Evaporation duct height region prediction method based on convolutional neural network

Abstract

The present application aims at the limited precision of the traditional evaporation waveguide theory model, the insufficient actual observation data of voyage and station, and the influence of the existing data-driven evaporation waveguide height prediction model on observation data, so that the prediction can only be made for a certain point or a certain route, and the regional distribution of the evaporation waveguide cannot be predicted. A regional prediction method of the evaporation waveguide height based on a convolutional neural network is provided, a numerical prediction product evaporation waveguide height regional prediction model (NWPP-EDH) based on the convolutional neural network is constructed by combining the grid data of the numerical prediction product, and a comparison experiment is conducted on the prediction precision of the NWPP-EDH model and multiple traditional theory models. The experimental results show that the NWPP-EDH model has a large improvement in the prediction precision compared with the traditional evaporation waveguide height prediction model.

Description

Technical Field

[0001] This invention relates to the field of evaporation waveguide height region prediction technology, and specifically to a method for predicting evaporation waveguide height region based on a convolutional neural network. Background Technology

[0002] Atmospheric waveguides are a special type of stratification in the atmospheric boundary layer that significantly influences the propagation path of electromagnetic waves. They correspond to the unique property of the atmospheric boundary layer where the refractive index sharply decreases with increasing altitude. Under this characteristic, electromagnetic waves are abnormally bent downwards during propagation, resulting in a curvature radius smaller than that of the Earth. Under suitable frequency and angle conditions, the energy of the electromagnetic waves is repeatedly refracted within the atmospheric stratification. At this point, the atmospheric stratification acts like a conductor, supporting electromagnetic waves to propagate beyond line of sight with low energy attenuation while restricting their transmission to the area within the waveguide's atmospheric stratification. The proper use of atmospheric waveguides can enhance communication capabilities between ships at sea and provide electromagnetic signal shielding for aerial communication equipment. Atmospheric waveguides include evaporation waveguides, surface waveguides, and suspended waveguides. Evaporation waveguides are a special type of surface waveguide; the rapid decrease in humidity above the sea surface with increasing altitude due to seawater evaporation is the primary cause of their formation. Evaporation ducts are frequently generated in marine environments, with an average occurrence rate of up to 60% in global ocean areas. In the South China Sea, the probability of usable evaporation ducts is as high as 80%, and in some special sea areas, permanent evaporation ducts may even form. Generally, evaporation ducts appear at altitudes below 40 meters above the sea surface, and this altitude range has a significant impact on the effectiveness of ship communication equipment at sea.

[0003] Evaporation duct height (EDH) directly reflects the stratification height of the evaporation duct and is a crucial characteristic affecting its intensity, consistently remaining a hot topic and key issue in evaporation duct research. Methods for obtaining EDH mainly include direct detection and indirect measurement. Direct detection involves using various radiosonde devices such as meteorological gradiometers, microwave refractometers, and low-altitude tethered balloons to directly or indirectly obtain atmospheric refractive index profiles, thereby acquiring characteristic values ​​such as EDH height. However, direct detection methods suffer from high cost and complex operation. Indirect measurement methods include theoretical modeling, inversion, and machine learning. Among these, theoretical modeling is the most commonly used method for obtaining EDH height, referring to the ability to calculate the EDH height using known meteorological data as input. In 1971, Jeske from the University of Hamburg, Germany, proposed the Jeske model. In 1985, Paulus further improved the Jeske model by studying more accurate data, which is called the Paulus-Jeske model, or simply the PJ model. In the 1980s, ASGavrilov and other scholars from the Russian State Meteorological and Hydrological University proposed the RSHMU evaporation duct height prediction model. In 1992, Musson-Genon, Gauthier, and Bruth jointly proposed the MGB model based on the Monin-Obukhov similarity theory. In 1997, Babin proposed the BYC model, which applied the advanced COARE air-sea flux algorithm, filling the gaps in waveguide height research under low wind speed conditions and correcting the influence of seawater salinity. In 2000, Frederickson et al. of the U.S. Naval Postgraduate School proposed the NPS model. In 2001, Liu Chengguo et al. proposed the pseudo-refractive index model by adding the concept of pseudo-refractive index to the PJ model. In 2002, Dai Fushan et al. derived the Local model by replacing the MO similarity theory with local similarity theory based on the BYC model. In 2015, Ding Juli et al. proposed the UED model. Although theoretical models are convenient, their accuracy and generalization are still lacking.

