A method and apparatus for forecasting strong winds at low altitudes
By fusing multi-source observation data using the U-Net deep learning model, the problem of large forecasting errors in low-altitude wind fields under strong wind conditions in numerical models has been solved, achieving accurate correction of low-altitude three-dimensional wind fields and sensitive forecasting of high-impact wind events.
Patent Information
- Application Number
- CN202610875222.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-17
- Publication Date
- 2026-11-17
AI Technical Summary
Existing numerical models struggle to capture dynamic nonlinear abrupt changes in low-level wind fields under strong winds and complex terrain, leading to increased forecast bias and error in wind speed inflection points and failing to accurately reflect the structure of low-level wind fields.
The U-Net deep learning model is used to fuse background field data from CMA-GD numerical model with observation data from wind profiler radar and ground automatic weather stations. A gridded real-time wind field is generated through linear interpolation and error calculation. A weighted mean square error loss function is designed to optimize the weight of wind speed levels and improve the sensitivity to high-impact wind events.
It has achieved accurate correction of low-altitude three-dimensional high-resolution wind fields, significantly enhanced the ability to capture winds of level 6 and above, improved the stability of wind power grid connection and low-altitude flight safety, and reduced wind field forecasting errors.
Smart Images

Figure CN123113053A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of meteorological forecasting technology, and in particular to a forecast correction method and apparatus for strong winds at low altitudes. Background Technology
[0002] With the intensification of atmospheric circulation caused by global climate change, the frequency and intensity of extreme wind events such as low-level wind shear, downbursts, and turbulence have increased significantly. Changes in wind speed and direction in low-level wind fields, particularly from near the ground to an altitude of 1000 meters, directly impact the low-altitude economy (drone logistics, urban air traffic), wind energy resource assessment and power generation forecasting, wind-resistant design of super high-rise buildings, and aviation takeoff and landing safety. Therefore, achieving refined forecasting and early warning of strong winds at different altitudes is of urgent practical significance for ensuring the safety of low-altitude aircraft, improving the stability of wind power grid connection, and preventing secondary wind disasters.
[0003] Existing numerical model boundary layer parameterization schemes are mostly based on flux-gradient relationships constructed under assumptions of stationarity, homogeneity, and quasi-steadiness. Under conditions of strong winds and complex terrain, the intermittent enhancement of turbulence and significant changes in the structural scale of gusts lead to a substantial reduction in the applicability of parameterization schemes. The models struggle to capture dynamic nonlinear abrupt changes during strong wind processes. For example, under the control of cold fronts or outflow boundaries during thunderstorms, wind shear at different altitudes in the lower atmosphere often exhibits non-monotonic changes. Numerical models have systematic biases or phase lags in their forecasts of such vertical wind speed inflection points, resulting in significantly increased forecast errors for strong winds and an inability to accurately reflect the structure of the lower-level wind field. Summary of the Invention
[0004] The purpose of this invention is to at least address one of the shortcomings of the prior art and provide a forecast correction method for strong low-altitude winds.
[0005] To achieve the above objectives, the present invention adopts the following technical solution: Specifically, a forecast correction method for strong low-altitude winds is proposed, including the following: Acquire model grid forecast data; Based on the model grid forecast data, the model background field is generated using linear interpolation. The wind profiler radar and ground automatic station observation data are acquired and linearly interpolated to each height layer of the background field of the model. The observation increment is calculated to obtain error information. The error information is propagated throughout the domain using the nearest grid method, and finally a gridded real-time wind field at each height layer is generated. The background field of the model is normalized, and the original values of the data in the background field of the model are linearly scaled to a fixed interval of [0, 1] based on the statistical extreme values of historical observation data within a pre-selected time range. The historical observation data within the pre-selected time range are divided into training and validation sets; Set up a weighted mean square error loss function for strong wind optimization, assign different weights to the four wind speed levels, and dynamically weight the loss contribution of each effective grid point according to the interval to which the actual wind speed belongs. The pre-constructed U-Net deep learning model is trained using the training set, validation set, and loss function to obtain the trained model; Based on the background field of the model, the gridded real wind field, and error information, a two-dimensional gridded wind field forecast correction result is generated by the trained model.
