WRF wind speed correction method fused with surface cover and time period encoding

Abstract

The application discloses a WRF wind speed revision method fusing surface coverage and time period coding, and is suitable for systematic error revision of WRF mode 10m wind speed in a complex terrain region, by fusing WRF mode data, ERA5-Land reanalysis data and site observation data, combining multi-source data space-time alignment, principal component factor screening, multi-module collaborative feature extraction and uncertainty weighted supervision method, to solve the problems of insufficient wind speed simulation precision, significant deviation and lack of targeted revision mechanism of the existing WRF mode in the complex terrain region and the region with strong surface heterogeneity. The application fuses multi-source heterogeneous data such as terrain, surface coverage and time periodicity information, introduces a deep learning method to construct a multi-module collaborative deviation correction framework, so as to significantly improve the spatial precision and regional adaptability of wind speed prediction.

Description

Technical Field

[0001] This invention relates to a WRF wind speed correction method that integrates land cover and time period coding, and belongs to the field of meteorological numerical forecasting technology. Background Technology

[0002] Wind speed, as a crucial parameter of atmospheric motion, has significant applications in meteorological forecasting, wind energy development, aviation and navigation, and urban environmental assessment. Currently, the Wind Response Model (WRF) is the mainstream wind speed forecasting tool, possessing a strong physical mechanism foundation and capable of providing spatially continuous and temporally dynamic wind speed predictions. However, the WRF model is essentially a physics-driven computational model, and its results are susceptible to the influence of initial conditions, physical parameterization schemes, resolution, and terrain processing errors. In practical applications, systematic biases often occur, especially in complex terrain areas and regions with significant surface heterogeneity, where wind speed errors are more pronounced.

[0003] To improve the accuracy of wind speed forecasts, researchers have recently attempted to correct WRF outputs using model post-processing techniques. Traditional post-processing methods include linear regression, Kalman filtering, and Bayesian correction. While these methods are effective in simple scenarios, they struggle to fully utilize multi-source heterogeneous data and fail to capture the nonlinear characteristics and spatial dependencies in wind speed fields. With the development of artificial intelligence, more and more research is incorporating deep learning methods into numerical forecast error correction. By constructing nonlinear mapping relationships to learn and correct model biases, deep learning has shown great potential in correcting variables such as wind speed, precipitation, and temperature.

[0004] In wind speed correction tasks, factors influencing wind field structure exhibit multi-source and regional variations. Based on systematic error analysis, it was found that wind speed errors show significant time dependence and underlying surface sensitivity. Specifically, in the time dimension, enhanced boundary layer stability at night leads to a significantly larger wind speed observation error compared to daytime, causing diurnal differences. In spring and winter, frequent cold air activity and the coupling of boundary layer dynamic-thermal processes result in significantly greater errors than in summer and autumn. In the underlying surface dimension, high-roughness surfaces, such as forests, shrublands, and urban built-up areas, experience enhanced turbulence dissipation, resulting in significantly greater wind speed simulation errors compared to flat surfaces. Low hills and high mountains are affected by topographic waves and mechanical turbulence, leading to significantly higher errors than river valleys, plains, and mid-altitude mountainous areas. Therefore, constructing an intelligent correction method that can integrate multi-source information and consider both physical background and data characteristics is of great significance for improving wind speed accuracy.

[0005] While current research has made some progress in single-factor modeling, it generally lacks the ability to collaboratively model multi-source factors such as topography, surface conditions, and time cycles, and has also failed to effectively achieve fine-grained regional correction of WRF model biases. Therefore, there is an urgent need for a wind speed correction method that integrates multi-source data and possesses spatial awareness and temporal periodicity encoding capabilities. This method should be able to utilize the relationship between high-resolution geographic information and meteorological variables, and also perform end-to-end learning through deep neural network models to achieve higher accuracy and stronger generalization capabilities in wind speed forecast correction. Summary of the Invention

[0006] The technical problem to be solved by this invention is to provide a WRF wind speed correction method that integrates land cover and time periodic coding. By integrating multi-source heterogeneous data such as topography, land cover and time periodic information, a deep learning method is introduced to construct a multi-module collaborative bias correction framework, thereby significantly improving the spatial accuracy and regional adaptability of wind speed forecasts.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0008] The WRF wind speed correction method integrating land cover and time period coding includes the following steps:

[0009] Step 1: Collect WRF meteorological data, ERA5-Land data, meteorological station observation data, DEM digital elevation data, and land cover data for the area to be corrected, construct a dataset, and divide it into training set and test set; use the 10m wind speed in the WRF meteorological data as the data to be corrected, and combine the 10m wind speed in the ERA5-Land data and the meteorological station observation data as the ground truth data.

[0010] Step 2: Construct a WRF wind speed correction model, comprising four branches. The first branch extracts meteorological features through a meteorological feature extraction module that integrates multi-scale convolutional structures and channel-spatial dual attention mechanisms. The second branch extracts temporal periodic features through a periodic information extraction module based on sine and cosine periodic coding. The third branch extracts topographic features through a geographic feature extraction module based on multi-scale convolution and adaptive pooling. The fourth branch extracts land cover features through a land cover feature extraction module that introduces land surface type embedding vectors and combines cross-scale attention. By concatenating and convolving the meteorological features, temporal periodic features, topographic features, and land cover features through channel concatenation, a multi-source heterogeneous fusion feature is obtained. This multi-source heterogeneous fusion feature is then input into a lightweight Unet codec structure to achieve nonlinear modeling and spatial information recovery of the multi-source heterogeneous fusion feature, outputting a 10m wind speed correction result.

[0011] Step 3: Train the WRF wind speed correction model constructed in Step 2 using the training set. During the training process, an uncertainty-weighted loss function is used to adaptively balance the two ground truth values ​​to guide the model training and obtain the trained WRF wind speed correction model.

[0012] Step 4: Input the WRF meteorological data, DEM digital elevation data and land cover data from the test set into the trained WRF wind speed correction model and output the wind speed correction results.

[0013] As a preferred embodiment of the present invention, in step 1, hourly WRF meteorological data, ERA5-Land data, meteorological station observation data, DEM digital elevation data, and land cover data of the area to be corrected over the past 5 years are collected. Among them, the WRF meteorological data includes: 10m meridional wind speed component, 10m zonal wind speed component, downward shortwave radiation, 2m temperature, cumulative gridded snowfall and ice, latent heat flux, upward heat flux, upward longwave radiation, topographic height, specific humidity, surface air pressure, snow water equivalent, land surface index, surface temperature, total cumulative cumulus convective precipitation, total gridded precipitation, instantaneous downward shortwave flux, instantaneous upward shortwave flux, instantaneous downward longwave flux, and instantaneous upward longwave flux; the ERA5-Land data includes 10m meridional wind speed component and 10m zonal wind speed component; the meteorological station observation data includes 10m wind speed.

[0014] In a preferred embodiment of the present invention, in step 2, before extracting meteorological features and time-cycle features, principal component analysis is used to perform dimensionality reduction on the remaining WRF meteorological data, excluding the 10m meridional wind speed component and the 10m zonal wind speed component. From the 18 meteorological elements affecting wind speed, the top four meteorological elements with the greatest impact on wind speed are selected as auxiliary meteorological elements, which, together with the 10m meridional wind speed component and the 10m zonal wind speed component, constitute the key features for wind speed correction. The key features are then subjected to coordinate transformation from Lambert projection to latitude and longitude projection, maintaining consistency with the ERA5-Land data coordinates and unifying the spatiotemporal resolution. The coordinate transformation formula is as follows:

[0015] ,

[0016] in, This represents the coordinates of the key feature points in the Lambert projection plane coordinate system. This represents the center point of the Lambert projection. Represents Euclidean distance. Point The polar angle relative to the center point of the projection, Point Latitude in radians This represents the scale factor of the Lambert projection. Denotes the conic constant of the Lambert projection. Point The radian system of longitude, This represents the central meridian of the Lambert projection.

