UKAN-based hour-by-hour maximum wind forecasting method within 24 hours

Abstract

The invention relates to the technical field of meteorology, in particular to a UKAN-based hour-by-hour maximum wind forecasting method within 24 hours, which comprises the following steps of: firstly, acquiring historical real-time maximum wind speed data and EC fine grid numerical model forecasting product data in an integral point period, respectively establishing ground data channel data and high-altitude data channel data, and carrying out scaling preprocessing on input data; and then constructing a deep learning maximum wind speed forecasting model based on the UKAN, determining a loss function of model training, training the model by using the data, and finally calculating by using the trained deep learning maximum wind speed forecasting model to obtain an hour-by-hour maximum wind speed forecasting field in 0-24 hours in the future. By designing the UKAN maximum wind speed forecasting model and constructing the site data loss function, the forecasting precision of 0-24 h hourly maximum wind speed is improved, the application prospect is good, and more accurate forecasting service is provided for disaster prevention and reduction.

Description

Technical Field

[0001] The invention relates to the field of meteorological technology, and in particular to a method for forecasting hourly extreme wind speeds within 24 hours based on UKAN. Background Art

[0002] Wind has a wide impact on human life. On the one hand, as a renewable energy source, it can make great contributions to human society through reasonable use and development. On the other hand, as a vital meteorological factor, wind is easily affected by natural factors such as topography, terrain, temperature, and air pressure, showing strong randomness, intermittency, and volatility. It has a very significant impact on production safety activities. Strong wind disasters are difficult to accurately predict and pose a huge threat to people's production and life. Understanding the changing laws and evolution trends of strong winds and making rapid and accurate forecasts of strong winds are of great significance for disaster prevention and relief work.

[0003] At present, wind speed prediction methods can be roughly divided into physical methods and statistical methods. Physical methods, such as numerical weather forecasting, mainly consider the physical factors that affect the wind field (such as terrain characteristics, air pressure and ambient temperature, etc.) and the physical interaction laws between them to predict the wind field. This type of method requires a clear understanding and reproduction ability of the actual physical process, but due to the imperfection of the model's physical parameterization scheme and the uncertainty of many parameters, there are large errors in the forecast of near-surface wind fields. Statistical methods usually use a large amount of historical data to construct a statistical model for prediction. In recent years, some scholars have tried to revise the results of numerical weather forecast models based on machine learning methods. The forecast model established by the revised model based on machine learning methods can capture nonlinear wind speed changes, show good performance in wind speed forecasting, and achieve a refined forecast of wind speed.

[0004] There are few research results on the forecast of extreme wind speed near the ground. Here, extreme wind refers to the maximum instantaneous wind speed value that occurs in a certain period of time, and instantaneous wind speed refers to the average wind speed of 3 seconds. The destructive power of extreme wind speed near the ground is very large, and it is very meaningful to accurately forecast it. There are many factors that affect the change of extreme wind speed, and its nonlinear change characteristics are very obvious. To accurately forecast the extreme wind speed, it is essential to reasonably select the forecasting factors with great influence and build a high-performance forecasting model. Summary of the invention

[0005] The purpose of the present invention is to provide a method for forecasting extreme wind hourly within 24 hours based on UKAN, which can accurately forecast the extreme wind speed by fixed point and quantitative forecast.

[0006] To achieve the above object, the present invention provides a method for forecasting hourly maximum wind speed within 24 hours based on UKAN, comprising the following steps:

[0007] Step 1: Obtain the historical real-time maximum wind speed data and EC fine grid numerical model forecast data at the ground meteorological station at every hour;

[0008] Step 2: Construct a sample dataset of hourly maximum wind speed and EC numerical model forecast;

[0009] Step 3: Construct multi-channel data of ground data and high-altitude data;

[0010] Step 4: Scaling the model input data;

[0011] Step 5: Build a deep learning maximum wind speed forecast model based on UKAN;

[0012] Step 6: Construct the loss function for model training;

[0013] Step 7: Training the deep learning maximum wind speed forecast model based on the data and the loss function;

[0014] Step 8: Use the trained deep learning maximum wind speed forecast model to calculate the hourly maximum wind speed forecast field for the next 0-24 hours.

[0015] Optionally, the historical actual maximum wind speed data of the ground meteorological station is the maximum wind speed data of the automatic weather station in the past hour; the EC numerical model forecast data is the rainfall forecast fields of 51 ensemble members of the EC ensemble forecast, and the U and V wind field data of the EC fine grid 500hPa, 850hPa and 925hPa levels.

