A large space range multi-station O3 concentration 24-hour accurate prediction method

Abstract

The application discloses a large-space-range multi-station O3 concentration 24-hour accurate prediction method, belongs to the field of atmospheric pollutant concentration prediction, and comprises the following steps: obtaining historical pollutant time series data, meteorological reanalysis data and aerosol optical depth data of multiple ground monitoring stations in a target area, and constructing a space-time matched input feature set; constructing a space graph structure representing the ozone transmission relationship according to the spatial position of the station and the meteorological driving factor; aggregating the neighborhood space information by using a graph attention network layer to generate space-time dependent features; inputting the features into a gated recurrent unit layer to extract long-term and short-term time series dynamic patterns; and finally generating the hourly ozone concentration prediction value of each station in the future 24 hours through a fully connected output layer. The application realizes the synchronous prediction of all stations in a large space range by using a single model, effectively represents the spatial transmission and chemical reaction process of the pollutants, and significantly improves the long-term prediction accuracy and peak capturing ability.

Description

Technical Field

[0001] This invention belongs to the field of atmospheric pollutant concentration prediction, and in particular relates to a method for accurate 24-hour prediction of O3 concentration at multiple sites over a large spatial area. Background Technology

[0002] Ground-level ozone (O3) is a potent oxidizing secondary pollutant formed by photochemical reactions of nitrogen oxides and volatile organic compounds, posing a significant threat to human health and ecosystems. Currently, O3 concentration prediction methods are mainly divided into deterministic methods and statistical models. Deterministic methods (such as WRF-chem and GOES-CMAQ) simulate the diffusion, deposition, and reaction processes of pollutants based on physicochemical mechanisms, but they face challenges in data acquisition and have relatively low prediction accuracy. Statistical models are data-driven, using neural networks (such as LSTM and GRU) to learn nonlinear characteristics from historical data, demonstrating higher accuracy in short-term predictions. In recent years, graph neural networks (GNNs) and graph attention networks (GATs) have attracted attention in air pollutant concentration prediction research due to their ability to extract neighborhood spatial information from asymmetric sites.

[0003] However, existing forecasting studies still suffer from the following technical limitations: First, most statistical models are limited to a single city or a few ground stations, failing to achieve synchronous generalized forecasting across all stations in a large spatial area; second, the transport processes of pollutants between regions are not well characterized, and the spatial migration of O3 due to wind and its synergistic chemical reaction mechanism with meteorological and precursor substances are not effectively incorporated into the forecasting models; third, as the forecast duration extends to 24 hours, uncertainties increase significantly, existing models are insufficient in capturing peak concentrations, and early forecast errors tend to amplify exponentially with time steps. Furthermore, the attenuation effect of aerosol optical thickness (AOD) on ultraviolet irradiance and its indirect impact on O3 formation are rarely included as input parameters in existing forecasting methods. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method for accurate 24-hour prediction of O3 concentration at multiple sites over a large spatial area, comprising: Based on historical pollutant time-series data, meteorological reanalysis data, and aerosol optical thickness data from multiple ground monitoring stations within the target area, a set of spatiotemporal matching input features is obtained. Based on the spatial location of each station and meteorological driving factors based on wind direction and wind speed, a spatial map structure is constructed to characterize the ozone transport relationship between stations. Using a graph attention network layer, based on the spatial graph structure and the input feature set, the neighborhood spatial information of each station is aggregated to generate spatiotemporal features containing spatial dependencies; The spatiotemporal features are input into the gated recurrent unit layer to extract the independent temporal dynamic patterns of each station, thereby obtaining long-term and short-term temporal dependence features. Based on the aforementioned long-term and short-term time-series dependence characteristics, the ozone concentration prediction values ​​for each site for the next 24 hours are generated through a fully connected output layer. Specifically, when generating hourly ozone concentration predictions for each site over the next 24 hours, the input feature set of six consecutive historical moments is used as the time step and input into the gated loop unit layer.

[0005] Optionally, obtaining the input feature set includes: performing linear interpolation to complete missing values ​​in the historical pollutant time series data, selecting pollutant data from the nearest station based on Euclidean distance, and performing auxiliary completion on station data with continuous missing values ​​exceeding a preset threshold.

