Pollutant assimilation method, device and equipment fusing three-dimensional variation and deep learning

Abstract

The invention relates to a pollutant assimilation method, a pollutant assimilation device and pollutant assimilation equipment fusing three-dimensional variation and deep learning. The method comprises the following steps: collecting pollutant observation data and a background field in a research area; the pollutant observation data and the background field are input into a three-dimensional variational assimilation system, and a pollutant hour-by-hour three-dimensional variational assimilation analysis field is generated; constructing a fusion HATNet assimilation model, and inputting the pollutant observation data, the background field and the three-dimensional variational assimilation analysis field into the HATNet assimilation model for training until a trained HATNet assimilation model is obtained; to-be-predicted pollutant observation data and a background field are input into the trained HATNet assimilation model, an hour-by-hour pollutant concentration analysis field is predicted and generated, and the pollutant concentration analysis field is applied to numerical forecasting. By adopting the method, the pollutant concentration can be efficiently and accurately forecasted.

Description

Technical Field

[0001] The present application relates to the technical field of atmospheric pollutant data assimilation and artificial intelligence fusion, and in particular to a pollutant assimilation method, device and equipment integrating three-dimensional variation and deep learning. Background Art

[0002] Atmospheric chemistry models are crucial for predicting and simulating air quality, estimating the three-dimensional distribution of gas and aerosol concentrations. However, accurate atmospheric chemistry prediction remains a challenge due to uncertainties in pollutant emissions, initial conditions (IC), boundary conditions (BC), and parameterization of chemical processes. Among these factors, IC provides key atmospheric and chemical variables and plays an important role in determining the accuracy of atmospheric chemistry forecasts. Data assimilation (DA) methods, such as variational assimilation and ensemble Kalman filter (EnKF) methods, are widely regarded as effective techniques to reduce IC uncertainties and improve the accuracy of atmospheric chemistry forecasts.

[0003] Among them, the three-dimensional variational assimilation (3Dvar) method is one of the most widely used variational assimilation methods, which provides detailed and accurate initial states and improves the ability to predict air pollution. The EnKF method achieves assimilation by generating a set of multi-model predictions and providing uncertainty quantification, which is suitable for solving nonlinear systems. However, both the 3Dvar and EnKF assimilation methods require a lot of computing resources and have slow execution times. In recent years, with the increase in the dimension of state variables and the number of samples, the computational complexity of these methods has also increased, resulting in longer calculation times and lower accuracy for pollutant concentration forecasts and air quality predictions. Summary of the invention

[0004] Based on this, it is necessary to provide a pollutant assimilation method, device and equipment that integrates three-dimensional variational and deep learning to address the above technical problems.

[0005] A pollutant assimilation method integrating three-dimensional variation and deep learning, the method comprising: Step 1, collect pollutant observation data and background fields in the study area; Step 2: Input the pollutant observation data and background field into the three-dimensional variational assimilation system to generate the hourly three-dimensional variational assimilation analysis field of pollutants; Step 3: Construct a HATNet (Hadamar Attention Transformer Network) assimilation model with a Unet structure that integrates the Hadamard attention module and the Transformer module, and input the pollutant observation data, background field, and three-dimensional variational assimilation analysis field into the HATNet assimilation model for training until a trained HATNet assimilation model is obtained; the Hadamard attention module extracts feature information of the model input data by weighting and highlighting key features, and the Transformer module further integrates and transforms features by capturing long-distance dependencies; Step 4: Input the pollutant observation data and background field to be predicted into the trained HATNet assimilation model, predict and generate hourly pollutant concentration analysis fields, and apply the pollutant concentration analysis fields to numerical forecasting.

[0006] In one embodiment, collecting pollutant observation data and background fields in a study area includes: Collect pollutant observation data from ground stations in the study area and conduct quality control; quality control includes: extreme value test and consistency test on pollutant observation data, filling missing values, eliminating outliers, and gridding pollutant observation data; The background field of hourly three-dimensional pollutant concentrations provided by the WRF-Chem (Weather Research and Forecasting Chemistry) numerical model running results is collected.

