Offshore atmospheric refractive index regional distribution deduction method based on single-station vertical observation

Abstract

The invention belongs to the technical field of marine environment monitoring, and relates to a marine atmospheric refractive index regional distribution deduction method based on single-station vertical observation. The method comprises the steps of 1, deducing sparse observation data fusion in a target area by using an improved COAWST-LES mode and a graph neural network algorithm; step 2, based on single-station atmospheric vertical observation data and the fusion data generated in the step 1, data spatial-temporal feature reconstruction is carried out, and a four-dimensional feature field is constructed; step 3, adopting a four-dimensional variational assimilation strategy to carry out numerical mode and AI coupling iteration solution to realize regional deduction estimation; and calculating the atmospheric refractive index by using the deduced and estimated air temperature, humidity and air pressure. According to the method, empirical correction is fused, a turbulence process is added in a traditional mesoscale numerical mode, and data without an observation area can be accurately estimated through manufacturing of a numerical mode dynamic downscaling training set and artificial intelligence area error correction.

Description

Technical Field

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

[0002] Atmospheric refractive index is a parameter that characterizes the bending and deflection of the propagation path of electromagnetic waves in the atmosphere when they propagate in the atmosphere. The calculation of atmospheric refractive index is related to the basic elements of the atmosphere and is usually obtained by calculation using an empirical model.

[0003] The monitoring of atmospheric refractive index usually uses equipment such as radiosondes and lidar to carry out single-station vertical observation and diagnosis of the atmosphere, and obtain the vertical profile of atmospheric refractive index above the ground / sea surface.

[0004] As the core input parameter of the electromagnetic wave propagation simulation model, the vertical profile of atmospheric refractive index is mainly used for the prediction and simulation of the radar power range. The electromagnetic wave simulation model needs the three-dimensional atmospheric refractive index environmental distribution within the radar detection range during calculation, but the current single-station vertical observation can only obtain local profile information and lacks effective regional observation means. Therefore, in actual simulation applications, it is usually assumed that the atmospheric refraction environment within the target range is uniform and unchanged, and the single-station observation information is copied to other locations to form regional three-dimensional environmental data to drive the operation of the electromagnetic wave simulation model.

[0005] This uniform assumption does not consider the spatial differences within the region, thus resulting in a large deviation in the simulation results of the electromagnetic wave propagation path.

[0006] Regarding the problem of deducing from single - station vertical observations at sea to three - dimensional spatial distribution in a region, there have been mainly three technical paths in the past. However, up to now, none of them have effectively solved this problem. The first path is the mesoscale numerical model prediction method. Based on limited observation points, data assimilation is used to introduce data into the numerical model to simulate three - dimensional spatial changes within the region. The drawback of this path is that the model scale is relatively large, and its physical process scheme does not include small - scale atmospheric processes. The simulation results are too smooth to simulate the fluctuations and mutations in real observation data caused by small - scale processes such as turbulence, resulting in low practical application value of the simulation results. The second path is the small - scale geophysical fluid dynamics simulation, which can consider processes such as turbulence. However, its disadvantages are huge computational requirements and it can only simulate fluid motions such as wind and current, and it cannot be used for calculating atmospheric refractive index for the time being. The third path is artificial intelligence algorithms such as machine learning or deep learning, which establish the mapping relationship between profiles at other locations and the observation points to achieve the deduction from a single point to a region. However, this path lacks long - term data in the target region for model training. The meteorological elements in remote sensing and re - analysis data, which are commonly used in past artificial intelligence research, cannot be used for training in this problem due to their too - coarse resolution, and can only be used for deduction and estimation from a single station to multiple limited locations. Therefore, it is temporarily impossible to estimate the accurate regional distribution of atmospheric refractive index at sea using limited station observations. Summary of the Invention

