Ozone concentration forecasting method

Abstract

The invention discloses an ozone concentration forecasting method, which utilizes air quality observation data of 326 cities throughout the country, processes abnormal and missing data, and reanalyzes relative humidity, precipitation, boundary layer height, short-wave radiation, sea level air pressure, wind direction, wind speed and dew point in the data in combination with ERA5. A high-quality database used for supporting deep learning is established, graph structures are established with cities and provinces as nodes respectively, and the spatial dependency relationship between the nodes is adjusted in a self-adaptive mode through a dynamic learnable edge weight mechanism; self-adaptively updating the edge weight according to the input spatio-temporal characteristics, and further quantifying and analyzing the propagation path of ozone among different regions; and key meteorology and precursor factors are dynamically identified through a channel attention mechanism. According to the invention, the accuracy and real-time performance of ozone forecasting are improved, and transmission paths and driving factors thereof between regions can be disclosed; and scientific basis and technical support are provided for ozone pollution prevention and control and regional joint prevention and joint control.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of environmental prediction, and particularly relates to an ozone concentration prediction method BACKGROUND

[0002] Ozone, as a typical photochemical pollutant, will not only harm human health, but also cause crop yield reduction and destroy ecological system stability due to its strong oxidizing property in the troposphere. In recent years, with the continuous promotion of air pollution prevention and control work in China, the concentration levels of main pollutants such as SO2, PM 2.5 , CO, PM 10 and NO2 have decreased significantly, but ozone pollution has shown a fluctuating upward trend and remains at a high level, which has become an important factor restricting the sustained improvement of air quality. The generation mechanism of ozone is complex, which is not only affected by local nitrogen oxides and volatile organic compounds photochemical reaction, but also driven by regional atmospheric transmission, so it is difficult to predict. Traditional ozone prediction methods mainly include statistical models and chemical transport models, among which statistical models cannot effectively describe the nonlinear relationship between ozone, precursors and meteorological factors, and chemical transport models are limited in business application due to large amount of calculation and poor real-time performance.

[0003] Therefore, in order to solve the above problems, an ozone concentration prediction method is proposed in the present application. SUMMARY

[0004] The present application aims to design an ozone concentration prediction method to solve the above technical problems.

[0005] In order to achieve the above technical effects, the present application is realized by the following technical scheme: an ozone concentration prediction method, characterized in that it comprises the following steps:

[0006] S1, the air quality observation data of 326 cities in China from 2015 to 2024 is obtained by network data collection, and the data is filled with null, normalized and divided into data sets, and combined with the relative humidity, precipitation, boundary layer height, shortwave radiation, sea level pressure, wind direction, wind speed and dew point in ERA5 reanalysis data to establish a high-quality database for supporting deep learning;

[0007] S2, a Transformer model based on multi-scale graph neural network and dynamic learnable edge weight mechanism is adopted to train the pollutant historical observation data and ERA5 reanalysis data of 326 cities in China, build an ozone prediction model covering 326 cities in China, and optimize the model; specifically as follows:

[0008] S2.1, a multi-scale based graph neural network model is used to train the historical data set, and ten years of pollutant data and numerical simulation data provided by ERA5 reanalysis data are used as the input of the graph neural network model, and the predicted ozone concentration is used as the output, to establish an hourly ozone prediction model for 326 cities nationwide;

[0009] S2.2, a "city-province" double-layer graph structure is constructed to realize multi-scale feature extraction: at the city level, a directed graph structure is constructed, and the edges in the graph represent the potential pollutant transmission paths between cities, which is based on geographical proximity and introduces a terrain perception filtering strategy to ensure the physical feasibility of the transmission path; at the provincial level, an undirected graph structure is constructed, and the node features are generated by average pooling aggregation of all city features in the province, and the edge weight is initialized based on the provincial geographical distance decay law to simulate the macro spatial interaction;

[0010] S2.3, a dynamic learnable edge weight mechanism is introduced in the modeling process. According to the input real-time spatio-temporal features, the mechanism adaptively adjusts the weight strength of the connection edges between nodes in the graph through a gated fusion network, realizing the smooth fusion of the static physical prior weight calculated based on the wind field dynamics formula and a dynamic residual learned by an attention network; so that the model can accurately quantify the dynamic transmission path and contribution of pollutants under different meteorological conditions, and identify potential pollution input relationships;

