A method for estimating the regional distribution of atmospheric refractivity at sea based on vertical observations at a single station

Abstract

The present invention belongs to the technical field of marine environmental monitoring, and relates to a method for deducing the regional distribution of atmospheric refractivity at sea based on single-station vertical observation. The method comprises: step one, using the improved COAWST-LES model and graph neural network algorithm to deduce the sparse observation data fusion in the target area; step two, based on the single-station atmospheric vertical observation data and the fused data generated in step one, reconstructing the spatiotemporal characteristics of the data and constructing a four-dimensional feature field; step three, using a four-dimensional variational assimilation strategy, performing numerical model and AI coupling iterative solution to achieve regional deduction estimation; using the deduced estimated temperature, humidity, and air pressure to calculate the atmospheric refractive index. The present invention integrates empirical correction, adds turbulent processes to the traditional mesoscale numerical model, and can accurately estimate the data in the unobserved area through the production of a numerical model dynamic downscaling training set and artificial intelligence regional error correction.

Description

Technical Field

[0001] The invention belongs to the technical field of marine environment monitoring and relates to a method for deducing regional distribution of marine atmospheric refractive index based on single-station vertical observation. Background Art

[0002] The atmospheric refractive index (ARI) is a parameter that characterizes the bending and deflection effects of the atmosphere on the propagation path of electromagnetic waves. The calculation of ARI is related to fundamental atmospheric elements and is typically derived using empirical models.

[0003] The monitoring of atmospheric refractivity usually uses sounding balloons, lidar and other equipment to carry out single-station atmospheric vertical observation and diagnosis to obtain the vertical profile of atmospheric refractivity at altitudes above the ground / sea surface.

[0004] The vertical profile of atmospheric refractivity is a core input parameter for electromagnetic wave propagation simulation models, primarily used to predict and simulate radar range. These electromagnetic wave simulation models require the three-dimensional distribution of the atmospheric refractivity within the radar's detection range. However, current vertical observations from a single station can only capture localized profile information, lacking effective regional observation methods. Therefore, in actual simulation applications, it is often assumed that the atmospheric refractivity within the target range is uniform and unchanging. This single-station observation information is replicated to other locations to generate regional three-dimensional environmental data to drive the electromagnetic wave simulation model.

[0005] This uniform assumption does not take into account the spatial differences within the region, which leads to large deviations in the simulation results of the electromagnetic wave propagation path.

[0006] In the past, there were three main technical paths to the problem of extrapolating vertical observations from a single offshore station to regional three-dimensional spatial distribution, but none of them has been able to effectively solve the problem so far. The first path is the mesoscale numerical model forecasting method, which is based on a limited number of observation points and is introduced into the numerical model through data assimilation to simulate three-dimensional spatial changes in the region. The disadvantage of this path is that the model scale is large, and its physical process scheme does not include small-scale atmospheric processes. The simulation results are too smooth and cannot simulate the fluctuations and mutations in real observation data due to small-scale processes such as turbulence, resulting in low practical application value of the simulation results. The second path is small-scale geofluid mechanics simulation, which can take into account processes such as turbulence, but its disadvantage is that the amount of calculation is huge, and it can only simulate fluid motion such as wind and flow, and cannot be used for atmospheric refractive index calculation for the time being. The third path is artificial intelligence algorithms such as machine learning or deep learning, which establish a mapping relationship between the profiles of other locations and observation points to achieve extrapolation from a single point to a region. However, this path lacks long-term data in the target area for model training. The meteorological elements of remote sensing and reanalysis data commonly used in past artificial intelligence research cannot be applied to training in this problem because their resolution is too coarse. Only single-station estimation can be performed from a single station to multiple limited locations. Therefore, it is temporarily impossible to use limited station observations to accurately estimate the regional distribution of marine atmospheric refractivity. Summary of the Invention

[0007] In response to the current problem, the present invention discloses a regional atmospheric refractivity deduction method and data production process for offshore data-free areas based on single-station atmospheric vertical observation. Taking the profile data of single-station atmospheric vertical observation as the evaluation benchmark, a new approach to estimating the regional atmospheric refractivity distribution is designed by combining numerical model modification, numerical forecasting and intelligent correction of regional errors, solving the problem that the current traditional technical path is difficult to accurately estimate the regional atmospheric refractivity distribution based on limited observation data.

