An evaporation waveguide prediction method based on data feature classification and neural network model

Abstract

The present application relates to an evaporation waveguide prediction method based on data feature classification and neural network model, wherein the training data is pre-classified according to time, space and atmospheric stratification state, the atmospheric correction refractive index calculated by using reanalysis meteorological data and an evaporation waveguide calculation model is used to train a neural network to obtain an evaporation waveguide prediction model of a long time and a large sea area. Compared with the traditional neural network model, the present application has higher prediction accuracy and accuracy, can be used for evaporation waveguide prediction of a large sea area, obtains the environmental characteristics of electromagnetic wave propagation at sea, and guarantees the over-the-horizon working demand of the electromagnetic system at sea.

Description

Technical Field

[0001] This invention belongs to the technical fields of near-sea surface evaporation waveguides, marine atmospheric environment, and marine physics. It relates to an evaporation waveguide prediction method based on data feature classification and neural network model, which can be used for evaporation waveguide prediction in large-area sea areas, obtain the environmental characteristics of marine electromagnetic wave propagation, and ensure the over-the-horizon operation requirements of marine electromagnetic systems. Background Technology

[0002] Evaporation waveguides are a common atmospheric waveguide phenomenon frequently observed in the troposphere. Their occurrence is related to the interaction between atmospheric meteorological factors. Due to seawater evaporation, sea-vapor interaction occurs at the sea surface. The vapor continuously diffuses, and as the diffusion altitude increases, the atmospheric refractive index decreases. At a certain altitude, the refractive curvature becomes less than the curvature of the Earth's sea surface, trapping electromagnetic waves within this layer and enabling beyond-line-of-sight propagation.

[0003] Currently, methods for obtaining the corrected refractive index profile of evaporation waveguides mainly include direct measurement, predictive modeling, and inversion methods. Predictive modeling is currently the primary method for obtaining the height of evaporation waveguides and their corrected refractive index. The evaporation waveguide calculation model employs a comprehensive algorithm for air-sea flux obtained from long-term marine surveys. It substitutes air temperature, relative humidity, wind speed, atmospheric pressure, and sea surface temperature at a specific sea altitude into the evaporation waveguide calculation model, and combines this with the formula for atmospheric corrected refractive index, obtaining the atmospheric corrected refractive index profile through several iterations. However, this method is slow and inefficient when calculating evaporation waveguide profiles over long periods or estimating the distribution of evaporation waveguides over large sea areas. Furthermore, a single neural network model requires a large amount of training data, and its applicability varies significantly depending on the month, sea area, and atmospheric stratification.

[0004] Because current methods for obtaining evaporation waveguide profiles have certain limitations, it is impossible to obtain high-precision real-time profiles of marine evaporation waveguides quickly, accurately, and over a wide area in a long-term manner. This makes it difficult to meet the practical needs of accurate prediction, data comparison, and inversion model optimization for evaporation waveguides. Summary of the Invention

[0005] Technical problems to be solved

[0006] To avoid the shortcomings of existing technologies, this invention proposes an evaporation waveguide prediction method based on data feature classification and neural network models. Based on the data feature classification method, a novel nonlinear mapping evaporation waveguide profile estimation model is established using a neural network model, which can be applied to overcome the drawbacks of the iterative process and feature classification process in existing models.

[0007] Technical solution

[0008] An evaporation waveguide prediction method based on data feature classification and neural network models, characterized by the following steps:

[0009] Step 1: Select the reanalysis meteorological data of a certain year as the training area, and select training points within the training area; extract k data items at the training points;

[0010] Step 2: For the k data items in Step 1, perform the first data division according to the month, numbered sequentially from 1 to p; then perform the second data division according to the latitude and longitude resolution of w*j in the regional grid, and number the resulting blocks from left to right and from top to bottom as 1 to q; then perform the third data division according to the atmospheric stratification stability, numbered sequentially from 1 to n, to obtain p*q*n meteorological datasets.

[0011] Step 3: Substitute the meteorological data from p*q*n meteorological datasets into the evaporation waveguide calculation model sequentially to calculate the atmospheric corrected refractive index values ​​from 0 to 50 meters; use the atmospheric refractive index to obtain the atmospheric corrected refractive index profile, and the height corresponding to the lowest point of the atmospheric corrected refractive index profile is the evaporation waveguide height; for any input or output data Xi, calculate the expected value Ei and standard deviation Si according to the standardization formula:

[0012] Z i =(X i -E i ) / S i

[0013] All training data were normalized.

