization processing on the multiple modal data, it is ensured that different modal data are processed on the same scale, avoiding the influence of some modal data with too large magnitude differences on the model performance.

[0067] According to an embodiment of the present invention, time feature data is extracted from the multiple modal data respectively corresponding to each station through the time feature extraction module. Specifically, TCN (Temporal Convolutional Network) can efficiently capture long-term dependencies in time series, enhancing the prediction ability for short-term and long-term trends. And compared with the traditional RNN structure, TCN has the advantage of parallel computing, improving the efficiency of processing long sequence data.

[0068] According to a specific implementation manner of an embodiment of the present invention, when performing VMD (Variational Mode Decomposition) on the historical pollutant concentration data of each site, each mode indeed represents a specific aspect or pattern of the original data. These modes can capture different frequency components, trends, periodic changes, and other hidden features in the data. Specifically, different modes may represent the following specific aspects or patterns: long-term trends (i.e., low-frequency modes usually reflect the long-term trends or slow-changing parts in the data. For pollutant concentration data, low-frequency modes may capture annual or seasonal change trends, such as the trend of higher pollution levels in winter and lower levels in summer), short-term fluctuations (i.e., high-frequency modes, which reflect the impacts of rapid changes or instantaneous events. This may include sharp increases or decreases in pollutant concentrations within a short period due to increased traffic flow, intensified industrial activities, or special weather phenomena such as sandstorms, rainfall, etc.), periodic changes (certain modes may show obvious periodic behaviors, and these periods may be at the daily, weekly, or monthly levels. For pollutant concentration data, it may be the impact of weekly traffic patterns on pollutant emissions), random noise or unexpected events (some modes may mainly consist of random noise or non-periodic unexpected events. This may include measurement errors of monitoring devices, sudden local pollution sources such as fires, accidental emissions, etc.), intermediate-frequency patterns (in addition to extreme high and low frequencies, there are also some intermediate-frequency modes, which are between long-term trends and short-term fluctuations. These modes may capture fluctuations on a relatively long time scale but not as long as inter-annual changes, such as the air quality change trend within a few months), spatial correlation patterns (when considering data from multiple sites, certain modes may reflect the correlations between geographical locations. If two adjacent sites are affected by the same pollution source, then a certain common mode of them may show similar time series characteristics).

[0069] According to the above specific implementation manner, the multiple mode data obtained by performing mode decomposition on the historical pollutant concentration data through VDM can be characterized as follows: (1) Low-frequency mode: Reflecting long-term trends or slow changes, which may reveal the long-term improvement effects after the implementation of environmental policies, or the impact of natural climate cycles on air pollution. (2) High-frequency mode: Reflecting short-term fluctuations or instantaneous events, which can help identify the impact of traffic emissions during daily commuting peak hours, or short-term pollution events under specific weather conditions. (3) Periodic mode: Reflecting periodic changes, such as daily, weekly, and monthly cycles; helping to understand how factors such as diurnal temperature differences and seasonal changes periodically affect air quality. (4) Random noise mode: Reflecting random noise or non-periodic events; providing a way to evaluate the stability and reliability of the monitoring system and excluding the influence of outliers. (5) Spatial correlation mode: Reflecting the influence of spatial positions, capable of showing the pollution propagation paths and interactions between different regions, which is very important for formulating cross-regional environmental protection measures).

[0070] Given the complexity of pollutant concentration data, pollutant concentration data usually has more complex temporal characteristics and higher volatility. These data may contain multiple frequency components (such as long-term trends, short-term fluctuations, periodic changes, etc.), so it is necessary to extract different modal characteristics through VMD decomposition. Through VMD decomposition, not only can complex data sets be simplified into a more understandable and processable form, but also potential important information can be mined, thus significantly improving the performance and accuracy of prediction models.

[0071] According to an embodiment of the present invention, an adjacency matrix integrating historical pollutant concentration data, historical meteorological data, and site location data in the input data is constructed by a spatial feature extraction module, and spatial feature data is extracted through a GCN. Specifically, the GCN fully considers the spatial correlation between different sites, improves the understanding of the overall pollution situation in the region, and the extraction of spatial features enables the prediction model to better handle the cross-site impact, thereby improving the accuracy of the prediction result.

[0072] According to a specific embodiment of an embodiment of the present invention, the time feature data and the spatial feature data can be directly input into the convolutional gated recurrent unit in the spatio-temporal fusion convolutional network module to fuse the time feature data and the spatial feature data, and an attention mechanism is introduced to obtain spatio-temporal features corresponding to different modalities; wherein, the fully connected layers of the spatio-temporal fusion convolutional network module are respectively connected to the spatio-temporal features of a single modality, and the predicted outputs of the air pollutant concentrations corresponding to each modality of each site are output.

[0073] According to another specific embodiment of an embodiment of the present invention, the output of the time convolutional network can be connected to the input of the graph convolutional network based on the spatio-temporal fusion convolutional network module, and then the time feature data and the spatial feature data are fused through the convolutional gated recurrent unit (ConvGRU) in the spatio-temporal fusion convolutional network, and an attention mechanism is introduced to obtain spatio-temporal features corresponding to different modalities. Specifically, by fusing the features in the time and space dimensions, a more comprehensive and in-depth feature representation is achieved, further improving the performance of the prediction model. The introduced attention mechanism can dynamically assign weights according to importance, ensuring that key information receives more attention and improving the prediction accuracy. The fully connected layers of the spatio-temporal fusion convolutional network module are respectively connected to the spatio-temporal features of a single modality, and the predicted outputs of the air pollutant concentrations corresponding to each modality of each site are output. By integrating the spatio-temporal features of different modalities through the fully connected layer, it is ensured that the information of each modality can be fully utilized.

[0074] According to a specific implementation manner of an embodiment of the present invention, after connecting the output of the temporal convolutional network with the input of the graph convolutional network, the spatio-temporal fusion is performed through a convolutional gated recurrent unit (ConvGRU) in the convolutional network module based on spatio-temporal fusion. This ensures that the time series features can be integrated into the spatial features in an orderly manner. Specifically, the TCN has extracted the local temporal features of each site from the time dimension, and these features can be regarded as the state representations of each node at different time points. When these state representations are passed as inputs to the GCN, the GCN can further capture the spatial dependencies between these states in the graph structure. Further, by using a gated recurrent unit such as ConvGRU, the spatial information can be effectively fused while maintaining the time continuity. The update gate and reset gate in the ConvGRU can control which time features should be retained or forgotten, and how to combine the new spatial features. This method helps to dynamically adjust the importance of spatio-temporal features and enhance the model's understanding of complex spatio-temporal patterns. In addition, introducing an attention mechanism can further improve the effect of spatio-temporal feature fusion. The attention mechanism allows the model to automatically adjust the weights at each position or time step according to the requirements of the current task, highlighting the information that is most helpful for prediction. By first letting the TCN process the time series data, then the GCN process the spatial associations, and finally applying the attention mechanism, the whole process can be made more orderly and targeted.

[0075] According to the embodiment of the present invention, by aggregating the prediction outputs corresponding to all modalities through the output module, the final predicted concentration output feature data is obtained, and the final predicted concentration output feature data is subjected to inverse normalization processing to obtain the final prediction result, which improves the comprehensiveness and accuracy of the prediction result. Further, the inverse normalization processing restores the data to the original scale, ensuring the practical interpretability and application value of the prediction result.