[0006] Optionally, constructing the spatial graph structure includes: calculating the Euclidean distance between stations based on the latitude and longitude coordinates of each station, establishing a connection between two stations whose distance is less than a preset threshold, and constructing a binary adjacency matrix based on the connection relationship.

[0007] Optionally, aggregating the neighborhood spatial information includes: calculating the ozone spatial transport contribution weight driven by meteorological factors based on the Euclidean distance between stations, the wind speed component of neighboring points, the angle between the wind direction and the edge connecting the stations, and the ozone transport distance over the predicted duration.

[0008] Optionally, aggregating the neighborhood spatial information further includes: using a nonlinear activation function to extract the nitrogen dioxide concentration, temperature, and relative humidity of neighboring stations, and calculating the synergistic contribution weights of ozone precursors and meteorological conditions to the chemical reactions that generate ozone.

[0009] Optionally, after constructing the spatial graph structure, the method further includes: weighting the adjacency matrix of the spatial graph structure according to the correlation coefficient of ozone concentration between stations or the inverse weighting coefficient of station distance to form a weighted adjacency matrix.

[0010] Optionally, the historical pollutant time-series data includes hourly concentration data of nitrogen dioxide, fine particulate matter, carbon monoxide, inhalable particulate matter, sulfur dioxide, and ozone; the meteorological reanalysis data includes air pressure, relative humidity, temperature, wind speed, wind direction, boundary layer height, and total precipitation.

[0011] Optionally, before generating the spatiotemporal matching input feature set, the method further includes: resampling the meteorological reanalysis data and the aerosol optical thickness data to a spatial resolution of 0.05°×0.05° that matches the ground monitoring station using bilinear interpolation.

[0012] On the other hand, the present invention also provides an electronic device including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0013] On the other hand, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0014] Compared with the prior art, the present invention has the following advantages and technical effects: To address the limitation of existing models in achieving simultaneous generalized prediction across all stations within a large spatial area, this application constructs a spatial map structure covering all ground monitoring stations nationwide and outputs prediction results for all stations simultaneously using a single model, significantly improving the model's spatial generalization capability. To address the insufficient characterization of pollutant regional transport processes, this application introduces meteorological-driven spatial transport weights and chemical reaction synergistic contribution weights, dynamically quantifying the pollution contribution of neighboring stations through a graph attention network, effectively simulating the spatial migration and generation mechanism of O3. To address the issues of long-term prediction error accumulation and weak peak capture capability, this application integrates the temporal memory capability of gated recurrent units, reducing the root mean square error by approximately 18.6% in 24-hour predictions and exhibiting a more sensitive response to sudden peaks. Furthermore, the inclusion of AOD data as an input parameter further enhances the model's accuracy in characterizing photochemical reaction processes. Attached Figure Description

[0015] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a prediction diagram according to an embodiment of the present invention; Figure 2 This is a schematic diagram comparing the prediction results of four different models for future O3 concentrations at all stations across the country, according to an embodiment of the present invention. Figure 3 This is a schematic diagram comparing the peak prediction results of different models in an embodiment of the present invention. Detailed Implementation

[0016] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0017] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0018] Example 1 This embodiment provides a method for accurate 24-hour prediction of O3 concentration at multiple sites over a large spatial area, including: In this embodiment, the data types used include three categories: aerosol optical thickness data (MERRA-2 AOD), ERA5 meteorological reanalysis data, and six major pollutants from ground monitoring stations. The remote sensing ultraviolet irradiance data did not meet the spatiotemporal resolution requirements of this embodiment and was therefore not used. The data for the six major air quality pollutants are sourced from ground monitoring stations and specifically include: NO2, PM2.5, CO, and PM2.5. 10 SO2 and O3. Complex chemical reactions and interconversions exist between different pollutants. Among them, NO2 and PM2.5 share a common origin with O3 concentrations, or act as precursors for O3 chemical reactions, generating O3 through secondary reactions; therefore, they are important input parameters.