[0007] In one embodiment, pollutant observation data and background fields are input into a three-dimensional variational assimilation system to generate hourly three-dimensional variational assimilation analysis fields of pollutants, including: A three-dimensional variational assimilation system is constructed based on the three-dimensional variational assimilation theory. The pollutant observation data and background field are input into the three-dimensional variational assimilation system. The pollutant observation data are used as constraints and the background field is used as a priori field. A three-dimensional variational assimilation loss function is constructed and the target functional is solved to generate a three-dimensional variational assimilation analysis field of pollutants hourly.

[0008] In one embodiment, the HATNet assimilation model consists of a symmetrical encoder and decoder, and the encoder and decoder are connected through a Transformer module; The encoder includes a convolution block and multiple Hadamard attention modules. The convolution block is used to extract the input pollutant observation data and the preliminary features of the pollutant concentration in the background field through convolution operations. The Hadamard attention module is used to perform weighted processing on the preliminary features of the pollutant concentration output by the convolution block at different levels. The Transformer module is used to further integrate and transform the encoded features finally output by the encoder, and output the processed features to the decoder; The decoder also contains a convolution block and multiple Hadamard attention modules, which are used to gradually decode and restore the features output by the Transformer module to generate a pollutant concentration analysis field. The number of Hadamard attention modules in the encoder and decoder is consistent and one-to-one corresponding, and the encoder and decoder use the DSB (Double-Scale Bridge) module to fuse the output of each layer of the encoder with the masks generated by each layer of the decoder to achieve multi-scale feature extraction.

[0009] In one embodiment, pollutant observation data, background field, and three-dimensional variational assimilation analysis field are input into a HATNet assimilation model for training until a trained HATNet assimilation model is obtained, including: After the pollutant observation data are interpolated into a grid form that is consistent with the study area and resolution of the background field, they are input into the HATNet assimilation model together with the background field and the three-dimensional variational assimilation analysis field. The three-dimensional variational assimilation analysis field is used as a label, and the HATNet assimilation model is trained using supervised learning until a trained HATNet assimilation model is obtained.

[0010] In one embodiment, the loss function of the HATNet assimilation model training is defined as ; in, represents the three-dimensional variational assimilation analysis field, represents the prediction result of the HATNet assimilation model, and MSE represents the mean square error.

[0011] A pollutant assimilation device integrating three-dimensional variation and deep learning, the device comprising: Data acquisition module, used to collect pollutant observation data and background fields in the study area; The variational assimilation module is used to input pollutant observation data and background fields into the three-dimensional variational assimilation system to generate hourly three-dimensional variational assimilation analysis fields of pollutants; A deep learning module is used to construct a HATNet assimilation model with a Unet structure that integrates a Hadamard attention module and a Transformer module, and input pollutant observation data, background fields, and three-dimensional variational assimilation analysis fields into the HATNet assimilation model for training until a trained HATNet assimilation model is obtained; the Hadamard attention module extracts feature information of the model input data by weighting and highlighting key features, and the Transformer module further integrates and transforms features by capturing long-distance dependencies; The prediction output module is used to input the pollutant observation data and background field to be predicted into the trained HATNet assimilation model, predict and generate hourly pollutant concentration analysis fields, and apply the pollutant concentration analysis fields to numerical forecasting.

[0012] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented: Step 1, collect pollutant observation data and background fields in the study area; Step 2: Input the pollutant observation data and background field into the three-dimensional variational assimilation system to generate the hourly three-dimensional variational assimilation analysis field of pollutants; Step 3: construct a HATNet assimilation model with a Unet structure that integrates a Hadamard attention module and a Transformer module, and input pollutant observation data, background fields, and three-dimensional variational assimilation analysis fields into the HATNet assimilation model for training until a trained HATNet assimilation model is obtained; wherein the Hadamard attention module extracts feature information of the model input data by weighting and highlighting key features, and the Transformer module further integrates and transforms features by capturing long-distance dependencies; Step 4: Input the pollutant observation data and background field to be predicted into the trained HATNet assimilation model, predict and generate hourly pollutant concentration analysis fields, and apply the pollutant concentration analysis fields to numerical forecasting.