[0076] Through the above training steps, it is achieved that through variational mode decomposition and particle swarm optimization algorithm, multiple modal data are finely extracted and optimized, enhancing the robustness and accuracy of the model. By efficiently capturing the long-term dependencies in the time series through the temporal convolutional network, the prediction ability for short-term and long-term trends is enhanced, improving the accuracy of the prediction. By fully considering the spatial correlations between different sites through the graph convolutional network, the understanding of the overall pollution situation in the region is improved, enhancing the accuracy of the prediction result. Through the convolutional gated recurrent unit and the attention mechanism, the feature fusion in the time and space dimensions is realized, further improving the performance of the prediction model and adapting to more diverse application scenarios. Through the fully connected layer and the inverse normalization processing, the comprehensiveness and practical application value of the prediction result are ensured, enhancing the practicality and reliability of the model. The technical effect of significantly improving the comprehensiveness and accuracy of the air pollutant concentration prediction and enhancing the robustness and practicality of the prediction model is achieved.

[0077] In some of these embodiments, after the step of obtaining the historical pollutant concentration data, historical meteorological data, and site geographical location data corresponding to multiple environmental monitoring sites respectively, the above method further includes: preprocessing the obtained historical pollutant concentration data and historical meteorological data corresponding to multiple environmental monitoring sites respectively through an input module; wherein, the preprocessing includes missing value processing and outlier processing; calculating the mutual information value between the historical pollutant concentration data and the historical meteorological data, sorting the mutual information values, and selecting historical pollutant concentration types and meteorological data types that meet the target quantity.

[0078] Among them, the missing value processing provided by the embodiments of the present invention refers to identifying and processing missing values in the data. Common methods include imputation methods (such as mean imputation, linear imputation) or deleting data records containing missing values. Through missing value processing, the data integrity is ensured, the model bias caused by missing values is avoided, and it helps to improve the stability and accuracy of subsequent model training.

[0079] Among them, the outlier processing provided by the embodiments of the present invention identifies and processes outliers (abnormal values). Common methods include statistical detection (such as Z-score, IQR), machine learning methods (such as Isolation Forest), etc. By processing outliers, the outliers are removed or corrected, ensuring the consistency and rationality of the data, reducing the interference of abnormal data on model training, and improving the robustness and generalization ability of the model.

[0080] According to the embodiments of the present invention, the mutual information method can be used to perform correlation analysis on the historical pollutant concentration data and historical meteorological data, and find the top N types of historical pollutant concentrations and historical meteorological data that are closely correlated with the air pollutant concentration to be predicted. Among them, mutual information (MI) is a statistic that measures the mutual dependence between two random variables, and it quantifies the amount of information that can be obtained about one variable by observing the other variable. The mutual information method is widely used in fields such as feature selection, pattern recognition, and signal processing, especially when it is necessary to evaluate the correlation between different data sets.

[0081] Among them, the mutual information value provided by the embodiments of the present invention can quantify the correlation between different features, helping to identify the features that have the greatest influence on the target variable (air pollutant concentration). By sorting the mutual information values, it can be intuitively understood which features have the greatest impact on the prediction result, improving the interpretability of the model. Furthermore, according to the sorting of the mutual information values, historical pollutant concentration types and meteorological data types that meet the target quantity are selected. It can select the most influential features, reduce the data dimension, reduce the computational complexity, and improve the model training efficiency. And while retaining the information most useful for prediction and removing redundant features, it helps to improve the accuracy and stability of the prediction.

[0082] In some of these embodiments, the above prediction model further includes a Mogrifier gating operation module. Before the step of extracting temporal feature data from the multiple modal data respectively corresponding to each site through the temporal convolutional network in the temporal feature extraction module, the training step of the prediction model further includes: optimizing the data of the multiple modal data respectively corresponding to each site through the Mogrifier gating operation module.

[0083] Based on the above settings, the Mogrifier gating operation can dynamically select and retain the information most useful for prediction, avoiding important information that may be lost in traditional preprocessing methods, and filtering out unimportant or redundant data through the gating mechanism, reducing the burden of subsequent processing and improving the computational efficiency.

[0084] According to a specific implementation manner of the embodiment of the present invention, the Mogrifier gating operation module can also perform adaptive adjustment according to the specific characteristics of the data. The Mogrifier gating mechanism automatically determines which parts of the data need to be retained and which can be appropriately modified by learning the gating parameters. This helps to avoid information loss caused by overly simplified or fixed-rule data preprocessing. And the interaction between the forward and backward gating enables the model to gradually optimize the data representation in multiple iterations, ensuring both the integrity of the data and enhancing the feature expression ability. That is, adaptive optimization is performed according to the data characteristics of different sites and different time periods, so that the prediction model can remain stable in the face of abnormal data or noise, not only enhancing the adaptability of the prediction model to complex environmental changes, but also enhancing the reliability of the prediction results. In some of these embodiments, the above method further includes an optimization step of the prediction model: using the particle swarm optimization algorithm to search for the optimal parameter configuration of the prediction model to achieve the optimization of the prediction model.

[0085] By using the particle swarm optimization algorithm (PSO, Particle Swarm Optimization) to optimize the prediction model, the present invention not only significantly improves the optimization level of the model parameter configuration, but also significantly improves the overall performance of the prediction model through means such as global optimal search, precise parameter configuration, improving model performance, and accelerating convergence.

[0086] In some of these embodiments, the optimization steps of the above prediction model further include: determining a loss function according to the predicted output result, the actual output result, and the predicted time step value; determining the initialization parameters and particle swarm parameters of the prediction model; wherein, the particle swarm parameters include the particle swarm size, the maximum number of iterations, the inertia weight, the acceleration factor, the initial velocity, and the initial position; iteratively updating the particle swarm to obtain the velocity and position of each particle, the individual extreme value of each particle, and the global extreme value of the particle swarm after each iterative update; using the position of the particle corresponding to the final global extreme value of the particle swarm obtained when the maximum number of iterations is satisfied to update the output parameters of the prediction model, so as to complete the optimization of the prediction model.

[0087] Through the above settings, the loss function can accurately quantify the difference between the prediction result and the actual result, providing a clear goal for model optimization; by introducing the predicted time step value, the loss function can dynamically reflect the prediction error in different time periods, ensuring that the optimization process takes into account the influence of the time dimension. The setting of the initialization parameters and particle swarm parameters improves the convergence speed and the stability of the optimization process. Then, through multiple iterative updates, the particle swarm gradually approaches the optimal solution, ensuring the progressiveness and effectiveness of the optimization process and avoiding falling into local optimal solutions. Finally, by updating the output parameters of the prediction model, it is ensured that the model performance reaches the best state, significantly improving the prediction accuracy and robustness and enhancing the generalization ability.

[0088] Step S103, determining the prediction result of the air pollutant concentration corresponding to the target environmental monitoring site according to the output result of the prediction model.

[0089] Through the processing of the input data by the foregoing prediction model, it is possible to comprehensively consider multiple factors such as pollutant concentration, meteorological conditions, and geographical location, and provide prediction results with high prediction accuracy, high reliability, and high timeliness. The obtained prediction results have high practical application value and can be used as data support for environmental protection and public health management.

[0090] The above-mentioned method for predicting air pollutant concentration provided by the embodiments of the present invention, by obtaining the current pollutant concentration data, current meteorological data and target site geographical location data corresponding to the target environmental monitoring site; using the current pollutant concentration data, current meteorological data and target site geographical location data as the target input data of the prediction model; wherein, the prediction model includes an input module, a time feature extraction module, a spatial feature extraction module, a spatio-temporal fusion convolutional network module and an output module; the input module is used to perform modal decomposition processing on the pollutant concentration data in the input data to obtain multiple modal data; the time feature extraction module is used to extract time feature data from the multiple modal data; the spatial feature extraction module is used to construct an adjacency matrix that fuses the pollutant concentration data, meteorological data and site location data in the input data, and extract spatial feature data by means of a graph convolutional network; the spatio-temporal fusion convolutional network module is used to extract spatio-temporal features corresponding to each modality; the output module is used to aggregate the prediction outputs corresponding to multiple modalities to output a prediction result; determining the prediction result of the air pollutant concentration corresponding to the target environmental monitoring site according to the output result of the prediction model, overcomes the problems in the related art that the prediction method of air pollutant concentration has poor comprehensiveness and accuracy of the prediction result and a narrow applicable scenario, realizes that through modal decomposition and feature extraction in the prediction model, key information is ensured to be retained during the prediction process, enhances the robustness of the model, and enables the prediction method to adapt to more diverse application scenarios based on the spatio-temporal feature fusion module of the prediction model, improves its versatility and flexibility, and finally reduces the deviation that may be brought by single-modal prediction by aggregating the prediction outputs of multiple modalities, improves the stability and reliability of the final prediction result, and through multi-modal data fusion, time feature extraction, spatial feature extraction and spatio-temporal feature fusion by the prediction model, achieves the technical effects of significantly improving the comprehensiveness and accuracy of the prediction result and expanding the applicable scenario of the prediction method.