[0019] Because ground-based pollutant monitoring stations are physical sensors, malfunctions are frequent, leading to anomalies and missing data. Some stations have significant, consecutive gaps in pollutant data. Predicting O3 concentrations over multiple consecutive moments requires continuous and complete time-series data; missing data prevents training and prediction. For stations with fewer missing values, linear interpolation is used to complete the data for six pollutants. However, when a pollutant has 1000 consecutive missing moments, directly using linear interpolation results in the same concentration at multiple moments, failing to reflect the increasing or decreasing trend of concentration, making the interpolation result unreasonable. To address this, this embodiment calculates the Euclidean distance of each station based on the first law of geography (closer distances result in higher correlations), selecting pollutant data from the nearest station to supplement the missing data.

[0020] The meteorological reanalysis data (ERA5) was sourced from the European Centre for Medium-Range Weather Forecasts (ECMWF) and included air pressure, relative humidity, temperature, wind speed (10 mV), wind direction (10 mU), boundary layer height (BLH), and total precipitation (Tp). ERA5 has a spatial resolution of 0.25° × 0.25°. Resampling was required for spatiotemporal matching with AOD data and surface station data. This embodiment employs bilinear interpolation to resample the meteorological reanalysis data to a spatial resolution of 0.05° × 0.05°.

[0021] Aerosol optical thickness (AOD) characterizes the degree of light attenuation by atmospheric aerosol particles along the satellite observation path and is a key parameter for atmospheric environmental quality monitoring. Due to its unique extinction properties, AOD can weaken the intensity of solar radiation and ultraviolet radiation, thus affecting near-surface ozone concentration. MERRA-2 AOD data possesses 24-hour and spatially seamless properties, meeting the time series integrity requirements of time-series prediction models; therefore, this AOD data was used as the input parameter. To achieve spatiotemporal matching with meteorological data and ground monitoring stations, the MERRA-2 AOD data was resampled to a resolution of 0.05° × 0.05° (approximately 5 km). Within the study area, a 5 km grid effectively distinguishes the spatial locations of different stations.

[0022] In this embodiment, the relative positions between stations are first determined. The first step in constructing the graph structure is to determine the number of vertices contained in the graph structure, which is determined based on the total number of ground monitoring stations. Each monitoring station has unique latitude and longitude coordinates in space, and the relative positions between stations can be represented based on the latitude and longitude coordinates of each station and their differences.

[0023] Secondly, rules for generating connections between stations are established. Based on the latitude and longitude coordinates of each station, the distance between any two stations is calculated. When the distance between two stations is less than a set distance threshold, a connection is established between them to indicate that O3 will transmit and influence each other between the two stations. In this embodiment, 150km is used as the distance threshold for generating connections between two stations, ensuring that all ground monitoring stations are connected to each other according to this rule (see...). Figure 2 Since the forecast period is 24 consecutive hours in the future, and taking into account factors such as wind speed, topographical obstruction, and possible chemical reactions of O3 gas during its spatial transport, 150km was determined as the appropriate distance threshold for establishing connections between stations.

[0024] Finally, an adjacency matrix representing spatial topological relationships is constructed. A binary adjacency matrix A∈{0,1}^(n×n) is used to represent the spatial topological relationships between stations, where n is the total number of stations. The matrix is ​​initialized to all zeros. Iterating through all stations, if there is an edge relationship between station i and station j, the element A_{i,j} at the corresponding position in the matrix is ​​set to 1. Furthermore, more weight constraints can be added, such as the correlation of O3 concentration between stations and the inverse weighting coefficient of station distance, thus forming a weighted adjacency matrix.

[0025] In this embodiment, the core of the constructed GAT_GRU prediction model consists of three hierarchical modules (see...). Figure 3The prediction model's network layers include a graph attention layer, a gated recurrent unit (GRU) layer, and a fully connected output layer, each with different feature extraction functions. The graph attention layer (GAT) serves as the core module for spatial feature aggregation, dynamically quantifying the pollution contribution weights of neighboring sites to the target site through an attention mechanism. This layer explicitly simulates the physical process of pollutant transport, optimizing the calculation of attention scores based on the spatial relationships between sites (i.e., the adjacency matrix), thus achieving spatial dependency modeling under physical constraints. The GRU layer receives the neighborhood spatial information output by the GAT layer, extracting independent temporal dynamic patterns for each site and capturing the long- and short-term temporal dependencies between input parameters. The fully connected output layer (FC) integrates spatiotemporal features to generate multi-step prediction results, with the output dimension being the number of sites × prediction step size × number of target variables.

