An evaporation waveguide diagnosis method based on EDH-SimUNet

Abstract

An evaporation duct diagnosis method based on EDH-SimUNet uses a variety of data collected within a region as input to the evaporation duct model and predicts the evaporation duct height above the sea surface in the region. The evaporation duct model is structured using EDH-SimUNet, which uses depthwise separable convolution for downsampling, transposed convolution for upsampling, and skip connections to mitigate gradient vanishing and information loss. EDH-SimUNet avoids the significant discrepancies in diagnostic results caused by the determination of universal functions, stability correction functions, and scale parameters in physical models under inhomogeneous and strongly stable atmospheric conditions. This diagnostic method effectively addresses the high time cost of iterative calculations in traditional physical evaporation duct models when diagnosing duct height. It can also adapt to changing data and environments, automatically adjusting parameters to optimize performance, and easily expand to new datasets and tasks.

Description

Technical Field

[0001] The present invention relates to the cross-integration application field of deep learning (DL) and evaporation waveguide modeling, and in particular to a diagnosis method for evaporation waveguide height based on deep learning. Background Art

[0002] An evaporation duct is a type of atmospheric stratification formed when water vapor evaporates, causing a sharp decrease in atmospheric humidity with altitude. This stratification exhibits an anomalous atmospheric refractive index structure. The evaporation duct model can be applied to the study and prediction of various evaporation-induced duct phenomena, such as lake evaporation, reservoir evaporation, and soil evaporation.

[0003] Based on the Monin-Obukhov similarity theory, researchers have proposed a series of physical models for diagnosing the evaporation duct height. Commonly used evaporation duct models include the PJ model, the Musson-Gauthier-Bruth (MGB) model, the Babin model, the Navy Postgraduate School (NPS) model, and the pseudo-refraction model.

[0004] These physical models based on similarity theory require the calculation of the characteristic length L and characteristic parameter θ in the atmospheric boundary layer. * and q * . On the one hand, some approximations are used in the calculation process of these parameters, and there is a certain degree of uncertainty; on the other hand, the determination of these parameters often requires the use of CPU for very time-consuming iterative calculations. For example, according to actual tests, when using the 24 cores of Intel Xeon Gold 5218 to calculate the evaporation duct height of the South China Sea near the sea surface for one year, it takes several hours of continuous calculation. In the past decade, due to the computing power requirements for matrix operations in fields such as computer graphics and deep learning, the growth rate of GPU computing power has been much faster than that of CPU. For example, the current most powerful civilian-grade GPU, NVIDIA RTX 4090, has a single-precision floating-point performance of about 83TFLOPs, while the most powerful civilian-grade CPU, AMD 7995WX, is about 12TFLOPs.

[0005] To more quickly calculate evaporation duct heights over large ocean areas, a logical research direction is to leverage the powerful computing power of GPUs. Thanks to the computing power provided by GPUs and the massive amounts of data generated by advances in sensor technology, data-driven modeling using deep neural networks has rapidly developed.

[0006] Specifically for near-sea evaporation ducts, advances in satellite assimilation technology have enabled the accumulation of vast amounts of weather data in various databases, such as the ERA5 dataset provided by the European Centre for Medium-Range Weather Forecasts (ECMWF) and reanalysis data from the National Center for Atmospheric Research (NCAR). This has made it possible to diagnose evaporation duct heights using deep neural networks. Using this technology, high-temporal and spatial resolution evaporation duct diagnosis can be completed in just seconds.

[0007] In the existing technology, the research on the prediction of evaporation duct height mainly includes:

[0008] 1. A Chinese patent application with publication number CN112711899A proposes a “fusion prediction method for evaporation duct height”, which includes the following steps: (1) obtaining the true value of the evaporation duct height; (2) predicting the evaporation duct model height; (3) model sensitivity analysis; (4) least squares support vector machine regression modeling: based on the sensitivity analysis results, in different environmental intervals, with the obtained meteorological and hydrological observation data and the evaporation duct model height prediction value as input and the evaporation duct height true value as output, regression is performed to obtain a nonlinear mapping relationship between the input and output, and a fusion prediction model of multiple evaporation duct model prediction heights is established; (5) evaporation duct height prediction.