[0014] Step 4: Construct the neural network: The input layer of the neural network has 5 neurons; the hidden layer of the neural network has m neurons, where m ranges from 5 to m and n; the output layer has n neurons, where n ranges from 24 to n and n to 501.

[0015] Step 5: Train the deep neural network separately for different datasets. Use 1000 random data points as a group of sampling samples and input them into the deep neural network to train a total of p*q*n neural network sub-models. If the error of a certain sub-model does not converge, return to step 4 and reconstruct the number of hidden layers and output layers of the neural network; otherwise, save the trained neural network model as a long-term, large-area sea evaporation waveguide prediction model.

[0016] Step 6: Predict the atmospheric corrected refractive index using the evaporation waveguide prediction model.

[0017] The k data items include sea surface pressure, sea surface temperature, air temperature at 2m, specific humidity at 2m, rainfall, tide level, evaporation, and wind speed at 10m.

[0018] In step 1, k items of data k ≥ 5.

[0019] The latitude and longitude of w*j are both greater than 0.5°.

[0020] The atmospheric refractive index values ​​in step 3 are taken at 0.1-meter intervals within the range of 0-50 meters.

[0021] The neural network model is validated by randomly selecting a point within the training area as a test point, and using the method in step 3 to obtain all meteorological data and atmospheric corrected refractive index of the test point throughout the year, thereby verifying the accuracy of the neural network method.

[0022] The atmospheric stratification stability includes stable conditions, neutral conditions, and unstable conditions, which correspond to the air-sea temperature difference being greater than zero, equal to zero, and less than zero, respectively. The air-sea temperature difference is the difference between the air temperature at 2m and the sea surface temperature.

[0023] The learning rate of the neural network should be less than 2; the activation function can be selected from sigmoid, tanh, ReLU, step function, etc.; the neural network structure adopts a fully linked form.

[0024] The neural network structure includes at least the activation function type, learning rate, weights of neurons in each layer, and biases.

[0025] The number of neurons in the output layer of the deep neural network is set according to engineering needs, and shall not exceed 501.

[0026] Beneficial effects

[0027] This invention proposes an evaporation waveguide prediction method based on data feature classification and a neural network model. The method pre-classifies training data according to time, space, and atmospheric stratification. Using reanalysis meteorological data and atmospheric corrected refractive index calculated by the evaporation waveguide computational model, the neural network is trained to obtain a long-term, large-area evaporation waveguide prediction model. Compared with traditional neural network models, this technology has higher prediction accuracy and precision, and can be used for evaporation waveguide prediction in large-area sea areas, obtaining environmental characteristics of marine electromagnetic wave propagation and ensuring the over-the-horizon operation requirements of marine electromagnetic systems. Attached Figure Description

[0028] Figure 1 This is a flowchart of the present invention.

[0029] Figure 2 This is a concept diagram for data feature classification.

[0030] Figure 3 This is a schematic diagram of a neural network structure.

[0031] Figure 4This is a schematic diagram of the root mean square error distribution of the model during neural network training.

[0032] Figure 5 This is an example of an evaporation waveguide profile predicted by this method.

[0033] Figure 6 The error distribution map of evaporation waveguide height calculated by the traditional neural network method and this method is shown. Detailed Implementation

[0034] The present invention will now be further described in conjunction with the embodiments and accompanying drawings:

[0035] This method is characterized by pre-classifying training data according to time, space, and atmospheric stratification, and using reanalysis meteorological data and atmospheric corrected refractive index calculated by the evaporation waveguide computational model to train a neural network to obtain a long-term, large-area evaporation waveguide prediction model. Figure 1 This is a flowchart of the method of the present invention.

[0036] The steps of this method embodiment are as follows:

[0037] Step 1: Download the 2021 reanalysis meteorological data, select the training area (8°N~17°N, 111.5°E~117.5°E), and select training points within the area; extract five data points from the reanalysis meteorological data: sea surface pressure, sea surface temperature, air temperature at 2m, specific humidity at 2m, and wind speed at 10m at the training points.

[0038] Step 2: Classify the reanalysis meteorological data. First, divide the data according to the month, numbering them from 1 to p. Second, divide the data according to the latitude and longitude resolution w*j in the regional grid, where w and j are both greater than 0.5°, and number the resulting blocks from left to right and top to bottom from 1 to q. Third, divide the data according to the atmospheric stratification stability, numbering them from 1 to n, resulting in a total of p*q*n datasets.

