A long-term prediction method and system for ozone concentration based on three-dimensional regional meteorological field

Abstract

The application provides a long-term ozone concentration prediction method and system based on a three-dimensional regional meteorological field, and relates to the technical field of environmental prediction. The method fuses a three-dimensional meteorological field composed of ground and high-altitude meteorological characteristics, an ozone precursor source intensity prior field constructed based on a regional emission inventory, and site O3 observation data, forms multi-source heterogeneous spatio-temporal characteristics, and inputs a 2D-CNN-Transformer model combined with weather forecast data to realize direct multi-time ozone concentration prediction in the future 1-14 days. The model uses CNN to extract spatial distribution characteristics, maps the source intensity prior field to a spatial attention bias item, uses the Transformer to learn the time sequence dependency relationship between the historical state, the future meteorological driving condition and the prediction time, and constructs a loss weight constraint model according to the source intensity prior field to train the model. The application improves the accuracy and stability of long-term regional ozone concentration prediction.

Description

Technical Field

[0001] This invention relates to the field of environmental prediction technology, and in particular to a method and system for long-term prediction of ozone concentration based on a three-dimensional regional meteorological field. Background Technology

[0002] Current ozone concentration prediction is mostly based on models constructed from near-surface meteorological factors, historical pollutant concentration sequences, and two-dimensional regional characteristics, which is the mainstream technical means for regional ozone pollution early warning.

[0003] However, existing methods struggle to characterize the complex mechanism by which ozone formation is influenced by the synergistic effects of multi-scale meteorological processes. Deep learning models also generally lack key information on precursor sources, resulting in insufficient feature representation completeness. Furthermore, in medium- to long-term forecasting scenarios, traditional daily recursive models accumulate significant errors over time, leading to decreased accuracy and stability in multi-day forecasts. Consequently, they cannot effectively support the early assessment and routine operational early warning applications of regional ozone pollution processes. Summary of the Invention

[0004] To address the aforementioned issues, this invention proposes a long-term ozone concentration prediction method and system based on a three-dimensional regional meteorological field, thereby improving the accuracy and stability of regional ozone concentration prediction.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a long-term ozone concentration prediction method based on a three-dimensional regional meteorological field, comprising: Based on regional emission inventory data, a strong prior field for ozone precursor sources is constructed. Acquire a three-dimensional meteorological feature field covering an area larger than the ozone concentration forecast target area, including upstream and surrounding background meteorological information that affects the ozone evolution in the target area; By integrating three-dimensional meteorological feature field, strong prior field of ozone precursor source and station O3 observation data, multi-source heterogeneous spatiotemporal characteristics are obtained; The multi-source heterogeneous spatiotemporal features and weather forecast data are input into a 2D-CNN-Transformer model to obtain ozone concentration prediction results. In the model training phase, spatial features are extracted from the multi-source heterogeneous spatiotemporal features based on a CNN network, and the source strength prior field is mapped as a spatial attention bias term to enhance the model's response to the spatial distribution of ozone precursor source strength and key affected areas. The correlation between different time steps and different prediction timelines is modeled based on a Transformer network to capture the temporal evolution characteristics of ozone concentration over the next 1–14 days. Loss weights are constructed based on the source strength prior field, and the model is trained by combining the loss function with the loss weights to obtain a trained 2D-CNN-Transformer ozone concentration prediction model.

[0006] Secondly, the present invention provides a long-term ozone concentration prediction system based on a three-dimensional regional meteorological field, comprising: The prior field construction module is configured to construct a strong prior field for ozone precursor sources based on regional emission inventory data. The three-dimensional meteorological feature construction module is configured to acquire a three-dimensional meteorological feature field covering an area larger than the ozone concentration forecast target area, including upstream and surrounding background meteorological information that affects the ozone evolution in the target area; The feature fusion module is configured to fuse three-dimensional meteorological feature fields, ozone precursor source strong prior fields, and station O3 observation data to obtain multi-source heterogeneous spatiotemporal features; The concentration prediction module is configured to input the multi-source heterogeneous spatiotemporal features and weather forecast data into a 2D-CNN-Transformer model to obtain ozone concentration prediction results; In the model training phase, spatial features are extracted from the multi-source heterogeneous spatiotemporal features based on a CNN network, and the source strength prior field is mapped as a spatial attention bias term to enhance the model's response to the spatial distribution of ozone precursor source strength and key affected areas. The correlation between different time steps and different prediction timelines is modeled based on a Transformer network to capture the temporal evolution characteristics of ozone concentration over the next 1–14 days. Loss weights are constructed based on the source strength prior field, and the model is trained by combining the loss function with the loss weights to obtain a trained 2D-CNN-Transformer ozone concentration prediction model.

[0007] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the long-term ozone concentration prediction method based on a three-dimensional regional meteorological field as described in the first aspect.

[0008] Fourthly, the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the long-term ozone concentration prediction method based on a three-dimensional regional meteorological field described in the first aspect.