[0091] The embodiments of the present invention also provide a specific embodiment for training and optimizing a prediction model, as Figure 2 shown, the training and optimization of this prediction model mainly includes the following steps:

[0092] Step S201, establish the input module of the prediction model, and collect the hourly historical pollutant concentration data, historical meteorological data and site geographical location data of environmental monitoring stations in various places.

[0093] Step S202: Preprocess the historical pollutant concentration data and historical meteorological data; use the mutual information method to perform a correlation analysis on the historical pollutant concentration data and historical meteorological data, and find the top N types of historical pollutant concentration data and historical meteorological data that are closely correlated with the air pollutant concentration to be predicted; use the variational mode decomposition algorithm and the particle swarm optimization algorithm to perform mode decomposition processing and optimization processing on the historical pollutant concentration data of each station, obtain multiple mode data, and perform standard normalization on each mode data.

[0094] Among them, the role of PSO-VMD (variational mode decomposition - particle swarm optimization algorithm) is to reduce the complexity of pollutant concentration data, improve the prediction accuracy of the prediction model, and enhance the adaptability of input data.

[0095] The specific steps of using PSO-VMD (variational mode decomposition - particle swarm optimization method) to perform mode decomposition processing and optimization processing on the historical pollutant concentration data of each station are as follows: use VMD (variational mode decomposition algorithm) through the input module to perform variational mode decomposition on the historical pollutant concentration data of each station, and use PSO (particle swarm optimization algorithm) to perform optimization processing on the variational mode decomposition results to obtain multiple mode data corresponding to each station respectively.

[0096] Among them, the characteristics of multiple mode data are expressed as: , where k is the number of modes; after performing standard normalization on each mode data, the normalized mode data is denoted as: , where k is the number of modes.

[0097] According to a specific embodiment of the present invention, PSO-VMD (variational mode decomposition - particle swarm optimization algorithm) is used to perform mode decomposition processing and optimization processing on the historical pollutant data of each station to obtain the characteristic representations of N different modes , where k is the number of modes, and specifically includes the following sub-steps:

[0098] Step S2021: Establish a VMD optimization problem and construct the following expression:

[0099] ,

[0100] (1)

[0101] In the above formula (1), is the mode component after the historical pollutant data is decomposed by VMD; is the center frequency of each mode component; is the impulse function; is the gradient calculation; is the original signal.

[0102] Step S2022, introduce a secondary penalty factor , transform the constrained variational problem into an unconstrained variational problem, and the expression is Equation (2):

[0103] (2).

[0104] Step S2023, use PSO to obtain the optimal solution , .

[0105] Step S2024, initialize the particle swarm parameters and optimization parameters , .

[0106] Step S2025, start updating and iterating the particle swarm from k = 1, obtain the updated velocity and position of each particle, and update the individual extreme value of the particle and the global extreme value of the particle swarm according to Step S2027 until the optimization residual , end the update iteration, and obtain the global extreme value of the final particle swarm.

[0107] Step S2026, update the input parameters of the reconstructed VMD model using the position of the particle corresponding to the global extreme value of the final particle swarm, and input the updated input parameters into the VDM model. At this time, the output of the VMD model is recorded as the best layer mode of the input original signal.

[0108] Step S2027, the process of updating and iterating the particle swarm in the above Step S2025 is as follows: k is the current iteration number of the particle swarm, and the current position of the i-th particle after the k-th iteration is denoted as , , the current velocity of the i-th particle after the k-th iteration is denoted as , ; the particle updates its own velocity and position according to the following formulas (3) and (4): (3)

[0109] (4)

[0110] In the above formulas (3) and (4), is the inertia weight; is the velocity of the particle; and are non-negative constants, called acceleration factors; and are random numbers distributed between [0, 1]; ; ; represents the minimum velocity of the i-th particle represents the maximum velocity of the i-th particle.

[0111] Step S2028, when the residual index is satisfied, the optimization terminates, and the hierarchical signals of different modalities are finally output , where k is the number of modalities.

[0112] Step S2029, on the premise that the decomposition layer number K is determined, the calculation process of the residual index is as follows:

[0113] (5)

[0114] In the above formula (5), are the decomposed modalities, is the original signal.

[0115] Step S203, establish a Mogrifier gating mechanism for the input data, and optimize the data of multiple modalities corresponding to each site through the Mogrifier gating operation module.

[0116] By establishing a Mogrifier mechanism for the input data information and establishing a Mogrifier gating operation, the information loss caused by the data preprocessing in Step S202 can be prevented.

[0117] According to a specific implementation manner of an embodiment of the present invention, the specific implementation steps of establishing a Mogrifier gating mechanism for the input data include:

[0118] Perform a Mogrifier gating operation on the input and the output of the previous hidden layer as follows:

[0119] , if t is odd (6)

[0120] , if t is even (7)

[0121] In the above formulas (6) and (7), and are the current input value and the intermediate value in the Mogrifier structure iteration respectively, , ; and are the output of the hidden layer and the intermediate value in the Mogrifier structure, ; is the sigmoid operation; t is a hyperparameter in the Mogrifier structure; 、 is to help , A matrix established additionally for interaction is the Hadamard product.

[0122] Step S204: Establish a time feature extraction module, and extract time feature data from multiple modal data respectively corresponding to each site through a temporal convolutional network in the time feature extraction module.

[0123] According to a specific implementation manner of an embodiment of the present invention, the steps of extracting time features of input data by using multi-scale TCN operations for different modalities k include:[[]]

[0124] (8)

[0125] (9)

[0126] In the above formulas (8) and (9), is the t-th element of the input sequence; is the dilation rate, which varies with the modality; is the convolutional kernel size; is a trainable parameter; is the tanh activation function; is the output of the time feature extraction module for the k-th modality; represents a one-dimensional convolutional layer with a convolutional kernel of 1×1.

[0127] Step S205: Establish a spatial feature extraction module, input the historical pollutant concentration data, historical meteorological data of the first N types and the corresponding site geographical location data obtained in step S202 into the spatial feature extraction module, construct an adjacency matrix in the graph convolutional network, and extract spatial feature data through the graph convolutional network.

[0128] The adjacency matrix is a very important concept in graph theory and graph convolutional networks (GCN), which is used to represent the connection relationship between nodes in a graph.

[0129] According to a specific implementation manner of an embodiment of the present invention, the steps of constructing an adjacency matrix and extracting spatial feature data through a graph convolutional network include:[[]]

[0130] Step S2051: Construct an adaptive graph convolutional network GCN and an adjacency matrix A according to the distribution of monitoring stations.

[0131] Step S2052: Represent all stations and their pairwise relationships as a weighted graph G=(V, E, A), where V is the set of stations, including N stations; E is the set of edges; is the adjacency matrix, where the weight represents the station Vi and site V j The spatio-temporal correlation strength between them, and the edges are connected according to the weights, , .