[0007] Based on the current problems, the present invention discloses a method for deducing the regional atmospheric refractive index and a data production process for areas without data at sea based on single - station atmospheric vertical observations. Taking the profile data of single - station atmospheric vertical observations as the evaluation benchmark, through a combination of numerical model transformation, numerical prediction, and intelligent regional error correction, a new idea for estimating the regional atmospheric refractive index distribution is designed to solve the problem that it is difficult for current traditional technical paths to accurately estimate the regional atmospheric refractive index 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 refractive index at sea based on single - station vertical observations, comprising the following steps: Step 1: Use the improved COAWST - LES model and the graph neural network algorithm to deduce the fusion of sparse observation data in the target region; the improved COAWST - LES model is to replace the original atmospheric module WRF in the COAWST air - sea coupling model with the WRF - LES model considering large - eddy simulation processes; Step 2: Based on the single - station atmospheric vertical observation data and the fusion data generated in Step 1, perform data spatio - temporal feature reconstruction to construct a four - dimensional feature field; among them, the vertical observation data includes temperature, humidity, and air pressure; Step 3: Adopt a four-dimensional variational assimilation strategy to perform coupled iterative solutions of the numerical model and AI, and achieve regional deduction estimation; calculate the atmospheric refractive index using the deduced and estimated air temperature, humidity, and air pressure.

[0009] Preferably, Step 2 specifically includes: constructing a fusion model with the Graph Neural Network (GNN) algorithm as the main framework; using the model grid points as the main nodes, supplemented by the positions of each observation point; calculating the dynamic weights between the position of the single-station vertical observation and other grid points within the region through the Transformer algorithm, and using the Coriolis force parameter as the attention bias term; 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 conduct long-term large-eddy simulations for the target region to generate a high-resolution training dataset; splitting the training dataset into different scenarios for separate training according to different weather scenarios.

[0010] Preferably, Step 3 specifically includes: parsing the single-station atmospheric vertical observation data; performing high-frequency time series analysis on the instantaneous observation data through wavelet transform to extract the fluctuation characteristics of minute-level changes and capture the turbulence signal; 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 for subsequent correction of the model output results; Based on the Generative Adversarial Network (GAN) algorithm, generate a three-dimensional perturbation field centered on the position of the single-station atmospheric vertical observation; use the generated fusion data as the input of the GAN generator; use the large-eddy simulation carried out by the improved COAWST-LES model for the target region to obtain the real turbulence distribution; based on the classical linear wave theory of boundary layer meteorology, construct a gravity wave parameterization scheme to quickly estimate the propagation direction of turbulent perturbations.

[0011] Preferably, Step 3 specifically includes: generating the background field for the current time 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 observation and the model background field for the current time period, and generating a small-scale perturbation field for the changes in the air temperature, humidity, and air pressure fields; using the background field added with perturbations as the driving force to carry out short-term forecasting of the coupled numerical model, and updating the GAN input with the new forecast field to form iterative optimization, so as to achieve the deduction estimation from the single-point atmospheric vertical observation to the target region.

[0012] Preferably, a mass conservation constraint is embedded in the GAN network loss function to add physical constraints to the GAN modeling process.

[0013] Preferably, at the position of the single-station vertical observation in the output layer, the single-point perturbation amount is forced to be 0.

[0014] Preferably, spectral normalization is performed on the perturbation intensity generated by the GAN.

[0015] The present invention discloses a method for deducing the regional distribution of marine atmospheric refractive index based on single-station vertical observation. Compared with the prior art, it integrates empirical correction, adds a turbulence process to the traditional mesoscale numerical model, and through the production of a numerical model dynamic downscaling training set and artificial intelligence regional error correction, can accurately estimate the data in the unobserved area. Description of the Drawings

[0016] Figure 1 It is the flow chart of data fusion based on the COAWST-LES model and the graph neural network algorithm provided by the present invention; Figure 2 It is the construction process of the four-dimensional feature field in the embodiment of the present invention; Figure 3 It is the process of realizing deduction and estimation based on the coupling iteration of the numerical model and AI in the embodiment of the present invention; Figure 4 It is the scatter plot of atmospheric refractive index in the embodiment of the present invention; Figure 5 It is the line graph of atmospheric refractive index in the embodiment of the present invention; Figure 6 It is the dot-line graph of atmospheric refractive index in the embodiment of the present invention; Figure 7 It is the root mean square error graph of different algorithms in the embodiment of the present invention. Detailed Embodiments

[0017] To facilitate the understanding of the present invention, the present invention will be described in more detail below with reference to the accompanying drawings and specific embodiments. The 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. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosed content of the present invention more thorough and comprehensive.