[0006] Furthermore, specifically, the model grid forecast data is obtained by collecting grid forecast products provided by the high-resolution regional numerical weather prediction model of CMA-GD. The grid forecast products have a grid resolution of 0.03°×0.03° and a time resolution of hourly. The variables include U-wind and V-wind, as well as 2m temperature, 10m wind speed, mean sea level pressure, and temperature and humidity wind at different isobaric surfaces.
[0007] Furthermore, specifically, the error information is obtained by calculating the observation increment, including: The actual wind speed value at each observation station is compared with the value of the neighboring grid point in the model background field, and the error at that point is calculated to obtain the error information.
[0008] Furthermore, specifically, the pre-selected timeframe is 2019–2023, with data from 2019–2022 used to create a training set for model training, and data from 2023 used to create a validation set for model validation.
[0009] Furthermore, specifically, the loss function is: ; in and Let be the actual and predicted wind speeds for the i-th grid point, respectively. The total number of valid samples, weight function Determine using the following formula: .
[0010] Furthermore, specifically, the specific structure of the U-Net deep learning model includes, The encoder consists of a 4-level downsampling module, which extracts multi-scale atmospheric motion features layer by layer through 3×3 convolution, ReLU activation, batch normalization, and 2×2 max pooling, and compresses them to obtain a global context feature map. A time embedding layer is added to the bottleneck layer to encode the forecast lead time into a 16-dimensional dense vector and fuse it with the feature residuals to explicitly model the spatiotemporal periodicity and geographical dependence of meteorological elements. The decoder is a 4-level upsampling module, which restores the spatial size through 2×2 transposed convolution and splices the encoder features at the same scale with skip connections to compensate for the loss of downsampling information and restore the fine structure of the wind field. The output layer is mapped to 40 output channels through 1×1 convolution to directly output the U / V wind components at each altitude level.
[0011] The present invention also proposes a forecast correction device for strong low-altitude winds, comprising the following: The data acquisition module is used to acquire model grid forecast data; The model background field generation module is used to generate the model background field based on the model grid forecast data by linear interpolation. The gridded real-time wind field generation module is used to acquire observation data from wind profiler radar and ground automatic stations, linearly interpolate them to each height layer of the background field of the model, calculate the observation increment to obtain error information, and propagate the error information throughout the domain using the nearest grid method to finally generate the gridded real-time wind field at each height layer. The normalization module is used to normalize the background field of the model, and linearly scales the original values of the data in the background field of the model to a fixed interval of [0, 1] based on the statistical extreme values of historical observation data within a pre-selected time range. The training data partitioning module is used to partition historical observation data within a pre-selected time range into training and validation sets. The loss function setting module is used to set the weighted mean square error loss function for strong wind optimization. Different weights are set for the four wind speed levels. For each effective grid point, its loss contribution is dynamically weighted according to the interval to which the actual wind speed belongs. The model training module is used to train the pre-built U-Net deep learning model using the training set, validation set, and loss function to obtain the trained model. The forecast correction result calculation module is used to generate two-dimensional gridded wind field forecast correction results based on the model background field, gridded real wind field and error information through the trained model.