[0039] The atmospheric stratification stability includes stable conditions, neutral conditions, and unstable conditions, which correspond to the air-sea temperature difference being greater than zero, equal to zero, and less than zero, respectively. The air-sea temperature difference is the difference between the air temperature at 2m and the sea surface temperature.

[0040] like Figure 2 As shown, the reanalysis meteorological data were classified. The first data division was performed according to the month, numbered 1 to 12. The second data division was performed according to the 3°*3° latitude and longitude resolution in the regional grid, and the resulting blocks were numbered 1 to 6 from left to right and top to bottom. The third data division was performed according to the atmospheric stratification stability, numbered 1 to 3, resulting in a total of 216 datasets.

[0041] Step 3: Substitute the meteorological data from all the above categories into the evaporation waveguide calculation model in sequence. The vertical profiles of wind speed, temperature, and specific humidity in the near-surface layer are represented as follows:

[0042]

[0043]

[0044]

[0045] In the formula, u(z), T(z), and q(z) represent the wind speed, air temperature, and specific humidity at height z, respectively. 0u z 0θ and z 0q The roughness heights, u, are respectively determined by wind speed, temperature, and specific humidity. * θ * and q * For the Morning-Opkhov scale parameters of wind speed, temperature, and humidity, z 0u z 0θ z 0q u * θ * and q * The value was calculated using the COARE 3.0 flux algorithm. κ is the Karman constant, and Γ... d Let ψ be the dry adiabatic lapse rate, and L be the similarity length. m and ψ h The stability correction functions for wind speed and temperature are used, with different forms chosen for different atmospheric stability states. ζ = z / L is the Moning-Opkhov parameter, used to represent the atmospheric stability state: when L→∞ and ζ=0, the atmosphere is in a neutral state; when L>0 and ζ>0, the atmosphere is in a stable state, the smaller L is and the larger ζ is, the more stable the atmosphere; when L<0 and ζ<0, the atmosphere is in an unstable state, the smaller |L| is and the larger |ζ| is, the more unstable the atmosphere. Additionally, the atmospheric stability state can also be determined by the sea-air temperature difference: a positive sea-air temperature difference indicates a stable atmosphere, a zero sea-air temperature difference indicates a neutral atmosphere, and a negative sea-air temperature difference indicates an unstable atmosphere.

[0046] The pressure profile is obtained by solving the static equations and the ideal gas law simultaneously and integrating:

[0047]

[0048] In the formula, z1 is the observation height, z2 is any height of the surface layer, g is the gravitational acceleration, R is the universal gas constant, approximately 287 J / kg / K, and T v This represents the average virtual temperature at heights z1 and z2, in K.

[0049] The vertical profile of the water vapor partial pressure is obtained from the following formula:

[0050]

[0051] In the formula, ε is a constant with a value of 0.622.

[0052] N = (n-1) × 10 6 (6)

[0053]

[0054]

[0055] In the formula, z represents the height in meters (m). The refractive index N and the corrected refractive index M are the most commonly used variables for quantifying the refractive index. They are dimensionless, but for ease of use, units without physical meaning have been artificially added: N-unit and M-unit, respectively.

[0056] The atmospheric corrected refractive index values ​​for 0-50 meters are obtained through the above calculations. The atmospheric corrected refractive index profile is obtained using the atmospheric refractive index. The height corresponding to the lowest point of the atmospheric corrected refractive index profile is the height of the evaporation waveguide. The atmospheric refractive index values ​​are taken at intervals of 0.1 meters between 0-50 meters.

[0057] For any input or output data Xi, calculate the expected value Ei and standard deviation Si according to the standardization formula.

[0058] Z i =(X i -E i ) / S i

[0059] All training data were normalized.

[0060] Step 4: Construct a neural network. The input layer of the neural network has 5 neurons; the hidden layer has 15 neurons; and the output layer has 24 neurons. The neural network structure is as follows: Figure 3 As shown;

[0061] The learning rate of the neural network should be less than 2; the activation function can be selected from sigmoid, tanh, ReLU, step function, etc.; the neural network structure adopts a fully linked form.

[0062] The neural network structure includes at least the activation function type, learning rate, weights of neurons in each layer, and biases.

[0063] The number of neurons in the output layer of the deep neural network is set according to engineering needs, and shall not exceed 501.