[0009] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention integrates near-surface meteorological data with upper-air meteorological data at different altitudes to construct a three-dimensional meteorological driving field that reflects ozone photochemical formation conditions, boundary layer mixing processes, vertical exchange processes, regional transport, and large-scale circulation background, providing dynamic meteorological constraints for the spatiotemporal evolution of ozone concentration in the model. It introduces emission inventory data for NOx, NMVOC, and CO to construct a source strength prior field, which is used as model input, spatial attention bias, and loss weighting to enhance the model's ability to learn features related to ozone formation. Simultaneously, it employs a CNN-Transformer structure to jointly model spatial features, long-term dependencies, and prediction timeliness differences, enabling direct multi-timeliness prediction of ozone concentration (daily maximum 8-hour average ozone concentration, MDA8O3) for the next 1–14 days. This reduces the error accumulation caused by daily recursion, improves prediction accuracy and stability, and enables the identification and early warning of ozone pollution events based on the prediction results.

[0010] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0011] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute a limitation thereof.

[0012] Figure 1 The main flowchart of a long-term ozone concentration prediction method based on a three-dimensional regional meteorological field provided in this embodiment of the invention; Figure 2 A flowchart of model prediction provided for embodiments of the present invention; Figure 3 The diagram illustrates the comparison of MDA8 O3 concentration predictions based on different schemes provided in this embodiment of the invention; where (a) represents the comparison of the coefficient of determination (R²) of the model under different forecast lead times; (b) represents the comparison of the hit rate of different forecast schemes for high ozone events (HPD); and (c) represents the comparison of the observation and the MDA8 O3 concentration time series of the two forecast schemes. Detailed Implementation

[0013] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0014] Currently, air pollutant concentration prediction mainly employs two methods: numerical models and data-driven approaches. Numerical simulation methods, based on atmospheric physicochemical mechanisms, couple meteorological fields, emission inventories, and chemical mechanisms to simulate pollutant concentration evolution. While offering strong mechanistic interpretation, they suffer from high computational costs and parameter uncertainties, leading to prediction biases and hindering long-term predictions of air pollution concentrations over large areas. Data-driven models, based on multi-source datasets, learn the complex nonlinear relationship between input features and pollutant concentrations, automatically extracting key features and parameters. They offer high efficiency and accuracy in air quality forecasting. In recent years, the development of deep learning architectures has enhanced the feature learning and spatiotemporal representation capabilities of data-driven methods, making them more suitable for pollution-meteorological coupled systems. Convolutional Neural Networks (CNNs), relying on local receptive fields and parameter sharing mechanisms, are suitable for efficiently modeling the multi-scale spatial structure of regular gridded meteorological and pollutant concentration fields, exhibiting strong computational parallelism and scalability. Transformers, relying on self-attention mechanisms, establish arbitrary time-step dependencies in the time dimension, avoiding gradient decay and error accumulation problems, and demonstrating stronger expressive power in long-term time-series modeling and prediction. Although CNN and Transformer are complementary, existing methods for long-term prediction of pollutants such as ozone mostly adopt a single structure, which makes it difficult to take into account both spatial feature extraction and long-term time-series dependency modeling, thus affecting the simulation and prediction of pollution evolution characteristics.

[0015] The formation of high concentrations of pollutants depends on emission intensity, with meteorological conditions playing a decisive modulating role in pollutant transport, distribution, and diffusion. Near-surface O3 pollution is closely related to the meteorological background of photochemical formation and precursor accumulation, such as high temperatures, strong solar radiation, near-surface wind convergence, and boundary layer stability. Local conditions are driven by various weather systems, including subtropical high pressure, typhoon peripheral circulation, and anticyclone control. Therefore, regional and upper-air meteorological conditions affect local ozone levels. Accurate O3 forecasting requires comprehensive consideration of these factors; however, existing studies often use only local meteorological elements and small-scale regional fields as model inputs, neglecting the importance of multi-scale regional meteorological fields, making it difficult to support high-precision, large-scale pollution forecasting and causal analysis.

[0016] To address the aforementioned issues, this invention provides a medium- to long-term prediction and early warning method for MDA8 O3 based on a three-dimensional regional meteorological field, prior constraints on ozone precursor source strength, and multi-time-dependent joint prediction using CNN-Transformer. By fusing near-surface meteorological fields and upper-level atmospheric fields at different altitudes, a three-dimensional regional meteorological driving field is constructed to reflect the meteorological conditions related to ozone photochemical generation, boundary layer diffusion, vertical exchange, and regional transport, providing meteorological driving information for the model to learn the spatiotemporal evolution characteristics of MDA8 O3 concentration. By introducing emission inventory data to construct prior constraints on source strength, the model's ability to learn features of high-source-strength ozone precursor regions and their surrounding influence areas is improved. Spatial structural features are extracted using CNN, and a long-term time-series dependency relationship is established between historical meteorological evolution, future meteorological forecast conditions, and different prediction time-series results using Transformer, enabling direct joint prediction of MDA8 O3 concentration over multiple days. This alleviates the problem of accumulated recursive prediction errors and improves the accuracy, stability, and reliability of medium- to long-term predictions, thereby providing technical support for regional ozone pollution process early warning, refined air quality management, and public health risk protection.