[0026] The GAT network layer calculates the attention score for each connection edge by using an attention mechanism to consider the neighboring sites, thus characterizing the importance of different neighboring points to the center point. For O3 concentration prediction in real-world scenarios, it's necessary to consider the spatial transport process of O3 gas due to wind, the attenuation characteristics of AOD on ultraviolet irradiance, and to characterize the chemical reaction mechanism of O3 concentration itself as much as possible. The GAT network layer aggregates spatial information of different locations and numbers of neighboring sites for each site using a specific calculation method. First, the meteorological factor-driven spatial transport process of O3 is characterized by the following formula: (1) In the formula, d i,j It is a website i (nearest points) and center point j The Euclidean distance between them, v i Neighboring points i The wind speed component, It is the angle between the wind direction and the edge connecting the site. When the wind is downwind, the effect on the prediction center point is positive, and when the wind is upwind, the effect is negative. k Let be the transport distance of O3 over the predicted duration. exp() represents the distance attenuation effect of O3 gas transport. Used to enhance the contribution of stations in the downwind direction.

[0027] Among them, the Euclidean distance between stations d i,j The calculation formula is as follows: (2) In the formula, i The point is a neighboring point. j To predict the center point, d i,j fori and j Euclidean distance between points.

[0028] Secondly, considering the chemical reaction mechanism of O3 concentration, it is necessary to quantify the synergistic effect of precursors and meteorological conditions on O3 generation, which can be characterized as follows: (3) In the formula, the NO2 concentration at neighboring stations is extracted using the nonlinear activation function ReLU, and the temperature ( T ) and relative humidity ( RH) The chemical reaction relationship with changes in O3 concentration. Of course, ultraviolet irradiance has a significant impact, but since this parameter is not currently included in the input parameters, temperature is used as a reflection of ultraviolet intensity.

[0029] In this embodiment, the hyperparameters for model training are set. The training process of a deep neural network includes several configurable hyperparameters, specifically including batch size (batch_size=4), number of iterations (epochs=40), and learning rate (lr=0.0001). Due to the massive amount of data and limited computer physical resources, reducing batch_size can alleviate the memory shortage problem. After several attempts, setting batch_size to 4 is found to be suitable. This setting saves a significant amount of memory space, but at the same time, it extends the model training time.

[0030] To address the potential overfitting issue during model training, this embodiment employs early stopping. This method monitors the training loss on the validation set. When the model's loss on the validation set no longer improves, training is stopped, the current model is saved, and it is applied to the test set for prediction, thereby improving the objective performance of the model's predictions.

[0031] In time series data prediction, using data from multiple consecutive historical moments can provide richer time series features, thereby improving the ability of time series prediction models such as GRU to extract these features. In this embodiment, the time step is set to 6, providing multiple input parameters from 6 consecutive historical moments to predict the future trend of O3 concentration changes.

[0032] The model's predictions need to be compared with measured data, and the comparison results are quantified using scientific evaluation metrics. This embodiment uses three different evaluation metrics to assess the accuracy of the model's predictions: the coefficient of determination (R²) for goodness of fit, used to evaluate the degree of fit between the predicted results and the measured data; and the root mean square error (RMSE), used to evaluate the square root of the ratio of the square of the deviation between the predicted value and the true value to the number of observations (n). Different evaluation metrics address different evaluation details, enabling a more objective assessment of the accuracy of the predictions.