[0011] S2.4, a channel attention mechanism is introduced in the graph neural network; the mechanism acts on the input feature level, dynamically calculates the importance score of each input feature channel through an attention network composed of fully connected layers, and uses the Softmax function to normalize to get the final channel attention weight, which reweights the original input features; This enables the model to automatically focus on key driving factors that significantly affect ozone prediction, thereby improving prediction accuracy while enhancing the model's explainability of ozone formation and transmission causes;

[0012] S2.5, realize the dynamic analysis function of inter-provincial pollution transmission path; after the model training is completed, the dynamic inter-provincial edge weight matrix learned by the model can be directly used to quantify and analyze the cross-regional transmission of pollutants; by analyzing the size and distribution of edge weights at a specific time step or time period, the key transmission path is dynamically identified, and the pollution contribution strength of the source province to the target province is evaluated, providing a scientific basis for accurate source tracing and joint prevention and control of regional ozone pollution;

[0013] S3, the evaluation results of the optimized national ozone prediction model on the validation set are used as the evaluation basis for model performance, the mean square error MSE is used as the evaluation index, the model with the lowest MSE value on the validation set is selected as the optimal model, and is used for fast and accurate prediction.

[0014] Further, in S1, a national pollutant and meteorological data large database is established as follows:

[0015] Through network data download, common pollutant observation data, ozone hourly concentration and daily maximum 8-hour moving average concentration data (MDA8 O3), AQI index and ERA5 reanalysis data of 326 cities in China from 2015 to 2024 are downloaded; the ERA5 reanalysis data includes relative humidity, precipitation, boundary layer height, shortwave radiation, sea level pressure, wind direction, wind speed and dew point; through data anomaly value processing, null value filling, data normalization and data set division, a high-quality database for supporting deep learning is established;

[0016] During the project implementation period, the ERA5 reanalysis data and pollutant observation data are continuously imported and data is arranged to realize dynamic updating of the large database.

[0017] Further, the common pollutants are NO2, SO2, CO, PM 2.5 and PM 10 .

[0018] Further, the data anomaly value processing, null value filling, data normalization and data set division are as follows:

[0019] Data anomaly value processing: the monitoring data is large or the equipment is not sensitive due to equipment damage abnormal reasons, so that the monitoring data is continuous zero in a period of time, and this part of data content in the conventional monitoring pollutant data and meteorological data is removed;

[0020] Null value filling: due to observation equipment or observation conditions, the observation data is obviously discontinuous, so it is very important to supplement the missing value, the linear interpolation method is adopted, and the missing value is filled by combining the historical average value of the same date and time in previous years;

[0021] Data normalization: data normalization is beneficial to improve the stability of training and the convergence speed of model training, and the normal distribution normalization method is adopted for all variables except wind direction, specifically, the feature attribute of data is subtracted by mean value and divided by variance; the standard normal distribution with mean value of 0 and variance of 1 is converted, and the expression is as follows:

[0022] x=(x-μ) / σ

[0023] In the formula, μ is the data mean value; σ is the data variance;

[0024] The mean and variance of the training set are recorded. After standardizing the training set, the test and validation data are standardized using the variance and mean from the training set. For wind direction, a periodic variable, a sine function is used for periodic feature encoding, as shown in the following expression:

[0025]

[0026] In the formula, θ represents the wind direction angle;

[0027] Dataset division: The data used are hourly observations of O3, MDA8, O3, NO2, SO2, CO, and PM from over 100 ground observation stations in 326 cities across China from 2015 to 2024. 2.5 PM 10 The data includes AQI data, as well as relative humidity, precipitation, boundary layer height, shortwave radiation, sea level pressure, wind direction, wind speed, and dew point from ERA5 reanalysis data.

[0028] The data from 2024 was selected as the test set to evaluate the model's capabilities. The remaining data was divided into training and validation sets with a 17:1 ratio. The training set was used to train the model parameters, and the validation set was used to select the model with the optimal number of training epochs.