[0012] The beneficial effects of this invention are as follows: This invention proposes a forecast correction method and apparatus for strong low-altitude winds. On one hand, it integrates background field data from the CMA-GD numerical model with observation data from wind profiler radar and automatic weather stations. By calculating the observation increment and using the nearest grid method, error information is fed back to the entire grid field, generating spatially continuous gridded real-time wind field analysis products at various altitude levels that reflect actual observation information. On the other hand, to improve sensitivity to high-impact wind events, a weighted mean square error loss function is designed, dividing wind speed into four levels: weak wind, strong wind, gale, and blizzard, and assigning corresponding... By sacrificing weights to avoid underfitting the model to extreme winds, the ability to capture winds of force 6 and above is significantly enhanced. Addressing the problems of large forecast errors, loss of local details, and insufficient fitting of extreme wind fields in traditional numerical models (CMA-GD), this invention integrates multi-source observation data and numerical forecast background fields to generate three-dimensional wind field reanalysis data. Based on the U-Net deep learning architecture, a low-altitude three-dimensional high-resolution wind field correction model for short-term 24-hour forecasts is developed, achieving accurate wind field correction at a horizontal resolution of 0.03° with 25m-level layering at a height of 1000 meters. In summary, the wind field forecast correction results obtained by the proposed scheme are closer to the actual wind field. Attached Figure Description
[0013] The above and other features of this disclosure will become more apparent from the detailed description of the embodiments illustrated in conjunction with the accompanying drawings. In the accompanying drawings, the same reference numerals denote the same or similar elements. Obviously, the drawings described below are merely some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained from these drawings without any creative effort. In the drawings: Figure 1 The diagram shows a flowchart of a forecast correction method for strong low-altitude winds according to the present invention. Figure 2 The diagram shown is a schematic diagram of a forecast correction method for strong low-altitude winds according to the present invention. Figure 3 The figure shown is a training curve of the loss function proposed in this invention in a 100-meter wind field. Figure 4 The figure shows the training curve of the loss function proposed in this invention in a 250-meter wind field. Detailed Implementation
[0014] The following will provide a clear and complete description of the concept, specific structure, and technical effects of the present invention in conjunction with embodiments and accompanying drawings, so as to fully understand the purpose, solution, and effects of the present invention. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The same reference numerals used throughout the accompanying drawings indicate the same or similar parts.
[0015] Example 1, referring to Figure 1 as well as Figure 2 This invention proposes a forecast correction method for strong low-altitude winds, comprising the following: Step 110: Obtain model grid forecast data; Step 120: Based on the model grid forecast data, generate the model background field by linear interpolation (the grid forecast data of each standard pressure layer output by the model are linearly interpolated to obtain the forecast data of each variable at the height layers of 25m, 50m, 75m, 100m...1000m, i.e., the model background field). Step 130: Acquire wind profiler radar and ground automatic station observation data, and linearly interpolate them to each height layer of the background field of the model. Calculate the observation increment to obtain error information, and propagate the error information throughout the domain using the nearest grid method to finally generate a gridded real-time wind field at each height layer. Step 140: Normalize the background field of the model by linearly scaling the original values of the data in the background field of the model to a fixed interval of [0, 1] based on the statistical extreme values of historical observation data within a pre-selected time range. Step 150: Divide the historical observation data within the pre-selected time range into training and validation sets; Step 160: Set up a weighted mean square error loss function for strong wind optimization. Set different weights for the four wind speed levels. For each effective grid point, its loss contribution is dynamically weighted according to the interval to which the actual wind speed belongs. Step 170: Train the pre-built U-Net deep learning model using the training set, validation set, and loss function to obtain the trained model; Step 180: Based on the background field of the model, the gridded real wind field, and the error information, generate a two-dimensional gridded wind field forecast correction result through the trained model.
[0016] In this embodiment 1, 1. A method for generating multi-source data from a 3D wind field gridded real-world scenario By integrating background field data from the CMA-GD numerical model with observation data from wind profiler radar and automatic weather stations, and by calculating observation increments and using the nearest grid method, error information is fed back to the entire grid field to generate gridded real-time wind fields at each altitude level. Specifically: Gridded forecast products from a 0.03°×0.03° high-resolution regional numerical weather prediction model were collected, including variables such as U-winds and V-winds, as well as 2m temperature, 10m wind speed, mean sea level pressure, and temperature, humidity, and wind at different isobaric layers. Linear interpolation was performed on the gridded forecast data for each standard pressure layer output by the model to obtain forecast data for each variable at altitudes of 25m, 50m, 75m, 100m...1000m, i.e., the model background field. Multi-level, multi-point real-time observation data from wind profiler radar and automatic weather stations at corresponding times were collected, subjected to quality control and linear interpolation, and then processed to the various altitude layers of the model background field. The actual wind speed value of each observation station is compared with the value of the neighboring grid point in the model background field to calculate the error of that point. The error information of these discrete stations is fed back to the background field by the neighboring grid point method. That is, the station error is subtracted from the original value of the grid point next to the station in the background field to generate a spatially continuous gridded actual wind field analysis product at each height level that can reflect the actual observation information.