[0008] The technical solution provided by the present invention is: a method for deducing the regional distribution of atmospheric refractivity at sea based on single-station vertical observation, comprising the following steps:

[0009] Step 1: Use the improved COAWST-LES model and graph neural network algorithm to deduce the sparse observation data fusion in the target area; the improved COAWST-LES model replaces the original atmospheric module WRF in the ocean-air coupled model COAWST with the WRF-LES model that considers the large eddy simulation process;

[0010] Step 2: Based on the single-station atmospheric vertical observation data and the fused data generated in step 1, reconstruct the spatiotemporal characteristics of the data and construct a four-dimensional feature field; the vertical observation data includes temperature, humidity, and air pressure;

[0011] Step 3: Use the four-dimensional variational assimilation strategy to perform iterative solution by coupling the numerical model with AI to achieve regional deduction estimation; use the deduced estimated temperature, humidity, and air pressure to calculate the atmospheric refractive index.

[0012] Preferably, step two specifically includes: building a fusion model with the graph neural network GNN algorithm as the main framework; the nodes are mainly pattern grid points, supplemented by the positions of each observation point; calculating the dynamic weights of the vertical observation location of a single station and other grid points in the area through the Transformer algorithm, and using the Coriolis force parameter as the attention bias item; setting weak supervision rules for multi-source data, and only imposing spatial correlation statistical feature constraints on the observation data; using the improved COAWST-LES model to carry out long-term large eddy simulation for the target area to generate a high-resolution training data set; splitting the training data set into different scenarios for different weather scenarios for separate training.

[0013] Preferably, step three specifically includes: parsing the single-station atmospheric vertical observation data; performing high-frequency time series analysis on the instantaneous observation data through wavelet transform, extracting the fluctuation characteristics of minute-level changes, and capturing turbulence signals; calculating the kinetic energy spectrum slope, quantifying the proportion of small-scale energy, and obtaining the magnitude of turbulent kinetic energy in the real data through energy spectrum diagnosis, which is used for subsequent correction of the model output result;

[0014] Based on the generative adversarial network (GAN) algorithm, a three-dimensional disturbance field centered on the vertical atmospheric observation position of a single station is generated; the generated fused data is used as the input of the GAN generator; a large eddy simulation is carried out on the target area using the improved COAWST-LES model to obtain the realistic turbulence distribution; based on the classic linear wave theory of boundary layer meteorology, a gravity wave parameterization scheme is constructed to quickly estimate the propagation direction of turbulent disturbances.

[0015] Preferably, step three specifically includes: generating a background field for the current period through the improved COAWST-LES model with a time resolution of once every 30 minutes; using the GAN algorithm, combining the single-station atmospheric vertical observations and the model background field for the current period, generating a small-scale disturbance field for changes in temperature, humidity, and pressure fields; using the disturbed background field as a driver, carrying out short-term forecasts of the coupled numerical model, and updating the GAN input with the new forecast field to form an iterative optimization, thereby realizing the deduction and estimation from single-point atmospheric vertical observations to the target area.

[0016] Preferably, mass conservation constraints are embedded in the GAN network loss function to add physical constraints to the GAN modeling process.

[0017] Preferably, at the single-station vertical observation position of the output layer, the single-point perturbation amount is forced to be 0.

[0018] Preferably, the perturbation intensity generated by GAN is spectrally normalized.

[0019] The present invention discloses a method for deducing the regional distribution of atmospheric refractivity at sea based on single-station vertical observations. Compared with the existing technology, this method integrates empirical corrections, adds turbulence processes to traditional mesoscale numerical models, and accurately estimates data in unobserved areas through the production of a dynamic downscaling training set for the numerical model and artificial intelligence regional error correction. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 The data fusion process based on the COAWST-LES model and graph neural network algorithm provided by the present invention;

[0021] Figure 2 The following is a process for constructing a four-dimensional feature field in an embodiment of the present invention;

[0022] Figure 3 This is a process for implementing deduction and estimation based on numerical model and AI coupling iteration in an embodiment of the present invention;

[0023] Figure 4 This is a scatter plot of atmospheric refractive index in an embodiment of the present invention;

[0024] Figure 5 This is a line graph of atmospheric refractive index in an embodiment of the present invention;

[0025] Figure 6 It is a dot-line diagram of the atmospheric refractive index in an embodiment of the present invention;

[0026] Figure 7 Root mean square error diagram of different algorithms in the embodiment of the present invention. DETAILED DESCRIPTION

[0027] To facilitate understanding of the present invention, the present invention is described in more detail below with reference to the accompanying drawings and specific embodiments. Preferred embodiments of the present invention are shown in the accompanying drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described in this specification. Rather, these embodiments are provided to provide a more thorough and comprehensive understanding of the present disclosure.