[0017] Example 1 like Figure 1 As shown in the figure, this embodiment discloses a long-term ozone concentration prediction method based on a three-dimensional regional meteorological field, including the following steps: S1: Construct a strong prior field for ozone precursor sources based on regional emission inventory data; S2: Obtain a three-dimensional meteorological feature field covering an area larger than the ozone concentration forecast target area, including upstream and surrounding background meteorological information that affects the ozone evolution in the target area; S3: By integrating three-dimensional meteorological feature field, strong prior field of ozone precursor source and station O3 observation data, multi-source heterogeneous spatiotemporal characteristics are obtained; S4: Input the multi-source heterogeneous spatiotemporal features and weather forecast data into the 2D-CNN-Transformer model to obtain ozone concentration prediction results; In the model training phase, spatial features are extracted from the multi-source heterogeneous spatiotemporal features based on a CNN network, and the source strength prior field is mapped as a spatial attention bias term to enhance the model's response to the spatial distribution of ozone precursor source strength and key affected areas. The correlation between different time steps and different prediction timelines is modeled based on a Transformer network to capture the temporal evolution characteristics of ozone concentration over the next 1–14 days. Loss weights are constructed based on the source strength prior field, and the model is trained by combining the loss function with the loss weights to obtain a trained 2D-CNN-Transformer ozone concentration prediction model.

[0018] Next, combined Figure 1This embodiment provides a detailed description of a long-term ozone concentration prediction method based on a three-dimensional regional meteorological field.

[0019] (a) Data processing Acquire three-dimensional meteorological characteristic fields, emission inventory data, and station O3 observation data for the target area.

[0020] 1. Three-dimensional meteorological characteristic field The three-dimensional meteorological feature field is a spatial three-dimensional forecast data field constructed based on three dimensions: longitude, latitude, and vertical altitude. It includes meteorological elements such as temperature, pressure, humidity, and wind. The horizontal coverage of the three-dimensional meteorological forecast field is larger than the ozone concentration forecast target area. In addition to covering the target area itself, it also includes the upstream and surrounding background meteorological influence areas, which is used to characterize the large-scale circulation background, regional transport conditions, and their impact on the ozone evolution process in the target area.

[0021] Using the start time of the weather forecast as the sample anchor point, the original three-dimensional meteorological feature field is processed by horizontal spatial grid unification, vertical layer matching, and forecast lead time alignment to obtain a three-dimensional meteorological feature field corresponding to the target forecast lead time. Specifically, horizontal spatial grid unification is used to transform meteorological variables with different spatial resolutions to a unified latitude and longitude grid for the target area; vertical layer matching is used to transform data from different pressure layers, altitude layers, or model native layers to the target vertical layer; and forecast lead time alignment is used to map future weather forecast data with different lead times to the target forecast date.

[0022] By introducing a three-dimensional meteorological forecast field with a coverage area larger than the target area, the model can simultaneously acquire meteorological conditions within the target area and large-scale background meteorological information from the outside, thereby enhancing its ability to characterize regional ozone transport, pollution accumulation, and meteorological-driven changes, and improving the accuracy and stability of long-term prediction results for regional MDA8 O3 concentration.

[0023] To ensure data quality, vector-preserving interpolation was used for near-surface wind field variables, while bilinear interpolation or area-weighted interpolation was used for continuous variables such as temperature, humidity, air pressure, and geopotential height. Conservative regrinding and time accumulation or splitting were used for flux or cumulative variables such as solar radiation and precipitation, and discrete interpolation was used for diagnostic variables such as boundary layer height.

[0024] 2. Emissions Inventory Data Emissions inventory data is a collection of systematically quantified records of the total amount of pollutants or greenhouse gases released into the atmosphere (or greenhouse gases) from various emission sources over a specific time span and spatial region.

[0025] The emission inventory data described in this embodiment includes one or more of ozone precursor emission components and particulate matter-related emission components. The ozone precursor emission components include one or more of NOx, CO, and NMVOC (non-methane volatile organic compounds).

[0026] The emission inventory data is processed by pollutant category analysis, mass conservation regrinding, time spectrum allocation and normalization to generate an emission feature field that corresponds to the three-dimensional meteorological feature field in terms of spatial grid and prediction time.

[0027] Based on the ozone precursor components such as NOx, NMVOC, and CO in the emission characteristic field, an ozone precursor source strength prior field is constructed to characterize the spatial distribution and temporal variation characteristics of ozone formation-related precursors in the target area.

[0028] Specifically, based on the normalized emission characteristic fields of ozone precursors such as NOx, NMVOC, and CO, a prior field of ozone precursor source strength is constructed: ; in, Represents the spatial grid under the h-th prediction lead time. The prior value of the ozone precursor source strength at the location, where K represents the set of ozone precursor components; This represents the emission field of the k-th type of precursor after mass conservation regrinding, time spectrum allocation, and normalization. This represents the contribution weights of different precursors to the prior source strength. b represents the bias term. This represents the Sigmoid function, used to compress the prior field to the range of 0-1.

[0029] In this embodiment, by constructing a strong prior field for ozone precursor sources, the spatiotemporal distribution patterns of ozone precursors can be accurately characterized, the input features of the prediction model can be improved, and the accuracy and spatiotemporal evolution fitting ability of the regional ozone concentration prediction results can be effectively enhanced.

[0030] 3. O3 observation data at the site Quality control and daily maximum 8-hour moving average calculations were performed on the hourly O3 observation data of the stations to obtain the station-scale MDA8 O3 labels. Based on the station's spatial location, terrain similarity, meteorological similarity, and observation integrity, label mapping weights were constructed to map the station MDA8 O3 labels to the target grid and generate a label credibility mask.