[0033] Example 2 This embodiment provides a method for accurate 24-hour prediction of O3 concentration at multiple sites over a large spatial area, including: like Figure 1 As shown, the objective of this embodiment is to construct a deep neural network prediction model to effectively predict the O3 concentration of all stations within a region over the next 24 hours. By aggregating spatial information from adjacent stations and integrating chemical reaction processes, the accuracy of peak and trough predictions is improved. Furthermore, by constructing three-dimensional data, the model can simultaneously acquire information from each station and share network weights during training, thereby enabling a single prediction model to simultaneously predict the O3 concentration of all spatial locations within the region over a longer period.

[0034] The data used in this embodiment includes MERRA-2 AOD data, meteorological reanalysis data, and data on six major air pollutants. These data are sourced from Earth Data, ECWMF, and environmental departments, covering the period from January 1, 2021 to December 31, 2023, with an hourly time resolution. The meteorological reanalysis data (ERA5) is from the European Centre for Medium-Range Weather Forecasts (ECMWF) and specifically includes: air pressure, relative humidity, temperature, wind speed (10mV), wind direction (10mU), boundary layer height (BLH), and total precipitation (Tp).

[0035] The process of classifying data types and datasets includes: (1) Hourly monitoring concentrations of six major air pollutants were collected from more than 2,000 ground monitoring stations across the country from January 1, 2021 to December 31, 2023. At the same time, meteorological reanalysis data and MERRA-2AOD data were collected for the corresponding time periods. All of the above data are important input parameters for the prediction model.

[0036] (2) Data Partitioning: To achieve simultaneous prediction of concentrations at all stations within the region, a dataset with a three-dimensional structure of x, y, and z needs to be constructed to support simultaneous training of all pixels and synchronous prediction of all pixels. The time span of the above-mentioned data types is from January 1, 2021 to December 31, 2023. The number of records for a single data type is approximately 50 million, and the total number of valid records for 14 input parameters, including O3 concentration, is approximately 700 million. All data are sorted by time and divided into training set, validation set, and test set required for deep neural network training. Among them, the data from January 1, 2021 to September 30, 2022 is used as the training set (accounting for approximately 60% of the total data), the data from October 1, 2022 to January 31, 2023 is used as the validation set (accounting for approximately 10% of the total data), and the data from February 1, 2023 to December 31, 2023 is used as the test set (accounting for approximately 30% of the total data).

[0037] The GAT_GRU model implements PM for all spaces within the region. 2.5 The concentration prediction process includes: (1) The model is trained, validated and tested using different pre-defined datasets. Early stopping is used to prevent overfitting in deep neural networks.

[0038] (2) The GAT_GRU model can predict the future concentration within all grid cells involved in the modeling. The prediction duration is expressed in the form of T+n: T+1 represents the prediction of the concentration in the next 1 hour, T+6 represents the prediction of the concentration in the next 6 hours, and so on. T+12 and T+24 represent the prediction of the O3 concentration in the next 12 hours and 24 hours, respectively. The specific process is as follows: Figure 1 As shown.

[0039] (3) The GAT network layer calculates the attention score between each prediction center site and other sites with which it has an edge, to quantify the importance of pollution contribution. The aggregated spatial information of neighboring sites is used as an additional feature of the GRU network layer, thereby more accurately predicting future concentration trends. The fully connected layer is used to optimize the prediction results of the GRU network and outputs them according to the specified dimensions.

[0040] Accurate long-term prediction of near-surface O3 concentration is of great significance in real life. This invention constructs a novel prediction model (GAT_GRU) to predict O3 concentrations at over 2000 ground monitoring stations nationwide for the next 24 hours. To explore the effect of aggregating neighborhood spatial information on improving the accuracy of the prediction results, this invention selects three other baseline models: Multilayer Perceptron Network (MLP), Long Short-Term Memory Neural Network (LSTM), and a single GRU model. The R², RMSE, and MAE levels of the prediction results for all stations over the next 1-24 hours are compared among the four different models. Specific results are as follows: Figure 2 As shown.