[0009] The main drawbacks of this technology are: the support vector machine algorithm used in this method has shortcomings, such as: (1) the support vector machine is not suitable for processing large-scale data. When the data has many features, the training time will increase significantly. (2) when faced with small data sets, the support vector machine is sensitive to noise, which makes the model prone to overfitting.

[0010] 2. Chinese patent application with publication number CN116559977A proposes "A method for predicting the height of an NPS-GRYANIK20 evaporation duct based on a BP neural network". It adopts a method of dividing the environmental intervals based on the NPS-GRYANIK20 evaporation duct model, constructs a BP neural network prediction model based on a predator algorithm in different environmental intervals, and realizes accurate prediction of the evaporation duct height. Specifically, the method includes: obtaining the true value of the evaporation duct height, dividing the environmental intervals based on the NPS-GRYANIK20 evaporation duct model, and predicting the height of the evaporation duct based on an improved BP neural network prediction model.

[0011] The main drawbacks of this technique are that the predator algorithm optimization process employed by this method is divided into multiple stages, involving complex computational steps. This can lead to high computational costs, especially when training neural networks on large datasets. Furthermore, the predator algorithm is not applicable to all types of problems, and the literature lacks the necessary experiments to verify its applicability.

[0012] 3. The Chinese patent application with publication number CN116579225A proposes a “method for regional prediction of evaporation duct height based on convolutional neural network.” This method combines the grid data of numerical forecast products to construct a regional prediction model of evaporation duct height based on convolutional neural network (NWPP-EDH).

[0013] The main drawback of this technology is that it uses classic 2×2 convolutional layers in all network modules. Smaller convolution kernels have smaller receptive fields, resulting in poor model fitting. This is demonstrated by the NWPP-EDH RMSE indicator exceeding 1 in the experiments. Furthermore, using a typical convolutional approach results in a larger network size, increasing the risk of overfitting, slowing down the model's output speed, and placing higher demands on the hardware of the device running the model. Summary of the Invention

[0014] To address the high time overhead caused by the iterative calculation required in the current evaporation waveguide model for diagnosing waveguide height, the present invention provides a diagnostic solution based on a UNet-type deep neural network, hereinafter referred to as EDH-SimUNet.

[0015] EDH-SimUNet is a UNet-type network consisting of downsampling, upsampling, and skip connections to mitigate vanishing gradients. EDH-SimUNet uses depthwise separable convolutions for downsampling and transposed convolutions for upsampling. Based on prior knowledge, distant weather conditions are not very helpful for diagnosing evaporation duct height, eliminating the need for a costly attention mechanism.

[0016] Experimental testing shows that EDH-SimUNet is an order of magnitude faster than traditional physical models in diagnosis speed. Furthermore, compared to traditional machine learning methods such as XGBoost and random forests, EDH-SimUNet also has advantages in diagnosis speed and performs better in terms of error.

[0017] The large-area evaporation waveguide diagnostic method based on EDH-SimUNet of the present invention comprises the following steps:

[0018] Step 1: Filter the temperature, pressure, humidity, wind speed, and sea surface temperature data near the sea surface in the area from the analysis data set as the features of the training data;

[0019] Step 2: Use the evaporation duct model to calculate the evaporation duct height corresponding to the air temperature, air pressure, humidity, wind speed, and sea surface temperature data, which is to create labels for the training data;

[0020] Step 3: Build the EDH-SimUNet neural network and divide the dataset into a training set and a test set. The training set is further randomly divided into a training set and a validation set. Adjust the parameters based on the performance of the validation set.

[0021] Step 4: Using the test set, calculate indicators such as RMSE, MAE, MedAE, and RS to evaluate and verify the model performance.

[0022] This method EDH-SimUNet uses depth-wise separable convolution operations for downsampling, transposed convolution for upsampling, and skip connections to alleviate the problems of gradient vanishing and information loss.