[0029] The beneficial effects of this invention are:

[0030] This invention utilizes observational data on pollutants from multiple cities, processes anomalies and missing data, and combines hourly-scale meteorological numerical simulation data on relative humidity, precipitation, boundary layer height, shortwave radiation, sea level pressure, wind direction, wind speed, and dew point from ERA5 reanalysis data to establish a high-quality database to support deep learning. Based on this, an ozone concentration prediction model based on a multi-scale graph neural network and a dynamically learnable edge weight mechanism is constructed. A multi-level graph structure is established with cities and provinces as nodes. The model can adaptively adjust the spatial dependencies between nodes according to the spatiotemporal characteristics of the input, thereby simultaneously capturing the atmospheric pollution transport characteristics at both local and regional scales. During training, the model incorporates multi-scale feature extraction and edge weight update mechanisms, enabling the quantification and analysis of ozone propagation paths across different regions and the identification of potential pollution input relationships. Simultaneously, a channel attention mechanism is used to dynamically identify key meteorological factors and precursor factors, improving prediction accuracy while enhancing model interpretability. Through these technical solutions, this invention achieves rapid and accurate prediction of ozone concentrations nationwide, providing technical support for regional joint prevention and control and ozone pollution source tracing analysis. Attached Figure Description

[0031] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 This is a flowchart of an ozone concentration forecasting method according to the present invention. Detailed Implementation

[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] Example 1

[0035] A method for predicting ozone concentration includes the following steps:

[0036] Step 1: Collect pollutant observation data and ERA5 reanalysis data from 326 cities across the country to establish a large ozone pollution database, as detailed below:

[0037] Common pollutants (NO2, SO2, CO, PM2.5) from 2015 to 2024 were obtained from 326 cities across China through methods such as online data download. 2.5 PM 10 The dataset includes observational data, hourly ozone concentration and daily maximum 8-hour moving average concentration data (MDA8O3), AQI index, and ERA5 reanalysis data. The ERA5 reanalysis data includes relative humidity, precipitation, boundary layer height, shortwave radiation, sea level pressure, wind direction, wind speed, and dew point. Through outlier handling, null value imputation, data normalization, and dataset partitioning, a high-quality database is established to support deep learning.

[0038] Data outlier handling: If equipment malfunction or abnormalities cause the monitoring data to be too large or the equipment to be insensitive, resulting in the monitoring data being continuously zero for a period of time, this part of the data in the routine monitoring pollutant data and meteorological data will be removed.

[0039] Missing value filling: Due to the limitations of observation equipment or conditions, the observation data is often discontinuous. Therefore, it is crucial to fill in the missing values. This invention uses a linear interpolation method and combines it with the historical average value of the same date and time over the years to fill in the missing values.

[0040] Data normalization: Data normalization helps improve training stability and model convergence speed. This invention uses a normal distribution normalization method for all variables except wind direction. Specifically, it subtracts the mean from the data's feature attributes and divides by the variance, transforming it into a standard normal distribution with a mean of 0 and a variance of 1.

[0041] x=(x-μ) / σ

[0042] In the formula, μ is the data mean; σ is the data variance;

[0043] The mean and variance of the training set are recorded. After standardizing the training set, the test and validation data are standardized using the variance and mean of the training set. For wind direction, a periodic variable, a sine function is used for periodic feature encoding, as shown in the following expression:

[0044]

[0045] In the formula, θ represents the wind direction angle.

[0046] Dataset division: The data used are hourly observations of O3, MDA8, O3, NO2, SO2, CO, and PM from over 100 ground observation stations in 326 cities across China from 2015 to 2024. 2.5 PM 10 The dataset included AQI data and ERA5 reanalysis data on relative humidity, precipitation, boundary layer height, shortwave radiation, sea level pressure, wind direction, wind speed, and dew point. Data from 2024 was selected as the test set for evaluating model performance. The remaining data were divided into training and validation sets at a 17:1 ratio; the training set was used to train the model parameters, and the validation set was used to select the model with the optimal number of training epochs.

[0047] Step 2: Using a Transformer model based on a multi-scale graph neural network and a dynamically learnable edge weight mechanism, historical pollutant observation data and ERA5 reanalysis data from 326 cities across China are trained to construct an ozone forecasting model covering all 326 cities. The ozone forecasting model is then optimized using deep learning, as detailed below:

[0048] The model uses cities and provinces as nodes, with the initial input features of each node consisting of the corresponding pollutant concentrations and meteorological data, thus constructing a two-layer "city-province" graph structure. A multi-scale feature extraction mechanism is introduced into the graph neural network, learning local transport features on the city-level map and regional transport features on the provincial map, enabling the model to simultaneously capture the transport patterns of atmospheric pollutants at different spatial scales.