[0017] 2. Loss Function Construction Method To enhance the sensitivity to high-impact wind events, a weighted mean squared error loss function was designed, dividing wind speeds into four levels: weak wind (<10.8 m / s), strong wind (≥10.8 m / s), gale (≥17.2 m / s), and blizzard (≥24.5 m / s), and assigning appropriate loss weights to each level. This avoids underfitting the model to extreme winds and significantly enhances its ability to capture winds of level 6 and above.
[0018] 3. This invention addresses the problems of large forecasting errors, loss of local details, and insufficient fitting of extreme wind fields in traditional numerical models (CMA-GD). It integrates multi-source observation data with numerical forecast background fields to generate three-dimensional wind field real-time reanalysis data. Based on the U-Net deep learning architecture, it develops a low-altitude three-dimensional high-resolution wind field correction model for short-term 24-hour forecasts. It achieves accurate wind field correction at a horizontal resolution of 0.03° with 25m-level layering at a height of 1000 meters. It significantly improves the forecasting capability of strong winds of level 6 and above, supporting operational applications such as low-altitude flight, wind power dispatching, and urban emergency response.
[0019] Validated in 2023, the U-Net low-level three-dimensional wind field forecasting model demonstrated comprehensive and significant performance advantages over the traditional CMA-GD numerical model. In terms of overall wind field forecast accuracy, the U-Net model significantly reduced forecast errors for different altitude levels and wind components. Regarding its ability to forecast severe gales, the U-Net model also showed a marked improvement in its ability to identify strong wind events of force 6 and above, with a significantly higher TS score than the traditional model. The comparison results are shown in Tables 2 and 3 below.
[0020] Table 2. Mean Absolute Error (MAE) of Wind Fields at Different Height Levels Model 100-meter U-wind 100m V wind 250m U-wind 250m V wind CMA-GD 2.006 1.969 1.946 1.975 UNET Model 1.114 1.039 1.265 1.264 Table 3. Wind speed TS score for winds of level 6 and above Model 100 meters wind speed 250 meters wind speed CMA-GD 0.267 0.262 UNET Model 0.409 0.347 In a preferred embodiment of the present invention, the model grid forecast data is obtained by collecting grid forecast products provided by the high-resolution regional numerical weather prediction model of CMA-GD. The grid forecast products have a grid resolution of 0.03°×0.03° and a time resolution of hourly. The variables include U-wind and V-wind, as well as 2m temperature, 10m wind speed, mean sea level pressure, and temperature and humidity wind at different isobaric surfaces.
[0021] In a preferred embodiment of the present invention, specifically, error information is obtained by calculating the observation increment, including... The actual wind speed value at each observation station is compared with the value of the neighboring grid point in the model background field, and the error at that point is calculated to obtain the error information.
[0022] In this preferred embodiment, multi-layer, multi-point real-time observation data from wind profiler radar and automatic weather stations at corresponding times are collected, and quality control and linear interpolation are performed to process the data to various height layers relative to the model background field. The real-time wind speed value of each observation station is compared with the values of neighboring grid points in the model background field to calculate the error at that point. This discrete station error information is then fed back and propagated throughout the entire grid area using the nearest grid point method. This generates spatially continuous, gridded real-time wind field analysis products at various height layers that reflect real-time observation information.
[0023] As a preferred embodiment of the present invention, specifically, the pre-selected year range is 2019–2023, wherein the data from 2019–2022 is used to create a training set for model training, and the data from 2023 is used to create a validation set for model validation.