[0013] Compared with the prior art, the pollutant assimilation method, device and equipment integrating three-dimensional variation and deep learning have the following beneficial effects: 1. By taking the three-dimensional variational assimilation analysis field output by the three-dimensional variational assimilation system as a constraint for deep learning model training, the fusion of three-dimensional variational assimilation and deep learning is realized, thereby effectively utilizing the fusion and calibration capabilities of three-dimensional variational assimilation for pollutant observation data and background fields, providing a more accurate and physically consistent data basis for the deep learning model, and improving the accuracy and reliability of the model output results.

[0014] 2. Construct a HATNet assimilation model based on the Unet structure that integrates the Hadamard attention module and the Transformer module. The Hadamard attention module and the Transformer module can enhance the learning ability of the spatial distribution characteristics of pollutant concentrations, and the symmetrical encoder and decoder in the Unet structure can capture spatial information more effectively, thereby improving the accuracy of pollutant concentration forecasting.

[0015] 3. The trained HATNet assimilation model is an end-to-end assimilation model that can directly assimilate the pollutant concentration analysis field based on the input pollutant observation data and background field data, avoiding the design of large matrices such as background errors and covariances and complex solution calculations. There is no iterative solution process in three-dimensional variational assimilation, and the demand for computing resources is small, which can greatly improve the speed of three-dimensional data assimilation. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 A schematic diagram of a process flow of a pollutant assimilation method integrating three-dimensional variational analysis and deep learning in one embodiment; Figure 2 A schematic diagram of the logical architecture of a pollutant assimilation method integrating three-dimensional variational analysis and deep learning in one embodiment; Figure 3 A schematic diagram of the structure of a HATNet assimilation model in one embodiment; Figure 4 This is the concentration spatial distribution map of a pollutant at 18:00 on November 19, 2023, taking SO2 as an example; Figure 4 (a) is the concentration distribution of SO2 analysis field after three-dimensional variational assimilation. Figure 4 (b) is the concentration distribution of SO2 analysis field after assimilation by HATNet assimilation model; Figure 5 Schematic diagram of the vertical incremental distribution of SO2 concentration at 18:00 on November 19, 2023 in an embodiment, where: Figure 5 (a) is the vertical increment distribution of SO2 concentration after three-dimensional variational assimilation. Figure 5 (b) is the vertical increment distribution of SO2 concentration after assimilation by the HATNet assimilation model; Figure 6 It is a scatter plot of the analysis field and the ground station observation concentration obtained by assimilating the SO2 concentration using three-dimensional variational assimilation and the method proposed in this application for the SO2 data in the study area at 18:00 on November 19, 2023 in an embodiment; Figure 7 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0018] In one embodiment, Figure 1 and Figure 2 As shown, a pollutant assimilation method integrating three-dimensional variation and deep learning is provided, including the following steps: Step 1: Collect pollutant observation data and background fields in the study area.

[0019] Specifically, step 1 includes: collecting pollutant observation data from ground stations in the study area and performing quality control; wherein, quality control includes: performing extreme value test and consistency test on pollutant observation data, filling missing data values ​​with the average value before and after the collection time, eliminating data with too large or too small pollutant concentration, and gridding pollutant observation data. Collect the background field of hourly three-dimensional pollutant concentration provided by the WRF-Chem numerical model operation results.

[0020] Step 2: Input the pollutant observation data and background field into the three-dimensional variational assimilation system to generate a three-dimensional variational assimilation analysis field of pollutants on an hourly basis.

[0021] Specifically, step 2 includes: constructing a three-dimensional variational assimilation system based on the three-dimensional variational assimilation theory, inputting the pollutant observation data and the background field into the three-dimensional variational assimilation system, wherein the pollutant observation data is used as a constraint and the background field is used as a priori field, constructing a three-dimensional variational assimilation loss function and solving the target functional, and generating a three-dimensional variational assimilation analysis field of pollutants hour by hour.

[0022] Step 3: Construct a HATNet assimilation model with a Unet structure that integrates the Hadamard attention module and the Transformer module, and input the pollutant observation data, background field, and three-dimensional variational assimilation analysis field into the HATNet assimilation model for training until a trained HATNet assimilation model is obtained; wherein, the Hadamard attention module extracts the feature information of the model input data by weighting and highlighting key features, and the Transformer module further integrates and transforms the features by capturing long-distance dependencies.