[0049] A dynamic, learnable edge weight mechanism is introduced during the modeling process. This mechanism adaptively adjusts the weight strength of the connecting edges between nodes in the graph based on the spatiotemporal characteristics of the input, such as real-time wind field, distance, and azimuth, thereby achieving dynamic modeling of spatial dependencies. Through this mechanism, the model can quantify and analyze the ozone propagation paths between different regions and identify potential pollution transport relationships.

[0050] Furthermore, a channel attention mechanism is introduced in the feature preprocessing stage of the graph neural network to evaluate the importance and dynamically weight the input multidimensional meteorological factors and precursor factors. This enables the model to automatically focus on key variables that have a significant impact on ozone prediction, thereby improving prediction accuracy and enhancing the interpretability of the causes of ozone pollution.

[0051] A. Model Selection Principles

[0052] To address the complexity of ozone prediction tasks, this invention employs a deep learning model combining a multi-scale graph neural network (MNN) and a Transformer architecture to adapt to the spatiotemporal correlation characteristics of multi-source heterogeneous data. The MNN can extract both local and regional pollutant transport features, and its dynamic edge weight mechanism adaptively captures spatial dependencies between different nodes. The Transformer's sequence modeling capabilities are suitable for predicting ozone concentrations over long time series. Through this structural combination, the model can balance spatial transport characteristics with temporal dynamics.

[0053] B. Model Establishment

[0054] In constructing the model, pollutant concentration data and meteorological reanalysis data from 326 cities across China were first used as input features to establish a graph structure, where nodes represent cities or provinces, and the initial weights of edges were determined based on geographical proximity and meteorological similarity. Multi-scale graph convolution operations were used to extract spatiotemporal features at different levels, and a Transformer module was embedded into the graph neural network to enhance its ability to capture time-series dependencies. Simultaneously, a dynamically learnable edge weight mechanism and a channel attention mechanism were introduced into each layer of computation, enabling the model to adaptively adjust edge weights and identify key factors during training, thereby improving prediction accuracy and interpretability.

[0055] C. Model Training Optimization

[0056] This invention will train and optimize the model through the following process:

[0057] I) First, the ozone monitoring dataset will be divided into a training set, a validation set, and a test set. To test the applicability of the model, the data from 2024 will be selected as the test set, and the data from other years will be divided into a training set and a validation set at a ratio of 17:1 for model training and validation.

[0058] II) Since meteorological data and air quality data have different units and large differences in magnitude, in order to reduce model training error, this invention adopts the normal distribution normalization method to eliminate the influence of the difference in the units of pollutants and meteorological data; at the same time, for the periodic variable wind direction, a sine function is used for periodic feature encoding.

[0059] III) During training, a batch size of 32 is selected, and the AdamW optimizer is used to update the model parameters. The learning rate is dynamically adjusted during training in combination with a learning rate scheduling strategy to improve the convergence efficiency and generalization ability of the model. At the same time, to ensure the randomness of the training samples, the training set is shuffled before each round of training.

[0060] IV) The model training judgment condition set by this invention is: when the maximum number of training rounds is reached or the loss function value of the validation set for 10 consecutive rounds does not decrease, it is considered that the condition has been met and training is stopped to save the parameters; otherwise, it returns to continue the next round of training.

[0061] V) Finally, to scientifically evaluate the accuracy of the model's predictions, this invention selects the mean square error (MSE) as the model's evaluation parameter;

[0062] Step 3: Use the evaluation results of the optimized national ozone forecast model on the validation set as the basis for evaluating the model performance. Use the mean squared error (MSE) as the evaluation index, select the model with the lowest MSE value on the validation set as the optimal model, and use it for rapid and accurate forecasting.

[0063] Example 2

[0064] The model construction process of an ozone concentration prediction method is as follows:

[0065] 1. Constructing the graph structure and input features

[0066] This invention constructs a multi-scale graph structure that can realistically reflect the physical mechanism of cross-regional transport of air pollutants, which includes a city-level directed graph and a provincial-level undirected graph.

[0067] 1.1 Constructing a city-level graph structure G city =(V city E city The 326 cities across the country are represented as nodes v in the graph. i ∈V city Each node is associated with a set of static attributes, including its latitude and longitude coordinates (lon). i ,lat i ) and altitude alt i These properties are used for graph structure construction and edge weight initialization. Edge e ij ∈E cityThis represents potential pollutant transport paths. Its construction is based on static node attributes: first, proximity filtering is performed by calculating Euclidean distance based on latitude and longitude; then, a terrain-aware filtering strategy is introduced, using the Bressenham line algorithm and altitude data for terrain feasibility filtering. For retained edges, their edge attribute vector e is defined. ij =[d ij ,θ ij ], where d ij To calculate the actual geographical distance between cities using the geodesic distance formula, θ ij The azimuth angle from the source city to the target city.