[0024] In this preferred embodiment, all input elements in the pattern background field are normalized. Based on the statistical extreme values of historical observation data from 2019 to 2023, the original values are linearly scaled to a fixed interval of [0,1], unifying the data scale and improving the stability of model training. Specifically, the normalized input elements are shown in Table 4 below. Table 4 elements Minimum value Maximum value t2m -25 43 u10m -54 55 v10m -53 55 mslp 933 1052 tcco 0 1 rh2m 0 100 rhum700 0 100 rhum850 0 100 rhum925 0 100 rhum1000 0 100 uwnd700 -65 66 uwnd850 -72 69 uwnd925 -69 71 uwnd1000 -60 61 vwnd700 -70 71 vwnd850 -71 75 vwnd925 -69 74 vwnd1000 -59 62 temp700 -18 31 temp850 -16 42 temp925 -16 47 temp1000 -29 51 hght1000 -634 376 hght700 2510 3236 hght850 821 1682 hght925 68 999 omeg1000 -3 2 omeg700 -11 17 omeg850 -10 11 omeg925 -6 6 In a preferred embodiment of the present invention, the loss function is specifically: ; in and Let be the actual and predicted wind speeds for the i-th grid point, respectively. The total number of valid samples, weight function Determine using the following formula: .
[0025] In this preferred embodiment, a weighted mean square error (MSE) loss function is designed for strong wind optimization. This loss function fully integrates meteorological operational standards and, based on key thresholds in China's "Wind Force Scale," classifies wind speeds into four levels: weak wind (<10.8 m / s), strong wind (≥10.8 m / s), gale (≥17.2 m / s), and blizzard (≥24.5 m / s), assigning loss weights of 1.0, 2.0, 5.0, and 10.0 respectively. For each effective grid point, its loss contribution is dynamically weighted according to the interval to which the actual wind speed belongs, referring to... Figure 3 as well as Figure 4 The effectiveness of the loss function proposed in this invention can be seen.
[0026] In a preferred embodiment of the present invention, the specific structure of the U-Net deep learning model includes: The encoder consists of a 4-level downsampling module, which extracts multi-scale atmospheric motion features layer by layer through 3×3 convolution, ReLU activation, batch normalization, and 2×2 max pooling, and compresses them to obtain a global context feature map. A time embedding layer is added to the bottleneck layer to encode the forecast lead time into a 16-dimensional dense vector and fuse it with the feature residuals to explicitly model the spatiotemporal periodicity and geographical dependence of meteorological elements. The decoder is a 4-level upsampling module, which restores the spatial size through 2×2 transposed convolution and splices the encoder features at the same scale with skip connections to compensate for the loss of downsampling information and restore the fine structure of the wind field. The output layer is mapped to 40 output channels through 1×1 convolution to directly output the U / V wind components at each altitude level.
[0027] In this preferred embodiment, multi-source meteorological element forecast data at different altitudes are input. The encoder is designed to consist of four cascaded downsampling modules. Each module includes two 3×3 convolution operations, followed by a ReLU activation function and batch normalization for feature extraction and stable training. Additionally, a 2×2 max-pooling layer is included to halve the spatial size of the feature map while doubling the number of feature channels, achieving a higher level of feature abstraction. The encoder outputs a highly compressed, deep feature map that carries the global contextual information of the input field.
[0028] The bottleneck layer is designed to encode forecast timeliness information into dense vectors through the Embedding layer, and then perform residual connections with the features of this layer. The role of the Embedding layer is to transform discrete or continuous spatiotemporal labels into low-dimensional dense vectors, enabling the model to explicitly perceive and model the periodic changes and geographical dependencies of meteorological elements, thereby significantly improving the physical rationality and spatiotemporal consistency of weather forecasts.
[0029] The decoder is also designed with four cascaded upsampling modules, symmetrical to the encoder. Each module contains: a 2×2 transposed convolution to upsample the feature map, doubling its spatial size; and a concatenation of feature maps of the same scale from the encoder path with the upsampled feature map along the channel dimension. This directly transmits the fine-grained information captured by the encoder to the decoder, effectively compensating for information loss during downsampling and ensuring that the generated wind field remains accurate at the grid scale.
[0030] At the decoder end, final mapping and adjustment through one or more fully connected layers helps the model integrate global information and fine-tune the output. The network ultimately maps the feature channel number to the target output dimension via a 1×1 convolutional layer. This dimension corresponds to the U and V horizontal wind components in the wind field forecast, thus generating a two-dimensional gridded wind field forecast correction product.