[0017] In a preferred embodiment of the present invention, in step 2, the meteorological feature extraction module includes three parallel convolutional branches, which extract local, small-scale, and mesoscale meteorological features, respectively. The kernel sizes of the three parallel convolutional branches are 3×3, 5×5, and 7×7, respectively. The input to each convolutional branch is a time series formed by key features after coordinate transformation. The meteorological features extracted by the three parallel convolutional branches are concatenated into intermediate features by channel splicing. The intermediate features are then compressed using a 1×1 convolution to obtain the compressed features. The specific formula is as follows:

[0018] ,

[0019] in, This represents the input to each convolutional branch. Represent real numbers, Indicates the training batch size. This indicates a modified linear unit activation function. These represent convolution and concatenation operations, respectively. and These represent local, small-scale, and mesoscale meteorological characteristics, respectively. This represents the intermediate features after splicing along the channel axis. This represents the compressed feature map;

[0020] For the compressed feature map Perform global average pooling along the spatial dimension to obtain the global description vector. Global description vector The query vector Q, key vector K, and value vector V are generated through linear mapping, and the channel weights are calculated using a multi-head attention mechanism and a sigmoid activation function. Channel weights Intermediate features after splicing along the channel axis Multiplying yields the channel enhancement features. The specific formula is as follows:

[0021] ,

[0022] in, This represents the channel dimension component of the global description vector z. These represent the compressed feature maps. Height and width, Representing the spatial location of the feature map Spatial pixel values, This represents the Sigmoid activation function. This represents the output linear layer weights of the multi-head attention mechanism. and These represent the three learnable weight matrices in the attention mechanism. This represents the dimension of the query vector Q. This represents the Hadamard product, and T represents the transpose.

[0023] Channel enhancement features Average pooling and max pooling are performed along the channel dimension to obtain two single-channel spatial feature maps. and ,Will and By concatenating along the channel dimension, a joint spatial graph is obtained, which is then sequentially processed through a 7×7 convolution and a sigmoid activation function to obtain a spatial attention graph. Spatial attention map With channel enhancement features Pixel-by-pixel multiplication yields spatially enhanced feature maps. This serves as the output of the meteorological feature extraction module.

[0024] In a preferred embodiment of the present invention, in step 2, the periodic information extraction module takes the time series formed by the key features after coordinate transformation as input, extracts the season, month, and day / night information corresponding to each moment in the time series, and calculates the season number. Day and night coding and month number ; , They correspond to the four seasons: spring (March-May), summer (June-August), autumn (September-November), and winter (December-February). , These correspond to the months from January to December respectively; Daytime is 1, nighttime is 0; 8:00-20:00 is daytime, 20:00-8:00 the next day is nighttime; based on seasonal numbering. Day and night coding and month number Sine and cosine codes are performed separately to capture the periodic features of time. The sine and cosine code results are then concatenated in sequence to form a periodic time feature vector of length 6, which serves as the output of the periodic information extraction module. The sine and cosine code formulas are as follows:

[0025] ,

[0026] in, and These represent the seasonal cycle encoding vector, the day-night cycle encoding vector, and the monthly cycle encoding vector, respectively.

[0027] In a preferred embodiment of the present invention, in step 2, the geographic feature extraction module takes DEM digital elevation data as input, processes the digital elevation data using ArcGIS spatial analysis tools, and extracts multi-source terrain parameters: elevation A, slope, aspect, terrain roughness R, and slope variability Sr, as shown in the following formula:

[0028] The formula is as follows:

[0029] ,

[0030] ,

[0031] ,

[0032] Among them, altitude A was obtained directly from the DEM. These represent the gradient calculation kernels in the horizontal and vertical directions, respectively. Represents the gradient components in the horizontal and vertical directions. Represents two-dimensional convolution. For slope, Slope direction, For grid points The terrain roughness, For grid points altitude value, Indicates The average elevation of the 3×3 window centered on the center. Set the effective number of grid cells within the window to 9. These represent the rates of spatial variation of slope in the horizontal and vertical directions, respectively. They are respectively The slope of the grid points directly in front and behind. They are respectively The slope of the grid points directly above and below, For digital elevation data grid resolution, For grid points The slope variability; the above multi-source terrain parameters are spliced ​​along the channel dimension to obtain multi-source terrain input features;

[0033] A dual-branch parallel coding structure is used to extract the initial terrain semantic representation F0. Specifically, it includes: using branch one to perform 3×3 convolution, batch normalization and ReLU activation on the multi-source terrain input features, using branch two to perform 1×1 convolution on the multi-source terrain input features, and summing the output features of branch one and branch two element by element to obtain the initial terrain semantic representation F0.

[0034] The terrain complexity index Tc is calculated based on terrain roughness R and slope variability Sr: Based on Tc, three terrain levels are defined, and corresponding binary masks are generated. For low-complexity terrain, the mask is Tc_low. The terrain is of medium complexity, and the mask is Tc_mid. For highly complex terrain, the mask is Tc_high;

[0035] Low-complexity, medium-complexity, and high-complexity features are extracted from the initial terrain semantic representation F0 using low-complexity, medium-complexity, and high-complexity branches, respectively. Each complexity feature is multiplied element-wise with its corresponding complexity mask and then summed to obtain adaptive dynamic feature extraction. The low-complexity, medium-complexity, and high-complexity branches have the same structure, all including a 3×3 dilated convolution, batch normalization, and GELU activation function connected in sequence. The dilation rates of the 3×3 dilated convolution in the low, medium, and high-complexity branches are 1, 3, and 5, respectively.

[0036] The adaptively dynamically extracted features are sequentially subjected to 1×1 convolution, batch normalization, GELU activation, and adaptive average pooling to obtain the output of the geographic feature extraction module, namely the terrain features.

[0037] As a preferred embodiment of the present invention, in step 2, the land cover feature extraction module takes the time series formed by the ESRI discrete classification land cover data of spring, summer, autumn and winter and the key features after coordinate transformation as input. The learnable embedding layer transforms the ESRI discrete classification land cover data of spring, summer, autumn and winter into continuous semantic vectors, and rearranges them through 1×1 convolution to obtain feature tensors. Then, through layer normalization, the feature tensors are split into four sub-features of spring, summer, autumn and winter. Each sub-feature is configured with a dedicated extraction branch. The structure of each extraction branch is uniformly 3×3 convolution, batch normalization and GELU activation function, thereby outputting the independent features F_spr, F_sum, F_aut and F_win of each season.

[0038] Seasonal labels are applied to the time series formed by key features after coordinate transformation. The seasonal labels are then converted into vectors through one-hot encoding, and then projected into low-dimensional seasonal semantic vectors through a single fully connected layer. Introducing a 4×4 learnable weight matrix Seasonal semantic vectors With weight matrix After multiplication, the dynamic weights are obtained by normalization using the Softmax function. and The seasonal perception fusion feature is obtained by multiplying the four independent seasonal features element by element and splicing them along the channel dimension.

[0039] The seasonal perception fusion features are sequentially subjected to 1×1 convolution, batch normalization, GELU activation, adaptive average pooling, and 3×3 convolution to obtain the output of the land cover feature extraction module, which is the land cover feature.

[0040] In a preferred embodiment of the present invention, in step 2, the meteorological features, time period features, topographic features, and land cover features are respectively subjected to layer normalization processing and then spliced ​​along the channel dimension to obtain the fused features. Then, the features are fused using two consecutive convolutional blocks. Compression and semantic fusion are performed to obtain the wind speed deviation representation vector. ; to integrate features After performing a 1×1 convolution, it is then compared with the wind speed deviation representation vector. Residual fusion is performed to obtain multi-source heterogeneous fusion features. ; Integrating multi-source heterogeneous features The input is a three-layer lightweight Unet codec structure, which is then upsampled and compressed to generate a 10m wind speed correction result.