[0016] Optionally, to forecast precipitation in the next 24 hours, use a 12-36h forecast field, i.e. the 12-hour forecast field of the model corresponds to the start time of the actual forecast.

[0017] Optionally, the multi-channel data of ground data and high-altitude data include a ground data input channel and an high-altitude data input channel. The ground data input channel includes two elements, the average wind at 10 meters and the gusts at 10 meters in the past 3 hours. First, the first three forecast fields are taken, namely 3h\6h\9h, and then the first two forecast time periods of the corresponding model forecast time period for the forecast period are taken; the high-altitude data input channel includes three levels of wind speed forecast fields calculated from U and V wind fields of 500hPa, 850hPa, and 925hPa, and each level takes the first two forecast time periods of the corresponding model forecast time period for the forecast period; a total of 16 channel data.

[0018] Optionally, the deep learning maximum wind speed forecast model described in step 5 is a neural network architecture based on Unet combined with a Kolmogorov-Arnold network.

[0019] Optionally, the loss function expression for model training is as follows:

[0020]

[0021] Among them, y s is the site observation value, is the site prediction value, w s is the weighting coefficient, and N is the number of sites in the area.

[0022] The present invention provides a method for forecasting hourly maximum wind speed within 24h based on UKAN. First, the historical actual maximum wind speed data and EC fine grid numerical model forecast product data of the hourly period are obtained, and the ground data channel data and the high-altitude data channel data are respectively established, and the input data is pre-processed for scaling; then a deep learning maximum wind speed forecast model based on UKAN is constructed, the loss function of the model training is determined, the model is trained using data, and finally the future 0-24h hourly maximum wind speed forecast field is calculated using the trained deep learning maximum wind speed forecast model. The present invention improves the forecast accuracy of the 0-24h hourly maximum wind speed by designing the UKAN maximum wind speed forecast model and the construction of the site data loss function, has good application prospects, and provides more accurate forecast services for disaster prevention and mitigation. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0024] Figure 1 The present invention is a flowchart of the steps of a method for forecasting extreme wind hourly within 24 hours based on UKAN.

[0025] Figure 2 It is a schematic diagram of the execution flow of a specific embodiment of the present invention. DETAILED DESCRIPTION

[0026] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and should not be construed as limiting the present invention.

[0027] See also Figure 1 The present invention provides a method for forecasting hourly maximum wind speed within 24 hours based on UKAN, comprising the following steps:

[0028] S1: Obtain the historical real-time maximum wind speed data of the ground meteorological station and the EC fine grid numerical model forecast data;

[0029] S2: Construct a sample dataset of hourly maximum wind speed and EC numerical model forecast;

[0030] S3: Construct multi-channel data of ground data and high-altitude data;

[0031] S4: Scaling the model input data;

[0032] S5: Construct a deep learning maximum wind speed forecast model based on UKAN;

[0033] S6: Construct loss function for model training;

[0034] S7: Training the deep learning maximum wind speed forecast model based on the data and the loss function;

[0035] S8: Use the trained deep learning maximum wind speed forecast model to calculate the hourly maximum wind speed forecast field for the next 0-24 hours.

[0036] The historical actual maximum wind speed data of the ground meteorological station are the maximum wind speed data of the automatic weather station in the past hour; the EC numerical model forecast data are the rainfall forecast fields of 51 ensemble members of the EC ensemble forecast, and the U and V wind field data of the EC fine grid 500hPa, 850hPa and 925hPa levels.

[0037] To forecast precipitation in the next 24 hours, a 12-36h forecast field is used. That is, the 12-hour forecast field of the model corresponds to the start time of the actual forecast.

[0038] The following is further described in conjunction with specific embodiments (the specific process is as follows Figure 2 shown):

[0039] 1. Numerical model data

[0040] EC fine grid 10m gust, hourly wind, 1000hPa, 850hPa, 925hPa UV wind field. Data time range: January 2020-April 2024.

[0041] 2. Site live data

[0042] Data spatial range: Maximum wind data of the past hour at weather stations within the range of 19.5-28°N, 104-117°E. Time range: January 2020-April 2024.