[0018] The method for deducing the regional distribution of marine atmospheric refractive index based on single-station vertical observations provided by the present invention essentially belongs to a numerical prediction method integrated with empirical correction, which can make up for the drawbacks of existing artificial intelligence algorithms and numerical prediction methods. Compared with the artificial intelligence deduction method driven solely by data, the deduction result of this method is based on numerical prediction, uses the dynamic downscaling simulation of the numerical model as the training set, can realize the deduction from a single station to all positions in the region, and the regional distribution deduction result is constrained by the physical framework of the numerical model, with higher interpretability. Compared with the traditional numerical prediction method, this method has been improved in physical processes and corrected by artificial intelligence, and its result adds the influence of small-scale physical processes and regional error correction, which can greatly improve its practical value. The core idea of the present invention is to add a turbulent process to the traditional mesoscale numerical model, and through the production of a numerical model dynamic downscaling training set and artificial intelligence regional error correction, to achieve accurate estimation of data in the unobserved area. The specific steps are as follows: 1. Modify the air-sea coupled model COAWST.

[0019] Replace the original atmospheric module WRF in the air-sea coupled model COAWST with the WRF-LES model considering large-eddy simulation process, so that the COAWST model has the ability to simulate atmospheric turbulence, denoted as the COAWST-LES model.

[0020] 2. Based on the modified COAWST-LES model and the graph neural network algorithm, realize the fusion of sparse observation data in the target deduction area. Collect the driving data set of atmospheric elements in the target area, and use the COAWST-LES model with large-eddy simulation process turned on to conduct long-term historical simulations for the target area to generate a high-resolution training data set. At the same time, collect the public multi-source observation data such as stations, buoys, radiosondes, and satellites in the target area during the historical period as a supplement to the regional fusion observation data.

[0021] Construct a fusion model: adopt the graph neural network algorithm GNN fusion method, with the network architecture based on GNN as the backbone network, the nodes being the model grid points, and the positions of the observation stations as the auxiliary. Calculate the dynamic weights between the position of the single-station vertical observation and other grid points in the region by introducing the Transformer attention mechanism and the Coriolis force bias term. Set the weak supervision rules for multi-source data, and only constrain the statistical features of the spatial correlation (such as the covariance matrix) of the observation data with lower accuracy.

[0022] During the formal model training, split the training data set into different scenarios for training respectively, and gradually add various unstable stratification scenarios based on the training results in the ideal weather or uniform stratification scenarios for parameter optimization and result correction, so as to improve the accuracy of multi-source data fusion in the target area.

[0023] III. Based on the single-station atmospheric vertical observation data, reconstruct its spatio-temporal characteristics, and construct a four-dimensional characteristic field. Analyze the single-station atmospheric vertical observation data (mainly temperature, humidity, and pressure elements). Perform high-frequency time series analysis on the instantaneous observation data through wavelet transform. Let the original signal be , and its wavelet transform is expressed as: , where is the scale parameter, is the translation parameter, is the mother wavelet function, is the time. By analyzing , extract the fluctuation characteristics of minute-level changes, calculate the kinetic energy spectrum slope, quantify the proportion of small-scale (<10km) energy, and obtain the magnitude of turbulent kinetic energy in the real data through energy spectrum diagnosis.

[0024] Subsequently, generate a three-dimensional perturbation field centered on the single-station atmospheric vertical observation position based on the GAN algorithm.

[0025] Take the fused data generated in the above step two as the input of the generator of GAN, and take the real turbulent statistical characteristics provided by COAWST-LES large eddy simulation (horizontal resolution of 200m - 500m) as the constraint. Generate a three-dimensional perturbation field centered on the single-station atmospheric vertical observation position through the GAN algorithm. Based on the classical linear wave theory of boundary layer meteorology, construct a gravity wave parameterization scheme, and obtain the buoyancy frequency through the single-station vertical observation data: ; where is the gravitational acceleration, is the reference potential temperature, is the potential temperature, is the height.

[0026] Subsequently, extract the dominant horizontal scale and vertical scale , and obtain the corresponding horizontal wave number and vertical wave number as . Further obtain the phase velocity and group velocity of the gravity wave, so as to quickly estimate the propagation direction of the turbulent perturbation and construct a four-dimensional characteristic field.