[0041] As a preferred embodiment of the present invention, in step 3, the uncertainty-weighted loss function is as follows:

[0042] ,

[0043] in, This represents the uncertainty-weighted loss function. and These represent the normalization coefficients of the ERA5-land data and the meteorological station observation data, respectively. and These represent the contributions of ERA5-land data and meteorological station observation data, respectively.

[0044] A computer device includes a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the computer program to implement the steps of the WRF wind speed correction method that integrates land cover and time periodic coding.

[0045] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects:

[0046] 1. This invention addresses the problem of large simulation errors in 10m wind speed by the WRF model under complex terrain and time-cycle variations. Through spatiotemporal error analysis, key error regions and time periods are identified. Principal component analysis is used for dimensionality reduction, and six key correction factors are selected. At the same time, meteorological factors, time cycles, terrain and surface information from multiple sources are integrated. Through multi-module collaborative extraction and feature fusion strategies, a unified wind speed error representation vector is formed in the neural network, which effectively improves the accuracy of WRF wind speed correction and has strong generalization ability.

[0047] 2. This invention constructs a meteorological feature extraction module (MSDNet) that integrates a dual attention mechanism and a multi-scale structure. This module can improve the ability to model the spatial features of wind speed errors. The dual attention mechanism captures the spatial correlation and important features of wind speed errors, while the multi-scale structure can better handle wind speed error features at different scales, thereby extracting meteorological features more accurately and providing richer information for wind speed correction.

[0048] 3. This invention introduces time-period sine and cosine encoding, which represents the time period variation based on sine and cosine period encoding. It guides a lightweight convolutional network to model the nonlinear relationship between wind speed and time period, effectively capturing the impact of time period variation on wind speed. This solves the problem of insufficient consideration of time variation in the prior art and improves the accuracy of wind speed correction in different seasons.

[0049] 4. This invention introduces a surface-embedded seasonal sensing modulator to enhance surface heterogeneity modeling; targeting the moderating effect of surface type on wind speed, it extracts surface cover features that are both seasonally specific and error-targeting. Simultaneously, by combining feature enhancement, it dynamically generates contribution weights for seasonal features to improve local and global structure modeling capabilities.

[0050] 5. This invention adopts dual-truth value joint supervision and uncertainty weighted loss function. In terms of supervision, ERA5-Land wind speed and station-observed wind speed are introduced as dual truth values. An uncertainty weighted loss function is introduced to dynamically adjust the weights of the two supervision signals, avoid bias, and improve the model's generalization ability in different regions and time periods. Attached Figure Description

[0051] Figure 1 This is an overall flowchart of the present invention;

[0052] Figure 2 This is an overall framework diagram of the present invention;

[0053] Figure 3 This is a chart showing the error analysis of wind speed between day and night;

[0054] Figure 4These are seasonal wind speed error maps; among them, (a) is the wind speed error distribution map for spring, (b) is the wind speed error distribution map for summer, (c) is the wind speed error distribution map for autumn, and (d) is the wind speed error distribution map for winter.

[0055] Figure 5 This is a simulation error diagram of wind speed for different land cover types;

[0056] Figure 6 This is a topographic wind speed error map;

[0057] Figure 7 This is the flowchart for principal component analysis;

[0058] Figure 8 The results are shown in the PCA feature dimensionality reduction diagram; where (a) is the correlation coefficient of principal component analysis, (b) is the heatmap of the first four principal components, and (c) is the interpretation of the cumulative variance of the principal components.

[0059] Figure 9 This is a structural diagram of the meteorological feature extraction module (MSDNet);

[0060] Figure 10 This is a structural diagram of the CycleNet module for extracting periodic information;

[0061] Figure 11 This is a structural diagram of the GeoNet module for extracting geographic features;

[0062] Figure 12 This is a structural diagram of the Land Cover Feature Extraction Module (LandNet);

[0063] Figure 13 This is a diagram of the lightweight UNet codec architecture;

[0064] Figure 14 These are comparison charts of the effects of wind speed correction before and after in different seasons; among them, (a) is a comparison chart of the error before and after wind speed correction in spring, (b) is a comparison chart of the error before and after wind speed correction in summer, (c) is a comparison chart of the error before and after wind speed correction in autumn, and (d) is a comparison chart of the error before and after wind speed correction in winter.

[0065] Figure 15 These are comparison charts showing the effects of wind speed correction before and after at different altitudes.

[0066] Figure 16 This is a comparison chart showing the effects of wind speed correction before and after different land cover types;

[0067] Figure 17 These are comparison charts showing the effects of wind speed correction before and after different time periods. Detailed Implementation

[0068] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0069] To address the issues of insufficient accuracy, significant bias, and lack of targeted correction mechanisms in 10m wind speed simulation by the WRF model in complex terrain and highly heterogeneous areas, this invention proposes a multi-module collaborative, feature-guided, lightweight, and efficient correction framework. By fusing multi-source heterogeneous data such as terrain, land cover, and temporal periodicity information, and introducing deep learning methods to construct a multi-module collaborative bias correction framework, this invention significantly improves the spatial accuracy and regional adaptability of wind speed forecasts.

[0070] We collected WRF meteorological data (to be corrected), ERA5-Land data, and station observation data from 2019 to 2023 for the study area (24.6°N–29.3°N, 105.5°E–110.2°E). Spatiotemporal error analysis was performed to identify key error regions and time periods. Principal component analysis was used for feature dimensionality reduction. We constructed a meteorological feature extraction module (MSDNet), a periodic information extraction module (CycleNet), a geographic feature extraction module (GeoNet), and a land cover feature extraction module (LandNet). Multi-source data features were concatenated, and wind speed maps were reconstructed using a decoder. Joint supervision was performed using ERA5-Land and station observations to output high-precision correction results. Data from 2019 to 2022 was used as the training set, and data from 2023 was used as the test set. Figure 1 and Figure 2 As shown, the method of the present invention mainly includes the following steps:

[0071] Step 1: Collect hourly resolution WRF meteorological data, ERA5-Land data, meteorological station observation data, digital elevation data, and land cover data for the area to be corrected (24.6–29.3°N, 105.5–110.2°E) from 2019 to 2023. The WRF meteorological data includes: 10m meridional wind speed component, 10m zonal wind speed component, downward shortwave radiation, 2m temperature, cumulative gridded snowfall and ice accumulation, latent heat flux, upward heat flux, upward longwave radiation, topographic height, specific humidity, surface air pressure, snow water equivalent, land surface indices, surface temperature, and total cumulus convection. Accumulated precipitation, total grid-point accumulated precipitation, instantaneous downwind shortwave flux, instantaneous upwind shortwave flux, instantaneous downwind longwave flux, and instantaneous upwind longwave flux, etc.; ERA5-Land data includes 10m meridional wind speed components and 10m zonal wind speed components. The 10m meridional and 10m zonal wind speed components from the WRF data are used as elements to be corrected, while the 10m meridional and 10m zonal wind speed components from the ERA5-Land data are used as ground truth data. The spatiotemporal distribution characteristics of wind speed errors are analyzed from the dimensions of time, topography, and land cover, identifying areas and time periods with significant errors. The main steps include:

[0072] Step 1.1, analyzing the distribution pattern of wind speed error from a time perspective, includes classifying the WRF data to be corrected by date into seasonal and diurnal categories: March to May is spring, June to August is summer, September to November is autumn, December to February is winter, 8:00 AM to 8:00 PM is daytime, and 8:00 PM to 8:00 AM the following day is nighttime. By comparing with ERA5-Land data and station observation data: the mean and variance of errors for each season in the region are calculated to analyze the distribution pattern of wind speed error from a seasonal perspective and identify seasons with significant errors; the average wind speed error for the region during the day and night is calculated. Specific analysis results are shown in Tables 1 and 2.