[0004] In 2018, Zhu Xiaoyu et al. attempted to combine traditional evaporation waveguide theoretical models and machine learning models, using Support Vector Regression (SVR) to optimize the PJ model. The resulting SVR_PJ model achieved higher accuracy compared to the traditional PJ model. Also in 2018, Zhu Xiaoyu et al. constructed a multilayer perceptron (MLP) model for predicting the height of evaporation waveguides in the South China Sea, SCS-MLP, based on South China Sea data, showing a significant improvement over the PJ model. In 2019, Zhao Wenpeng et al. applied backpropagation neural networks to predict evaporation waveguide height, achieving good results. Subsequently, they proposed a purely data-driven gradient boosting tree (GBDT) model for predicting evaporation waveguide height, PDD_GBR, which showed significant improvements in both accuracy and regional generalization ability. In 2020, Zhao Wenpeng used the extreme gradient boosting (XGBoost) method to construct the XGB model. Theoretically, XGBoost, as an improved version of the GBDT algorithm, has good performance in the application of evaporating waveguide height diagnosis. Experimental results show that the XGB model has a significant improvement in accuracy compared with the PJ theoretical model, and its accuracy and generalization performance are significantly improved compared with the SCS-MLP and PDD_GBR models.

[0005] The methods described above, when using machine learning to predict evaporation duct height, all exhibit higher accuracy and stronger generalization ability compared to traditional models. However, their data sources are meteorological and hydrological observation data from flight routes or a single meteorological station, corresponding to a specific point or flight route. This approach lacks consideration of spatial regional correlation and cannot reflect the distribution of evaporation duct height over a given sea area. Furthermore, the actual amount of observation data from point and flight route observation equipment is relatively small, potentially leading to insufficient sample size and inadequate model training. This paper addresses the problem that traditional evaporation duct theoretical models have limited accuracy, and the lack of sufficient actual observation data from cruises and stations prevents machine learning-based evaporation duct prediction models from performing regional predictions of evaporation duct height. Summary of the Invention

[0006] To address the aforementioned problems, this invention utilizes the Yin-He global spectral model (YHGSM) forecast product dataset for a portion of the South China Sea to construct a regional prediction model for evaporation waveguide height (NWPP-EDH) based on a deep learning convolutional neural network.

[0007] The technical solution to achieve the purpose of this invention is as follows: A method for predicting the height region of evaporating waveguides based on convolutional neural networks is characterized by the following steps: Step 1: Obtain two-dimensional numerical forecast grid data of the sea area to be predicted, clean and normalize the data to obtain the feature dataset; Step 2: Construct a numerical weather prediction model for evaporation waveguide height region based on a convolutional neural network, and train the evaporation waveguide height region prediction model on the feature dataset; Step 3: Input the two-dimensional numerical prediction grid data into the evaporation waveguide height region prediction model for prediction; Step 4: Output the evaporation waveguide height value for this sea area.

[0008] Furthermore, the evaporation waveguide height region prediction model established in step 2 uses the U-Net network structure as its basic framework, transforming the three-layer downsampling into a two-layer downsampling. It includes an encoding module, a bridge module, and a decoding module. The encoding module is used to convert the data matrix into feature codes; the decoding module is used to restore the feature codes into matrix data; and the bridge module is used to connect the encoding module and the decoding module.

[0009] Furthermore, the encoding module includes three convolutional layers, with a downsampling operation between the first and second layers, and a downsampling operation between the second and third layers. Each convolutional layer includes two 2*2 convolutional modules, each containing a BN layer to alleviate overfitting, and employing the ReLU function as the activation layer.

[0010] Furthermore, the bridge module includes one layer, which comprises two 2*2 convolutional modules, each containing a BN layer and using the ReLU function as the activation layer.

[0011] Furthermore, the decoding module includes three layers, each layer including two 2*2 convolutional modules. Each convolutional module contains a BN layer to alleviate overfitting and uses the ReLU function as the activation layer. An upsampling operation module is set in front of each convolutional module.

[0012] Furthermore, the evaporation waveguide height region prediction model uses mean square error as the loss function, and its loss function is:

[0013] The above formula is the MSE obtained from the model's output value and label value, where y_pred refers to the model's output value and y_real refers to the evaporation waveguide height label value.