[0023] This method effectively addresses the high time cost of iterative calculations in traditional physical evaporation duct models when diagnosing duct height. Furthermore, the method adapts to changing data and environments, automatically adjusting parameters to optimize performance, and easily scales to new datasets and tasks. Therefore, EDH-SimUNet has the potential to avoid the significant discrepancies in diagnostic results caused by the determination of universal functions, stability correction functions, and scaling parameters in physical models under inhomogeneous and strongly stable atmospheric conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 Schematic diagram of the large-area evaporation waveguide diagnosis process based on EDH-SimUNet;

[0025] Figure 2 Schematic diagram of the comparison between EDH-SimUNet waveguide diagnosis results and true values ​​(the first row is the true value, the second row is the EDH-SimUNet diagnosis results);

[0026] Figure 3 It is the top-level structure of EDH-SimUNet;

[0027] Figure 4 The underlying structure of the input and output processing module in EDH-SimUNet;

[0028] Figure 5 This is the underlying structure of the downsampling module in EDH-SimUNet;

[0029] Figure 6 This is the underlying structure of the upsampling module in EDH-SimUNet; DETAILED DESCRIPTION

[0030] The present invention provides an evaporation duct diagnosis method based on EDH-SimUNet, which uses the collected data of air temperature, air pressure, humidity, wind speed and sea surface temperature in the region as the input of the evaporation duct model to predict the evaporation duct height on the sea surface in the region.

[0031] The structure of the evaporation waveguide model is an EDH-SimUNet neural network. The steps of constructing the EDH-SimUNet neural network model include:

[0032] Step 1: Filter the air temperature, pressure, humidity, wind speed, and sea surface temperature data near the sea surface in the region from the reanalysis dataset as the features of the training data;

[0033] Step 2: Use the evaporation duct model to calculate the evaporation duct height corresponding to the air temperature, air pressure, humidity, wind speed, and sea surface temperature data, which is to create labels for the training data;

[0034] Step 3: Build the EDH-SimUNet neural network; divide the dataset into a training set and a test set; the training set is further randomly divided into a training set and a validation set; adjust the parameters of the EDH-SimUNet neural network based on the performance of the validation set;

[0035] Step 4: Evaluate and verify the performance of the evaporation duct model using the test set;

[0036] In step 3, the framework of the EDH-SimUNet neural network is as follows in the order of data processing: input processing module, downsampling module A, downsampling module B, downsampling module C, downsampling module D, upsampling module D, upsampling module C, upsampling module B, upsampling module A and output processing module; each module is jump-connected, that is:

[0037] The output of the input processing module is used as the input of the downsampling module A and the upsampling module A;

[0038] The output of downsampling module A is used as the input of downsampling module B and upsampling module B;

[0039] The output of downsampling module B is used as the input of downsampling module C and upsampling module C;

[0040] The output of the downsampling module C is used as the input of the downsampling module D and the upsampling module D;

[0041] The output of the downsampling module D is used as the input of the upsampling module D;

[0042] The output of the upsampling module A is used as the input of the output processing module;

[0043] The input of the input processing module is the collected air temperature, pressure, humidity, wind speed and sea surface temperature data;

[0044] The output of the output processing module is the prediction result;

[0045] The input processing module maps the input features to the shape required by the downsampling module;

[0046] The output processing module only contains a 1×1 convolution layer to ensure that the final output of the evaporation waveguide model has an appropriate number of channels;

[0047] In the four downsampling modules, each time the data passes through a downsampling module, the number of channels, i.e., the feature dimension, doubles, and the length and width are halved;

[0048] In the four upsampling modules, each time the data passes through an upsampling module, the number of channels, i.e., the feature dimension, is halved, and the length and width are doubled.

[0049] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0050] The process of the large-area evaporation waveguide diagnostic method based on EDH-SimUNet in this example is as follows Figure 1 shown.