[0073] Table 1. Simulation error of 10m wind speed in WRF mode in different seasons

[0074]

[0075] Table 2. Simulation error of 10m wind speed during the day and night in WRF mode.

[0076]

[0077] Step 1.2, analyzing the wind speed error distribution from a topographic perspective, includes classifying the WRF area to be corrected according to altitude: from ground level to 500m is river valley plain, from 500m to 1000m is low mountain and hilly area, from 1000m to 1500m is mid-altitude mountain, and from 1500m to 3000m is high-altitude mountain. The hourly average error of wind speed at 10m at each altitude is calculated. The specific analysis results are shown in Table 3.

[0078] Table 3. Simulation error of 10m wind speed in WRF mode under different terrains.

[0079]

[0080] Step 1.3: Analyze the distribution pattern of wind speed error from the perspective of land cover. Classify the WRF area to be corrected according to land cover type, and statistically analyze the hourly average error of 10m wind speed under various land cover types. The results are shown in Table 4.

[0081] Table 4. Simulation error of 10m wind speed under different land cover types in WRF mode.

[0082]

[0083] Based on the above analysis of time period, topography, and land cover dimensions, the 10m wind speed error in the WRF model exhibits a significant spatiotemporal distribution pattern, primarily influenced by three factors: time period, topography, and land cover. The main conclusions are as follows:

[0084] The error on the time scale at a wind speed of 10m is as follows: Figure 3 , 4 As shown, Figure 4 (a)-(d) in the table correspond to the wind speed error distribution in spring, summer, autumn, and winter, respectively. The conclusion is that spring has the worst performance, followed by winter, at approximately 1.09 m / s and 0.97 m / s, respectively. Summer and autumn perform relatively better, at 0.67 m / s and 0.64 m / s, respectively. The daytime error is higher than the nighttime error, with an average error of 1.18 m / s, 28.3% higher than the nighttime error. The peak occurs at 10:00 AM (1.82 m / s), nearly twice the nighttime average. The nighttime error is lower, with an average error of 0.92 m / s, and the lowest value occurs at 0:00 AM (0.52 m / s).

[0085] The spatial error at a wind speed of 10m is as follows: Figure 5 and Figure 6 As shown, the 10m wind speed error is significant in high-altitude mountainous areas and low-mountain and hilly areas, reaching 1.29 m / s and 1.23 m / s respectively. River valleys and plains are next, with an error of 1.03 m / s. Mid-altitude mountainous areas perform relatively well, with an error of 0.78 m / s. Areas with forest and shrub cover perform the worst, with average errors of 1.9 m / s and 1.25 m / s respectively. Areas with grassland and building cover perform the next best, both with average errors of 1.2 m / s. The average wind speed error for other cover types is within 1 m / s.

[0086] Based on the above error analysis, the 10m wind speed error in WRF simulations exhibits significant differences under various terrain, land cover, seasonal, and diurnal conditions. To improve correction accuracy, the synergistic effects of multi-source data need to be comprehensively considered. By fusing the aforementioned heterogeneous multi-source information through feature extraction modules (such as MSDNet, GeoNet, CycleNet, and LandNet), a mapping relationship between error and geographical, temporal, and surface features can be established, achieving high-precision post-correction of WRF wind speed simulations.

[0087] Step 2: Use principal component analysis to perform feature dimensionality reduction on WRF meteorological elements, such as... Figure 7 As shown, the top four contributing factors are used as auxiliary meteorological factors, which together with the 10m meridional wind speed component and the 10m zonal wind speed component constitute the key features for wind speed correction. At the same time, the key features are transformed from Lambert projection to equal latitude and longitude projection, and the coordinates are kept consistent with the ERA5-Land grid data. The spatiotemporal resolution is unified as hourly, 0.1°×0.1°, and the data size is 48×48.

[0088] Step 2.1: 8760 samples of hourly WRF data are collected, each containing G=18 meteorological characteristic variables (downward shortwave radiation SWDOWN, 2m temperature T2, cumulative gridded snowfall and ice accumulation SNOWNC, latent heat flux LH, upward heat flux HFX, upward longwave radiation OLR, topographic height HGT, specific humidity Q2, surface pressure PSFC, snow water equivalent SNOW, land surface index LANDMASK, surface temperature TSK, total cumulative cumulus convective precipitation RAINC, total gridded precipitation RAINNC, instantaneous downward shortwave flux SWDNB, instantaneous rising shortwave flux SWUPB, instantaneous downward longwave flux LWDNB, and instantaneous rising longwave flux LWUPB) to form a data matrix. Each feature is standardized to have a mean of 0 and a variance of 1, resulting in the standardized matrix Z. The specific formula is as follows:

[0089] ,

[0090] in, It is The matrix represents the original WRF meteorological data, where E=8760 corresponds to hourly data over one year, and G=18 corresponds to the number of meteorological elements in each sample. This represents the g-th feature value of the e-th sample. This represents the sample mean of the g-th feature. This represents the sample standard deviation of the g-th feature. This represents the standardized eigenvalues, with a mean of 0 and a variance of 1, eliminating dimensional differences.

[0091] Based on the standardized matrix Z, calculate the covariance matrix C, eigenvalues, and eigenvectors. Sort the eigenvalues ​​from largest to smallest. The first four eigenvectors form the projection matrix P. Project the standardized data Z onto the first four principal components, as follows: Figure 8 As shown in (a)-(c), four high-impact wind speed factors were finally obtained. These factors, combined with the elements to be corrected, serve as meteorological characteristics for subsequent correction tasks. The calculation formula is as follows:

[0092] ,

[0093] Where C represents the covariance matrix, represents the covariance relationship between features, and Z and These represent the standardized data matrix and the transpose matrix, respectively. This indicates degree-of-freedom correction, ensuring an unbiased estimate of the covariance matrix. This represents the variance contribution of the corresponding principal component. The feature vector represents the projection direction of the data along the principal axis. This represents the projection matrix, which is constructed by selecting the eigenvectors corresponding to the four largest eigenvalues ​​and arranging them column-wise, thus preserving most of the variance information. This represents the data matrix after dimensionality reduction.

[0094] Step 2.2 involves projecting the WRF mode data to be corrected. The WRF mode defaults to Lambert projection, while the ERA5-Land data uses an equal latitude and longitude projection. This inconsistency in projection can lead to spatial misalignment and affect data matching. To unify the spatial resolution, the projected WRF data is resampled to a 0.1°×0.1° equal latitude and longitude grid using bilinear interpolation, ensuring precise alignment with the ERA5-Land grid. All data is then uniformly cropped to the target area (24.6°N-29.3°N, 105.5°E-110.2°E), corresponding to a 48×48 grid size. The coordinate transformation formula is as follows:

[0095] ,

[0096] in, This represents the coordinates of the WRF output data in the Lambert projection plane coordinate system. This represents the coordinates of the projection center point of the Lambert projection. This represents the Euclidean distance from the point to the center of the projection. This represents the polar angle of the point relative to the center of the projection. Point Latitude in radians This represents the scale factor of the Lambert projection. Point The radian system of longitude, This represents the central meridian of the Lambert projection, corresponding to STAND_LON in the WRF parameters. Through this conversion, WRF data can be precisely aligned with ERA5-Land on the same grid, avoiding wind speed correction deviations caused by projection differences.

[0097] Step 3: Construct a meteorological feature extraction module (MSDNet) that integrates multi-scale perception and dual attention mechanisms, such as... Figure 9 As shown, the module receives meteorological data of size [B,6,48,48]. Through multi-scale convolutional feature extraction and channel-space dual attention mechanism, it effectively enhances the model's ability to respond to meteorological factors under different scales and regional conditions, and provides discriminative and in-depth meteorological representations.