[0041] like Figure 2 As shown, the accuracy of all models decreased with increasing prediction time (from 1 hour to 24 hours) because the uncertainty of PM2.5 concentration prediction increased significantly over time, making prediction more difficult. The GAT_GRU model aggregates neighborhood spatial information between stations, effectively resisting prediction errors caused by uncertainties over time. The GAT_GRU model has the most accurate prediction accuracy and the lowest error. The mean R² of the predicted values ​​and observed values ​​for all stations over 24 hours is 0.617, while the mean R² of the other three models are 0.492, 0.561, and 0.578, respectively, with ΔR² values ​​of 0.125, 0.056, and 0.039, all showing some improvement. The mean RMSE decreased by approximately 2-6 µg / m³, and the mean MAE decreased by 1.6-5.1 µg / m³. These quantitative indicators show that the GAT model's aggregation of neighborhood spatial information significantly improves the prediction results. Furthermore, the accuracy of the four models decreased significantly at time T+6, which is the next 6 hours. This may be because the prediction models rely on previous predictions as input, and the early errors amplify exponentially with the prediction step size. The O3 photochemical generation rate is synergistically regulated by precursor concentrations and radiation intensity during the day, and the peak capture bias significantly affects the prediction accuracy at subsequent times. The decrease in accuracy is relatively stable between 12 and 24 hours, possibly because the diurnal cycle of meteorological and photochemical parameters allows the model to reduce uncertainty by learning periodic characteristics. The GAT_GRU model, which integrates the temporal memory capability of GRU with the spatial dependency resolution of GAT, reduced the RMSE error by approximately 18.6% in the T+24 prediction compared to the baseline models (LSTM and GRU). These results indicate that coupled spatiotemporal feature modeling can significantly improve the robustness of regional O3 prediction, providing technical support for joint prevention and control of complex pollution. The newly constructed model in this invention provides more accurate 24-hour prediction results for O3 concentrations at stations nationwide, and is an effective prediction tool. like Figure 3As shown, on the other hand, accurate prediction of peak concentration is also an important aspect of evaluating the performance of prediction models. It is essential to better predict pollution events that will occur in the future, which can effectively reduce the more serious harm caused to the human body by high concentrations of O3. This invention compares the performance of three models—MLP, LSTM, and GAT_GRU—in predicting O3 concentrations over the next 1 hour and 6 hours, focusing on their ability to capture peak concentrations, such as... Figure 3 As shown in the figure, in the prediction at time T+1, all three models can fit the actual concentration curve well, but the GAT_GRU model is more sensitive to sudden peaks (e.g., the prediction error of the peak at time T+1 is the lowest among the four models), while the MLP and LSTM models show peak lag and peak underestimation, respectively. This indicates that the traditional MLP model, due to its lack of time memory, has difficulty capturing rapidly changing characteristics; although the LSTM model improves time-dependent modeling through gating mechanisms, it is insufficient in extracting features of spatial heterogeneity (such as spatial diffusion of pollution sources). The prediction results show that the GAT_GRU model has the advantage in peak prediction in both short-term and longer-term predictions.

[0042] As the prediction timeframe was extended to 6 hours into the future, the performance differences between the models became more pronounced. The LSTM model's predictions showed a smoothing trend, with a decreased ability to capture consecutive peaks and troughs. In contrast, the GAT_GRU model, through its graph attention mechanism, dynamically weighted the spatial node correlations, maintaining high accuracy in complex spatiotemporally coupled scenarios. Notably, GAT_GRU significantly outperformed the LSTM and MLP models in predicting the peak occurrence time in the T+6 prediction, thanks to its ability to aggregate contamination transmission through its fused graph structure. Conversely, the MLP model, lacking spatiotemporal prior knowledge, exhibited significant bias in its peak predictions.

[0043] In summary, the superiority of the GAT_GRU model in terms of peak emission prediction capability is mainly reflected in two aspects. First, the graph attention layer effectively identifies key pollution source nodes and their spatial interactions, enhancing the characterization of sudden emission events. Second, the GRU unit, through adaptive gating, balances the weights of historical information and real-time observation data, avoiding the gradient decay problem in long-term prediction. These results demonstrate that the GAT_GRU model, which integrates spatiotemporal features, can provide reliable technical support for accurate regional O3 concentration control.