[0132] Step S2053, establish an adjacency matrix A ( ), and the matrix A integrates pollutant concentration data, meteorological data and site location data.

[0133] Step S2054, define the temporal data adjacency matrix of pollutant concentration data as:

[0134] (10)

[0135] In the above formula (10), represents the adjacency matrix of the temporal similarity graph of particulate pollutants; respectively represent the temporal data of the concentrations of N pollutants at the environmental monitoring station ; represents the average DTW distance between the temporal data of the pollutant concentrations at the environmental monitoring station ; represents the variance of the Gaussian distribution of the site pollutant data; represents the distance threshold of the set time pattern of the pollutant concentration.

[0136] Step S2055, define the adjacency matrix of meteorological information between sites as:

[0137] (11)

[0138] In the above formula (11), represents the adjacency matrix of the temporal similarity graph of meteorological information; are respectively the meteorological data of the environmental monitoring station ; represents the average DTW distance between the temporal data of the meteorology at the environmental monitoring station ; represents the variance of the Gaussian distribution of the site meteorological data; represents the distance threshold of the set time pattern of the meteorological data.

[0139] Step S2056, define the adjacency matrix of spatial information:

[0140] The geographical location of site is represented as , and the geographical location of site is represented as , where respectively represent the latitude and longitude of site ; similarly, respectively represent the latitude and longitude of the site , and the geographical distance between sites is:

[0141] (12)

[0142] (13)

[0143] (14)

[0144] (15)

[0145] (16)

[0146] In the above formulas (12)-(16), represents the conversion function between degrees and radians, and R is the radius of the Earth's equator.

[0147] The spatio-temporal adaptive matrix E that fuses the temporal information of pollutant concentration data, meteorological data, and the spatial information between sites:

[0148] (17)

[0149] In the above formula (17), , , , are trainable parameters, is the activation function.

[0150] Take the adjacency matrix A as:

[0151] (18)

[0152] In the above formula (18), E represents the spatio-temporal adaptive matrix, and A represents the spatio-temporal adaptive adjacency matrix.

[0153] Step S2057, for each modality k, establish a graph convolutional network GCN expressed as:

[0154] (19)

[0155] can be decomposed into the following operation rules:

[0156] (20)

[0157] (21)

[0158] (22)

[0159] In the above formulas (20)-(22), A is the adjacency matrix of the graph structure; is the identity matrix; is the adjacency matrix with self-connections; represents the degree matrix of the adjacency matrix with self-connections; represents the eigenvector of the k-th mode and the l-th layer of the graph nodes; is the eigenvector of the nodes in the (l + 1)-th layer after convolution of the k-th mode; when l = 0, ; represents the learnable parameters of the l-th layer convolution; is the sigmoid activation function.

[0160] Step S206: Input the time feature data and the spatial feature data into the convolutional gated recurrent unit in the spatio-temporal fusion convolutional network module to fuse the time feature data and the spatial feature data, and introduce the attention mechanism to obtain the spatio-temporal features corresponding to different modes; among them, the fully connected layers of the spatio-temporal fusion convolutional network module are respectively connected to the spatio-temporal features of a single mode, and the predicted outputs of the air pollutant concentrations corresponding to each mode of each station are output.

[0161] Specifically, in the above steps, a spatio-temporal fusion convolutional network model, the TSNN model, needs to be established. That is, the time convolutional network (TCN) and the graph convolutional network (GCN) are fused through the update gate and the reset gate, and the spatio-temporal features of a single mode are connected through the attention transfer mechanism and the fully connected layer FC to obtain the convolutional network TSNN model for predicting the air pollutant concentration with spatio-temporal feature coupling.

[0162] According to a specific implementation manner of an embodiment of the present invention, a spatio-temporal fusion convolutional network model, the TSNN model, is established, which specifically includes the following sub-steps:

[0163] Step S2061: The spatio-temporal convolutional network TSNN model for the air quality pollutant concentration obtains the outputs of the time feature extraction convolutional network TCN and the graph convolutional network GCN, inputs them into the convolutional gated recurrent unit ConvGRU model for spatio-temporal fusion, and processes the data output by the GCN using the update gate in the structure of the convolutional gated recurrent unit ConvGRU model according to the following formula (23) to extract the temporal features of the historical data:

[0164] (23)

[0165] In the above formula (23), , are trainable weights; is the bias term; is the sigmoid function; is the state of a hidden layer in the graph convolutional network. When t = 0, is the output data of the input module after the Mogrifier gating operation.

[0166] Step S2062, use the reset gate in the convolutional gated recurrent unit (ConvGRU) model structure according to the following formula (24) to further process the data transmitted from the previous moment and further extract the temporal features of historical data:

[0167] (24)

[0168] In the above formula (24), 、 are trainable weights; is the bias term; is the sigmoid function; is the output of the k-th modal time feature extraction module; is the state of a hidden layer in the graph convolutional network. When t = 0, is the output data of the input module after the Mogrifier gating operation.

[0169] Step S2063, process the data transmitted from the previous moment and to obtain the current memory content in the ConvGRU. The calculation formula is as follows:

[0170] (25)

[0171] In the above formula (25), 、 、 are trainable weights; is the state of a hidden layer in the graph convolutional network. When t = 0, is the output data of the input module after the Mogrifier gating operation; is the Hadamard product; tanh is a hyperbolic tangent activation function.

[0172] Step S2064, integrate the calculation results of formulas (24) to (25) to obtain the final hidden layer state information at the current moment. The calculation formula is as follows:

[0173] (26)

[0174] In the above formula (26), is a trainable parameter; is the output feature of spatio-temporal fusion.

[0175] Step S2065, by introducing the following attention mechanisms as shown in equations (27)-(29), the global spatio-temporal features of multiple environmental monitoring stations are obtained:

[0176] (27)

[0177] (28)

[0178] (29)

[0179] In the above equations (27)-(29), 、 are weight vectors; is the attention coefficient; is the global spatio-temporal feature of the k-th modality.

[0180] Step S2066, the fully connected output module in Step S206 connects the global spatio-temporal features output by TSNN through the fully connected layer FC to output the predicted output of the pollutant concentration of the k-th modality of the site 。

[0181] Step S207, the predicted outputs corresponding to all modalities are aggregated through the output module to obtain the predicted concentration output feature data, and the predicted concentration output feature data is de-normalized to obtain the prediction result.

[0182] Specifically, the above steps need to first construct the output layer of the prediction model, then aggregate the predicted outputs of the k modalities to predict the concentration output feature data of all sites at T time steps (for example, in the next 24 hours, one time step per hour, i.e., T = 24), and de-normalize the output feature data to obtain the prediction result.

[0183] According to a specific embodiment of the present invention, the output layer of the prediction model is established. Aggregate the predicted outputs of the k modalities: (30)

[0184] In the above equation (30), represents the predicted results of the air pollutant concentrations of the k modalities, is the weight aggregation function.

[0185] For perform de-normalization to obtain the predicted output of the air pollutant concentration 。

[0186] Step S208, the particle swarm optimization algorithm is used to search for the optimal parameter configuration of the prediction model to optimize the prediction model.

[0187] Using the Particle Swarm Optimization (PSO) algorithm to search for the uncertain parameters in the prediction model helps improve the prediction accuracy, prevent falling into local optima, and obtain a prediction model that meets the requirements.

[0188] According to a specific embodiment of the present invention, the training of the uncertain parameters of the prediction model composed of an input module, a TCN module, a GCN module, and a TSNN module using the Particle Swarm Optimization (PSO) algorithm includes the following steps:

[0189] Step S2081, determine the loss function:

[0190] (31)

[0191] In the above formula (31), represents the predicted output; represents the actual output; represents the predicted time step.