[0098] Step 3.1: The input data is processed through three parallel convolutional branches to extract local, small-scale, and mesoscale meteorological response features, respectively. The convolutional kernel sizes are 3×3, 5×5, and 7×7, with a padding stride of 1 for each branch. Each branch has 16 output channels, corresponding to an output size of [B, 16, 48, 48]. Then, a channel concatenation operation is performed to form an intermediate feature map with dimensions [B, 48, 48, 48]. To reduce computational complexity while maintaining semantic richness, a 1×1 convolution is used to compress the 48-channel features, resulting in a feature map with 16 channels and a spatial size of 48×48. The formula for multi-scale perceptual convolution is as follows:

[0099] ,

[0100] in, Indicates input data, This indicates the batch size for training the neural network. This indicates a modified linear unit activation function. These represent convolution and concatenation operations, respectively. and These represent local, small-scale, and mesoscale meteorological characteristics, respectively. This represents the feature map after being spliced ​​along the channel axis, with dimensions [B, 48, 48, 48]. This represents the fused features after feature compression.

[0101] Step 3.2, merging the compressed features Global average pooling is performed in the spatial dimension to obtain a global description vector for each channel. This description vector is then used through two linear mapping layers to generate a query vector Q, a key vector K, and a value vector V, respectively. Each vector retains the original channel dimension of 16. An attention mechanism is then used to calculate the channel attention matrix. Specifically, if the number of attention heads is h, then each head has 16 / h input channels. The attention outputs of all heads are concatenated and mapped back to the original channel dimension through a linear layer. Finally, the channel weight vector is generated by activation using the Sigmoid function. This weight is related to Multiply each channel sequentially to perform a weighted adjustment of the channel dimensions and generate channel-enhanced features. The formula for calculating channel attention is as follows:

[0102] ;

[0103] Step 3.3, Enhance channel features Average pooling and max pooling are performed along the channel dimension to obtain two single-channel spatial feature maps. and Both have dimensions [B, 1, 48, 48]. Concatenating them along the channel dimension yields a joint spatial map, which is then fed into a 7×7 convolutional layer with padding, outputting a dimension of [B, 1, 48, 48]. This result, activated by the sigmoid function, serves as the spatial attention map. Used with Pixel-wise multiplication generates spatially enhanced feature maps. The spatial attention is calculated using the dimension [B, 16, 48, 48], and represents a high-order meteorological feature that integrates multi-scale semantics and dual attention enhancement. The spatial attention calculation formula is as follows:

[0104] .

[0105] Step 4, construct as follows Figure 10 The CycleNet module shown acquires time series information and extracts the season, month, and day / night information corresponding to each moment from the WRF pattern and ERA5-Land reanalysis data, and further converts it to obtain the season number. These correspond to the four seasons: spring, summer, autumn, and winter, and are used to express macro-seasonal information; month information is an integer value, ranging from 1 to 12; day and night information is an integer value. , used to express fine-grained time periods.

[0106] The periodic information extraction module includes: based on seasonal number Day and night coding and month number Sine and cosine codes are performed separately to capture the periodic features of time, including the seasons, day and night, and monthly cycles. Sine and cosine transforms are then performed on these cycles. The six encoded results are concatenated in sequence to form a periodic time feature vector of length 6 with dimensions [6, 48, 48], which is used to express the temporal semantic information of wind speed correction. The specific calculation formula is as follows:

[0107] .

[0108] Step 5, construct as follows Figure 11 The GeoNet geographic feature extraction module shown employs the Sobel operator to calculate terrain slope and aspect, utilizes ArcGIS window statistics to calculate terrain roughness R, and calculates slope variability Sr using a quadratic gradient method. It then uses the concatenation of elevation, slope, aspect, terrain roughness, and slope variability along the channels as input. Through multi-source terrain parameter fusion and adaptive dynamic extraction based on terrain complexity, it achieves accurate geographic feature capture and outputs a 48×48 scale DEM feature map. This module includes the following steps:

[0109] Step 5.1: Terrain Information Extraction and Feature Construction. Using ArcGIS spatial analysis tools, digital elevation data (DEM) is processed to directly obtain elevation A. The Sobel operator is used to perform horizontal and vertical convolution operations on the DEM to obtain the gradient components dx and dy for each pixel. Slope and aspect are calculated, and elevation A, slope, and aspect are normalized to the [0,1] interval. Terrain roughness R is calculated using ArcGIS window statistics, and slope variability Sr is calculated using a quadratic gradient averaging method. This completes the full extraction of multi-source terrain parameters (A, Slope, Aspect, R, Sr). The above five types of terrain parameters are concatenated along the channel dimension to obtain the multi-source terrain input feature [B,5,256,256]. The specific calculation formula is as follows:

[0110] ,

[0111] ,

[0112] ,

[0113] in, This represents the average elevation of a 3×3 window, and S represents the slope.

[0114] Step 5.2, basic feature encoding, uses a dual-branch parallel encoding structure to extract the initial terrain semantic representation F0. Specifically, it includes: Branch 1 performs a 3×3 convolution on the concatenated [B,5,256,256] features, followed by batch normalization and ReLU activation function to enhance the extraction of local terrain spatial correlation features; Branch 2 performs a 1×1 convolution on the same concatenated features with 32 kernels, a stride of 1, and zero padding, and achieves semantic fusion of terrain parameters through linear transformation of channel dimensions; The output features of Branch 1 and Branch 2 are summed element-wise to obtain the initial terrain semantic representation F0 with dimensions [B,32,256,256], preserving local spatial details and enhancing semantic correlation between channels.

[0115] Step 5.3, Terrain Complexity Classification and Mask Generation: Based on terrain roughness R and slope variability Sr, a terrain complexity index Tc is constructed. The calculation formula is as follows:

[0116] ,

[0117] The weighting coefficients of 0.4 and 0.6 are set based on physical mechanisms. The slope variability Sr reflects the severity of terrain undulation, and the terrain roughness R reflects the surface friction coefficient. Three terrain levels are divided according to the range of Tc, and corresponding binary masks are generated. Tc < 0.2 is low-complexity terrain, with a mask of Tc_low; 0.2 ≤ Tc ≤ 0.6 is medium-complexity terrain, with a mask of Tc_mid; and Tc > 0.6 is high-complexity terrain, with a mask of Tc_high. All three types of masks maintain the [B, 1, 256, 256] dimension, which is completely aligned with the F0 spatial scale.

[0118] Step 5.4, terrain-adaptive dynamic extraction, dynamically matches dilated convolutions of different receptive fields based on terrain complexity levels to achieve accurate capture of multi-scale geographic features, specifically including:

[0119] The low-complexity branch inputs F0 into a 3×3 dilated convolution with a dilation rate of 1 and padding of 1, followed by batch normalization and the GELU activation function. This results in a small receptive field focusing on the local microscopic details of low-lying terrain. The medium-complexity branch inputs F0 into a 3×3 dilated convolution with a dilation rate of 3 and padding of 3, followed by batch normalization and the GELU activation function. This results in a medium receptive field balancing the local details of hills and low-altitude mountains. The high-complexity branch inputs F0 into a 3×3 dilated convolution with a dilation rate of 5 and padding of 5, followed by batch normalization and the GELU activation function. This results in a large receptive field capturing the global terrain structure of medium-to-high-altitude mountains and steep slopes. The output features of the three branches are then element-wise multiplied with the corresponding complexity mask to filter out only the effective features of the corresponding complexity region. The filtered features are then element-wise summed to obtain the adaptive dynamic feature extraction [B, 32, 256, 256]. The dilated convolution calculation formula mentioned above is as follows:

[0120] ,

[0121] in, This represents the output feature of the basic feature encoding. The kernel represents the dilated convolution kernel, obtained through model initialization and training. t represents the dilation rate, which can be 1, 3, or 5, corresponding to three terrain complexities: Tc_low, Tc_mid, and Tc_high, respectively. This indicates the output feature of the corresponding branch.