[0014] Furthermore, the specific operational steps of step 3 include: Step 31: Input two-dimensional numerical weather prediction grid data to obtain the grid data data matrix; Step 32: The data matrix is ​​converted into feature codes by the encoding module, and the feature codes are input into the bridge module; Step 33: The bridge module inputs the feature encoding into the decoding module, the decoding module downsamples the feature encoding, and restores the feature encoding into a data matrix and outputs it; Step 34: Calculate the predicted evaporation waveguide height based on the data matrix.

[0015] Beneficial effects: The NWPP-EDH model proposed in this invention is a deep learning model based on a convolutional neural network structure. This model structure is designed for mesh structures and features encoding, bridging, and decoding functions. It can predict the height of evaporating waveguides in a specified region and has good ground truth fitting capabilities. Finally, compared with traditional BYC, MGB, and NPS evaporating waveguide height prediction models, the NWPP-EDH model shows a significant improvement in prediction accuracy. Attached Figure Description

[0016] Figure 1 This is a structural diagram of the NWPP-EDH model proposed in this invention; Figure 2 This is a graph showing the prediction results of using the NWPP-EDH model in a specified latitude and longitude region; Figure 3 Comparison chart of RMSE for NWPP-EDH, BYC, MGB and NPS models; Figure 4 The error distribution plots for different models are shown below. Figure 4 In the middle (a)-(d), the error distribution diagrams of the NWPP-EDH, MGB, BYC and NPS models are respectively. Detailed Implementation

[0017] To enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions of the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0018] This invention constructs an evaporation waveguide height region prediction model (NWPP-EDH) based on deep learning convolutional neural networks, which can predict the region of evaporation waveguide height based on this model.

[0019] 1. Basic U-Net Network Structure Convolutional Neural Networks (CNNs) are neural networks specifically designed to process data with grid-like structures. Thanks to the use of convolution operations to replace some ordinary matrix multiplications, CNNs exhibit excellent performance in extracting information from grid data. CNNs can reduce the dimensionality of data while preserving the original information within the data, transforming high-dimensional, large-volume data into low-dimensional, smaller-volume data, thereby improving data processing efficiency. This invention uses meteorological and hydrological numerical weather prediction product data as a two-dimensional grid data type; therefore, a CNN is chosen to construct an evaporation duct height regional prediction model. Considering the specific requirements of the practical application, a U-Net network was selected as the baseline network during the model construction process. By adjusting the layers of the baseline U-Net network, a CNN-based numerical weather prediction product evaporation duct height regional prediction model was constructed.

[0020] The U-Net network is a network proposed based on fully connected convolutional neural networks. It was initially designed for segmentation learning of medical images. Medical image datasets typically have a small total number of images but large individual images. To address these characteristics, U-Net incorporates a clever data augmentation scheme, splitting large individual images into sub-images, directly solving the problem of insufficient training samples. The two-dimensional grid data used in this study also features large individual data grids and a small number of data grids. Building a model based on the U-Net network can solve the problem of the limited number of data grids. The U-Net network structure not only incorporates data augmentation but also optimizes the training process. U-Net employs a novel feature fusion method, using a concatenation approach to ensure that features transmitted within the U-shaped network structure do not suffer from edge feature loss, and it also considers the learning of both deep and shallow features. These characteristics enable U-Net to comprehensively and efficiently learn the features of the sample data. The meteorological and hydrological numerical forecast grid data used in this invention requires a deep learning network capable of comprehensively covering the data features. Therefore, a deep learning model that meets the requirements can be built based on the U-Net network structure.

[0021] 2. NWPP-EDH Model Structure (1) Basic structure of neural networks Traditional evaporation waveguide height theoretical models describe the nonlinear relationship between the input vector and the evaporation waveguide height. This relationship is a function of the input meteorological and hydrological parameters, which are functions of sea surface temperature, air temperature, relative humidity, and wind speed. The function value is the evaporation waveguide height (EDH). The mathematical relationship is as follows:

[0022] Where SST is sea surface temperature, Ta is air temperature at 10m above sea level, RH is relative humidity, U is wind speed at 10m, and P is atmospheric pressure. Traditional theoretical models of evaporation waveguide height are semi-empirical physical models built on the Monin-Obukhov similarity theory. These models describe the complex physical relationship between basic meteorological elements and evaporation waveguide height.