[0192] Step S2082, determine the optimization initialization parameters and the PSO initialization parameters.

[0193] Step S2083, start updating and iterating the particle swarm from k = 1, obtain the updated velocity and position of each particle, and update the individual extreme value of the particle and the global extreme value of the particle swarm until , end the update iteration, and obtain the global extreme value of the final particle swarm.

[0194] Step S2084, update the input parameters of the reconstructed prediction model using the position of the particle corresponding to the global extreme value of the final particle swarm, and input the updated input parameters into the prediction model. At this time, the output of the prediction model is the predicted output of the input pollutant concentration.

[0195] The process of updating and iterating the particle swarm in the above step S2083 is specifically as follows: k is the current iteration number of the particle swarm. Denote the current position of the i-th particle after the k-th iteration as , , denote the current velocity of the i-th particle after the k-th iteration as , ; The particle updates its own velocity and position according to the following formulas (32), (33): (32)

[0196] (33)

[0197] In the above formulas (32), (33), is the inertia weight; the maximum value is , is the velocity of the particle; , is a non - negative constant, called the acceleration factor; , is a random number distributed between [0, 1]; ; ; represents the minimum velocity of the i - th particle, represents the maximum velocity of the i - th particle.

[0198] To solve the above problems, an embodiment of the present invention also provides a prediction device for air pollutant concentration, as Figure 3 shown. The prediction device 300 for air pollutant concentration mainly includes:

[0199] A data acquisition module 301, configured to acquire current pollutant concentration data, current meteorological data, and target site geographical location data corresponding to a target environmental monitoring site.

[0200] By simultaneously collecting pollutant concentration, meteorological conditions (such as temperature, humidity, wind speed, etc.) and geographical information (such as site location), the integrity and diversity of the input data are ensured, providing a richer information basis for subsequent feature extraction and prediction.

[0201] In some embodiments, when the prediction device for air pollutant concentration provided by the embodiment of the present invention is applied to predict the air pollutant concentration in a target area, the target environmental monitoring site is multiple environmental monitoring sites within the target area; when the prediction device is applied to predict the air pollutant concentration of a single or multiple specified environmental monitoring sites, the target environmental monitoring site is a single or multiple specified environmental monitoring sites.

[0202] The prediction model adopted by the prediction device for air pollutant concentration provided by the embodiment of the present invention has wide applicability and high prediction accuracy, can meet diverse prediction requirements, and ensures the accuracy and comprehensiveness of the prediction results by comprehensively using data of different scales and types.

[0203] The model processing module 302 is configured to use the current pollutant concentration data, the current meteorological data, and the target site geographical location data as the target input data of the prediction model. Among them, the prediction model includes an input module, a time feature extraction module, a spatial feature extraction module, a spatio-temporal fusion convolutional network module, and an output module. The input module is configured to perform modal decomposition processing on the pollutant concentration data in the input data to obtain multiple modal data. The time feature extraction module is configured to extract time feature data (time series features) from the multiple modal data. The spatial feature extraction module is configured to construct an adjacency matrix that fuses the pollutant concentration data, meteorological data, and site location data in the input data, and extract spatial feature data by means of a graph convolutional network. The spatio-temporal fusion convolutional network module is configured to extract spatio-temporal features corresponding to each modality. The output module is configured to aggregate the prediction outputs corresponding to multiple modalities to output a prediction result.

[0204] In some of the embodiments, the above air pollutant concentration prediction device 300 further includes a prediction model training module, which is configured to: obtain the historical pollutant concentration data, historical meteorological data, and site geographical location data respectively corresponding to multiple environmental monitoring stations; perform variational modal decomposition on the historical pollutant concentration data of each station by using the variational modal decomposition algorithm through the input module of the prediction model, perform optimization processing on the variational modal decomposition result by using the particle swarm optimization algorithm, obtain multiple modal data respectively corresponding to each station, and perform data normalization processing on the multiple modal data; extract time feature data from the multiple modal data respectively corresponding to each station through the temporal convolutional network in the time feature extraction module; input the historical pollutant concentration data, historical meteorological data, and site geographical location data of each station into the spatial feature extraction module to construct an adjacency matrix in the graph convolutional network, and extract spatial feature data through the graph convolutional network; input the time feature data and the spatial feature data into the convolutional gated recurrent unit in the spatio-temporal fusion convolutional network module to fuse the time feature data and the spatial feature data, and introduce an attention mechanism to obtain spatio-temporal features corresponding to different modalities. Among them, the fully connected layers of the spatio-temporal fusion convolutional network module are respectively connected to the spatio-temporal features of a single modality, and output the prediction outputs of the air pollutant concentration corresponding to each modality of each station; perform aggregation processing on the prediction outputs corresponding to all modalities through the output module to obtain prediction concentration output feature data, and perform anti-normalization processing on the prediction concentration output feature data to obtain a prediction result.

[0205] Through the above prediction model training module, by means of variational mode decomposition and particle swarm optimization algorithm, multiple modal data are finely extracted and optimized, enhancing the robustness and accuracy of the model. By means of the temporal convolutional network, long-term dependencies in the time series are efficiently captured, enhancing the prediction ability for short-term and long-term trends and improving the prediction accuracy. By means of the graph convolutional network, the spatial correlation between different stations is fully considered, improving the understanding of the overall regional pollution situation and enhancing the accuracy of the prediction results. Through the convolutional gated recurrent unit and attention mechanism, feature fusion in the time and space dimensions is achieved, further enhancing the performance of the prediction model and adapting to more diverse application scenarios. Through the fully connected layer and inverse normalization processing, the comprehensiveness and practical application value of the prediction results are ensured, enhancing the practicality and reliability of the model. The technical effect of significantly enhancing the comprehensiveness and accuracy of air pollutant concentration prediction and enhancing the robustness and practicality of the prediction model is achieved.

[0206] In some of the embodiments, the above air pollutant concentration prediction device 300 further includes a data processing module. After the step of obtaining the historical pollutant concentration data, historical meteorological data, and station geographical location data respectively corresponding to multiple environmental monitoring stations, the above data processing module is configured to: preprocess the obtained historical pollutant concentration data and historical meteorological data respectively corresponding to multiple environmental monitoring stations through an input module; wherein, the preprocessing includes missing value processing and outlier processing; calculate the mutual information value between the historical pollutant concentration data and the historical meteorological data, sort the mutual information values, and select the historical pollutant concentration types and meteorological data types that meet the target quantity.

[0207] Among them, the mutual information value provided by the embodiments of the present invention can quantify the correlation between different features, helping to identify the features that have the greatest influence on the target variable (air pollutant concentration). By sorting the mutual information values, it is possible to intuitively understand which features have the greatest influence on the prediction results, improving the interpretability of the model. Furthermore, according to the sorting of the mutual information values, the historical pollutant concentration types and meteorological data types that meet the target quantity are selected. The most influential features can be selected, reducing the data dimension, reducing the computational complexity, and improving the model training efficiency. And while retaining the information most useful for prediction and removing redundant features, it helps to improve the accuracy and stability of the prediction.

[0208] In some of the embodiments, the above prediction model further includes a Mogrifier gating operation module. Before the step of extracting temporal feature data from the multiple modal data respectively corresponding to each station through the temporal convolutional network in the temporal feature extraction module, the above prediction model training module is further configured to: optimize the multiple modal data respectively corresponding to each station through the Mogrifier gating operation module.

[0209] Based on the above settings, through the Mogrifier gating operation, it is possible to dynamically select and retain the information that is most useful for prediction, avoiding important information that may be lost in traditional preprocessing methods, and filtering out unimportant or redundant data through the gating mechanism, reducing the burden of subsequent processing and improving the computing efficiency.