[0068] 1.2 Constructing the provincial graph structure G prov =(V prov E prov In the provincial map, each node represents a provincial administrative division, and its node features are aggregated by average pooling of the feature data of all cities in the province.

[0069] 1.3 The model input consists of multi-source spatiotemporal observation data for each city at time step t. This data has been standardized and includes concentration data for ozone, daily maximum 8-hour ozone, nitrogen dioxide, sulfur dioxide, carbon monoxide, fine particulate matter, inhalable particulate matter, and air quality index, as well as relative humidity, precipitation, boundary layer height, shortwave radiation, sea level pressure, wind speed, and dew point data obtained from ERA5 reanalysis data, totaling 15 features, denoted as x. i , which serves as the input to the graph neural network.

[0070] 2. Multi-scale map feature extraction

[0071] This invention employs a hierarchical graph feature extraction architecture, which uses graph neural networks at both the city and provincial levels to transform the input observation data into high-order feature representations containing complex spatial dependencies.

[0072] 2.1. Input data x of the model at time step t i The input is a city-level graph neural network, which performs message passing based on a city-level directed graph structure. The core function is to generate edge messages using a message function φ(·), which incorporates features from the source nodes. Target node features edge attribute e ij Wind field characteristics and dynamically calculated edge weights Using these as inputs, a high-dimensional side message vector m is generated through a nonlinear transformation. ij :

[0073]

[0074] In the formula, MLP is a multilayer perceptron consisting of linear layers and nonlinear activation functions. Target node v j The update characteristics are obtained through a symmetric message passing mechanism. This mechanism aggregates all incoming messages and all outgoing messages separately, and calculates their net effect:

[0075]

[0076] In the formula, N in (j) and N out (j) represent nodes v j The set of incoming and outgoing edge neighbors. During this process, the dynamic edge weights... Key contextual information is encoded into the message vector m. ij This process guides the model to generate more physically meaningful message content by internalizing the node embedding features. It incorporates spatial information from neighboring cities.

[0077] 2.2 Embedding city-level nodes Average pooling is performed based on provincial relationships to obtain the initial features of provincial nodes.

[0078]

[0079] In the formula, C p Let p represent the set of all cities belonging to province p; subsequently, Input a provincial-level graph neural network; this network is based on a provincial-level undirected graph structure and adopts a message-passing mechanism similar to that of the city-level network, but its edge weights β... pq The initialization focuses on geographic distance decay to simulate macroscopic interactions between provinces, ultimately outputting provincial node embedding features.

[0080] 2.3 To achieve cross-scale feature fusion, provincial nodes are embedded into the features. Broadcast to all city nodes based on city-province affiliation. Subsequently, a learnable, city-specific gating parameter γ is used. j Adaptive weighted fusion is performed on city-level embedded features and received provincial-level embedded features, while γ j By constraining the nodes within the range [0,1] using the Sigmoid function, the final enhanced multi-scale node embedding features are generated.

[0081]

[0082] In the formula, p(j) represents city v j The province to which it belongs;

[0083] This feature contains both local details and regional macro information, providing a powerful input for subsequent time series prediction modules.

[0084] 3. Dynamically learnable edge weight mechanism

[0085] This invention introduces a dynamic learnable edge weight mechanism, which enables the model to adaptively adjust the connection strength between nodes in the graph, thereby more accurately reflecting the actual transport potential of pollutants under different meteorological conditions.

[0086] 3.1 To ensure the scientific validity and convergence efficiency of the model, the edge weights are not learned from scratch, but are initialized based on physical priors.

[0087] For edges in a city-level graph, their initial static prior weights Calculated using the following physical formula:

[0088]

[0089] In the formula, v represents the wind speed at the source node, γ represents the azimuth angle from the source node to the target node, β represents the wind direction at the source node, and d represents the distance between the two nodes. This is then mapped to the Logit space using the Logit function, serving as the initial values ​​for the city-level edge weights.