[0028] The method for deducing the regional distribution of atmospheric refractive index at sea based on vertical observations of a single station provided by the present invention is essentially a numerical prediction method that integrates empirical corrections, and can make up for the shortcomings of existing artificial intelligence algorithms and numerical prediction methods. Compared with the purely data-driven artificial intelligence deduction method, the deduction results of this method are based on numerical forecasts, and the dynamic downscaling simulation of the numerical model is used as a training set, which can realize the deduction from a single station to all locations in the region, and the regional distribution deduction results are constrained by the physical framework of the numerical model and are more interpretable. Compared with traditional numerical prediction methods, this method has been improved through physical processes and corrected by artificial intelligence, and the results have increased the influence of small-scale physical processes and regional error corrections, which can greatly improve its practical value. The core idea of the present invention is to achieve accurate estimation of data in unobserved areas by adding turbulent processes to traditional mesoscale numerical models, and by making dynamic downscaling training sets of numerical models and correcting regional errors through artificial intelligence. The specific steps are as follows:

[0029] 1. Modify the ocean-atmosphere coupled model COAWST.

[0030] The original atmospheric module WRF in the ocean-air coupled model COAWST is replaced by the WRF-LES model that considers the large eddy simulation process, so that the COAWST model has the ability to simulate atmospheric turbulence, which is recorded as the COAWST-LES model.

[0031] Second, based on the modified COAWST-LES model and graph neural network algorithms, sparse observation data fusion is achieved within the target area. A dataset of atmospheric element drivers within the target area is collected. A long-term historical simulation of the target area is conducted using the COAWST-LES model with large eddy simulation enabled to generate a high-resolution training dataset. Furthermore, publicly available multi-source observation data from stations, buoys, soundings, and satellites within the target area over the historical period is collected to supplement the regional fusion observation data.

[0032] Constructing a fusion model: A graph neural network (GNN) fusion approach was employed. The network architecture utilizes a GNN as the backbone, with pattern grid points as the primary nodes, and observation station locations as auxiliary nodes. The Transformer attention mechanism and Coriolis force bias were introduced to calculate the dynamic weights between the vertical observation location of a single station and other grid points within the region. Weak supervision rules were established for multi-source data, restricting only the low-precision spatial correlation statistical features of the observation data (e.g., the covariance matrix).

[0033] During formal model training, the training data set is split into different scenarios for separate training. Based on the training results of ideal weather or uniform stratification scenarios, various unstable stratification scenarios are gradually added for parameter optimization and result correction to improve the accuracy of multi-source data fusion in the target area.

[0034] 3. Based on the single-station atmospheric vertical observation data, the temporal and spatial characteristics of the data are reconstructed to construct a four-dimensional feature field. The single-station atmospheric vertical observation data (mainly temperature, humidity, and air pressure elements) are analyzed. The instantaneous observation data are analyzed for high-frequency time series using wavelet transform. Assume that the original signal is , its wavelet transform is expressed as: ,in, is the scale parameter, is the translation parameter, is the mother wavelet function, For the moment. The analysis extracts the fluctuation characteristics of minute-level changes, calculates the slope of the kinetic energy spectrum, quantifies the proportion of small-scale (<10km) energy, and obtains the magnitude of turbulent kinetic energy in real data through energy spectrum diagnosis.

[0035] Subsequently, a three-dimensional disturbance field centered on the single-station atmospheric vertical observation position is generated based on the generative adversarial network (GAN) algorithm.

[0036] The fused data generated in step 2 above is used as the input to the GAN generator. Using the actual turbulence statistics provided by the COAWST-LES large eddy simulation (with a horizontal resolution of 200-500 m) as constraints, the GAN algorithm generates a three-dimensional disturbance field centered on the vertical observation location of a single atmospheric station. Based on the classical linear wave theory of boundary layer meteorology, a gravity wave parameterization scheme is constructed, and the buoyancy frequency is obtained from the vertical observation data of the single station: ;in, is the acceleration due to gravity, is the reference potential temperature, is the potential temperature, For height.

[0037] Then, the dominant level scale of the disturbance is extracted and vertical scale , and the corresponding horizontal and vertical wave numbers are The phase velocity and group velocity of gravity waves are further obtained, thereby quickly estimating the propagation direction of turbulent disturbances and constructing a four-dimensional characteristic field.