[0210] In some of these embodiments, the above-mentioned prediction device 300 for air pollutant concentration further includes an optimization module for the prediction model, which is used to: search for the optimal parameter configuration of the prediction model by using the particle swarm optimization algorithm to achieve the optimization of the prediction model.

[0211] By using the particle swarm optimization algorithm (PSO, Particle Swarm Optimization) to optimize the prediction model, the present invention not only significantly improves the optimization level of the model parameter configuration, but also significantly improves the overall performance of the prediction model by means of global optimal search, accurate parameter configuration, improving model performance, and accelerating convergence.

[0212] In some of these embodiments, the above-mentioned optimization module for the prediction model is further used to: determine the loss function according to the prediction output result, the actual output result, and the prediction time step value; determine the initial parameters and particle swarm parameters of the prediction model; where the particle swarm parameters include the particle swarm size, the maximum number of iterations, the inertia weight, the acceleration factor, the initial velocity, and the initial position; perform iterative updates on the particle swarm to obtain the velocity and position of each particle, the individual extreme value of each particle, and the global extreme value of the particle swarm after each iterative update; use the position of the particle corresponding to the final global extreme value of the particle swarm obtained when the maximum number of iterations is satisfied to update the output parameters of the prediction model to complete the optimization of the prediction model.

[0213] Through the above settings, based on the loss function, the difference between the prediction result and the actual result can be accurately quantified, providing a clear goal for model optimization; by introducing the prediction time step value, the loss function can dynamically reflect the prediction error in different time periods, ensuring that the optimization process takes into account the influence of the time dimension. The setting of the initial parameters and particle swarm parameters improves the convergence speed and the stability of the optimization process. Then, through multiple iterative updates, the particle swarm gradually approaches the optimal solution, ensuring the gradualness and effectiveness of the optimization process and avoiding falling into local optimal solutions. Finally, by updating the output parameters of the prediction model, the model performance is ensured to reach the best state, significantly improving the prediction accuracy and robustness and enhancing the generalization ability.

[0214] The prediction result determination module 303 is used to determine the prediction result of the air pollutant concentration corresponding to the target environmental monitoring site according to the output result of the prediction model.

[0215] Through the processing of the input data by the aforementioned prediction model, it is possible to comprehensively consider multiple factors such as pollutant concentration, meteorological conditions, and geographical location, and provide prediction results with high prediction accuracy, high reliability, and high timeliness. The obtained prediction results have high practical application value and can be used as data support for environmental protection and public health management.

[0216] The above prediction device for air pollutant concentration provided by the embodiments of the present invention includes a data acquisition module for acquiring current pollutant concentration data, current meteorological data, and target site geographical location data corresponding to a target environmental monitoring site; a model processing module for using the current pollutant concentration data, current meteorological data, and target site geographical location data as target input data of a prediction model. The prediction model includes an input module, a time feature extraction module, a spatial feature extraction module, a spatio-temporal fusion convolutional network module, and an output module. The input module is used to perform modal decomposition processing on the pollutant concentration data in the input data to obtain multiple modal data. The time feature extraction module is used to extract time feature data from the multiple modal data. The spatial feature extraction module is used to construct an adjacency matrix that fuses the pollutant concentration data, meteorological data, and site location data in the input data, and extract spatial feature data by means of a graph convolutional network. The spatio-temporal fusion convolutional network module is used to extract spatio-temporal features corresponding to each modality. The output module is used to aggregate the prediction outputs corresponding to multiple modalities to output a prediction result. The prediction result determination module is used to determine the prediction result of the air pollutant concentration corresponding to the target environmental monitoring site according to the output result of the prediction model. Through the modal decomposition and feature extraction in the prediction model, it is ensured that key information is retained during the prediction process, enhancing the robustness of the model. The spatio-temporal feature fusion module based on the prediction model enables the prediction method to adapt to more diverse application scenarios, improving its versatility and flexibility. Finally, by aggregating the prediction outputs of multiple modalities, the deviation that may be brought by single-modal prediction is reduced, and the stability and reliability of the final prediction result are improved. Through multi-modal data fusion, time feature extraction, spatial feature extraction, and spatio-temporal feature fusion by the prediction model, the comprehensiveness and accuracy of the prediction result are significantly improved, and the technical effect of expanding the applicable scenarios of the prediction method is achieved.

[0217] The embodiments of the present invention also provide a non-transitory machine-readable medium storing a computer program, wherein the computer program, when executed by a processor of a computer, is used to cause the computer to execute the method of the embodiments of the present invention.

[0218] The embodiments of the present invention also provide a computer program product, including a computer program, wherein the computer program, when executed by a processor of a computer, is used to cause the computer to execute the method of the embodiments of the present invention. The computer program product should be understood as a software product that mainly implements the above method of the present application through the computer program.

[0219] An embodiment of the present invention further provides an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor. The memory stores a computer program executable by the at least one processor, and when the computer program is executed by the at least one processor, it is used to cause the electronic device to execute the method of the embodiment of the present invention.

[0220] Refer to Figure 4 , and now a block diagram of an electronic device that can be a server or a client as an embodiment of the present invention will be described. It is an example of a hardware device that can be applied to various aspects of the present invention. The electronic device is intended to represent various forms of digital electronic computer devices, such as, a laptop computer, a desktop computer, a workbench, a personal digital assistant, a server, a blade server, a mainframe computer, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as, a personal digital processor, a cellular phone, a smart phone, a wearable device, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present invention described herein and / or claimed.

[0221] As Figure 4 shown, the electronic device includes a computing unit 401, which can execute various appropriate actions and processes according to a computer program stored in a read-only memory ROM 402 or a computer program loaded from a storage unit 408 into a random access memory RAM 403. In the RAM 403, various programs and data required for the operation of the electronic device can also be stored. The computing unit 401, the ROM 402, and the RAM 403 are connected to each other through a bus 404. An input / output I / O interface 405 is also connected to the bus 404.

[0222] Multiple components in the electronic device are connected to the I / O interface 405, including: an input unit 406, an output unit 407, a storage unit 408, and a communication unit 409. The input unit 406 can be any type of device capable of inputting information into the electronic device. The input unit 406 can receive input digital or character information, and generate key signal inputs related to the user settings and / or function controls of the electronic device. The output unit 407 can be any type of device capable of presenting information, and can include but is not limited to a display, a speaker, a video / audio output terminal, a vibrator, and / or a printer. The storage unit 408 can include but is not limited to a magnetic disk, an optical disk. The communication unit 409 allows the electronic device to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks, and can include but is not limited to a modem, a network card, an infrared communication device, and / or a wireless communication transceiver, such as a Bluetooth device, a WiFi device, a WiMax device, a cellular communication device, and / or the like.

[0223] The computing unit 401 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 401 include but are not limited to a CPU, a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing units, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 401 executes the various methods and processes described above. For example, in some embodiments, the method embodiments of the present invention can be implemented as a computer program, which is tangibly contained in a machine-readable medium, such as the storage unit 408. In some embodiments, part or all of the computer program can be loaded and / or installed onto the electronic device via the ROM 402 and / or the communication unit 409. In some embodiments, the computing unit 401 can be configured to execute the above methods in any other suitable manner (e.g., by means of firmware).

[0224] The computer program for implementing the method of the embodiments of the present invention can be written in any combination of one or more programming languages. These computer programs can be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing devices, such that when the computer program is executed by the processor or controller, the functions / operations specified in the flowchart and / or block diagram are implemented. The computer program can be executed entirely on the machine, partially on the machine, executed partially on the machine and partially on a remote machine as an independent software package, or executed entirely on a remote machine or server.