[0031] The specific hyperparameter settings of the U-Net deep learning model are shown in Table 1 below. Table 1 Hyperparameter Configuration Hyperparameter categories Configuration value illustrate Number of input channels 30 Corresponding to 30 forecast element files Number of output channels 40 Corresponding to 40 target wind field elements (U / V components at 20 heights). Time Embedding Dimension 16 Embedded vector dimension of timeliness features Encoder Channel [30, 30, 64, 128, 256] Number of output channels in each convolutional layer Decoder Channel [128, 64, 40] Number of output channels in each upsampling layer Batch size 64 Number of samples per training step Initial learning rate 0.001 Initial learning rate of AdamW optimizer Weight decay 1e-4 L2 regularization coefficient Number of training cycles 200 Maximum number of training rounds Stop early and be patient 10 Number of early shutdowns when verification of losses shows no improvement Gradient accumulation steps 4 Simulating gradient accumulation in larger batches Gradient clipping 1.0 Maximum norm constraint of gradient Optimizer AdamW lr=0.001, weight_decay=1e-4 Learning rate scheduling ReduceLROnPlateau mode='min', factor=0.1, patience=10, threshold=1e-4, min_lr=1e-6 Mixed precision training Autocast radScaler dtype=torch.float16 Specific implementation examples of the present invention are as follows: During the cold air event at 09:00 on October 21, 2025, the wind field forecast distribution of the U-Net 3D wind field forecast model was closer to the actual wind field than that of the CMA-GD model.
[0032] The present invention also proposes a forecast correction device for strong low-altitude winds, comprising the following: The data acquisition module is used to acquire model grid forecast data; The model background field generation module is used to generate the model background field based on the model grid forecast data by linear interpolation. The gridded real-time wind field generation module is used to acquire observation data from wind profiler radar and ground automatic stations, linearly interpolate them to each height layer of the background field of the model, calculate the observation increment to obtain error information, and propagate the error information throughout the domain using the nearest grid method to finally generate the gridded real-time wind field at each height layer. The normalization module is used to normalize the background field of the model, and linearly scales the original values of the data in the background field of the model to a fixed interval of [0, 1] based on the statistical extreme values of historical observation data within a pre-selected time range. The training data partitioning module is used to partition historical observation data within a pre-selected time range into training and validation sets. The loss function setting module is used to set the weighted mean square error loss function for strong wind optimization. Different weights are set for the four wind speed levels. For each effective grid point, its loss contribution is dynamically weighted according to the interval to which the actual wind speed belongs. The model training module is used to train the pre-built U-Net deep learning model using the training set, validation set, and loss function to obtain the trained model. The forecast correction result calculation module is used to generate two-dimensional gridded wind field forecast correction results based on the model background field, gridded real wind field and error information through the trained model.
[0033] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of the solution in this embodiment, depending on actual needs.
[0034] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated modules described above can be implemented in hardware or as software functional modules.
[0035] If the integrated module is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or system capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.
[0036] Although the description of the invention has been quite detailed and particularly of several described embodiments, it is not intended to limit it to any of these details or embodiments or any particular embodiment, but should be considered as providing a broad possible interpretation of the claims by referring to the appended claims and taking into account the prior art, thereby effectively covering the intended scope of the invention. Furthermore, the invention has been described above with respect to embodiments foreseeable by the inventors in order to provide a useful description, and non-substantial modifications to the invention that have not yet been foreseen may still represent equivalent modifications.
[0037] The above description is merely a preferred embodiment of the present invention. The present invention is not limited to the above-described embodiments. Any embodiment that achieves the technical effects of the present invention using the same means should fall within the protection scope of the present invention. Within the protection scope of the present invention, various modifications and variations can be made to the technical solutions and / or implementation methods.