[0090]

[0091] The physical prior initialization of provincial graph edge weights focuses on the geographical distance decay effect, following the general rule that spatial interactions weaken with increasing distance. Its initial static prior weights... Calculated using the following physical formula:

[0092]

[0093] In the formula, d e Let be the distance between province p and province q. Then, the distance is mapped to the Logit space using the Logit function, serving as the initial value for the provincial edge weights.

[0094]

[0095] 3.2 The static prior Logit value is modulated by a dynamic attention network. This network is based on a multilayer perceptron, whose input is a concatenated vector of source node features, target node features, normalized edge attributes, and wind field features, and outputs a dynamic residual.

[0096]

[0097] In the formula, and These represent the feature vectors of multi-source observation data from the source node and the target node, respectively. It is the normalized edge attribute vector. This is a wind field feature vector calculated from wind speed and direction. The MLP network learns the complex nonlinear relationships between these features and outputs a dynamic residual value to adjust the prior weights. The final learnable edge weights are represented by the Logit value. ij The static prior Logit and dynamic residuals are smoothly fused using a learnable global gating parameter λ, while λ is constrained to the range [0,1] by the Sigmoid function:

[0098]

[0099] 3.3 Finally, the fused Logit value is activated by the Sigmoid activation function σ(·). Conversion back to probabilistic form of edge weights

[0100]

[0101] Meanwhile, to characterize the normalized attention distribution of the target node to all incoming edge information, the Softmax function is used at the target node v j Normalize all incoming edges to obtain the final edge weights:

[0102]

[0103] Through the aforementioned operations and gating fusion mechanism in the Logit space, the model can effectively embed physical prior knowledge into learnable edge weights and allow them to be adaptively fine-tuned according to real-time meteorological conditions and pollutant distribution, thereby significantly improving the accuracy and reliability of simulations of complex atmospheric transport processes.

[0104] 4. Channel attention mechanism

[0105] To dynamically identify and enhance key factors that contribute significantly to ozone concentration prediction, this invention introduces a channel attention mechanism in the time-series prediction module. This mechanism adaptively learns the importance weights of different input feature channels, thereby highlighting key driving factors, suppressing redundant or noisy information, and improving prediction accuracy while enhancing the model's interpretability.

[0106] 4.1 First, embed the multi-scale node features obtained after multi-scale graph feature extraction. Perform dimensional transformation and recombination, merging batch dimension B and city dimension C to obtain the encoder input. Then, X encThe data is fed into a data embedding layer, where value embedding linearly projects the data into a high-dimensional space and incorporates positional encoding to capture sequence information, ultimately generating a semantically rich temporal feature representation. This lays the foundation for the subsequent encoder to extract global dependencies. In the above, B is the batch size, C is the number of cities, T is the length of the historical sequence, and D is the feature dimension. model This represents the feature dimension of the Transformer model.

[0107] Subsequently, the temporal features were represented. Perform pooling operations. Global average pooling is performed along the feature embedding dimension, compressing the high-dimensional embedding representation of each feature channel into a summary scalar, thereby generating a globally average pooled vector representing the global information of each feature channel.

[0108]

[0109] The vector g aggregates the comprehensive information of each feature channel after embedding and mapping. Each element corresponds to the global importance of an original feature channel, providing input for subsequent calculation of channel attention weights.

[0110] 4.2 Input the obtained vector g into an attention network consisting of fully connected layers. By projecting the input into a high-dimensional space to enhance its expressive power, it is then mapped back to the original dimension and the normalized attention score is output:

[0111] s=W2·δ(W1·g+b1)+b2

[0112] In the formula, W1, W2, b1, and b2 are learnable parameters, and δ is the ReLU activation function. Then, the Softmax function is used for normalization to obtain the final channel attention weights.

[0113] α = Softmax(s)

[0114] Normalization ensures that the sum of all weights is 1, making it easier to intuitively explain the relative importance of each channel.

[0115] 4.3 Finally, to apply the channel attention weights to the entire time series, the following steps are taken: Extending the time dimension, we generate time-series channel attention weights. The learned channel attention weights α are then used as scaling factors, and combined with the original multi-scale node embedding features. Channel-by-channel multiplication is performed to recalibrate the features:

[0116]

[0117] This operation enhances the characteristic channels that are crucial for ozone prediction, while mitigating the impact of irrelevant or redundant channels.