[0031] In addition, it also includes further calculation of the mechanism constraints related to ozone generation and transport based on the three-dimensional meteorological feature field, including boundary layer mixing characteristics, wind field transport characteristics, solar radiation accumulation characteristics, and high temperature and low humidity photochemical enhancement characteristics.

[0032] The three-dimensional meteorological feature field, emission feature field, ozone precursor source strength prior field, mechanism constraint features, and data quality mask are combined according to spatial grid, vertical layer, prediction timeliness, and variable channel to obtain a multi-source heterogeneous spatiotemporal feature tensor for input to the CNN-Transformer model; gridded MDA8 O3 labels and their credibility mask are then used. It serves as a supervisory signal for model training.

[0033] (ii) Model Prediction like Figure 2 As shown, a 2D-CNN-Transformer multi-time-dependent ozone concentration prediction model is constructed. The model receives two types of input data: one type is historical state data from the past N days, including a three-dimensional regional meteorological field, emission inventory characteristics, strong prior fields of ozone precursor sources, mechanistic constraint characteristics, and other auxiliary prediction factors. Simultaneously, a data quality mask is introduced during the training phase to identify the validity and reliability of multi-source input data and gridded MDA8 O3 labels. As one implementation method, other service prediction factors include time variables. The other type is meteorological forecast data for the next 14 days, used to provide meteorological driving conditions for future ozone changes.

[0034] Based on the above input data, the model learns the relationship between ozone concentration in the target area and changes in meteorological conditions, emission source strength and time, and generates daily grid-scale MDA8 O3 concentration prediction results for the next 1–14 days in one go.

[0035] Specifically, the model adopts a structure of "historical state encoding - future weather driving - multi-time-effect joint decoding".

[0036] First, the meteorological field, emission characteristics, and ozone precursor source strength prior field of the past N days are used as historical state inputs. Local spatial features at each time step are extracted by 2D-CNN to obtain spatial embedding features that reflect the historical pollution background, meteorological evolution, and emission source strength distribution of the target area.

[0037] Specifically, spatial attention bias terms and training loss weights are generated based on the strong prior field of ozone precursor sources to guide the neural network model to enhance its feature learning ability of ozone generation source regions and their surrounding influence areas.

[0038] Specifically, the strong prior field of ozone precursor sources is mapped to a spatial attention bias term and superimposed on the original query-key attention logits: ; in, For spatial attention bias, Used to control the moderating power of source strength priors on attention allocation. This indicates the target spatial location where the ozone concentration to be predicted is located. This indicates the spatial location or potential source region of influence involved in attention calculation. Indicates the target location With source strength location Spatial association weights between them are constructed based on source strength priors; Indicates a query. Indicates key-value pair, This indicates the feature dimension. This term can improve the attention score of the spatial location corresponding to the high source intensity of ozone precursors before attention normalization, thereby guiding the model to enhance its feature learning ability of relevant emission source areas and their potential impact areas.

[0039] Meanwhile, during model training, training loss weights are constructed based on the source strength prior field: ; in, Represents the spatial grid under the h-th prediction lead time. Loss weight at the point, It serves as a strong a priori source for ozone precursors. This is the weighting adjustment coefficient. For the h-th prediction time, the spatial grid Values ​​are taken from the strong prior field of ozone precursor sources; and These represent the rows and columns of the spatial grid, respectively. This weighting, by increasing the error penalty in high-intensity precursor source regions, guides the model to enhance its learning ability regarding ozone generation-related source regions and their potential impact areas.

[0040] Furthermore, combining label credibility masks Constructing a weighted prediction loss: ; in, Indicates the h-th prediction lead time, ... The actual ozone concentration on each spatial grid. This indicates that the model predicts ozone concentration. It is represented as a very small positive number to avoid denominators of 0 and to improve the stability of numerical calculations.

[0041] This weighting mechanism is a soft constraint. It guides the model to enhance its learning ability on high source intensity regions of ozone precursors and their potential impact regions by adjusting the error penalty intensity for different spatial locations and prediction time, rather than directly limiting the monotonic relationship between ozone concentration and precursor emission intensity.

[0042] Then, the 14-day weather forecast field is organized into future driving inputs according to the forecast lead time h=1,2,…,14, and a corresponding lead time code is constructed for each forecast lead time. This code is used to identify the forecast result for the current input on which day of the future, and is input into the Transformer module together with the weather forecast features, enabling the model to identify the differences in the impact of meteorological factors on the changes in MDA8 O3 concentration within different forecast lead times.

[0043] In the Transformer module, a self-attention mechanism is used to learn the dependencies between historical spatiotemporal windows, future weather forecast windows, and different prediction timeframes. Simultaneously, a strong prior field of ozone precursor sources is used as input features, spatial attention bias terms, and training loss weights to guide the model in enhancing its ability to learn features of ozone-generating source regions and their surrounding influence areas. Finally, a multi-timeframe prediction decoder simultaneously outputs grid-scale MDA8 O3 concentration predictions for days 1 to 14 of the future during a single forward propagation.

[0044] Unlike the daily recursive forecasting method, the model adopts a multi-time-leader direct forecasting approach. Instead of using the forecast results of the previous day as the input for the next day, it directly generates daily forecast results for the next 1-14 days using the 14-day weather forecast sequence and forecast time-leader codes. This reduces the problem of gradual accumulation of errors during the recursive forecasting process and improves the stability of medium- and long-term ozone concentration forecasts.