Claims
1. A forecast correction method for strong low-altitude winds, characterized in that, Including the following: Acquire model grid forecast data; Based on the model grid forecast data, the model background field is generated using linear interpolation. The wind profiler radar and ground automatic station observation data are acquired and linearly interpolated to each height layer of the background field of the model. The observation increment is calculated to obtain error information. The error information is propagated throughout the domain using the nearest grid method, and finally a gridded real-time wind field at each height layer is generated. The background field of the model is normalized, and the original values of the data in the background field of the model are linearly scaled to a fixed interval of [0, 1] based on the statistical extreme values of historical observation data within a pre-selected time range. The historical observation data within the pre-selected time range are divided into training and validation sets; Set up a weighted mean square error loss function for strong wind optimization, assign different weights to the four wind speed levels, and dynamically weight the loss contribution of each effective grid point according to the interval to which the actual wind speed belongs. The pre-constructed U-Net deep learning model is trained using the training set, validation set, and loss function to obtain the trained model; Based on the background field of the model, the gridded real wind field, and error information, a two-dimensional gridded wind field forecast correction result is generated by the trained model.
2. The forecast correction method for strong low-altitude winds according to claim 1, characterized in that, Specifically, the model grid forecast data is obtained by collecting grid forecast products provided by the CMA-GD high-resolution regional numerical weather prediction model. The grid forecast products have a grid resolution of 0.03°×0.03° and a time resolution of hourly. The variables include U-wind and V-wind, as well as 2m temperature, 10m wind speed, mean sea level pressure, and temperature and humidity wind at different isobaric surfaces.
3. The forecast correction method for strong low-altitude winds according to claim 1, characterized in that, Specifically, the error information is obtained by calculating the observation increment, including: The actual wind speed value at each observation station is compared with the value of the neighboring grid point in the model background field, and the error at that point is calculated to obtain the error information.
4. The forecast correction method for strong low-altitude winds according to claim 1, characterized in that, Specifically, the pre-selected timeframe is 2019–2023, with data from 2019–2022 used to create a training set for model training and data from 2023 used to create a validation set for model validation.
5. The forecast correction method for strong low-altitude winds according to claim 1, characterized in that, Specifically, the loss function is: ; in and Let be the actual and predicted wind speeds for the i-th grid point, respectively. The total number of valid samples, weight function Determine using the following formula: 。 6. The forecast correction method for strong low-altitude winds according to claim 1, characterized in that, Specifically, the specific structure of the U-Net deep learning model includes, The encoder consists of a 4-level downsampling module. It extracts multi-scale atmospheric motion features layer by layer through 3×3 convolution, ReLU activation, batch normalization and 2×2 max pooling, and compresses them to obtain a global context feature map. A time embedding layer is added to the bottleneck layer to encode the forecast lead time into a 16-dimensional dense vector and fuse it with the feature residual to explicitly model the spatiotemporal periodicity and geographical dependence of meteorological elements. The decoder is a 4-level upsampling module that restores the spatial dimensions through 2×2 transposed convolution. It is paired with skip connections to stitch together features of the same scale as the encoder to compensate for the loss of downsampling information and restore the fine structure of the wind field. The output layer is mapped to 40 output channels through 1×1 convolution, directly outputting the U / V wind components of each height layer.
7. A forecast correction device for strong low-altitude winds, characterized in that, Including the following: The data acquisition module is used to acquire model grid forecast data; The model background field generation module is used to generate the model background field based on the model grid forecast data by linear interpolation. The gridded real-time wind field generation module is used to acquire observation data from wind profiler radar and ground automatic stations, linearly interpolate them to each height layer of the background field of the model, calculate the observation increment to obtain error information, and propagate the error information throughout the domain using the nearest grid method to finally generate the gridded real-time wind field at each height layer. The normalization module is used to normalize the background field of the model, and linearly scales the original values of the data in the background field of the model to a fixed interval of [0, 1] based on the statistical extreme values of historical observation data within a pre-selected time range. The training data partitioning module is used to partition historical observation data within a pre-selected time range into training and validation sets. The loss function setting module is used to set the weighted mean square error loss function for strong wind optimization. Different weights are set for the four wind speed levels. For each effective grid point, its loss contribution is dynamically weighted according to the interval to which the actual wind speed belongs. The model training module is used to train the pre-built U-Net deep learning model using the training set, validation set, and loss function to obtain the trained model. The forecast correction result calculation module is used to generate two-dimensional gridded wind field forecast correction results based on the model background field, gridded real wind field and error information through the trained model.