[0023] The structure of the HATNet assimilation model is as follows Figure 3 As shown in Figure 1, the HATNet assimilation model consists of a symmetrical encoder and decoder, and the encoder and decoder are connected through a Transformer module.

[0024] Among them, the encoder contains a convolution block and multiple Hadamard attention modules. The convolution block is used to extract the input pollutant observation data and the preliminary characteristics of the pollutant concentration in the background field through convolution operations. The Hadamard attention module is used to weightedly process the preliminary characteristics of the pollutant concentration output by the convolution block at different levels.

[0025] The Transformer module is used to further integrate and transform the encoded features finally output by the encoder, and output the processed features to the decoder. This module can further enhance the ability of the assimilation model to capture and express important feature information.

[0026] The decoder also contains a convolution block and multiple Hadamard attention modules, which are used to gradually decode and restore the features output by the Transformer module to generate a pollutant concentration analysis field; among them, the number of Hadamard attention modules in the encoder and decoder is consistent and one-to-one corresponding, and the encoder and decoder use the DSB module to fuse the output of each layer of the encoder with the mask generated by each layer of the decoder to achieve multi-scale feature extraction.

[0027] Specifically, the HATNet assimilation model training in step 3 includes: interpolating the pollutant observation data into a grid form consistent with the research area and resolution of the background field, inputting it into the HATNet assimilation model together with the background field and the three-dimensional variational assimilation analysis field, and using the three-dimensional variational assimilation analysis field as a label to train the HATNet assimilation model using supervised learning until a trained HATNet assimilation model is obtained. The loss function during HATNet assimilation model training is defined as ; in, represents the three-dimensional variational assimilation analysis field, represents the prediction result of the HATNet assimilation model, and MSE represents the mean square error.

[0028] Step 4: Input the pollutant observation data and background field to be predicted into the trained HATNet assimilation model, predict and generate hourly pollutant concentration analysis fields, and apply the pollutant concentration analysis fields to numerical forecasting.

[0029] Through the above four steps, this method can integrate three-dimensional variational assimilation and deep learning, and use pollutant observation data, background field and three-dimensional variational assimilation analysis field to train a HATNet assimilation model. The HATNet assimilation model can directly assimilate the pollutant concentration analysis field based on the input pollutant observation data and background field data, avoiding the complex calculation process of three-dimensional variational assimilation, improving the efficiency of pollutant assimilation calculation, and based on the Hadamard attention module and Transformer module in the model, it can more effectively capture the spatial distribution characteristics of pollutant concentration and improve the accuracy of pollutant concentration prediction. And experiments have confirmed that the HATNet assimilation model constructed in this application can assimilate the output pollutant concentration analysis field at a speed 34 times that of the three-dimensional variational assimilation system.

[0030] In a specific embodiment, taking the optimization of SO2 concentration at 18:00 on November 19, 2023 as an example, the longitude range is from 76.72° east longitude to 151.28° east longitude, the latitude range is from 11.78° north latitude to 57.16° north latitude, the resolution is 12km, and the study area includes most of China. Specifically, the following steps are included: S1: Collect SO2 observation data from 1641 ground stations at 18:00 on November 19, 2023, perform quality control on the observation data, fill in missing values ​​and remove outliers. Collect the forecast field of WRF-Chem model at 18:00 on November 19, 2023 as the SO2 concentration background field.

[0031] S2: Based on the 3Dvar theory, a 3Dvar assimilation system is constructed. The SO2 observation data and the SO2 concentration background field are input into the three-dimensional variational assimilation system. The SO2 observation data are used as constraints and the SO2 concentration background field is used as a priori field. After constructing the three-dimensional variational assimilation loss function and solving the target functional, the three-dimensional variational assimilation field of SO2 concentration hour by hour is obtained. S3: Construct the HATNet assimilation model, interpolate the SO2 observation data into a grid form consistent with the corresponding background field research area and resolution, then use the SO2 observation data, the three-dimensional SO2 concentration background field and the three-dimensional variational assimilation field output in step S2 as training data, and use supervised learning to train the HATNet assimilation model. The loss function in the HATNet assimilation model is defined as follows: ; After the training is completed, the trained HATNet assimilation model is obtained; S4: The SO2 observation data and background field to be predicted are input into the trained HATNet assimilation model to generate hourly SO2 concentration analysis fields, and the optimization results of the SO2 concentration analysis fields can be used for subsequent numerical forecasts.