[0045] Based on the daily grid-scale MDA8 O3 concentration predictions for the next 1-14 days output by the model, an ozone pollution event identification and early warning module is constructed. For day h... Predicted concentration per grid This is compared with the ozone pollution threshold. Compare; when When the ozone level exceeds the standard, the grid is determined to be in an ozone-exceeding state under the corresponding prediction time.

[0046] As one implementation method, the 2D-CNN-Transformer model mainly consists of five functional modules: a meteorological feature extraction module, a historical meteorological sequence encoder, a decoder driven by a meteorological forecast sequence, an output smoothing module, and a final prediction module. The meteorological feature extraction module takes as input the three-dimensional meteorological field, emission inventory features, ozone precursor source strength prior field, and other predictive factors for each time step, and outputs spatial embedding features. The historical meteorological sequence encoder takes as input the spatial embedding feature sequence of the past N days, and outputs historical state memory features. The meteorological forecast sequence driven decoder takes as input the meteorological forecast features for the next 14 days, prediction lead time encoding, and historical state memory features, and outputs the decoded features corresponding to each future prediction lead time. The output smoothing module takes as input multi-lead time decoded features or a preliminary prediction sequence, and outputs smoothed features with enhanced temporal consistency. The final prediction module takes as input the smoothed multi-lead time features, and outputs the daily MDA8 O3 concentration prediction results for the next 1–14 days.

[0047] During the model input phase, the input meteorological field data is normalized or standardized to improve the stability of the training process and accelerate convergence. In the spatial feature modeling phase, a two-dimensional convolutional neural network is used to extract the spatial features of the meteorological field through multi-layer 3×3 convolution operations and the ReLU activation function. Furthermore, a channel attention module is introduced during the downsampling process to adaptively weight different feature channels, thereby enhancing the feature representation capabilities relevant to the prediction task.

[0048] Historical meteorological features are embedded into the Transformer encoder to capture long-term temporal dependencies. Simultaneously, 14-day weather forecast data is used to extract features from a CNN network with the same structure as the historical branch, and then fused with temporal location encoding before being input into the Transformer decoder. The decoder interacts with the encoder output through a cross-attention mechanism, thereby fusing historical meteorological information with future forecast features. Finally, a three-layer fully connected network is used to generate daily, grid-scale pollutant concentration predictions.

[0049] Furthermore, a duration constraint is introduced: when a preset proportion of grids within the same grid or target area meet the ozone exceedance criteria for D consecutive days, an ozone pollution event is determined. Based on the prediction lead time for the first fulfillment of the pollution event determination criteria, the early warning time for the pollution event is determined, and the occurrence date, duration, affected area, peak exceedance value, and warning level of the pollution event are output.

[0050] To verify the effectiveness of this embodiment, the following specific implementation method is provided.

[0051] The method for predicting O3 in the North China region using the MDA8 grid for the next 14 days includes the following steps: Step 1: Obtain meteorological forecast data, ground monitoring data, and emission inventory data; Step 2: Preprocess the acquired multi-source data by unifying the horizontal spatial grid, matching the vertical layers, aligning the prediction time, resampling and interpolation, etc., to unify the data from different sources to a spatial resolution of 0.1°×0.1° and a daily time resolution, and normalize each input feature; extract a three-dimensional meteorological forecast field with a coverage area larger than the forecast target area, the three-dimensional meteorological forecast field including upstream and surrounding background meteorological information that affects the ozone evolution in the target area; Step 3: Match multidimensional meteorological data, emission inventory data, ozone precursor source strong prior field and other predictive factors with the corresponding MDA8 O3 observation data to construct a dataset with grid-scale MDA8 O3 as the prediction target, and divide it into model training dataset and prediction dataset. Step 4: Input the constructed model training dataset into the CNN-Transformer model to train and optimize the model parameters; Step 5: Generate daily grid-scale MDA8O3 concentration predictions for the next 1–14 days using the output smoothing module and the final prediction module.

[0052] The meteorological forecast data in step 1 includes near-surface meteorological elements and upper-air meteorological elements at one or more altitude layers, used to construct a three-dimensional regional meteorological field reflecting relevant meteorological conditions such as ozone formation, diffusion, vertical exchange, regional transport, and large-scale circulation background. Optionally, near-surface meteorological elements include 2m air temperature, relative humidity, surface-to-surface solar radiation, surface air pressure, and 10m wind field component; upper-air meteorological elements include air temperature, wind field component, relative humidity, and geopotential height at a selected pressure layer, which may include the 850hPa layer and the 500hPa layer. Emission inventory data includes pollutant component data such as BC, OC, NOx, SO2, CO, NH3, PM2.5, PM10, and NMVOC.

[0053] The data preprocessing in step 2 includes: performing hourly O3 concentration quality control on ground monitoring stations and calculating MDA8 O3; spatially interpolating and gridding the station-scale MDA8 O3 to a spatial resolution of 0.1°, which is then used as the model training label. Horizontal spatial grid unification, vertical layer matching, and prediction timeline alignment are performed on near-surface and upper-level meteorological field data to unify them to the target spatial grid, target vertical layer, and target prediction timeline. Pollutant category analysis, mass conservation regrinding, time allocation, and normalization are performed on the emission inventory data to generate an emission characteristic field corresponding to the target spatial grid and prediction timeline; and a strong prior field of ozone precursor sources is constructed based on ozone precursor components such as NOx, NMVOC, and CO.