[0043] 3. Forecast model design

[0044] Since the time resolution of the EC fine grid model forecast field is 3 hours, in order to reduce the time resolution to 1 hour, the 0-24h hourly revised forecast adopts a strategy of segmented and mixed construction of training samples, that is, 8 time periods (0-3h\4-6h\7-9h\10-12\13-15h\16-18h\19-21h\22-24h) are used, and 3 1-hour revised forecasts are completed in each time period. At the same time, since the EC fine grid model's forecast of wind within 3 days does not show a significant decrease with the increase of forecast time, the samples of 8 time periods are mixed to form a large training sample, which can also increase the sample size of model training.

[0045] 4. Model channel construction

[0046] Due to the lag of EC model data, in actual forecasting, the precipitation conditions for the next 24 hours are often predicted using the 12-36h forecast field. That is, the 12-hour forecast field of the model corresponds to the start time of the actual forecast. Therefore, the model forecast timeliness corresponding to the 8 forecast periods is:

[0047] 8 forecast time periods 0-3h 4-6h 7-9h 10-12h Model forecast timeliness 15h 18h 21h 24h 8 forecast time periods 13-15h 15-18h 19-21h 22-24h Model forecast timeliness 27h 30h 33h 36h

[0048] (1) Ground data input channel

[0049] The ground data includes two elements: the average wind of 10 meters (recorded as: 10m_wind, the same below) and the gust of 10 meters in the past 3 hours (recorded as: 10m_gust, the same below). First, take the first three forecast fields (3h\6h\9h), and then take the first two forecast time periods corresponding to the model forecast time period to be forecasted. For example, the forecast period to be forecasted is 7-9h, and the model forecast time period corresponding to this period is 21h. The selected model forecast time periods are 18h and 21h as the input of this forecast period. In summary, the ground input channel contains 2*(3+2)=10 channels in total.

[0050] (2) Input channel for high-altitude data construction

[0051] The high-altitude data includes three levels of wind speed forecast fields calculated from the U and V wind fields of 500hPa, 850hPa, and 925hPa. Each level takes the first two forecast time periods of the corresponding model forecast time period to be forecasted. For example, if the forecast period is 7-9h, the model forecast time period corresponding to this period is 21h, and the selected model forecast time period is 18h and 21h as the input of this forecast period. Therefore, a total of 6 channel data at three levels are used as the input of the model.

[0052] (3) Model input preprocessing

[0053] In the input of the model, the range of the ground data channel and the high-altitude data channel are both intercepted from the data in the range of 19.5-28°N and 104-117°E of the model. Since the resolutions of the ground and high-altitude data of the EC model are (0.125*0.125 and 0.25*0.25) respectively, and the spatial resolution of the expected output of the model is 0.05*0.05, the input of the model needs to be interpolated to 0.05°×0.05° grid data. The size of the resulting model input matrix is ​​(b,16,180,261), where b is the batch size of the training, 16 is 16 channels, and 180 and 261 are the height and width of the grid data respectively. It should be noted that the values ​​180 and 261 may cause problems in the convolution process (i.e., it is easy to cause the height and width of the grid data to be unequal). Therefore, the input data of the model was scaled before model training, that is, (b, 16, 180, 261) was scaled to (b, 16, 256, 256), and then scaled back to the original (b, 22, 180, 161) after the model training was completed.

[0054] 5. Model selection

[0055] A deep learning network model based on UKAN is used, where U-KAN (U-Shape Kolmogorov-Arnold Network) is a neural network architecture based on the classic U-Net model and combined with the theory of Kolmogorov-Arnold Networks (KANs). The design of KANs is inspired by the Kolmogorov-Arnold theorem, which states that any continuous function can be represented by a finite number of simple function combinations, making KANs perform well in processing high-dimensional complex functions.

[0056] 6. Model custom loss function construction

[0057] The study uses site data to calculate the loss function and performs weighted calculations based on the actual wind speed:

[0058]

[0059] Among them, y s is the site observation value, is the site prediction value, w s is the weighting coefficient, and N is the number of stations in the area (including automatic stations).

[0060] 7. Dataset Division

[0061] The samples from January 2020 to April 2023 are used as training samples, and the samples from May 2023 to April 2024 are used as independent samples for testing.