[0032] Furthermore, the SO2 data in the study area at 18:00 on November 19, 2023 were assimilated using the method proposed in this application, and the three-dimensional variational assimilation results were compared and verified. The comparison results are as follows: Figure 4-Figure 6 As shown. Figure 4-Figure 6 It can be seen that compared with three-dimensional variational assimilation, the method proposed in this application has better assimilation analysis results for SO2 concentration. Based on this method, the accuracy of the initial field of pollutant concentration forecast can be effectively improved, which has important scientific significance and application value for improving the accuracy of pollutant forecast. It is an effective supplement to traditional meteorological data assimilation, avoiding large matrices such as background errors and covariances and complex solution calculations, and also avoiding additional errors introduced by matrix design. It can directly perform end-to-end data assimilation, which is simple and easy, with small calculation amount and low financial investment, and is of great significance to meteorological data assimilation and pollutant forecasting.

[0033] In one embodiment, a pollutant assimilation device integrating three-dimensional variation and deep learning is provided, comprising: Data acquisition module, used to collect pollutant observation data and background fields in the study area; The variational assimilation module is used to input pollutant observation data and background fields into the three-dimensional variational assimilation system to generate hourly three-dimensional variational assimilation analysis fields of pollutants; A deep learning module is used to construct a HATNet assimilation model with a Unet structure that integrates a Hadamard attention module and a Transformer module, and input pollutant observation data, background fields, and three-dimensional variational assimilation analysis fields into the HATNet assimilation model for training until a trained HATNet assimilation model is obtained; the Hadamard attention module extracts feature information of the model input data by weighting and highlighting key features, and the Transformer module further integrates and transforms features by capturing long-distance dependencies; The prediction output module is used to input the pollutant observation data and background field to be predicted into the trained HATNet assimilation model, predict and generate hourly pollutant concentration analysis fields, and apply the pollutant concentration analysis fields to numerical forecasting.

[0034] For the specific limitations on the pollutant assimilation device that integrates three-dimensional variational and deep learning, please refer to the limitations on the pollutant assimilation method that integrates three-dimensional variational and deep learning in the above text, which will not be repeated here. Each module in the above-mentioned pollutant assimilation device that integrates three-dimensional variational and deep learning can be implemented in whole or in part through software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0035] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 7 As shown. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a pollutant assimilation method that integrates three-dimensional variation and deep learning is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covering the display screen, or a button, trackball or touchpad set on the computer device housing, or an external keyboard, touchpad or mouse, etc.

[0036] Those skilled in the art will understand that Figure 7The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0037] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented: Step 1, collect pollutant observation data and background fields in the study area; Step 2: Input the pollutant observation data and background field into the three-dimensional variational assimilation system to generate the hourly three-dimensional variational assimilation analysis field of pollutants; Step 3: construct a HATNet assimilation model with a Unet structure that integrates a Hadamard attention module and a Transformer module, and input pollutant observation data, background fields, and three-dimensional variational assimilation analysis fields into the HATNet assimilation model for training until a trained HATNet assimilation model is obtained; wherein the Hadamard attention module extracts feature information of the model input data by weighting and highlighting key features, and the Transformer module further integrates and transforms features by capturing long-distance dependencies; Step 4: Input the pollutant observation data and background field to be predicted into the trained HATNet assimilation model, predict and generate hourly pollutant concentration analysis fields, and apply the pollutant concentration analysis fields to numerical forecasting.

[0038] The technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0039] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A pollutant assimilation method integrating three-dimensional variational and deep learning, characterized in that: The method comprises: Step 1: Collect pollutant observation data and background fields in the study area; Step 2, inputting the pollutant observation data and background field into a three-dimensional variational assimilation system to generate a three-dimensional variational assimilation analysis field of pollutants on an hourly basis; Step 3, constructing a HATNet assimilation model of a Unet structure integrating a Hadamard attention module and a Transformer module, and inputting the pollutant observation data, background field, and three-dimensional variational assimilation analysis field into the HATNet assimilation model for training until a trained HATNet assimilation model is obtained; wherein the Hadamard attention module extracts feature information of the model input data by weighting and highlighting key features, and the Transformer module further integrates and transforms features by capturing long-distance dependencies; Step 4: Input the pollutant observation data and background field to be predicted into the trained HATNet assimilation model, predict and generate hourly pollutant concentration analysis fields, and apply the pollutant concentration analysis fields to numerical forecasting.