[0118] 5. Encoder section

[0119] The multi-scale node embedding feature H is reweighted using a channel attention mechanism and incorporates city-level and provincial-level graph features. fused After reshaping, the input is placed into the encoder embedding layer to incorporate time and position information, resulting in the encoder input E. enc :

[0120] E enc =Embedding enc (H fused ,T enc )

[0121] In the formula, T enc The time stamp for the input sequence.

[0122] Subsequently, the embedded features are input into an encoder based on the Transformer architecture to capture the complex global spatiotemporal dependencies in the input sequence:

[0123] Z enc =TransformerEncoder(E enc )

[0124] The encoder of this invention consists of a single stacked layer, incorporating a multi-head self-attention mechanism and a feedforward neural network, and employs residual connections and layer normalization to ensure training stability. Its output Z... enc A deep feature representation that contains global contextual information.

[0125] 6. Decoder section

[0126] The decoder's input consists of concatenating the end fragment of the historical sequence with placeholders from the future sequence, and is also embedded.

[0127] E dec =Embedding enc (X dec ,T enc )

[0128] In the formula, X dec For the decoder input sequence, T enc The time stamp for the predicted sequence.

[0129] Combine the decoder embedding features with the encoder output context features Z decThe input is taken together with the Transformer-based decoder to generate a prediction sequence in an autoregressive manner:

[0130] Z dec =TransformerDecoder(E dec Z enc )

[0131] The decoder employs a masked self-attention mechanism to ensure the autoregressiveness of the predictions, and dynamically focuses relevant information in the encoder output through an encoder-decoder attention mechanism. Its output Z... dec This represents the hidden state of the future sequence.

[0132] 7. Prediction and Optimization

[0133] A linear projection layer maps the decoder's output state to the final ozone concentration.

[0134]

[0135] In the formula, This refers to the ozone concentration prediction sequence generated by the model. The model minimizes the predicted values. The mean square error (MSE) between the true value Y and the actual value Y is optimized.

[0136] 8. Analysis of inter-provincial pollution transmission paths

[0137] The provincial graph neural network provided by this invention can output a dynamic inter-provincial edge weight matrix that changes over time. matrix elements It is a function of time t, used to dynamically characterize the real-time pollutant transport contribution intensity from source province q to target province p under specific meteorological conditions;

[0138] Based on the dynamic edge weight matrix A t The method for analyzing the impact of inter-provincial pollution transmission is as follows: direct observation, statistics, and analysis of the dynamic edge weights. The numerical magnitude, temporal variation, and spatial distribution patterns of pollutants are used to quantify and analyze the dynamic evolution of pollutant transport intensity between different provincial regions; among them, the weight values... The magnitude of the value directly characterizes the strength of the transmission contribution, and its change over time reflects the dynamic response of the transmission relationship to meteorological conditions.

[0139] Using the above methods, the dynamic weights output by the model can be used to quantitatively analyze the dynamic transmission patterns of pollutants between provinces under different meteorological conditions, providing a scientific basis for accurate source tracing and joint prevention and control of regional ozone pollution.

Claims

1. A method for predicting ozone concentration, characterized in that, Includes the following steps: S1. Collect air quality observation data from 326 cities across the country from 2015 to 2024 through network data collection, perform null value filling, data normalization and dataset division, and combine the relative humidity, precipitation, boundary layer height, shortwave radiation, sea level pressure, wind direction, wind speed and dew point in the ERA5 reanalysis data to establish a high-quality database to support deep learning. S2. A Transformer model based on a multi-scale graph neural network and a dynamically learnable edge weight mechanism was used to train historical pollutant observation data and ERA5 reanalysis data from 326 cities across China to construct an ozone forecasting model covering all 326 cities. The model was then optimized. Details are as follows: S2.

1. A multi-scale graph neural network model is used to train the historical dataset. Ten years of pollutant data and numerical simulation data provided by ERA5 reanalysis data are used as input to the graph neural network model, and the predicted ozone concentration is used as output to establish an hourly ozone forecast model for 326 cities across the country. S2.2 Constructing a "city-province" two-layer graph structure to achieve multi-scale feature extraction: At the city level, a directed graph structure is constructed, where edges represent potential pollutant transport paths between cities. Its construction is based on geographical proximity and introduces a terrain-aware filtering strategy to ensure the physical feasibility of transport paths; At the provincial level, an undirected graph structure is constructed, where node features are generated by average pooling aggregation of features from all cities within the province, and edge weights are initialized based on the inter-provincial geographical distance decay law to simulate macroscopic spatial interactions; S2.