[0038] 4. Using a four-dimensional variational assimilation approach, we coupled numerical models with AI for iterative solutions, enabling regional simulations. Based on the spatiotemporal characteristics of single-station vertical observation data and combined with the perturbation spatial field generated by GAN, we established a closed-loop system: "Numerical models provide background fields → AI generates small-scale perturbation fields → numerical models dynamically respond." This approach enables an iterative solution method coupled with numerical models and AI.

[0039] The specific steps are as follows: A background field for the current time period is generated using the modified COAWST-LES model, with a temporal resolution of 30 minutes. Subsequently, a GAN algorithm is used to combine single-station atmospheric vertical observations for the current time period with the model background field, focusing on variations in temperature, humidity, and pressure to generate a small-scale perturbation field. Using this perturbed background field as a driver, a short-term forecast is performed using the coupled numerical model. The new forecast field is then used to update the GAN algorithm input, resulting in an iterative optimization process.

[0040] In this process, it is necessary to embed mass conservation constraints in the GAN network loss function and add physical constraints to the GAN modeling process to prevent GAN from destroying the overall conservation characteristics of the region when generating small-scale disturbance fields. The mass conservation formula is as follows: ;in, is the fluid density, is the velocity vector, is the divergence operator.

[0041] At the same time, at the single-station vertical observation position in the output layer, the single-point perturbation is forced to 0 to ensure strict matching with the observation. Spectral normalization is used to control the perturbation intensity generated by the GAN to not exceed the physically reasonable range. Spectral normalization is a weight normalization used to stabilize the training of the discriminator. The core idea of spectral normalization is to perform singular value decomposition on the weight matrix and use the largest singular value obtained to scale the weight matrix to ensure the continuity of the Lipschitz constant. The formula is as follows: ;in, W Represents the weight matrix of each layer in the neural network, Represents the largest singular value.

[0042] The maximum singular value can be obtained by singular value decomposition (SVD): ;in, 、 is an orthogonal matrix, Is a diagonal matrix, where the values on the diagonal are The singular values of .

[0043] Through iterative simulation, the vertical observation of the atmosphere at a single point is finally deduced and estimated to the target area. Finally, according to the Babin empirical formula, the atmospheric refractive index is calculated using the deduced and estimated temperature, humidity, air pressure and other factors. The formula is as follows:

[0044] ;

[0045] in, Represents the refractive index (unit: N-units), Indicates temperature (unit: K), Indicates air pressure (unit: hPa), represents the water vapor pressure, which can be calculated using the following equation:

[0046] ;

[0047] in, Indicates specific humidity (unit: g / kg), is a constant (0.622). Based on the calculation results, the regional atmospheric refractivity deduction results are finally generated.

[0048] To evaluate the technical effectiveness of the method for estimating the regional distribution of atmospheric refractivity at sea based on single-station vertical observations, the following experiment was conducted: Hangzhou and Hongjia stations were selected within the range of 120°E–122.5°E and 28.5°N–30.5°N. Hongjia station served as a known observation point, while Hangzhou station served as a verification site for testing.

[0049] ERA5 reanalysis data were collected and organized as driving data. At the same time, multi-source public observation data from stations, buoys, soundings, and satellites in the study area during the historical period were collected to supplement the regional fusion observation data.

[0050] To demonstrate the effectiveness of the proposed method, three different methods were used to compare and calculate the atmospheric refractive index. Specifically, they include: 1) directly using the temperature, humidity, and air pressure elements in the ERA5 reanalysis data and substituting them directly into the Babin formula to calculate the atmospheric refractive index; 2) using the same formula to calculate the atmospheric refractive index using a mesoscale numerical model that does not consider turbulence processes; and 3) using the method proposed in this invention to calculate the atmospheric refractive index. The three methods were used to calculate and plot scatter plots, line plots, and point-line plots of the atmospheric refractive index at different altitudes. The comparison results are shown in Figure 2. Figure 4 、 Figure 5 、 Figure 6 As shown in the figure, the black line represents the atmospheric refractivity calculated based on the radiosonde observation data from the Hangzhou station, which is used as a reference for real data. The green line represents the atmospheric refractivity calculated from the ERA5 reanalysis data. The blue line represents the atmospheric refractivity profile calculated using the traditional numerical model without considering the turbulence process. The red line represents the experimental result obtained using the method of the present invention.