[0064] Step 5: Train the neural network. Train the deep neural network separately for different datasets. Use 1000 random data points as a group of samples and input them into the deep neural network. This will train a total of 216 neural network sub-models. An example of the root mean square error distribution during the neural network training process is shown below. Figure 4 As shown;

[0065] Step 6: Randomly select February, Region 1, and unstable conditions as test points for the neural network training. Use 2022 data as test data, and obtain all meteorological data and atmospheric corrected refractive index for the test points throughout the year using the method in Step 3. Figure 5 This is an example of an evaporation waveguide profile obtained through training; to verify the accuracy of this method, a comparison is made using a traditional unclassified neural network method. Figure 6 The error distribution of evaporation waveguide height calculated by the traditional neural network method and the method proposed in this paper is shown. The error of the proposed method is concentrated within 3 meters, while the error distribution of the traditional method above 3 meters is much larger than that of the proposed method.

Claims

1. A prediction method for evaporating waveguides based on data feature classification and neural network models, characterized in that... The steps are as follows: Step 1: Select a specific year's reanalysis meteorological data as the training area, and select training points within the training area; extract the data from the training points. k Item data; The k The data include sea surface pressure, sea surface temperature, air temperature at 2m, specific humidity at 2m, rainfall, tide level, evaporation, and wind speed at 10m. Step 2: For Step 1 k The data items were first divided by month, numbered sequentially from 1 to p; then further divided by region grid. w * The data is divided a second time based on the latitude and longitude resolution of j, and the resulting blocks are numbered from left to right and top to bottom as follows: 1~q; then, according to the stability of the atmospheric stratification, the data is divided a third time, numbered sequentially from 1 to n, to obtain p. * q *n meteorological datasets; The atmospheric stratification stability includes stable conditions, neutral conditions, and unstable conditions, which correspond to the air-sea temperature difference being greater than zero, equal to zero, and less than zero, respectively. The air-sea temperature difference is the difference between the air temperature at 2m and the sea surface temperature. Step 3: Substitute the meteorological data from p*q*n meteorological datasets into the evaporation waveguide calculation model sequentially to calculate the atmospheric corrected refractive index values ​​from 0 to 50 meters; use the atmospheric corrected refractive index to obtain the atmospheric corrected refractive index profile, and the height corresponding to the lowest point of the atmospheric corrected refractive index profile is the evaporation waveguide height; for any input or output data Xi Find the expected value. Ei and standard deviation Si According to the standardized formula: All training data were normalized. Step 4: Construct the neural network: The input layer of the neural network has 5 neurons; the hidden layer of the neural network has [number missing] neurons. m indivual, m The range of values ​​is The number of neurons in the output layer is n indivual, n The range of values ​​is ; Step 5: Train the deep neural network separately for different datasets. Use 1000 random data points as a group of sampling samples and input them into the deep neural network to train a total of p*q*n neural network sub-models. If the error of a certain sub-model does not converge, return to step 4 and reconstruct the number of hidden layers and output layers of the neural network; otherwise, save the trained neural network model as the evaporation waveguide prediction model for the sea area. Step 6: Predict the atmospheric corrected refractive index using the evaporation waveguide prediction model.

2. The evaporation waveguide prediction method based on data feature classification and neural network model according to claim 1, characterized in that: In step 1 k Item data k ≥5.

3. The evaporation waveguide prediction method based on data feature classification and neural network model according to claim 1, characterized in that: The w * j The latitude and longitude are both greater than 0.5º.

4. The evaporation waveguide prediction method based on data feature classification and neural network model according to claim 1, characterized in that: The atmospheric corrected refractive index value in step 3 is taken at a point interval of 0.1 meters between 0 and 50 meters.

5. The evaporation waveguide prediction method based on data feature classification and neural network model according to claim 1, characterized in that: The neural network model is validated by randomly selecting a point within the training area as a test point, and using the method in step 3 to obtain all meteorological data and atmospheric corrected refractive index of the test point throughout the year, thereby verifying the accuracy of the neural network method.

6. The evaporation waveguide prediction method based on data feature classification and neural network model according to claim 1 or 5, characterized in that: The learning rate of the neural network should be less than 2; the activation function can be sigmoid, tanh, ReLU, or step function; the neural network structures are fully linked.

7. The evaporation waveguide prediction method based on data feature classification and neural network model according to claim 1 or 5, characterized in that: The neural network structure includes at least the activation function type, learning rate, weights of neurons in each layer, and biases.

8. The evaporation waveguide prediction method based on data feature classification and neural network model according to claim 1 or 5, characterized in that: The number of neurons in the output layer of the deep neural network is set according to engineering needs, and shall not exceed 501.

Images