[0122] Step 5.5, Feature Refinement and Scale Adaptation: The adaptively dynamically extracted features are input into the feature refinement unit. The feature refinement unit consists of a 1×1 convolution with 32 kernels, a stride of 1, zero padding, batch normalization, and a GELU activation function. The core geographic features are enhanced through feature refinement. Finally, the 256×256 spatial scale is compressed to 48×48 through adaptive average pooling, and the final output is a DEM feature map of [B,32,48,48].

[0123] Step 6, construct the land cover feature extraction module (LandNet), such as... Figure 12 As shown, the ESRI discrete classification land cover data for spring, summer, autumn, and winter, along with the corresponding actual seasonal labels of the samples, are used as dual inputs. The core consists of two parts: a season-aware modulator and feature compression. It extracts land cover features that are both season-specific and error-targeting, and outputs a land cover feature map. The specific steps include:

[0124] Step 6.1, Season-Aware Modulator: This step enables dynamic seasonal adaptation and feature enhancement. The ESRI-classified land cover data for each of the four seasons are sequentially concatenated into a tensor of dimensions [B, 4, 256, 256]. Simultaneously, a sample seasonal label tensor of dimension [B, 1] is constructed, with discrete values ​​1-4 corresponding to spring, summer, autumn, and winter, respectively, marking the actual meteorological season for each input sample. A learnable embedding layer transforms the discrete classification values ​​of each season into a continuous semantic vector of dimension [B, 4, 256, 256, 16]. This vector is then subjected to 1×1 convolution and rearrangement to obtain a [B, 12]... A tensor of [8, 256, 256] is used to split the features into four sub-features of [B, 32, 256, 256] through layer normalization, corresponding to the four seasons: spring, summer, autumn, and winter. Each sub-feature is configured with a dedicated extraction branch. Each branch has a unified structure of 3×3 convolution, batch normalization, and GELU activation function, and the weights of the convolution kernels in each branch are initialized and updated independently. The final output is the independent features F_spr, F_sum, F_aut, and F_win for the four seasons, all with dimensions of [B, 32, 256, 256]. The following is the layer normalization process (…). The formula for calculating ) is:

[0125] ,

[0126] in, Represents the characteristic mean. Represents the characteristic variance. For learnable scaling and offset parameters, To avoid tiny values ​​with a denominator of 0.

[0127] Step 6.2: Dynamically generate the contribution weights of the four seasons' features based on the seasonal identifiers of the input samples. Transform the [B,1] seasonal identifiers into a [B,4] vector through one-hot encoding, and then project it into a low-dimensional [B,4] seasonal semantic vector through a one-layer fully connected network. Introducing a 4×4 learnable weight matrix Seasonal semantic vectors With weight matrix After multiplication, the dynamic weights are obtained by normalization using the Softmax function. and The data is then broadcast to [B,4,32,256,256], and multiplied element-wise with the four independent seasonal features along the channel dimension to obtain a seasonal perception fusion feature of size [B,128,256,256].

[0128] Step 6.3, Feature Compression: This step adapts dimensions and refines core features. A 1×1 convolution with 64 kernels, a stride of 1, and zero padding is performed on the seasonally perceived fusion features, followed by batch normalization and GELU activation, outputting a simplified feature set of [B, 64, 256, 256]. Further, adaptive average pooling and a 3×3 convolutional layer are used to compress the high-resolution features into mesoscale features of [B, 32, 48, 48], preserving regional land cover structure information. The module takes high-resolution land cover classification data and meteorological data from four seasons as input.

[0129] Step 7: Construct a multi-source data fusion module to integrate heterogeneous feature information from multiple sources, including meteorological features, temporal cycle features, geographical features, and land cover features, and to uniformly model the spatial distribution pattern of wind speed deviation. This specifically includes the following steps:

[0130] Step 7.1, Multi-source Feature Standardization and Size Alignment. Meteorological features, temporal periodic features, geographical features, and land cover features are concatenated along the channel dimension to form the fused input feature tensor. Each branch's output features undergo layer normalization before input to ensure consistent numerical distribution and avoid gradient shift during fusion.

[0131] ;

[0132] Step 7.2, Channel Compression and Semantic Fusion. The concatenated 86-channel high-dimensional features are compressed using two consecutive ConvBlock convolutional blocks. Compression and semantic fusion are performed. Each module contains a 3×3 convolutional layer, a batch normalization layer, and a ReLU activation function to extract the joint representation relationship between cross-modal features. Finally, the compression is reduced to 64 channels to obtain a unified wind speed deviation representation vector. :

[0133] ;

[0134] Step 7.3 introduces residual connections and attention enhancement. High-dimensional splicing features are then... Compressed to 64 channels using 1×1 convolution, and then... The summation forms a unified multi-source heterogeneous fusion characteristic output by residual fusion:

[0135] .

[0136] Step 8, construct as follows Figure 13 The lightweight UNet encoding / decoding structure shown takes multi-source heterogeneous fused features as input and uses an encoder-decoder structure to achieve nonlinear modeling and spatial information recovery of the fused features, outputting a corrected high-precision 10m wind speed correction result. It is trained using a dual-ground value joint supervision mechanism and an uncertainty-weighted loss function, specifically including the following steps:

[0137] Step 8.1: The encoder uses 3 layers of convolution + ReLU blocks for feature compression and abstract representation, with the number of channels successively increasing from 64 to 128 to 256. The convolution kernel size is 3×3, and the stride is 2, achieving spatial downsampling. The bottleneck module uses two lightweight convolutional residual blocks, which consist of two 3×3 convolutional layers and a short-circuit connection, effectively enhancing nonlinear modeling capabilities while maintaining the feature size and number of channels unchanged. The decoder uses 2 layers of deconvolution for upsampling, with the number of channels gradually decreasing from 256 to 128 to 64, and the size gradually recovering from 6×6 to 12×12 to 24×24. Skip connections are used to fuse features with the corresponding layers of the encoder to preserve edge and detail information. Finally, upsampling and channel compression are used to restore the original resolution, resulting in the final wind speed correction map [B,48,48].

[0138] Step 8.2, Design of Joint Supervision Mechanism and Uncertainty Weighted Loss Function. To simultaneously utilize ERA5-Land reanalysis data and ground station observation information, a dual-truth value joint supervision mechanism is constructed, and a weighted loss function based on uncertainty modeling is introduced to dynamically balance the influence of the two types of supervision signals. Learnable uncertainty parameters are set for the error branches of ERA5-Land and ground station observations respectively. and The total loss function is derived using maximum likelihood estimation, enabling the model to adaptively adjust the weights of the two loss classes during training. This mechanism effectively mitigates the training imbalance problem caused by data heterogeneity and improves the overall performance of wind speed correction results across large-scale structures and local accuracy. The final total loss function is as follows:

[0139] ;

[0140] Step 8.3 involves model optimization to achieve WRF wind speed correction. The entire model is trained end-to-end using a backpropagation algorithm based on the Adam optimizer. A weighted loss function constructed with a joint supervision mechanism is used as the optimization objective, guiding the model to simultaneously learn the correction relationship between large-scale wind field structure and local fine-grained errors. Based on the principle of minimizing the overall loss on the validation set, the optimal model parameter file is automatically saved. This achieves high-precision post-processing correction of 10m wind speed in WRF simulations for any region.

[0141] This invention proposes a WRF wind speed correction method that integrates land cover and time period coding, aiming to improve the accuracy of 10m wind speed simulation in WRF models. The method collects nearly five years of meteorological data to be corrected, ERA5-Land data, and station observation data for the study area, performs spatiotemporal error analysis, and identifies key error regions and time periods. Principal component analysis is used for feature dimensionality reduction, and meteorological feature extraction module MSDNet, periodic information extraction module CycleNet, geographic feature extraction module GeoNet, and land cover feature extraction module LandNet are constructed respectively. Multi-source data features are concatenated, and the corrected wind speed is output through a lightweight UNet network. Joint supervision is performed by combining ERA5-Land and station observations to output high-precision correction results.