3. A dynamic learnable edge weight mechanism is introduced during the modeling process. Based on the real-time spatiotemporal characteristics of the input, this mechanism adaptively adjusts the weight strength of the connecting edges between nodes in the graph through a gated fusion network. This achieves a smooth fusion of static physical prior weights calculated based on wind field dynamics formulas and dynamic residuals learned by an attention network. This enables the model to accurately quantify the dynamic transport paths and contribution levels of pollutants under different meteorological conditions and identify potential pollution input relationships. S2.

4. Introduce a channel attention mechanism in the graph neural network. This mechanism operates at the input feature level. Through an attention network composed of fully connected layers, it dynamically calculates the importance score of each input feature channel and uses the Softmax function to normalize and obtain the final channel attention weight, thereby reweighting the original input features. S2.

5. Realize the dynamic analysis function of inter-provincial pollution transmission paths; after the model is trained, the dynamic inter-provincial edge weight matrix it learns can be directly used to quantify and analyze the cross-regional transmission relationship of pollutants; by analyzing the size and distribution of edge weights within a specific time step or time period, key transmission paths can be dynamically identified and the pollution contribution intensity of source provinces to target provinces can be assessed, providing a scientific basis for the accurate source tracing and joint prevention and control of regional ozone pollution. S3. The evaluation results of the optimized national ozone forecast model on the validation set are used as the basis for evaluating the model performance. The mean square error (MSE) is used as the evaluation index. The model with the lowest MSE value on the validation set is selected as the optimal model and used for rapid and accurate forecasting.

2. The ozone concentration forecasting method according to claim 1, characterized in that, In S1, the establishment of a national database of pollutant and meteorological data is detailed as follows: By downloading data from the internet, we collected observational data of common pollutants from 2015 to 2024 in 326 cities across China, hourly and daily maximum 8-hour moving average ozone concentration data (MDA8 O3), AQI index, and ERA5 reanalysis data. The ERA5 reanalysis data included relative humidity, precipitation, boundary layer height, shortwave radiation, sea level pressure, wind direction, wind speed, and dew point. Through outlier handling, null value imputation, data normalization, and dataset partitioning, we established a high-quality database to support deep learning. During the project execution period, ERA5 reanalysis data and pollutant observation data were continuously entered into the database and organized to achieve dynamic updates of the big data database.

3. The ozone concentration forecasting method according to claim 2, characterized in that, The common pollutants mentioned are NO2, SO2, CO, and PM2.

5. 2.5 and PM 10 .

4. The ozone concentration forecasting method according to claim 2, characterized in that, The specific details of outlier handling, null value imputation, data normalization, and dataset partitioning are as follows: Data outlier handling: If equipment malfunction or abnormalities cause the monitoring data to be too large or the equipment to be insensitive, resulting in the monitoring data being continuously zero for a period of time, this part of the data in the routine monitoring pollutant data and meteorological data will be removed. Missing value filling: Due to the limitations of observation equipment or conditions, the observation data is often discontinuous. Therefore, it is crucial to fill in the missing values. This invention uses a linear interpolation method and combines it with the historical average value of the same date and time over the years to fill in the missing values. Data normalization: Data normalization helps improve training stability and model convergence speed. This invention uses a normal distribution normalization method for all variables except wind direction. Specifically, the mean of the data's feature attributes is subtracted, and then the result is divided by the variance; this transforms the data into a standard normal distribution with a mean of 0 and a variance of 1, as shown in the following expression: x=(x-μ) / σ In the formula, μ is the data mean; σ is the data variance; The mean and variance of the training set are recorded. After standardizing the training set, the test and validation data are standardized using the variance and mean from the training set. For wind direction, a periodic variable, a sine function is used for periodic feature encoding, as shown in the following expression: In the formula, θ represents the wind direction angle; Dataset division: The data used are hourly observations of O3, MDA8, O3, NO2, SO2, CO, and PM from over 100 ground observation stations in 326 cities across China from 2015 to 2024. 2.5 PM 10 The data includes AQI data, as well as relative humidity, precipitation, boundary layer height, shortwave radiation, sea level pressure, wind direction, wind speed, and dew point from ERA5 reanalysis data. The data from 2024 was selected as the test set to evaluate the model's capabilities. The remaining data was divided into training and validation sets with a 17:1 ratio. The training set was used to train the model parameters, and the validation set was used to select the model with the optimal number of training epochs.

Images