[0225] In the context of embodiments of the present invention, a machine-readable medium can be a tangible medium that can contain or store a program for use by or in connection with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable signal medium can include, but is not limited to, an electronic, magnetic, optical, electromagnetic, or infrared system, apparatus, or device, or any suitable combination of the foregoing. More specific examples of a machine-readable storage medium would include an electrical connection based on one or more wires, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0226] It should be noted that the term "comprising" and its variations used in embodiments of the present invention are open-ended, that is, "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". The modifications of "one" and "a plurality" mentioned in embodiments of the present invention are illustrative rather than restrictive. Those skilled in the art should understand that, unless clearly specified otherwise in the context, it should be understood as "one or more".

[0227] The information / data involved in embodiments of the present invention (including but not limited to information / data for analysis, stored information / data, displayed information / data, etc.) are all information / data authorized by the user or fully authorized by all parties. Moreover, the collection, use, and processing of the relevant information / data need to comply with the relevant laws, regulations, and standards of the relevant countries and regions, and corresponding operation entrances are provided for the user to select authorization or rejection.

[0228] The various steps recorded in the method embodiments provided by embodiments of the present invention can be executed in different orders and / or executed in parallel. In addition, the method embodiments can include additional steps and / or omit the steps shown. The protection scope of the present invention is not limited in this regard.

[0229] As used herein, the term "embodiment" means that the specific features, structures, or characteristics described in connection with an embodiment may be included in at least one embodiment of the present invention. The phrase appears in various places in the specification and does not necessarily mean the same embodiment, nor does it mean independence or alternative to other embodiments that are mutually exclusive. The embodiments in this specification are described in a related manner, and the same or similar parts between the embodiments are cross-referenced. In particular, for the device, equipment, and system embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts refer to the partial description of the method embodiments.

[0230] The above-described embodiments merely represent several implementation manners of the present invention. The description is relatively specific and detailed, but it should not be construed as a limitation on the protection scope. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the appended claims.

Claims

1. A method for predicting air pollutant concentration, characterized in that: include: Obtain the current pollutant concentration data, current meteorological data and geographic location data of the target environmental monitoring site; The current pollutant concentration data, the current meteorological data and the target site geographic location data are used as target input data of the prediction model; wherein the prediction model includes an input module, a time feature extraction module, a spatial feature extraction module, a time-space fusion convolutional network module and an output module; the input module is used to perform modal decomposition processing on the pollutant concentration data in the input data to obtain multiple modal data; the time feature extraction module is used to extract time feature data from the multiple modal data; the spatial feature extraction module is used to construct an adjacency matrix that integrates the pollutant concentration data, meteorological data and site location data in the input data, and extracts spatial feature data by a graph convolutional network; the time-space fusion convolutional network module is used to extract the time-space features corresponding to each modality; the output module is used to aggregate the prediction outputs corresponding to multiple modalities to output the prediction results; Determine the predicted result of the air pollutant concentration corresponding to the target environmental monitoring site according to the output result of the prediction model; The training and optimization of the prediction model includes the steps of: Step S201, establishing an input module of a prediction model, collecting hourly historical pollutant concentration data, historical meteorological data, and station geographic location data from environmental monitoring stations in various places; Step S202, pre-processing the historical pollutant concentration data and the historical meteorological data; using the mutual information method to perform correlation analysis on the historical pollutant concentration data and the historical meteorological data, and finding the top n types of historical pollutant concentration data and historical meteorological data that are closely correlated with the predicted air pollutant concentration; using the variational mode decomposition algorithm and the particle swarm optimization algorithm to perform mode decomposition and optimization processing on the historical pollutant concentration data of each station, to obtain multiple mode data, and standardize each mode data; Step S203, establishing a Mogrifier gating mechanism for the input data, and optimizing the multiple modal data corresponding to each site through the Mogrifier gating operation module; Step S204, establishing a time feature extraction module, and extracting time feature data from a plurality of modal data corresponding to each site through a time convolution network in the time feature extraction module; Step S205, establishing a spatial feature extraction module, inputting the historical pollutant concentration data, historical meteorological data and corresponding site geographic location data of the first n types obtained in step S202 into the spatial feature extraction module, constructing an adjacency matrix in the graph convolution network, and extracting spatial feature data through the graph convolution network; specifically including: Step S2051, constructing an adaptive graph convolutional network GCN and an adjacency matrix A according to the distribution of monitoring stations; Step S2052: All sites and their pairwise relationships are represented as a weighted graph G=(V,E,A), where V is a site set including N sites; E is an edge set; is the adjacency matrix, where the weights Represents the spatiotemporal correlation strength between site Vi and site Vj. The edges are connected according to the weights. , ; Step S2053, establish an adjacency matrix , the matrix A integrates pollutant concentration data, meteorological data and station location data; Step S2054, define the time series data adjacency matrix of pollutant concentration data as: （10） In the above formula (10), an adjacency matrix representing the temporal similarity graph of particle contaminants; Environmental Monitoring Station Time series data of n pollutant concentrations; Environmental Monitoring Station The average DTW distance between the time series data of pollutant concentrations; represents the variance of Gaussian distribution of pollutant data at the site; Indicates the distance threshold of the set pollutant concentration set time mode; Step S2055, define the adjacency matrix of meteorological information between sites as: （11） In the above formula (11), The adjacency matrix representing the temporal similarity graph of meteorological information; Environmental Monitoring Station Meteorological data; Environmental Monitoring Station The average DTW distance between the time series data of the weather; Indicates the variance of Gaussian distribution of meteorological data at the site; Indicates the distance threshold of the set time mode of the meteorological data; Step S2056, define the adjacency matrix of spatial information: Site The geographical location is represented by , site The geographical location is represented by ,in, Respectively represent sites The latitude and longitude of Respectively represent sites Latitude and longitude of the site The geographical distance between for: （12） （13） （14） （15） （16） In the above formulas (12)-(16), Represents the conversion function between angle and radian, R is the equatorial radius of the earth; The spatiotemporal adaptive matrix E that integrates the temporal information of pollutant concentration data, meteorological data, and spatial information between stations: （17） In the above formula (17), , , , is a trainable parameter, is the activation function; Take the adjacency matrix A as: （18） In the above formula (18), E represents the spatiotemporal adaptation matrix, and A represents the adjacency matrix; Step S2057, for each modality k, a graph convolutional network GCN is established and represented as: （19） Decomposed into the following operating rules: （20） （21） （22） In the above equations (20)-(22), A is the adjacency matrix; I is the identity matrix; is an adjacency matrix with self-connection; Degree matrix representing the adjacency matrix with self-connection; The feature vector of the kth mode and the lth layer of the graph node; That is, it is the feature vector of the node in the l+1 layer after the k-mode is convolved; when l=0, ; Represents the learnable parameters of the l-th layer convolution; is the sigmoid activation function; Step S206, inputting the temporal feature data and the spatial feature data into the convolution gated recurrent unit in the spatiotemporal fusion convolutional network module to fuse the temporal feature data with the spatial feature data, and introducing an attention mechanism to obtain spatiotemporal features corresponding to different modalities; wherein the fully connected layer of the spatiotemporal fusion convolutional network module respectively connects the spatiotemporal features of a single modality, and outputs the predicted output of the air pollutant concentration corresponding to each modality of each station; Step S207, aggregate the predicted outputs corresponding to all modes through the output module to obtain predicted concentration output feature data, and perform denormalization on the predicted concentration output feature data to obtain a prediction result.

2. The method for predicting air pollutant concentration according to claim 1, characterized in that: After the step of obtaining historical pollutant concentration data, historical meteorological data, and station geographic location data corresponding to a plurality of environmental monitoring stations, the method further includes: Preprocessing the historical pollutant concentration data and historical meteorological data corresponding to the plurality of environmental monitoring sites respectively obtained by the input module; wherein the preprocessing includes missing value processing and outlier processing; The mutual information values ​​of the historical pollutant concentration data and the historical meteorological data are calculated, the mutual information values ​​are sorted, and the historical pollutant concentration types and meteorological data types that meet the target quantity are selected.