[0051] First, a training dataset must be prepared. This dataset must include near-surface air temperature, pressure, humidity, wind speed, and sea surface temperature data. This example uses the EAR5 reanalysis data from the European Centre for Medium-Range Weather Forecasts (ECMWF) as an example for demonstration and experimentation. EAR5 meteorological data includes 10-meter wind speed, 2-meter dew point temperature, 2-meter temperature, mean sea surface pressure, and sea surface temperature data. The temporal resolution is 1 hour, and the spatial resolution is 0.25 degrees (approximately 30 km). Since the evaporation duct physical model requires specific humidity, the 2-meter specific humidity data is calculated based on the mean sea surface pressure and 2-meter dew point temperature. Next, the evaporation duct physical model calculates the corresponding duct height based on the known mean sea surface pressure, 10-meter wind speed, 2-meter temperature, sea surface temperature, and 2-meter specific humidity. This provides both the dataset features and the corresponding labels. To ensure that EDH-SimUNet has excellent generalization capabilities and can make more accurate diagnoses, the dataset should, in principle, be as large as possible. In the experiment, we prepared a total of 31 years of meteorological data, with a specific time span from 1993 to 2023, and a spatial range from E107° to E122.75° east-west and N4.25° to N20° north-south.

[0052] Before starting training, the data is preprocessed.

[0053] First, we need to remove outliers from the dataset. Even when using reanalysis data, a small number of NaN values ​​are unavoidable. Second, to ensure efficient model training, we need to normalize the data. This means processing the data at each grid point according to the following formula.

[0054]

[0055] Here, x, μ, and σ are the features and their corresponding mean and variance, respectively. It should be noted that preserving μ and σ is necessary. When EDH-SimUNet makes predictions, the meteorological elements at the time of prediction should be processed according to the μ and σ of the training samples, rather than according to the mean and variance of the meteorological elements themselves at the time of testing.

[0056] After preparing the data set, we can start building the network structure.

[0057] like Figure 3 As shown in the figure, the overall framework of EDH-SimUNet mainly includes input and output processing modules, downsampling modules and upsampling modules.

[0058] The design of the input processing module is as follows Figure 4 As shown in Figure 2, its main function is to map the input features to the shape required by the downsampling module. The output processing module only contains a 1×1 convolution layer to ensure that the final output of the model has an appropriate number of channels.

[0059] According to experimental tests, the number and structure of downsampling and upsampling modules are as follows Figure 5 and Figure 6 The design will enable EDH-SimUNet to have faster reasoning speed and more accurate diagnosis results.

[0060] The downsampling part contains four downsampling modules. Each time the data passes through a downsampling module, the number of channels (feature dimension) doubles and the length and width are halved.

[0061] The upsampling part is the inverse process of downsampling and includes four upsampling modules. Each time the data passes through an upsampling module, the number of channels (feature dimensions) is halved and the length and width are doubled.

[0062] Figure 5 The specific design of each downsampling module is shown. In general, the downsampling module needs to include a maximum pooling layer to reduce the resolution of the data, a depth-wise separable convolution layer to extract features, a batch normalization layer to speed up the model convergence, and a ReLU layer for nonlinear mapping. There is no special arrangement for the data flow in the downsampling part, that is, the data flows through each layer in order from top to bottom. Figure 3 As shown, starting from the input processing module and ending at the downsampling module C, after each module processes the data, it is necessary to temporarily store an output result so as to be subsequently input into the corresponding upsampling module.

[0063] The structure of the upsampling module is as follows Figure 6As shown, the design of each module is very similar to the downsampling part. The difference is that the upsampling part replaces the maximum pooling layer with a transposed convolution layer, so that the width and height of the data are gradually restored. The convolution layer of the upsampling part gradually restores the feature dimension of the data. Because the upsampling module has two inputs, the data flow at this time requires special arrangement. Taking upsampling module D as an example, the output from downsampling module D is processed by the transposed convolution layer, and the number of channels is halved while the width and height are doubled. Next, the input of the transposed convolution layer is spliced ​​with the result temporarily stored in the downsampling module C in the channel dimension and input into the first convolution layer. There is no special arrangement for the subsequent data flow; it can flow through each layer from top to bottom.