[0023] To address the practical problem this research aims to solve—namely, to achieve regional prediction of evaporation duct height in a certain sea area—a convolutional neural network deep learning model is constructed by combining two-dimensional grid meteorological and hydrological numerical forecast data. Analysis shows that the U-Net neural network structure is suitable for two-dimensional grid meteorological and hydrological numerical forecast data; therefore, this model will adopt the U-Net structure as its basic framework. Through modifications and adjustments, a regional prediction model for evaporation duct height based on numerical forecast products using convolutional neural networks will be constructed. The modifications made to the U-Net structure are as follows: This invention uses the U-Net network structure as its basic framework. Since the two-dimensional grid data of meteorological and hydrological numerical forecasting data is a 41*49 matrix, the grid size is insufficient to support good training results under three-layer data downsampling. Therefore, the model structure was improved by changing the three-layer downsampling to a two-layer downsampling. The basic framework of the proposed model is a 7-layer neural network structure, containing three functional modules corresponding to encoding, bridging, and decoding functions. The model parameter settings are shown in Table 1. It can be seen that the encoding module is divided into three layers, with one downsampling operation between the first and second layers, and another between the second and third layers, meaning that the data is compressed twice in the encoding module.

[0024] Table 1 NWPP-EDH Model Parameter Table

[0025] The encoding module transforms the input data matrix into feature codes, while the decoding module restores these feature codes to an element-wise regression problem within the matrix. The bridge layer connects the encoding and decoding modules. Each convolutional module contains a Batch Normalization (BN) layer to mitigate overfitting, using ReLU as the activation layer. The specific model structure is as follows: Figure 1 As shown.

[0026] (2) Define the loss function After providing two-dimensional numerical weather prediction grid data for a specific region and determining the neural network structure, the next goal is to adjust the network parameters so that it can produce more accurate predictions of evaporation duct height. To obtain accurate predictions, it is necessary to minimize the difference between the model's predicted evaporation duct height and the actual value; therefore, mean squared error is used as the loss function.

[0027] The loss function for this model is defined as follows:

[0028] The above formula is the MSE obtained from the model's output value and label value, where y_pred refers to the model's output value and y_real refers to the evaporation waveguide height label value.

[0029] Example To verify the accuracy of the regional prediction model for evaporation waveguide height based on convolutional neural networks proposed in this invention, numerical forecast product data from a portion of the South China Sea was selected as the test dataset in this embodiment, and a comparative experiment was conducted with traditional evaporation waveguide theoretical models: NPS, MGB, and BYC.

[0030] 1. Dataset We selected data from the YHGSM (Yellow Galaxy Global Spectrum Model) numerical weather prediction product for a portion of the South China Sea from November 16 to November 20, 2020 (a total of five days).

[0031] The Galaxy Global Spectrum Model is a numerical weather prediction model developed by the National University of Defense Technology. Its product data has relatively stable systematic errors. This data has been compiled into a dataset for machine learning research in the meteorological field.

[0032] The model performance test of this invention selected uniform latitude and longitude grid data with a grid spacing of 0.125 degrees from the region of 109°-115°E and 15°-22°N in the Milky Way global spectral model. The data included wind speed (U), air temperature (Ta), air pressure (P), relative humidity (RH), and sea surface temperature (SST) at a height of 10 meters.

[0033] 2. Calculate the label value for the height of the evaporating waveguide. Because large-scale direct observation of the true value of evaporation waveguides is currently very difficult, the algorithm used by the US Navy to diagnose evaporation waveguide height using basic meteorological and hydrological elements such as temperature, pressure, humidity, and wind was adopted. The obtained data was used as labels during model training. Specifically, given the above meteorological elements at each point, the label values ​​of evaporation waveguide height were calculated using the method described in the paper "Patterson WL and Hattan, CP Engineer's Refractive Effects Prediction System (EREPS); Naval Command, Control and Ocean Surveillance Center: San Diego, CA, USA, 1994." The calculation steps are as follows: Step 1: Determine if the wind speed is below 0.01 knots / hour (0.005144 m / s). If it is below this wind speed, record the evaporation waveguide height as 0. If it is above this wind speed, continue.

[0034] Step 2: Calculate the overall Richardson number.

[0035] Step 3: Use the overall Richardson number and empirical formulas to determine the Monin-Obukhov length.