[0142] To verify the actual role of each module in the correction process, different modules were combined to process the WRF data to be corrected, and ablation experiments were conducted. Seven control experiments were set up with AG, and epoch=20. The training loss and validation loss of each experiment were observed. The specific experimental settings and performance are shown in Table 5.

[0143] Table 5 Control Experiment Setup

[0144]

[0145] This experiment systematically evaluated the impact of multi-module combinations on the wind speed correction model. The baseline model MSDNet (A) had a validation loss of 0.3284. Adding CycleNet (B), GeoNet (C), or LandNet (D) modules individually improved performance, with GeoNet showing the best results and a validation loss of 0.2964. Among the two-module combinations, MSDNet+CycleNet+GeoNet (E) performed best, with a validation loss of 0.2850. The complete correction model (G), which integrates all modules, achieved the best results, with a validation loss of 0.2817, validating the effectiveness of multi-feature fusion.

[0146] To verify the correction effectiveness of this invention across seasonal, diurnal, topographic, and land cover dimensions, a systematic correction verification of the 2023 WRF model 10m wind speed simulation results was performed using the complete correction model (G). The experiment used year-round data and analyzed according to season, diurnal, topographic, and land cover. The model's correction effectiveness across each dimension was systematically evaluated. The results, shown in Tables 6-9 and related figures, demonstrate significant improvements in all dimensions. Tables 6 and 9... Figure 17 The results show the diurnal correction performance; Table 7 and Figure 14 To correct the model's performance in the seasonal dimension, Figure 14 In the table, (a)-(d) correspond to the four seasons: spring, summer, autumn, and winter, respectively; Table 8 and Figure 15 For different terrains, the performance is corrected, as shown in Table 9 and Figure 16 The wind speed performance before and after correction is shown for different land cover types.

[0147] Table 6. Day and Night Correction Performance

[0148]

[0149] Table 7 Correction performance on the seasonal dimension

[0150]

[0151] Table 8 Correction performance for different terrains

[0152]

[0153] Table 9. Wind speed performance before and after correction under different land cover types.

[0154]

[0155] This invention significantly improves the accuracy and reliability of wind speed prediction through comprehensive, multi-level data processing and innovative modeling methods, and has broad application prospects and practical value.

[0156] Based on the same inventive concept, embodiments of this application provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the aforementioned WRF wind speed correction method that integrates land cover and time periodic coding.

[0157] Based on the same inventive concept, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the aforementioned WRF wind speed correction method that integrates land cover and time periodic coding.

[0158] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0159] This invention is described with reference to flowchart illustrations of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each step in the flowchart, and combinations of steps in the flowchart, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing device, generate instructions for implementing the steps in the flowchart. Figure 1 A device for a function specified in one or more processes.

[0160] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 The function specified in one or more processes.

[0161] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 Steps of a specified function in one or more processes.

[0162] The above embodiments are merely illustrative of the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solution based on the technical concept proposed in this invention shall fall within the scope of protection of this invention.

Claims

1. A WRF wind speed correction method that fuses land surface cover and time period encoding, characterized in that, The method comprises the following steps: Step 1, collect WRF meteorological data, ERA5-Land data, meteorological station observation data, DEM digital elevation data and land cover data of the area to be corrected, construct a data set, and divide it into a training set and a test set; the 10m wind speed in the WRF meteorological data is taken as the data to be corrected, and the 10m wind speed in the ERA5-Land data and the meteorological station observation data is taken as the true value data; Step 2, construct a WRF wind speed correction model, including four branches, the first branch extracts meteorological features through a meteorological feature extraction module that fuses a multi-scale convolution structure and a channel-space dual attention mechanism, the second branch extracts time period features through a period information extraction module based on sine and cosine period encoding, the third branch extracts terrain features through a geographic feature extraction module based on multi-scale convolution and adaptive pooling, and the fourth branch extracts land cover features through a land cover feature extraction module that introduces a land type embedding vector and combines cross-scale attention; the meteorological features, time period features, terrain features and land cover features are concatenated and convolved to obtain multi-source heterogeneous fusion features; the multi-source heterogeneous fusion features are input into a lightweight Unet encoder-decoder structure to realize nonlinear modeling and spatial information recovery of the multi-source heterogeneous fusion features, and output 10m wind speed correction results; In step 2, the geographic feature extraction module takes the DEM digital elevation data as input, processes the digital elevation data using the ArcGIS spatial analysis tool, extracts multi-source terrain parameters: elevation A, slope S, aspect, terrain roughness R and slope rate Sr, concatenates the above multi-source terrain parameters along the channel dimension to obtain multi-source terrain input features; An initial terrain semantic representation F0 is extracted using a double-branch parallel encoding structure, specifically including: using branch one to perform 3x3 convolution, batch normalization and ReLU activation on the multi-source terrain input features, using branch two to perform 1x1 convolution on the multi-source terrain input features, and adding the output features of branch one and branch two element by element to obtain the initial terrain semantic representation F0; Based on the terrain roughness R and the slope variability Sr, the terrain complexity index Tc is calculated: According to Tc, three terrain grades are divided, and corresponding binary masks are generated, For low complexity terrain, the mask is Tc_low, For medium complexity terrain, the mask is Tc_mid, For high complexity terrain, the mask is Tc_high; Low-complexity features, medium-complexity features and high-complexity features are extracted from the initial terrain semantic representation F0 using low-complexity branches, medium-complexity branches and high-complexity branches respectively, and each complexity feature is multiplied by the corresponding complexity mask element by element and then added to obtain adaptive dynamic extraction features; the low-complexity branches, the medium-complexity branches and the high-complexity branches have the same structure, and each includes a 3x3 atrous convolution, a batch normalization and a GELU activation function connected in sequence, and the dilation rates of the 3x3 atrous convolution in the low-complexity branches, the medium-complexity branches and the high-complexity branches are 1, 3 and 5 respectively; The adaptive dynamic extraction features are sequentially subjected to 1x1 convolution, batch normalization, GELU activation and adaptive average pooling to obtain the output of the geographic feature extraction module, i.e., the terrain features; Step 3, train the WRF wind speed correction model constructed in step 2 using the training set, adaptively balance the double true values during the training process using an uncertainty weighted loss function to guide the model training, and obtain the trained WRF wind speed correction model; Step 4, input the WRF meteorological data, DEM digital elevation data and land cover data in the test set into the trained WRF wind speed correction model, and output the wind speed correction result.

2. The method of claim 1, wherein the WRF wind speed is corrected by fusing the land surface cover and the time period encoding. In step 1, the WRF meteorological data, ERA5-Land data, meteorological station observation data, DEM digital elevation data and land cover data of the area to be corrected in the past 5 years are collected; wherein the WRF meteorological data includes: 10m meridional wind speed component, 10m latitudinal wind speed component, downward shortwave radiation, 2m temperature, cumulative grid snow and ice amount, latent heat flux, upward heat flux, upward longwave radiation, terrain height, specific humidity, surface pressure, snow water equivalent, land surface index, surface temperature, total cumulus convective precipitation cumulative, total grid precipitation cumulative, instantaneous downward shortwave flux, instantaneous upward shortwave flux, instantaneous downward longwave flux and instantaneous upward longwave flux; the ERA5-Land data includes 10m meridional wind speed component and 10m latitudinal wind speed component; the meteorological station observation data includes 10m wind speed.