[0064] In order to demonstrate the feasibility of EDH-SimUNet in waveguide diagnosis, the present invention trained a SOTA (State of the Art) model according to the above process for evaluation. The training set used in training is the reanalysis data from 1993 to 2022 provided by ERA5. Specifically, during training, the training set is randomly divided into a validation set and a training set with a ratio of 20% as the validation set. In order to measure the generalization ability of the model, the reanalysis data of 2023 is used as a test set. In addition, traditional machine learning algorithms, such as Random Forest (RF) and XGBoost, were selected as control groups for evaluation. The indicators involved in the evaluation are Root Mean Square Error (RMSE), Mean Absolute Error (Meanabsolute Error), Median Absolute Error (MedAE) and Coefficient of Determination (RS). The test platform was an AMD Ryzen 5900X CPU, an RTX 3080 laptop GPU, 32GB of DDR4 3200MHz dual-channel memory, and Windows 11 23H2 operating system. Table 1 summarizes the performance of the four models on the test set. As can be seen from the table, EDH-SimUNet significantly outperforms traditional machine learning models in all metrics. EDH-SimUNet achieves extremely low RMSE and RS, indicating high accuracy in waveguide height diagnosis and full suitability for practical production applications. Furthermore, the time column in the table represents the time required for each model to diagnose all the data in the test set. Although EDH-SimUNet is a deep neural network, its inference speed is comparable to traditional machine learning methods and, thanks to hardware and software optimizations, is even faster than these models. Compared to the classic UNet, which uses a general convolutional approach, EDH-SimUNet has fewer parameters, faster computation speed, and more accurate diagnostic results. In order to intuitively understand the diagnostic effect of EDH-SimUNet, the present invention draws Figure 2 . Figure 2 The first row shows the waveguide distribution calculated using the evaporation waveguide model, and the second row shows the waveguide distribution diagnosed by EDH-SimUNet. Comparing the two rows, we can see that the waveguide distribution obtained by EDH-SimUNet is almost exactly the same as the result calculated by the evaporation waveguide model.

[0065]

[0066] Table 1 Model diagnostic performance evaluation

[0067] (↓ means the smaller the better, ↑ means the bigger the better)

Claims

1. An evaporation duct diagnosis method based on EDH-SimUNet, which uses the collected data of air temperature, pressure, humidity, wind speed and sea surface temperature in the region as the input of the evaporation duct model to predict the evaporation duct height on the sea surface in the region; The structure of the evaporation waveguide model is an EDH-SimUNet neural network. The steps of constructing the EDH-SimUNet neural network model include: Step 1: Filter the air temperature, pressure, humidity, wind speed, and sea surface temperature data near the sea surface in the region from the reanalysis dataset as the features of the training data; Step 2: Use the evaporation duct model to calculate the evaporation duct height corresponding to the air temperature, air pressure, humidity, wind speed, and sea surface temperature data, which is to create labels for the training data; Step 3: Build the EDH-SimUNet neural network; divide the dataset into a training set and a test set; the training set is further randomly divided into a training set and a validation set; adjust the parameters of the EDH-SimUNet neural network based on the performance of the validation set; Step 4: Evaluate and verify the performance of the evaporation duct model using the test set; In step 3, the framework of the EDH-SimUNet neural network is as follows in the order of data processing: input processing module, downsampling module A, downsampling module B, downsampling module C, downsampling module D, upsampling module D, upsampling module C, upsampling module B, upsampling module A and output processing module; The order of data processing in the input processing module is as follows: a 3×3 depth-wise separable convolutional layer with 6 input channels, 64 output channels, and 1 padding; a batch normalization layer; a ReLU layer; a 3×3 depth-wise separable convolutional layer with 64 input channels, 64 output channels, and 1 padding; a batch normalization layer; and a ReLU layer. The order of data processing in the output module is as follows: a 1×1 convolutional layer with 64 input channels and 1 output channel; The data processing order in the downsampling module A is as follows: 2×2 max pooling layer; 3×3 depthwise separable convolution layer with 64 input channels, 128 output channels and 1 padding; batch normalization layer; ReLU layer; 3×3 depthwise separable convolution layer with 128 input channels, 128 output channels and 1 padding; batch normalization layer; ReLU layer; The data processing order in the downsampling module B is as follows: 2×2 max pooling layer; 3×3 depthwise separable convolution layer with 128 input channels, 256 output channels and 1 padding; batch normalization layer; ReLU layer; 3×3 depthwise separable convolution layer with 25 input channels, 256 output channels and 1 padding; batch normalization layer; ReLU layer; The data processing order in the downsampling module C is as follows: 2×2 max pooling layer; input channel is 256; 3×3 depthwise separable convolution layer with 512 output channels and 1 padding; batch normalization layer; ReLU layer; input channel is 512; 3×3 depthwise separable convolution layer with 512 output channels and 1 padding; batch normalization layer; ReLU layer; The data processing order in the downsampling module D is as follows: 2×2 max pooling layer; 3×3 depthwise separable convolution layer with 512 input channels, 1024 output channels and 1 padding; batch normalization layer; ReLU layer; 3×3 depthwise separable convolution layer with 1024 input channels, 1024 output channels and 1 padding; batch normalization layer; ReLU layer; The data processing order in the upsampling module A is as follows: 2×2 transposed convolution layer with 128 input channels, 64 output channels, and stride 2; 3×3 depthwise separable convolution layer with 128 input channels, 64 output channels, and padding 1; batch normalization layer; ReLU layer; 3×3 depthwise separable convolution layer with 64 input channels, 64 output channels, and padding 1; batch normalization layer; ReLU layer; The data processing order in the upsampling module B is as follows: 2×2 transposed convolution layer with 256 input channels, 128 output channels, and stride 2; 3×3 depthwise separable convolution layer with 256 input channels, 128 output channels, and padding 1; batch normalization layer; ReLU layer; 3×3 depthwise separable convolution layer with 128 input channels, 128 output channels, and padding 1; batch normalization layer; ReLU layer; The data processing order in the upsampling module C is as follows: 2×2 transposed convolution layer with 512 input channels, 256 output channels, and stride 2; 3×3 depthwise separable convolution layer with 512 input channels, 256 output channels, and padding 1; batch normalization layer; ReLU layer; 3×3 depthwise separable convolution layer with 256 input channels, 256 output channels, and padding 1; batch normalization layer; ReLU layer; The data processing order in the upsampling module D is as follows: 2×2 transposed convolution layer with an input channel of 1024, an output channel of 512, and a stride of 2; 3×3 depth-wise separable convolution layer with an input channel of 1024, an output channel of 512, and padding of 1; batch normalization layer; ReLU layer; 3×3 depth-wise separable convolution layer with an input channel of 512, an output channel of 512, and padding of 1; batch normalization layer; ReLU layer.