[0054] In step 3, the preprocessed 3D meteorological feature field, emission inventory features, ozone precursor source strength prior field, and other prediction factors are matched with the corresponding grid scale MDA8 O3 labels according to the time dimension and spatial grid location to construct a supervised learning dataset. Optionally, a five-dimensional input tensor is constructed: X∈R H×W×L×T×C Where H and W represent the horizontal spatial grid size, L represents the number of vertical layers, T represents the historical time step and / or future prediction lead time, and C represents the number of variable channels. The variable channels include one or more of the following: meteorological variables, emission variables, prior characteristics of ozone precursor source strength, and ozone mechanism constraint characteristics. The dataset is divided into a model training dataset and a prediction dataset according to chronological order.

[0055] In step 4, the 2D-CNN-Transformer model includes a meteorological feature extraction module, a historical meteorological sequence encoder, a decoder driven by a meteorological forecast sequence, an output smoothing module, and a final prediction module, which are connected sequentially according to the data flow of "spatial feature extraction - historical state encoding - future meteorological-driven decoding - prediction result smoothing - concentration result output". The meteorological feature extraction module takes the three-dimensional meteorological field, emission inventory features, ozone precursor source strength prior field, and other predictive factors at each time step as input, and extracts spatial embedding features through two-dimensional convolution; the historical meteorological sequence encoder takes the spatial embedding feature sequence of the past N days as input and outputs historical state memory features; the meteorological forecast sequence driven decoder takes the meteorological forecast features of the next 14 days, the prediction time-leading encoding, and the historical state memory features as input, and outputs the decoded features corresponding to each prediction time-leading period in the future; the output smoothing module performs time consistency processing on the multi-time-leading decoded features or preliminary prediction sequences; the final prediction module converts the smoothed multi-time-leading features into daily grid-scale MDA8 O3 concentration prediction results for the next 1–14 days.

[0056] The optimal training settings determined experimentally in step 4 include: using the SmoothL1 Loss function and the Adam optimizer (initial learning rate 3×10). -4 The ReduceLROnPlateau learning rate scheduler is used, and the learning rate is reduced by a factor of 0.3 if the validation set loss does not decrease for three consecutive rounds. The batch size is set to 32 and the number of training rounds is set to 100. Dropout layers are embedded in the CNN, Transformer, and output modules to alleviate overfitting, and an Early Stopping strategy (patience=30) is used to further control the model complexity and stabilize the training process.

[0057] Optionally, the model uses the ReLU function as the activation function: ; In the formula: x represents the linear transformation output value of a certain layer in the neural network, that is, the feature value before the activation function is applied.

[0058] ; In the formula: u represents the output feature vector of the previous network layer, w represents the corresponding weight parameter, b represents the bias term, and T represents the transpose.

[0059] The loss function is trained using the mean squared error loss function (MSE): ; In the formula: These are the model's predicted values. For the true value, This represents the sample size.

[0060] The ozone concentration predicted in step 5 is evaluated using the coefficient of determination R², mean absolute error (MAE), and root mean square error (RMSE). The formula for calculating the coefficient of determination is: ; ; ; In the formula, N is the sample size. For the i-th observation, Let y be the i-th predicted value, and y be the average of all observed values.

[0061] The predicted frequency of O3 pollution events represents the proportion of days when both the predicted and observed O3 values ​​exceed 160 μg / m³ out of all days when the observed values ​​exceed this threshold.

[0062] This embodiment uses a 2D-CNN-Transformer model that integrates multi-source meteorological field and emission inventory data to accurately predict gridded MDA8 O3 concentrations in North China, thereby improving the stability of long-term forecasts. The prediction yields daily 14-day spatial distribution results of MDA8 O3 from January 1, 2024 to December 31, 2024. The coefficient of determination R0 for predicted MDA8 O3 on days 1, 7, and 14 is compared with observed values. 2 The coefficients of determination (R²) were 0.877, 0.886, and 0.881, respectively; corresponding to prediction frequencies of 83%, 74.5%, and 66% for O3 pollution events. Table 1 shows a comparison of the ozone concentration prediction results of this embodiment with other models. Table 1 shows the coefficients of determination (R²) for predicting ozone concentrations over 1 day and 7 days in this embodiment. 2The scores were 0.877 and 0.886 respectively, both outperforming existing comparative models such as GCN-LSTM, CNN-LSTM, and 3D-CNN, with higher prediction accuracy and a further widening performance advantage over the comparative models in long-term 7-day predictions.