[0027] 4. Adopt the four-dimensional variational assimilation idea to perform numerical model and AI coupled iterative solution to achieve regional deduction calculation. Based on the spatio-temporal characteristic analysis of single-station vertical observation data, combined with the perturbation space field generated by GAN, establish a closed-loop system of "the numerical model provides the background field → artificial intelligence generates the small-scale perturbation field → the numerical model dynamically responds", and realize the numerical model and artificial intelligence coupled iterative solution method.

[0028] The specific steps are as follows: Generate the background field for the current time period through the modified COAWST-LES model, with a time resolution of once every 30 minutes; Subsequently, use the GAN algorithm, combine the single-station atmospheric vertical observations and the model background field for the current time period, and focus on the changes in temperature, humidity, and pressure fields to generate a small-scale perturbation field. Use the background field with added perturbations as the driving force to conduct short-term forecasts of the coupled numerical model, and update the GAN algorithm input with the new forecast field to form iterative optimization.

[0029] During this process, it is necessary to embed the mass conservation constraint in the GAN network loss function, add physical constraints to the GAN modeling process to prevent the GAN from destroying the overall conservation characteristics of the region when generating the small-scale perturbation field. The mass conservation formula is as follows: ; where is the fluid density, is the velocity vector, is the divergence operator.

[0030] At the same time, at the single-station vertical observation position of the output layer, force the single-point perturbation amount to be 0 to ensure strict matching with the observation, and control the perturbation intensity generated by the GAN not to exceed the physically reasonable range through spectral normalization. Spectral normalization is a type of 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 scale the weight matrix with the largest singular value obtained to ensure the continuity of the Lipschitz constant. The formula is as follows: ; where W represents each layer of the weight matrix in the neural network, represents the largest singular value.

[0031] The largest singular value can be obtained through singular value decomposition (SVD): ; where and are orthogonal matrices, is a diagonal matrix, and the values on the diagonal are the singular values of

[0032] Through iterative simulation, finally realize the deduction and estimation from single-point atmospheric vertical observations to the target area. Finally, according to the Babin empirical formula, use the deduced and estimated elements such as temperature, humidity, and pressure to calculate the atmospheric refractive index. The formula is as follows: ; where represents the refractive index (unit: N-units), represents the temperature (unit: K), represents the pressure (unit: hPa), represents the water vapor pressure, which can be calculated through the following equation: ; wherein, represents the specific humidity (unit: g / kg), is a constant (0.622). Based on the calculation results, the regionalized atmospheric refractive index deduction results are finally generated.

[0033] To evaluate the technical effect of the method for deducing the regional distribution of the marine atmospheric refractive index based on single-station vertical observations provided by the present invention, the following experiments are carried out by the present invention: Two stations, namely Hangzhou Station and Hongjia Station, are selected within the range of 120°E–122.5°E and 28.5°N–30.5°N. Among them, Hongjia Station is used as a known observation point, while Hangzhou Station is used as a verification station for verification.

[0034] Collect and organize ERA5 reanalysis data as driving data. At the same time, collect multi-source public observation data such as stations, buoys, radiosondes, and satellites in the study area during the historical period as a supplement to the regional fusion observation data.

[0035] To prove the effectiveness of the method proposed by the present invention, three different methods are set to calculate the atmospheric refractive index by comparison, specifically including: ① directly use the elements such as temperature, humidity, and air pressure in the ERA5 reanalysis data and directly substitute them into the Babin formula to calculate the atmospheric refractive index; ② use a mesoscale numerical model without considering the turbulence process to calculate the atmospheric refractive index through the same formula; ③ use the method proposed by the present invention to calculate the atmospheric refractive index. Scatter plots, line charts, and dot-line charts of the atmospheric refractive index at different heights are calculated and drawn by the three methods, and the comparison results are as shown in Figure 4 、 Figure 5 、 Figure 6 shown, wherein, black represents the atmospheric refractive index calculated based on the radiosonde observation data of Hangzhou Station and is used as a reference for real data. Green is the atmospheric refractive index calculated from the ERA5 reanalysis data, the blue line is the atmospheric refractive index profile calculated using the traditional numerical model without considering the turbulence process, and the red line is the result obtained by using the method of the present invention for testing.