2. The evaporation waveguide diagnostic method based on EDH-SimUNet according to claim 1 is characterized by In the step 3, before the training set is used for evaporation waveguide model training, the training set is first standardized and the mean and variance are saved; The test set is normalized by mean and variance before being used for testing.

3. The evaporation duct diagnostic method based on EDH-SimUNet according to claim 2 is characterized in that the training data is standardized, that is, the data at each grid point is processed according to the following formula: ， in, 、 and are the features and their corresponding means and variances, respectively.

4. The evaporation duct diagnostic method based on EDH-SimUNet according to claim 1, characterized in that in step 4, the evaporation duct model performance is evaluated and verified by calculating the root mean square error (RMSE), mean absolute error (MAE), median absolute error (MedAE), and / or relative strength index (RS).

5. The evaporation waveguide diagnostic method based on EDH-SimUNet according to claim 1 is characterized in that The output of the input processing module is used as the input of the downsampling module A and the upsampling module A; The output of downsampling module A is used as the input of downsampling module B and upsampling module B; The output of downsampling module B is used as the input of downsampling module C and upsampling module C; The output of the downsampling module C is used as the input of the downsampling module D and the upsampling module D; The output of the downsampling module D is used as the input of the upsampling module D; The output of the upsampling module A is used as the input of the output processing module; The input of the input processing module is the collected air temperature, pressure, humidity, wind speed and sea surface temperature data; The output of the output processing module is the prediction result; The input processing module maps the input features to the shape required by the downsampling module; The output processing module only contains a 1×1 convolution layer to ensure that the final output of the evaporation waveguide model has an appropriate number of channels; In the four downsampling modules, each time the data passes through a downsampling module, the number of channels, i.e., the feature dimension, doubles, and the length and width are halved; In the four upsampling modules, each time the data passes through an upsampling module, the number of channels, i.e., the feature dimension, is halved, and the length and width are doubled.

Images