[0051] It can be concluded that the results of meteorological elements calculated from reanalysis data are quite different from the actual results because the resolution is too coarse. In comparison, the overall trend of the simulation results of the blue and red lines is closer to reality. However, by adopting the method proposed in this invention, the large eddy simulation process is introduced into the numerical model, and the regional error correction is performed by combining the high-resolution training set generated based on numerical simulation and the artificial intelligence algorithm, which can capture the actual disturbance and fluctuation characteristics in the atmosphere. Then, based on the sounding data of Hangzhou Station, the root mean square error diagram of the above three methods and the sounding data was drawn as shown below. Figure 7As shown in the figure, the green line represents the RMS error between the ERA5 reanalysis data and the sounding data, the blue line represents the RMS error between the traditional numerical model and the sounding data, and the red line represents the RMS error between the method proposed in this paper and the sounding data. It can be seen that the RMS error of the reanalysis data is generally larger. The RMS error of the proposed method is mostly lower than that of the traditional numerical model, closer to the real data, verifying the applicability and advantages of the proposed method in complex atmospheric environments.

Claims

1. A method for estimating the regional distribution of atmospheric refractivity at sea based on vertical observations at a single station, characterized by: The following steps are involved: Step 1: Use the improved COAWST-LES model and graph neural network algorithm to deduce the sparse observation data fusion in the target area; the improved COAWST-LES model replaces the original atmospheric module WRF in the ocean-air coupling model COAWST with the WRF-LES model that considers the large eddy simulation process; the fusion model is constructed with the graph neural network (GNN) algorithm as the main framework; the nodes are mainly model grid points, supplemented by the locations of various observation points; the dynamic weights of the vertical observation location of a single station and other grid points in the area are calculated using the Transformer algorithm, and the Coriolis force parameter is used as the attention bias term; Set weak supervision rules for multi-source data and only impose spatial correlation statistical feature constraints on the observed data; The improved COAWST-LES model is used to conduct long-term large eddy simulations in the target area to generate a high-resolution training dataset. According to different weather scenarios, the training data set is split into different scenarios and the fusion model is trained separately; the fusion model is used to generate fusion data; Step 2: Based on the single-station atmospheric vertical observation data and the fused data generated in step 1, reconstruct the spatiotemporal characteristics of the data and construct a four-dimensional feature field. The vertical observation data includes temperature, humidity, and air pressure. Specifically: Analyze single-station atmospheric vertical observation data; perform high-frequency time series analysis on instantaneous observation data through wavelet transformation, extract minute-level fluctuation characteristics, and capture turbulence signals; calculate the slope of the kinetic energy spectrum, quantify the proportion of small-scale energy, and obtain the magnitude of turbulent kinetic energy in the real data through energy spectrum diagnosis, which is used to subsequently correct the model output results; Based on the generative adversarial network (GAN) algorithm, a three-dimensional disturbance field centered on the vertical atmospheric observation position of a single station is generated. The generated fused data is used as the input of the GAN generator. A large eddy simulation is carried out on the target area using the improved COAWST-LES model to obtain the realistic turbulence distribution. Based on the classical linear wave theory of boundary layer meteorology, a gravity wave parameterization scheme is constructed to quickly estimate the propagation direction of turbulent disturbances and construct a four-dimensional characteristic field. Step 3: Adopt the four-dimensional variational assimilation strategy to perform iterative solution of numerical model and AI coupling to achieve regional deduction estimation; use the deduced estimated temperature, humidity, and air pressure to calculate the atmospheric refractive index, specifically: generate the background field for the current period through the improved COAWST-LES model with a time resolution of once every 30 minutes; use the GAN algorithm to combine the single-station atmospheric vertical observations and model background fields for the current period to generate a small-scale disturbance field for changes in temperature, humidity, and air pressure fields; use the disturbed background field as the driving force to carry out short-term forecasts of the coupled numerical model, and update the GAN input with the new forecast field to form an iterative optimization, realizing the deduction estimation from single-point atmospheric vertical observations to the target area.

2. The method for deducing regional distribution of atmospheric refractivity at sea based on single-station vertical observation according to claim 1, characterized in that: Mass conservation constraints are embedded in the GAN network loss function to add physical constraints to the GAN modeling process.

3. The method for deducing regional distribution of atmospheric refractive index at sea based on single-station vertical observation according to claim 1, characterized in that: At the single-station vertical observation position in the output layer, the single-point perturbation amount is forced to be 0.

4. The method for deducing regional distribution of atmospheric refractive index at sea based on single-station vertical observation according to claim 1, characterized in that: Perform spectral normalization on the perturbation strength generated by GAN.

Images