[0044] This invention proposes a prediction model with excellent spatial generalization performance, capable of accurately predicting the O3 concentration change trend of all ground monitoring stations deployed in China up to 2022 for the next 24 consecutive hours. The newly constructed GAT_GRU prediction model integrates the characteristics and advantages of GAT and GRU networks, respectively used to extract the time series features of O3 time series data and other input parameters, as well as the potential neighborhood spatial information between stations. In comparative experiments, the GAT_GRU model shows better overall prediction accuracy than the other three baseline models, and achieves better prediction results for high values. It can effectively simulate the peak of pollutants, enabling timely early warning and reducing the harm of exposure to high concentrations of ozone.

[0045] On the other hand, this embodiment also provides an electronic device, including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0046] On the other hand, this embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0047] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for accurate 24-hour prediction of O3 concentration at multiple sites over a large spatial area, characterized in that, include: Based on historical pollutant time-series data, meteorological reanalysis data, and aerosol optical thickness data from multiple ground monitoring stations within the target area, a set of spatiotemporal matching input features is obtained. Based on the spatial location of each station and meteorological driving factors based on wind direction and wind speed, a spatial map structure is constructed to characterize the ozone transport relationship between stations. Using a graph attention network layer, based on the spatial graph structure and the input feature set, the neighborhood spatial information of each station is aggregated to generate spatiotemporal features containing spatial dependencies; The spatiotemporal features are input into the gated recurrent unit layer to extract the independent temporal dynamic patterns of each station, thereby obtaining long-term and short-term temporal dependence features. Based on the aforementioned long-term and short-term time-series dependence characteristics, the ozone concentration prediction values ​​for each site for the next 24 hours are generated through a fully connected output layer. Specifically, when generating hourly ozone concentration predictions for each site over the next 24 hours, the input feature set of six consecutive historical moments is used as the time step and input into the gated loop unit layer.

2. The method according to claim 1, characterized in that, Obtaining the input feature set includes: performing linear interpolation to complete missing values ​​in the historical pollutant time series data, selecting the pollutant data of the nearest station based on Euclidean distance, and performing auxiliary completion on station data with continuous missing values ​​exceeding a preset threshold.

3. The method according to claim 1, characterized in that, Constructing the spatial graph structure includes: calculating the Euclidean distance between stations based on the latitude and longitude coordinates of each station, establishing a connection between two stations whose distance is less than a preset threshold, and constructing a binary adjacency matrix based on the connection relationship.

4. The method according to claim 1, characterized in that, Aggregating the neighborhood spatial information includes: calculating the ozone spatial transport contribution weight driven by meteorological factors based on the Euclidean distance between stations, the wind speed component of neighboring points, the angle between the wind direction and the edge connecting the stations, and the ozone transport distance over the predicted duration.

5. The method according to claim 1, characterized in that, Aggregating the neighborhood spatial information also includes: using a nonlinear activation function to extract the nitrogen dioxide concentration, temperature, and relative humidity of neighboring stations, and calculating the synergistic contribution weights of ozone precursors and meteorological conditions to the chemical reaction of ozone generation.

6. The method according to claim 1, characterized in that, After constructing the spatial graph structure, the method further includes: weighting the adjacency matrix of the spatial graph structure according to the correlation coefficient of ozone concentration between stations or the inverse weighting coefficient of station distance to form a weighted adjacency matrix.

7. The method according to claim 1, characterized in that, The historical pollutant time series data includes hourly concentration data for nitrogen dioxide, fine particulate matter, carbon monoxide, inhalable particulate matter, sulfur dioxide, and ozone; the meteorological reanalysis data includes air pressure, relative humidity, temperature, wind speed, wind direction, boundary layer height, and total precipitation.

8. The method according to claim 1, characterized in that, Before generating the spatiotemporal matching input feature set, the method further includes: resampling the meteorological reanalysis data and the aerosol optical thickness data to a spatial resolution of 0.05°×0.05° that matches the ground monitoring station using bilinear interpolation.

9. An electronic device comprising a memory, a processor, and a computing program stored in the memory and executable on the processor, characterized in that, When the processor executes the computing program, it implements the method of any one of claims 1-8.

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

Images