[0063] Table 1. Comparison of ozone concentration prediction results from different models;

[0064] In addition, such as Figure 3 As shown, Figure 3 The R² results in (a) show that, under the three forecast lead times of T+1, T+7, and T+14, the determination coefficient of the model that integrates upper-air meteorological elements and emission inventory on the basis of surface meteorological elements is always higher than that of the model that uses only surface meteorological elements and the model that uses surface and emission inventory data. This indicates that the introduction of the three-dimensional meteorological field can effectively improve the fitting degree of MDA8 O3 concentration prediction. Figure 3 The hit rate of high ozone events (HPD) in (b) also showed the same trend. The hit rate of the model integrating upper-air meteorological elements and emission inventory in the three time periods (83%, 74.5%, and 66%) was significantly higher than that of the other two schemes, which verified that the three-dimensional regional meteorological field also improved the forecasting ability of high concentration ozone events. Figure 3 The time series comparison in (c) further shows that the model integrating upper-air meteorological elements and emission inventory data captures the changing trends and peak values ​​of observed values ​​(blue solid line) better than the model using only surface meteorological elements, especially during periods of high ozone concentration, thus more closely reflecting actual observations. In summary, the 2D-CNN-Transformer model integrating multi-source meteorological fields and emission inventory data proposed in this embodiment can significantly improve the overall accuracy of ozone prediction and the ability to identify high-concentration events.

[0065] This specific embodiment effectively integrates multi-source data, including regional emission inventories, 3D meteorological fields, station observations, and weather forecasts, to construct a complete spatiotemporal feature system. Spatial feature extraction is performed using a CNN model, transforming the source strength prior field into a spatial attention bias term. This enhances the model's ability to perceive the spatial distribution of precursor emissions and key impact areas, improving the accuracy of spatial dimension modeling. Simultaneously, a Transformer model is used to uncover the correlation patterns between different time steps and prediction lead times, accurately capturing the long-term evolution characteristics of ozone concentration over the next 1-14 days. Furthermore, loss weights are set based on the source strength prior field to optimize the training process and specifically constrain the model's learning direction. The overall method fully leverages the advantages of both types of network structures, taking into account spatial heterogeneity and temporal dynamic changes, effectively improving the accuracy and stability of medium- and long-term ozone concentration predictions, and providing reliable technical support for regional air pollution forecasting.

[0066] Example 2 This embodiment provides a long-term ozone concentration prediction system based on a three-dimensional regional meteorological field, including: The prior field construction module is configured to construct a strong prior field for ozone precursor sources based on regional emission inventory data. The three-dimensional meteorological feature construction module is configured to acquire a three-dimensional meteorological feature field covering an area larger than the ozone concentration forecast target area, including upstream and surrounding background meteorological information that affects the ozone evolution in the target area; The feature fusion module is configured to fuse three-dimensional meteorological feature fields, ozone precursor source strong prior fields, and station O3 observation data to obtain multi-source heterogeneous spatiotemporal features; The concentration prediction module is configured to input the multi-source heterogeneous spatiotemporal features and weather forecast data into a 2D-CNN-Transformer model to obtain ozone concentration prediction results; In the model training phase, spatial features are extracted from the multi-source heterogeneous spatiotemporal features based on a CNN network, and the source strength prior field is mapped as a spatial attention bias term to enhance the model's response to the spatial distribution of ozone precursor source strength and key affected areas. The correlation between different time steps and different prediction timelines is modeled based on a Transformer network to capture the temporal evolution characteristics of ozone concentration over the next 1–14 days. Loss weights are constructed based on the source strength prior field, and the model is trained by combining the loss function with the loss weights to obtain a trained 2D-CNN-Transformer ozone concentration prediction model.

[0067] Example 3 This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in the long-term ozone concentration prediction method based on a three-dimensional regional meteorological field as described in Embodiment 1 above.

[0068] Example 4 This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the long-term ozone concentration prediction method based on a three-dimensional regional meteorological field as described in Embodiment 1 above.

[0069] The steps or modules involved in Embodiments 2 to 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0070] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for long-term prediction of ozone concentration based on a three-dimensional regional meteorological field, characterized in that, include: Based on regional emission inventory data, a strong prior field for ozone precursor sources is constructed. Acquire a three-dimensional meteorological feature field covering an area larger than the ozone concentration forecast target area, including upstream and surrounding background meteorological information that affects the ozone evolution in the target area; By integrating three-dimensional meteorological feature field, strong prior field of ozone precursors and station O3 observation data, multi-source heterogeneous spatiotemporal characteristics are obtained; The multi-source heterogeneous spatiotemporal features and weather forecast data are input into a 2D-CNN-Transformer model to obtain ozone concentration prediction results. In the model training phase, spatial features are extracted from the multi-source heterogeneous spatiotemporal features based on a CNN network, and the source strength prior field is mapped as a spatial attention bias term to enhance the model's response to the spatial distribution of ozone precursor source strength and key affected areas. The correlation between different time steps and different prediction timelines is modeled based on a Transformer network to capture the temporal evolution characteristics of ozone concentration over the next 1–14 days. Loss weights are constructed based on the source strength prior field, and the model is trained by combining the loss function with the loss weights to obtain a trained 2D-CNN-Transformer ozone concentration prediction model.

2. The method for long-term ozone concentration prediction based on a three-dimensional regional meteorological field as described in claim 1, characterized in that, The construction of a strong prior field for ozone precursor sources based on regional emission inventory data specifically includes: ； in, Represents the spatial grid under the h-th prediction lead time. The prior value of the ozone precursor source strength at the location, where K represents the set of ozone precursor components; This represents the emission field of the k-th type of precursor after mass conservation regrinding, time spectrum allocation, and normalization. represents the contribution weights of different precursors to the prior source strength; b represents the bias term. This represents the Sigmoid function, used to compress the prior field to the range of 0-1.