3. The method of claim 2, wherein the WRF wind speed is corrected by fusing the land surface cover and the time period encoding. In step 2, before performing meteorological feature and time period feature extraction, the remaining WRF meteorological data except 10m meridional wind speed component and 10m latitudinal wind speed component are processed by dimension reduction using principal component analysis method, and the top 4 meteorological factors with the greatest impact on wind speed are selected from the 18 meteorological factors affecting wind speed as auxiliary meteorological factors, together with 10m meridional wind speed component and 10m latitudinal wind speed component to form the key features of wind speed correction; and the key features are converted from Lambert projection to latitude and longitude projection, which is consistent with the coordinate of ERA5-Land data, and the spatial and temporal resolution is unified, and the coordinate conversion formula is as follows: ， in, This represents the coordinates of the key feature points in the Lambert projection plane coordinate system. This represents the center point of the Lambert projection. This represents the Euclidean distance from the point to the center of the projection. Point The polar angle relative to the center point of the projection, Point Latitude in radians This represents the scale factor of the Lambert projection. Denotes the conic constant of the Lambert projection. Point The radian system of longitude, This represents the central meridian of the Lambert projection.

4. The method of claim 3, wherein the WRF wind speed is corrected by fusing the land surface cover and the time period encoding. In step 2, the meteorological feature extraction module includes three parallel convolution branches, which respectively extract local, small range and mesoscale meteorological features, and the convolution kernel sizes of the three parallel convolution branches are 3x3, 5x5 and 7x7 respectively, and the input of each convolution branch is the time series formed by the key features after coordinate conversion, and the meteorological features extracted by the three parallel convolution branches are spliced into intermediate features through channel splicing, and the intermediate features are compressed by using 1x1 convolution to obtain compressed features. The specific formula is as follows: ， wherein, denotes the input to each convolution branch, denotes a real number, denotes the training batch size, denotes a rectified linear unit activation function, denotes a convolution, respectively a concatenation operation, and denote local, small-scale and mesoscale meteorological features, respectively, denotes the intermediate feature after concatenation along the channel axis, denotes the compressed feature map; compressed feature map Global average pooling is performed in the spatial dimension to obtain a global description vector Global description vector Generate query vector Q, key vector K and value vector V through linear mapping, and calculate channel weight through multi-head attention mechanism and Sigmoid activation function , the channel weight is multiplied by the intermediate feature after splicing along the channel axis to obtain a channel enhanced feature ; The specific formula is as follows: ， in, This represents the channel dimension component of the global description vector z. These represent the compressed feature maps. Height and width, Representing the spatial location of the feature map Spatial pixel values, This represents the Sigmoid activation function. This represents the output linear layer weights of the multi-head attention mechanism. and These represent the three learnable weight matrices in the attention mechanism. This represents the dimension of the query vector Q. This represents the Hadamard product, and T represents the transpose. Channel enhancement feature Average pooling and max pooling are performed along the channel dimension respectively to obtain two single-channel spatial feature maps and , and Concatenate along the channel dimension to obtain a joint spatial map, and then sequentially pass through a 7x7 convolution and a Sigmoid activation function to obtain a spatial attention map Multiply the spatial attention map with the channel enhancement feature pixel by pixel to obtain a spatial enhancement feature map as the output of the meteorological feature extraction module.

5. The method of claim 3, wherein the WRF wind speed is corrected by fusing the land surface cover and the time period encoding. In step 2, the period information extraction module takes the time sequence formed by the coordinate-converted key features as input, extracts the season, month and diurnal information corresponding to each time point in the time sequence, and converts to obtain the season number , diurnal code and month number ; , Corresponding to spring, summer, autumn and winter respectively, spring is from March to May, summer is from June to August, autumn is from September to November, and winter is from December to February; , Corresponding to 1 to 12 months respectively; , 1 for daytime and 0 for night, 8-20 for daytime and 20-8 the next day for night; Based on the season number , diurnal code and month number , respectively, the sine and cosine coding is carried out to capture the periodic characteristics of time, and the sine and cosine coding results are spliced in order to form a periodic time feature vector with a length of 6, which is taken as the output of the period information extraction module. The sine and cosine coding formula is as follows: ， wherein, and denote a seasonal periodicity encoding vector, a diurnal periodicity encoding vector and a monthly periodicity encoding vector, respectively.

6. The method of claim 1, wherein the method is a method of fusing land surface cover and time period encoded WRF wind speed corrections, and In step 2, the geographical feature extraction module takes DEM digital elevation data as input, processes the digital elevation data by using ArcGIS spatial analysis tool, extracts multi-source terrain parameters: elevation A, slope S, aspect, terrain roughness R and slope rate Sr, and the formula is as follows: ， ， ， Among them, altitude A was obtained directly from the DEM. These represent the gradient calculation kernels in the horizontal and vertical directions, respectively. Represents the gradient components in the horizontal and vertical directions. Represents two-dimensional convolution. For slope, Slope direction, For grid points The terrain roughness, For grid points altitude value, Indicates The average elevation of the 3×3 window centered on the center. Set the effective number of grid cells within the window to 9. These represent the rates of spatial variation of slope in the horizontal and vertical directions, respectively. They are respectively The slope of the grid points directly in front and behind. They are respectively The slope of the grid points directly above and below, For digital elevation data grid resolution, For grid points The slope variation rate.

7. The method of claim 3, wherein the WRF wind speed is corrected by fusing the land surface cover and the time period encoding. In step 2, the land cover feature extraction module takes the time series formed by the ESRIs discrete classification land cover data in spring, summer, autumn and winter and the key features after coordinate conversion as input, converts the ESRIs discrete classification land cover data in spring, summer, autumn and winter into continuous semantic vectors through a learnable embedding layer, rearranges through 1x1 convolution to obtain a feature tensor, and then splits the feature tensor into four sub-features of spring, summer, autumn and winter through layer normalization, configures a dedicated extraction branch for each sub-feature, and unifies the structure of each extraction branch as 3x3 convolution, batch normalization and GELU activation function, so as to output independent features F_spr, F_sum, F_aut and F_win of each season; The time series formed by the coordinate-converted key features are seasonally marked, the season marks are converted into vectors through one-hot encoding, and then projected into low-dimensional seasonal semantic vectors through 1-layer full connection , a 4*4 learnable weight matrix is introduced , the seasonal semantic vector is multiplied by the weight matrix , and the dynamic weight is obtained by normalizing through the Softmax function , and the four independent seasonal features are element-wise multiplied and concatenated along the channel dimension to obtain the seasonal perception fusion feature The season-aware fusion features are sequentially subjected to 1x1 convolution, batch normalization, GELU activation, adaptive average pooling and 3x3 convolution to obtain the output of the land cover feature extraction module, i.e., the land cover features.

8. The method of claim 1, wherein the WRF wind speed is corrected by fusing the land surface cover and the time period encoding. In step 2, the meteorological features, time period features, terrain features, and surface cover features are respectively subjected to layer normalization processing and then spliced in the channel dimension to obtain fused features ; Then, the features are fused using two consecutive convolutional blocks. Compression and semantic fusion are performed to obtain the wind speed deviation representation vector. ; to integrate features After performing a 1×1 convolution, it is then compared with the wind speed deviation representation vector. Residual fusion is performed to obtain multi-source heterogeneous fusion features. ; Integrating multi-source heterogeneous features The input is a three-layer lightweight Unet codec structure, which is then upsampled and compressed to generate a 10m wind speed correction result.

9. The method of claim 1, wherein the method is a method of fusing land surface cover and time period encoded WRF wind speed corrections. In step 3, the uncertainty weighted loss function is as follows: ， wherein, represents an uncertainty weighted loss function, and represents a normalization coefficient for the ERA5-land data and the weather station observation data, respectively, and represents a contribution amount for the ERA5-land data and the weather station observation data, respectively.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, When the processor executes the computer program, the steps of the WRF wind speed revision method for fusing land cover and time period coding are realized as claimed in any one of claims 1 to 7.

Images