[0062] The performance of the test set and completely independent samples is shown in the following table:

[0063] Test of the maximum wind speed in the past 3 hours from May 2023 to April 2024

[0064]

[0065]

[0066] The present invention uses EC numerical model forecast data before the start time as EC fine grid 10-meter gust, hourly wind, 1000hPa, 850hPa, 925hPa U-V wind field data and ground data including 10-meter average wind and 10-meter gust in the past 3 hours. Through the processing and calculation of the time and space resolution of each element consistent with the 1-hour maximum wind speed forecast requirement, the first two forecast time effectiveness of the corresponding model forecast time effectiveness of each layer to be forecasted are taken to obtain a preliminary numerical model forecast product data set. The obtained data set is used to construct the multi-channel input data of the model, including 10 new ground data and 6 wind speed forecast fields, a total of 16 channel elements, which are used as the input of the Unet variant module together, and finally the output of the entire generator is obtained. The forecasting ability of the forecast model for gale above level 6 is more than 1 times higher than the current maximum wind speed forecast level, and can effectively improve the forecast accuracy of 0-24h hourly maximum wind, has good application prospects, and provides more accurate forecast services for disaster prevention and mitigation.

[0067] In view of the uneven distribution of forecast objects, the present invention designs a weighted loss function, which is the sum of the weighted losses of the weighted mean absolute error and the weighted mean square error. The weighting coefficient here is to assign different weights according to the actual wind speed of the site, that is, the larger the wind speed value, the greater the weight assigned. The above-mentioned weight assignment rule is used to increase the information feedback capacity of the wind speed area above level 6, so that it will not be averaged out by a large area of ​​wind speed area below level 6, thereby achieving high-precision forecast of extremely large wind speeds above level 6, and having higher forecast accuracy in early warning of disaster prevention and mitigation.

[0068] What is disclosed above is only a preferred embodiment of the present invention, and it certainly cannot be used to limit the scope of rights of the present invention. Ordinary technicians in this field can understand that all or part of the processes of the above embodiment and equivalent changes made according to the claims of the present invention still fall within the scope of the invention.

Claims

1. A method for forecasting hourly extreme winds within 24 hours based on UKAN, characterized in that: The following steps are involved: Step 1: Obtain the historical real-time maximum wind speed data and EC fine grid numerical model forecast data at the ground meteorological station at every hour; Step 2: Construct a sample dataset of hourly maximum wind speed and EC numerical model forecast; Step 3: Construct multi-channel data of ground data and high-altitude data; Step 4: Scaling the model input data; Step 5: Build a deep learning maximum wind speed forecast model based on UKAN; Step 6: Construct the loss function for model training; Step 7: Training the deep learning maximum wind speed forecast model based on the data and the loss function; Step 8: Use the trained deep learning maximum wind speed forecast model to calculate the hourly maximum wind speed forecast field for the next 0-24 hours.

2. The method for forecasting extreme wind hourly within 24 hours based on UKAN according to claim 1, characterized in that: The historical actual maximum wind speed data of the ground meteorological station are the maximum wind speed data of the automatic weather station in the past hour; the EC numerical model forecast data are the rainfall forecast fields of 51 ensemble members of the EC ensemble forecast, and the U and V wind field data of the EC fine grid 500hPa, 850hPa and 925hPa levels.

3. The method for forecasting extreme wind hourly within 24 hours based on UKAN according to claim 2, characterized in that: To forecast precipitation in the next 24 hours, a 12-36h forecast field is used. That is, the 12-hour forecast field of the model corresponds to the start time of the actual forecast.

4. The method for forecasting extreme wind hourly within 24 hours based on UKAN according to claim 3, characterized in that: The multi-channel data of ground data and high-altitude data include a ground data input channel and an high-altitude data input channel. The ground data input channel includes two elements, the average wind at 10 meters and the gusts at 10 meters in the past 3 hours. First, the first three forecast fields, namely 3h\6h\9h, are taken, and then the first two forecast time periods of the corresponding model forecast time period for the forecast period are taken; the high-altitude data input channel includes three levels of wind speed forecast fields calculated from the U and V wind fields of 500hPa, 850hPa, and 925hPa. Each level takes the first two forecast time periods of the corresponding model forecast time period for the forecast period; a total of 16 channel data.

5. The method for forecasting extreme wind hourly within 24 hours based on UKAN according to claim 4, characterized in that: The deep learning maximum wind speed forecast model described in step 5 is a neural network architecture that combines the Kolmogorov-Arnold network on the basis of Unet.

6. The method for forecasting extreme wind hourly within 24 hours based on UKAN according to claim 5, characterized in that: The loss function expression of the model training is as follows: Among them, y s is the site observation value, is the site prediction value, w s is the weighting coefficient, and N is the number of sites in the area.

Images