3. The method for predicting air pollutant concentration according to claim 2, characterized in that: The prediction model further includes a Mogrifier gating operation module. Before the step of extracting time feature data from a plurality of modal data corresponding to each site respectively through the time convolution network in the time feature extraction module, the training step of the prediction model further includes: The Mogrifier gating operation module is used to optimize the multiple modal data corresponding to each site.

4. The method for predicting air pollutant concentration according to claim 1, characterized in that: The method also includes the step of optimizing the prediction model: A particle swarm optimization algorithm is used to search for the optimal parameter configuration of the prediction model to optimize the prediction model.

5. The method for predicting air pollutant concentration according to claim 4, characterized in that: The optimization step of the prediction model also includes: Determine the loss function based on the predicted output result, the actual output result and the predicted time step value; Determine the initialization parameters and particle swarm parameters of the prediction model; wherein the particle swarm parameters include particle swarm size, maximum number of iterations, inertia weight, acceleration factor, initial velocity, and initial position; Iteratively update the particle swarm to obtain the speed and position of each particle, the individual extreme value of each particle, and the global extreme value of the particle swarm after each iterative update; The output parameters of the prediction model are updated according to the position of the particle corresponding to the final global extreme value of the particle swarm obtained when the maximum number of iterations is met, so as to complete the optimization of the prediction model.

6. The method for predicting air pollutant concentration according to claim 1, characterized in that: When the method is applied to predict the concentration of air pollutants in a target area, the target environmental monitoring site is a plurality of environmental monitoring sites in the target area; When the method is applied to predict the concentration of air pollutants at a single or multiple designated environmental monitoring sites, the target environmental monitoring site is the single or multiple designated environmental monitoring sites.

7. A device for predicting air pollutant concentration, characterized in that: include: A data acquisition module is used to obtain the current pollutant concentration data, current meteorological data and geographical location data of the target site corresponding to the target environmental monitoring site; A model processing module, used to use the current pollutant concentration data, the current meteorological data and the target site geographic location data as target input data of the prediction model; wherein the prediction model includes an input module, a time feature extraction module, a spatial feature extraction module, a time-space fusion convolutional network module and an output module; the input module is used to perform modal decomposition processing on the pollutant concentration data in the input data to obtain multiple modal data; the time feature extraction module is used to extract time feature data from the multiple modal data; the spatial feature extraction module is used to construct an adjacency matrix that integrates the pollutant concentration data, meteorological data and site location data in the input data, and extracts spatial feature data by a graph convolutional network; the time-space fusion convolutional network module is used to extract the time-space features corresponding to each modality; the output module is used to aggregate the prediction outputs corresponding to multiple modalities to output the prediction results; A prediction result determination module, used to determine the prediction result of the air pollutant concentration corresponding to the target environmental monitoring site according to the output result of the prediction model; The training and optimization of the prediction model includes the steps of: Step S201, establishing an input module of a prediction model, collecting hourly historical pollutant concentration data, historical meteorological data, and station geographic location data from environmental monitoring stations in various places; Step S202, pre-processing the historical pollutant concentration data and the historical meteorological data; using the mutual information method to perform correlation analysis on the historical pollutant concentration data and the historical meteorological data, and finding the top n types of historical pollutant concentration data and historical meteorological data that are closely correlated with the predicted air pollutant concentration; using the variational mode decomposition algorithm and the particle swarm optimization algorithm to perform mode decomposition and optimization processing on the historical pollutant concentration data of each station, to obtain multiple mode data, and standardize each mode data; Step S203, establishing a Mogrifier gating mechanism for the input data, and optimizing the multiple modal data corresponding to each site through the Mogrifier gating operation module; Step S204, establishing a time feature extraction module, and extracting time feature data from a plurality of modal data corresponding to each site through a time convolution network in the time feature extraction module; Step S205, establishing a spatial feature extraction module, inputting the historical pollutant concentration data, historical meteorological data and corresponding site geographic location data of the first n types obtained in step S202 into the spatial feature extraction module, constructing an adjacency matrix in the graph convolution network, and extracting spatial feature data through the graph convolution network; specifically including: Step S2051, constructing an adaptive graph convolutional network GCN and an adjacency matrix A according to the distribution of monitoring stations; Step S2052: All sites and their pairwise relationships are represented as a weighted graph G=(V,E,A), where V is a site set including N sites; E is an edge set; is the adjacency matrix, where the weights Represents the spatiotemporal correlation strength between site Vi and site Vj. The edges are connected according to the weights. , ; Step S2053, establish an adjacency matrix , the matrix A integrates pollutant concentration data, meteorological data and station location data; Step S2054, define the time series data adjacency matrix of pollutant concentration data as: （10） In the above formula (10), an adjacency matrix representing the temporal similarity graph of particle contaminants; Environmental Monitoring Station Time series data of n pollutant concentrations; Environmental Monitoring Station The average DTW distance between the time series data of pollutant concentrations; represents the variance of Gaussian distribution of pollutant data at the site; Indicates the distance threshold of the set pollutant concentration set time mode; Step S2055, define the adjacency matrix of meteorological information between sites as: （11） In the above formula (11), The adjacency matrix representing the temporal similarity graph of meteorological information; Environmental Monitoring Station Meteorological data; Environmental Monitoring Station The average DTW distance between the time series data of the weather; Indicates the variance of Gaussian distribution of meteorological data at the site; Indicates the distance threshold of the set time mode of the meteorological data; Step S2056, define the adjacency matrix of spatial information: Site The geographical location is represented by , site The geographical location is represented by ,in, Respectively represent sites The latitude and longitude of Respectively represent sites Latitude and longitude of the site The geographical distance between for: （12） （13） （14） （15） （16） In the above formulas (12)-(16), Represents the conversion function between angle and radian, R is the equatorial radius of the earth; The spatiotemporal adaptive matrix E that integrates the temporal information of pollutant concentration data, meteorological data, and spatial information between stations: （17） In the above formula (17), , , , is a trainable parameter, is the activation function; Take the adjacency matrix A as: （18） In the above formula (18), E represents the spatiotemporal adaptation matrix, and A represents the adjacency matrix; Step S2057, for each modality k, a graph convolutional network GCN is established and represented as: （19） Decomposed into the following operating rules: （20） （21） （22） In the above equations (20)-(22), A is the adjacency matrix; I is the identity matrix; is an adjacency matrix with self-connection; Degree matrix representing the adjacency matrix with self-connection; The feature vector of the kth mode and the lth layer of the graph node; That is, it is the feature vector of the node in the l+1 layer after the k-mode is convolved; when l=0, ; Represents the learnable parameters of the l-th layer convolution; is the sigmoid activation function; Step S206, inputting the temporal feature data and the spatial feature data into the convolution gated recurrent unit in the spatiotemporal fusion convolutional network module to fuse the temporal feature data with the spatial feature data, and introducing an attention mechanism to obtain spatiotemporal features corresponding to different modalities; wherein the fully connected layer of the spatiotemporal fusion convolutional network module respectively connects the spatiotemporal features of a single modality, and outputs the predicted output of the air pollutant concentration corresponding to each modality of each station; Step S207, aggregate the predicted outputs corresponding to all modes through the output module to obtain predicted concentration output feature data, and perform denormalization on the predicted concentration output feature data to obtain a prediction result.

8. An electronic device comprising: A processor, and a memory storing a program, wherein the program comprises instructions, which, when executed by the processor, cause the processor to perform the method for predicting the concentration of air pollutants according to any one of claims 1 to 6.

9. A non-transitory machine-readable medium storing computer instructions, characterized in that: The computer instructions are used to enable the computer to execute the method for predicting the air pollutant concentration according to any one of claims 1-6.

Images