[0036] Step 4: Based on the conditions for determining air-sea stability, select a suitable formula for calculating the evaporation waveguide height (EDH) to determine the evaporation waveguide height.

[0037] After calculating the evaporation waveguide height using the above method, it is used as a tag value and bound to meteorological elements such as temperature, air pressure, relative humidity, wind speed, and sea surface temperature at that point to form a data sample.

[0038] Calculations show that the uniform latitude and longitude grid with a spacing of 0.125 degrees within the region of 109°–115°E and 15°–22°N is a 41*49 matrix, meaning we have a total of 41*49 grid points. Each grid point has 5 days of 72-hour numerical weather forecast data, with two forecasts made each day, one starting at 00:00 and the other at 12:00. Therefore, our 5-day dataset contains 2*5*41*49*73 data samples.

[0039] To accommodate the network structure of the model, the length and width of the grid should be numbers divisible by 2. Therefore, we actually cropped the dataset, and the actual data input to the model was a 40*48 matrix concatenated together.

[0040] 3. Regional Prediction Model Accuracy Evaluation Experiment To evaluate the regional prediction performance of the numerical weather prediction product evaporation duct height prediction model based on convolutional neural networks, this invention uses the BYC model, MGB model, and NPS model as benchmark models for comparative experiments. 80% of the total data is used as the training set, and 20% as the test set to test the regional prediction model of the numerical weather prediction product evaporation duct height based on convolutional neural networks.

[0041] Using 80% of the dataset as the training set, the trained model was tested on the remaining 20% ​​of the data. The prediction performance of the NWPP-EDH model is as follows: Figure 2 As shown. Figure 2 The image shows the prediction results of using the NWPP-EDH model in a specified latitude and longitude region. It can be clearly seen that the evaporation waveguide height distribution in this region is relatively uniform, mostly within the range of 10 to 20 meters. Under the influence of special natural conditions, there are some areas where the evaporation waveguide height is 0, that is, there is no evaporation waveguide at these locations.

[0042] This invention uses SCC as an index to evaluate the goodness of fit between the predicted results and the label values. SCC is a value located in the interval (0, 1), with a value closer to 1 indicating a better goodness of fit. A higher degree of explanation by the independent variable for the dependent variable indicates a stronger ability of the model to fit the trend of the true value. The formula for calculating SCC is:

[0043] Where n represents the number of samples, f(x) i ) represents the model's predicted value, y i This represents the label value in the data.

[0044] To improve the generalization ability of the trained model, the data is shuffled during the training and prediction processes, so the fitted curve will not be exactly the same each time. To more reasonably verify the model's fitting ability, this invention conducted three experiments and took the average value of SCC. The results are shown in Table 2.

[0045] Table 2. Experimental data on the fitting ability of the NWPP-EDH model.

[0046] As shown in Table 3, the average SCC of the NWPP-EDH model was 0.903 in the three experiments, thus proving that the NWPP-EDH model has a good ability to fit the true value. The model has been preliminarily proven to be reasonable and effective.

[0047] The root mean square error (RMSE) is used as the evaluation metric for model accuracy. RMSE reflects the accuracy of the model's prediction of the evaporation waveguide height; a lower RMSE value indicates better prediction accuracy. The formula for calculating RMSE is:

[0048] 80% of the data in the dataset was used as the training set, and 20% as the test set. On the test set, the scatter RMSE of the four models (NWPP-EDH, BYC, MGB, and NPS) was calculated using the prediction results at each grid point and the corresponding grid point label value. The scatter RMSE values ​​for each model are shown in Table 3.

[0049] Table 3 RMSE of NWPP-EDH, BYC, NPS and MGB models

[0050] like Figure 3 As shown, compared to the three theoretical models of evaporation waveguide height—BYC, MGB, and NPS—the NWPP-EDH model has a significantly smaller RMSE. Table 3 shows that the NWPP-EDH model exhibits a 71.8% decrease in RMSE compared to the BYC model, an 87% decrease compared to the MGB model, and a 60.9% decrease compared to the NPS model.

[0051] from Figure 4 It can be seen that the error of the MGB model is particularly large in the target sea area, while the NPS and BYC models have relatively normal prediction effects. This is because the MGB model, as a model widely used in Europe, mainly derives its empirical parameters from actual observations in the European sea area. Meanwhile, the BYC and NPS models, which combine the advanced COARE air-sea algorithm, are weaker in terms of regional generalization and ability to describe actual physical phenomena.