[0036] It can be concluded that the results calculated from the meteorological elements of the reanalysis data have a large gap with the actual results due to the too coarse resolution. In contrast, the overall trends of the simulation results of the blue line and the red line are closer to the real situation. However, by using the method proposed by the present invention, the large-eddy simulation process is introduced into the numerical model, and combined with the high-resolution training set generated based on numerical simulation and the artificial intelligence algorithm for regional error correction, the disturbance fluctuation characteristics actually existing in the atmosphere can be captured. Then, based on the radiosonde data of Hangzhou Station, the root-mean-square error diagrams of the above three methods and the radiosonde data are drawn as shown in Figure 7As shown, where the green line is the root mean square error between ERA5 reanalysis data and radiosonde data, the blue line is the root mean square error between the traditional numerical model and radiosonde data, and the red line is the root mean square error using the method of the present invention and radiosonde data. It can be concluded that the overall root mean square error of the reanalysis data is relatively large. The root mean square error of the method proposed by the present invention is mostly lower than that of the traditional numerical model and is closer to the real data, verifying the applicability and advantages of the method of the present invention in a complex atmospheric environment.

Claims

1. A method for deducing the regional distribution of the marine atmospheric refractive index based on single-station vertical observations, characterized in that It includes the following steps: Step 1: Using the improved COAWST-LES model and the graph neural network algorithm to deduce the fusion of sparse observational data in the target area; the improved COAWST-LES model replaces the original atmospheric module WRF in the COAWST air-sea coupling model with the WRF-LES model considering the large eddy simulation process; Step 2: Based on the single-station atmospheric vertical observational data and the fusion data generated in Step 1, conduct data spatio-temporal feature reconstruction to construct a four-dimensional feature field; among them, the vertical observational data includes temperature, humidity, and air pressure; Step 3: Adopt a four-dimensional variational assimilation strategy to conduct coupled iterative solution of the numerical model and AI to achieve regional deduction estimation; calculate the atmospheric refractive index using the deduced and estimated temperature, humidity, and air pressure.

2. The method for deducing the regional distribution of the atmospheric refractive index at sea based on single-station vertical observation according to claim 1, wherein Step 2 specifically includes: constructing a fusion model with the graph neural network GNN algorithm as the main framework; the nodes are mainly mode grid points, supplemented by the positions of each observation point; calculating the dynamic weights between the position of the single-station vertical observation and other grid points in the area through the Transformer algorithm, and using the Coriolis force parameter as the attention bias term; setting weak supervision rules for multi-source data, and only imposing spatial correlation statistical feature constraints on the observational data; using the improved COAWST-LES model to conduct long-term large eddy simulation for the target area to generate a high-resolution training dataset; splitting the training dataset into different scenarios for separate training according to different weather scenarios.

3. The method for deducing the regional distribution of the atmospheric refractive index at sea based on single-station vertical observation according to claim 1, wherein, Step 3 specifically includes: analyzing the single-station atmospheric vertical observational data; conducting high-frequency time series analysis on the instantaneous observational data through wavelet transform to extract the fluctuation characteristics of minute-level changes and capture the turbulence signal; 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 for subsequent correction of the model output results; Based on the generative adversarial network GAN algorithm, generate a three-dimensional perturbation field centered on the single-station atmospheric vertical observation position; use the generated fusion data as the input of the GAN generator; use the large eddy simulation carried out by the improved COAWST-LES model for the target area to obtain the real turbulence distribution; based on the classical linear wave theory of boundary layer meteorology, construct a gravity wave parameterization scheme to quickly estimate the propagation direction of turbulent perturbations.

4. The method for deducing the regional distribution of the atmospheric refractive index at sea based on single-station vertical observation according to claim 1, wherein Step 3 specifically includes: generating the background field of the current time 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 observation and the model background field of the current time period, to generate a small-scale perturbation field for the changes in the temperature, humidity, and air pressure fields; using the background field added with perturbations as the driving force to conduct short-term forecasting of the coupled numerical model, and updating the GAN input with the new forecasting field to form iterative optimization to achieve the deduction estimation from the single-point atmospheric vertical observation to the target area.

5. The method for deducing the regional distribution of the atmospheric refractive index at sea based on single-station vertical observation according to claim 4, wherein Embed the mass conservation constraint in the GAN network loss function to add physical constraints to the GAN modeling process.

6. The method for deducing the regional distribution of the marine atmospheric refractive index based on single-station vertical observation according to claim 4, wherein At the position of the single-station vertical observation in the output layer, force the single-point perturbation amount to be 0.

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

Images