Supervised learning method based on improved wind field direction loss function in deep learning

Abstract

The invention relates to a supervised learning method based on an improved wind field direction loss function in deep learning, and the method comprises the steps: redefining a wind direction error between a correction result and reference data, and controlling the range of the wind direction error between-180 degrees and 180 degrees; and meanwhile, the wind direction data is normalized before the model is input, and the normalized error range is redefined to be between [-1, 1]. The loss function is used for calculating the redefinition error of the normalized data. Experiments show that the improved loss function can better reflect the real error between the correction data and the reference data, thereby ensuring that the correction result of training quickly approaches the reference data.

Description

Technical Field

[0001] Sea surface wind field correction based on deep learning Background Art

[0002] Extreme winds at sea are closely related to the atmosphere and ocean, and their formation and development have a potential correspondence with atmospheric and oceanic variables. Atmospheric and oceanic variables such as temperature and humidity can reflect the characteristics of extreme winds, and the characteristics of extreme winds themselves are often defined by these variables. With the occurrence of extreme winds, these variables and the characteristics of winds above force 8 exhibit inherent patterns. However, the current understanding of these patterns remains inaccurate, leading to inaccurate monitoring of offshore winds. For many years, extreme winds have had a significant impact on coastal countries, attracting widespread attention, and the accumulation of historical data has provided a foundation for research. Although direct observation of atmospheric and oceanic variables related to extreme winds is difficult, the development of observational methods, such as satellite remote sensing data, has become a powerful data support, providing a rich source of information for extreme wind monitoring research.

[0003] At the same time, the powerful advantage of deep learning and machine learning lies in discovering patterns in historical data, providing a key approach to extreme wind monitoring methods. Geographic location has a significant relationship between wind direction and extreme winds. For example, high wind speeds in Gansu Province are often accompanied by northwesterly winds. To improve the accuracy of offshore wind field information, it is important to correct the data based on model wind direction factors. By integrating historical extreme wind pattern data with ocean observation data, building monitoring models can help improve the accuracy of extreme winds. Using machine learning techniques, we can mine correlations from multi-dimensional data to provide more accurate prediction and correction methods for extreme wind monitoring. Summary of the Invention

[0004] The technical problem solved by the present invention is that when the root mean square error formula is directly used to train the wind direction, the correction error of the trained wind direction will increase. For example, the wind direction data corrected by the training network is 355°, and the actual reference value is 0°. If the MSE formula is directly used, the value calculated by the error function will be very large, resulting in incorrect network parameter updates.

[0005] The technical solution of the present invention is characterized in that it comprises the following steps:

[0006] 1) The wind direction error range is [0°~359°). The constraint condition is to redefine the wind direction error between the correction result and the reference data and control it within the range of [-180°~180°).

[0007] 2) The wind direction data is normalized, and the normalized error range is redefined to between -1 and 1. The loss function is to calculate the error between the normalized reference wind direction and the output value of the wind direction correction model.

[0008] 3) Load the constraints into the loss function to optimize the objective function. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 Step Flowchart (with attached abstract)

[0010] Figure 2 Comparison of Correction Effects of Offshore Wind Direction Data DETAILED DESCRIPTION

[0011] Constraints are designed by analyzing the characteristics of wind direction data. The key to these constraints is to ensure the correct correspondence between the corrected data and the actual reference data, so that the wind direction correction loss function can effectively reflect the actual error between the corrected data and the reference data, thereby ensuring that the training correction results quickly approach the reference data.

[0012] The technical solution of the present invention is characterized in that it comprises the following steps:

[0013] 1) The wind direction error in the numerical model data ranges from 0° to 359°. The constraint condition is to redefine the wind direction error ΔS between the correction result and the reference data and control it within the range of -180° to 180°.

[0014]

[0015] Among them, S NL-Unet is the corrected wind direction of the network model, S o is the actual reference wind direction data. ΔS is the error between the corrected and reference data.

[0016] 2) Before inputting the data into the model, the wind direction data must be normalized to between [0, 2]. The normalization formula is as follows:

[0017]

[0018] Where S is the wind direction data, S min is the minimum value in the wind direction data, S max is the maximum value in the wind direction data.

[0019] The wind direction data is normalized, and the normalized error S d The range is redefined to [-1, 1]:

[0020]

[0021] in is the normalized reference wind direction at time t, is the output value of the wind direction correction model at time t, and the loss function is to calculate the error between the normalized reference wind direction and the output value of the wind direction correction model:

[0022]

[0023] 3) Data correction is performed based on the wind direction data in the numerical model data. Using the U-Net network as the backbone network, the model loads the constraints into the loss function to optimize the objective function.

[0024]

[0025] 4) Use the wind direction data from January 1, 2021 to December 31, 2021 to create training and test sets, and preprocess the data. The main steps of preprocessing include outlier processing and data normalization.

[0026] The backbone network is built and trained under the torch framework. The backbone network of the model is U-Net network, and the initial learning rate of the model is 1*10 -3 , BatchSize is 40, training Epoch is 60, SGD optimization algorithm is used, execute the above steps 1) 2) 3) in sequence and save the trained weight parameters.

[0027] 5) Use the test set to test the accuracy of the network output corrected wind direction data. The test results of the wind direction correction for each month of the year are as follows: Figure 2 Experiments show that the proposed wind direction data constraints and loss function optimization significantly reduce the error of wind direction pattern data.

Claims

1. The supervised learning method based on the improved wind field direction loss function in deep learning is characterized by The following steps are involved: 1) The wind direction error range is [0°~359°), and the constraint condition is to redefine the wind direction error between the correction result and the reference data and control it within the range of [-180°~180°); 2) The wind direction data is normalized, and the normalized error range is redefined to [-1, 1]. The loss function is to calculate the error between the normalized reference wind direction and the output value of the wind direction correction model; 3) Load the constraints into the loss function to optimize the objective function.

2. The supervised learning method based on the improved wind field direction loss function in deep learning according to claim 1, characterized in that: Step 1) The wind direction error range is [0°~359°), and the constraint condition is to redefine the wind direction error between the correction result and the reference data and control it within the range of [-180°~180°).

3. The supervised learning method based on the improved wind field direction loss function in deep learning according to claim 1, characterized in that: In step 2), the wind direction data is normalized, and the normalized error range is redefined to [-1, 1]. The loss function is to calculate the error between the normalized reference wind direction and the output value of the wind direction correction model.

Images