2. The method according to claim 1, characterized in that Collect pollutant observation data and background fields in the study area, including: Collect pollutant observation data from ground stations in the study area and perform quality control; wherein the quality control includes: performing extreme value test and consistency test on pollutant observation data, filling missing values, removing outliers, and gridding pollutant observation data; The WRF-Chem numerical model running results provide a background field of hourly three-dimensional pollutant concentrations.

3. The method according to claim 2, characterized in that The pollutant observation data and background field are input into the three-dimensional variational assimilation system to generate a three-dimensional variational assimilation analysis field of pollutants every hour, including: A three-dimensional variational assimilation system is constructed based on the three-dimensional variational assimilation theory, and the pollutant observation data and the background field are input into the three-dimensional variational assimilation system, wherein the pollutant observation data are used as constraints and the background field is used as a priori field. A three-dimensional variational assimilation loss function is constructed and the target functional is solved to generate a three-dimensional variational assimilation analysis field of pollutants hour by hour.

4. The method according to claim 1, characterized in that: The HATNet assimilation model consists of a symmetrical encoder and decoder, and the encoder and decoder are connected through a Transformer module; The encoder includes a convolution block and multiple Hadamard attention modules, wherein the convolution block is used to extract the input pollutant observation data and the preliminary features of the pollutant concentration in the background field through convolution operation, and the Hadamard attention module is used to perform weighted processing on the preliminary features of the pollutant concentration output by the convolution block at different levels; The Transformer module is used to further integrate and transform the encoding features finally output by the encoder, and output the processed features to the decoder; The decoder also includes a convolution block and multiple Hadamard attention modules, which are used to gradually decode and restore the features output by the Transformer module to generate a pollutant concentration analysis field; wherein, the number of Hadamard attention modules in the encoder and decoder is consistent and one-to-one corresponding, and the output of each layer of the encoder is fused with the mask generated by each layer of the decoder through the DSB module between the encoder and the decoder to achieve multi-scale feature extraction.

5. The method according to claim 4, characterized in that The pollutant observation data, background field and three-dimensional variational assimilation analysis field are input into the HATNet assimilation model for training until a trained HATNet assimilation model is obtained, including: After the pollutant observation data is interpolated into a grid form that is consistent with the study area and resolution of the background field, it is input into the HATNet assimilation model together with the background field and the three-dimensional variational assimilation analysis field. The three-dimensional variational assimilation analysis field is used as a label, and the HATNet assimilation model is trained using supervised learning until a trained HATNet assimilation model is obtained.

6. The method according to claim 5, characterized in that The loss function of the HATNet assimilation model training is defined as ； in, represents the three-dimensional variational assimilation analysis field, represents the prediction result of the HATNet assimilation model, and MSE represents the mean square error.

7. A pollutant assimilation device integrating three-dimensional variation and deep learning, characterized in that: The device comprises: Data acquisition module, used to collect pollutant observation data and background fields in the study area; A variational assimilation module, used for inputting the pollutant observation data and background field into a three-dimensional variational assimilation system to generate a three-dimensional variational assimilation analysis field of pollutants on an hourly basis; A deep learning module, used to construct a HATNet assimilation model of a Unet structure integrating a Hadamard attention module and a Transformer module, and input the pollutant observation data, background field, and three-dimensional variational assimilation analysis field into the HATNet assimilation model for training until a trained HATNet assimilation model is obtained; wherein the Hadamard attention module extracts feature information of the model input data by weighting and highlighting key features, and the Transformer module further integrates and transforms features by capturing long-distance dependencies; The prediction output module is used to input the pollutant observation data and background field to be predicted into the trained HATNet assimilation model, predict and generate hourly pollutant concentration analysis fields, and apply the pollutant concentration analysis fields to numerical forecasting.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

Images