3. The long-term ozone concentration prediction method based on a three-dimensional regional meteorological field as described in claim 1, characterized in that, The specific process for obtaining the three-dimensional meteorological feature field is as follows: The target area for ozone concentration forecasting is determined, and based on the target area and its upstream and surrounding potential transport influence areas, the coverage area of ​​the meteorological field is determined, wherein the coverage area of ​​the meteorological field is larger than the target area; Within the coverage area of ​​the meteorological field, an original three-dimensional meteorological feature field is constructed based on longitude, latitude, and vertical altitude; so that it simultaneously contains local meteorological information of the target area and large-scale background meteorological information that affects the ozone evolution of the target area. Using the start time of the weather forecast as the sample anchor point, the original three-dimensional meteorological feature field is processed by horizontal spatial grid unification, vertical layer matching and prediction time alignment to obtain a three-dimensional meteorological feature field corresponding to the target prediction time.

4. The long-term ozone concentration prediction method based on a three-dimensional regional meteorological field as described in claim 1, characterized in that, The fusion of three-dimensional meteorological feature field, ozone precursor source strong prior field, and station O3 observation data yields multi-source heterogeneous spatiotemporal characteristics, specifically including: Based on the calculation of the mechanistic constraints related to ozone generation and transport using a three-dimensional meteorological characteristic field; The emission inventory data is processed to generate an emission characteristic field; Quality control and daily maximum 8-hour moving average calculations were performed on the O3 observation data at the stations to obtain station-scale MDA8 O3 labels; a label confidence mask was generated based on the station-scale MDA8 O3 labels. By combining the three-dimensional meteorological feature field, the strong prior field of ozone precursor sources, the mechanism constraint features, the emission feature field, and the confidence mask according to the spatial grid, vertical layer, prediction time and variable channel, multi-source heterogeneous spatiotemporal features are obtained.

5. The method for long-term prediction of ozone concentration based on a three-dimensional regional meteorological field as described in claim 1, characterized in that, The step of inputting the multi-source heterogeneous spatiotemporal features and weather forecast data into the CNN-Transformer model to obtain ozone concentration prediction results specifically includes: By inputting historical multi-source heterogeneous spatiotemporal features into a CNN, and extracting local spatial features at each time step, spatial embedding features reflecting the historical pollution background and emission source intensity distribution of the target area are obtained. The system acquires future weather forecast data and corresponding time-lead time codes, inputs the weather forecast features and time-lead time codes into the Transformer, learns the dependencies between historical spatiotemporal windows, future weather forecast windows, and different forecast time leads through a self-attention mechanism, and guides feature learning with a spatial attention bias term generated by a strong prior field of gas precursor sources to obtain the ozone concentration prediction results at the daily grid scale.

6. The method for long-term ozone concentration prediction based on a three-dimensional regional meteorological field as described in claim 1, characterized in that, The mapping of the source strength prior field to a spatial attention bias term specifically includes: ； in, For spatial attention bias, Used to control the moderating power of source strength priors on attention allocation. This indicates the target spatial location where the ozone concentration to be predicted is located. This indicates the spatial location or potential source region of influence involved in attention calculation. Indicates the target location With source strength location Spatial association weights between them are constructed based on source strength priors; Indicates a query. Indicates key-value pair, Indicates the feature dimension.

7. The method for long-term prediction of ozone concentration based on a three-dimensional regional meteorological field as described in claim 1, characterized in that, The construction of loss weights based on the source strength prior field specifically includes: ； in, Represents the spatial grid under the h-th prediction lead time. Loss weight at the point, and The spatial grids are respectively for the h-th prediction time. , The ozone precursor source has a strong a priori field. and These represent the rows and columns of the spatial grid, respectively. This is the weighting adjustment coefficient.

8. A long-term ozone concentration prediction system based on a three-dimensional regional meteorological field, characterized in that, include: The prior field construction module is configured to construct a strong prior field for ozone precursor sources based on regional emission inventory data. The three-dimensional meteorological feature construction module is configured to acquire a three-dimensional meteorological feature field covering an area larger than the ozone concentration forecast target area, including upstream and surrounding background meteorological information that affects the ozone evolution in the target area; The feature fusion module is configured to fuse three-dimensional meteorological feature fields, ozone precursor source strong prior fields, and station O3 observation data to obtain multi-source heterogeneous spatiotemporal features; The concentration prediction module is configured to input the multi-source heterogeneous spatiotemporal features and weather forecast data into a 2D-CNN-Transformer model to obtain ozone concentration prediction results; In the model training phase, spatial features are extracted from the multi-source heterogeneous spatiotemporal features based on a CNN network, and the source strength prior field is mapped as a spatial attention bias term to enhance the model's response to the spatial distribution of ozone precursor source strength and key affected areas. The correlation between different time steps and different prediction timelines is modeled based on a Transformer network to capture the temporal evolution characteristics of ozone concentration over the next 1–14 days. Loss weights are constructed based on the source strength prior field, and the model is trained by combining the loss function with the loss weights to obtain a trained 2D-CNN-Transformer ozone concentration prediction model.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the long-term ozone concentration prediction method based on a three-dimensional regional meteorological field as described in any one of claims 1-7.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the long-term ozone concentration prediction method based on a three-dimensional regional meteorological field as described in any one of claims 1-7.

Images