[0052] Figure 4Figures (a)-(d) show the error distribution plots for the NWPP-EDH, MGB, BYC, and NPS models, respectively, representing the error distribution of each model within a 40*48 data grid. The specific calculation method involves averaging the absolute values ​​of the differences between the predicted and true values ​​of each element in each 40*48 matrix in the test set. This error distribution plot reflects the accuracy of the model's predictions; the darker the blue, the lower the model's error. It's clear that the NWPP-EDH model's error distribution plot is almost entirely dark blue, indicating its relatively small prediction error. The NPS model performs well next, with most of the blue being blue and a small portion green. However, NPS exhibits extremely large errors at individual grid points, indicating that while it has some accuracy, its prediction performance is unstable. The BYC model performs only moderately well; a significant portion of the green in the error distribution plot indicates that the model has some error. Consistent with the conclusions drawn from the scatter plot RMSE calculation process above, the MGB model has a large error in this sea area. The large areas of yellow and green in the error distribution map indicate that the model has poor regional prediction performance.

[0053] 4. Experimental Results Compared to traditional BYC, MGB, and NPS evaporation waveguide height prediction models, the NWPP-EDH model of this invention shows a decrease in RMSE of 71.8%, 87%, and 60.9%, respectively. As RMSE is an important indicator for measuring model accuracy, this demonstrates that the NWPP-EDH model proposed in this invention has outstanding performance in predicting the evaporation waveguide height region.

[0054] In this embodiment, a regional prediction model for evaporation duct height in the South China Sea was proposed and verified for the first time by combining numerical weather prediction product data, proving the accuracy of the model.

[0055] Contents not described in detail in this specification are existing technologies known to those skilled in the art. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the height region of evaporating waveguides based on convolutional neural networks, characterized in that, Includes the following steps: Step 1: Obtain two-dimensional numerical forecast grid data of the sea area to be predicted, clean and normalize the data to obtain the feature dataset; Step 2: Construct a numerical weather prediction model for evaporation waveguide height region based on a convolutional neural network, and train the evaporation waveguide height region prediction model on the feature dataset; Step 3: Input the two-dimensional numerical prediction grid data into the evaporation waveguide height region prediction model for prediction; Step 4: Output the evaporation waveguide height value for this sea area; In step 2, the evaporation waveguide height region prediction model is based on the U-Net network structure, which transforms the three-layer downsampling into a two-layer downsampling. It includes an encoding module, a bridge module, and a decoding module. The encoding module is used to convert the data matrix into feature codes. The decoding module is used to recover feature encoding into matrix data; the bridge module is used to connect the encoding module and the decoding module. The encoding module includes three convolutional layers. There is a downsampling operation between the first and second layers, and a downsampling operation between the second and third layers. Each convolutional layer includes two 2*2 convolutional modules. Each convolutional module contains a BN layer to alleviate overfitting and uses the ReLU function as the activation layer. The bridge module includes one layer, which consists of two 2*2 convolutional modules. Each convolutional module contains a BN layer and uses the ReLU function as the activation layer. The decoding module includes three layers, each layer including two 2*2 convolutional modules. Each convolutional module contains a BN layer to alleviate overfitting and uses the ReLU function as the activation layer. An upsampling operation module is set in front of each convolutional module.

2. The method for predicting the height region of evaporating waveguides based on convolutional neural networks as described in claim 1, characterized in that, The evaporation waveguide height region prediction model uses mean square error as the loss function, and the loss function is as follows: ； The above formula is the MSE obtained from the model's output value and label value, where y_pred refers to the model's output value and y_real refers to the evaporation waveguide height label value.

3. The method for predicting the height region of evaporating waveguides based on convolutional neural networks as described in claim 1, characterized in that, Step 3 includes the following specific steps: Step 31: Input two-dimensional numerical weather prediction grid data to obtain the grid data data matrix; Step 32: The data matrix is ​​converted into feature codes by the encoding module, and the feature codes are input into the bridge module; Step 33: The bridge module inputs the feature encoding into the decoding module, the decoding module downsamples the feature encoding, and restores the feature encoding into a data matrix and outputs it; Step 34: Calculate the predicted evaporation waveguide height based on the data matrix.

Images