WRF wind speed correction method fusing land surface coverage and time period coding

Abstract

The invention discloses a WRF wind speed correction method fusing land surface coverage and time period coding, and the method comprises the steps: fusing WRF mode data, ERA5-Land reanalysis data and station observation data, and combining a multi-source data space-time alignment method, a principal component factor screening method, a multi-module cooperative feature extraction method and an uncertainty weighting supervision method. The method is suitable for systematic error correction of 10m wind speed in a WRF mode in a complex terrain area, and solves the problems that the existing WRF mode is insufficient in wind speed simulation precision, obvious in deviation, lack of a targeted correction mechanism and the like in the complex terrain area and an area with strong surface heterogeneity. According to the method, multi-source heterogeneous data such as terrain, earth surface coverage and time periodicity information are fused, and a deep learning method is introduced to construct a multi-module collaborative deviation correction framework, so that the spatial precision and regional adaptability of wind speed forecasting are remarkably improved.

Description

TECHNICAL FIELD

[0001] The application relates to a WRF wind speed correction method fusing surface coverage and time period coding, and belongs to the technical field of meteorological numerical prediction. BACKGROUND

[0002] As one of important parameters of atmospheric movement, wind speed has important application value in the fields of meteorological prediction, wind energy development, aviation and navigation, and city environment evaluation. At present, the WRF model is a mainstream wind speed prediction tool, has a strong physical mechanism basis, and can provide spatially continuous and temporally dynamic evolution wind speed prediction results. However, the WRF model is essentially a physically driven calculation model, and its results are easily affected by initial conditions, physical parameterization schemes, resolutions and terrain processing errors. In actual application, systematic deviations often occur, especially in complex terrain regions and areas with significant surface heterogeneity, and wind speed errors are more obvious.

[0003] In order to improve the accuracy of wind speed prediction, in recent years, researchers have tried to use post-processing techniques to correct the WRF output. Traditional post-processing methods include linear regression, Kalman filtering and Bayesian correction, which have certain effects in simple scenarios, but it is difficult to fully utilize multi-source heterogeneous data, and it is difficult to capture the nonlinear characteristics and spatial dependence structure in the wind speed field. With the development of artificial intelligence technology, more and more researches introduce deep learning methods into numerical prediction error correction, learn and correct the model bias by constructing a nonlinear mapping relationship, and show good potential in wind speed, precipitation, temperature and other variable correction tasks.

[0004] In the wind speed correction task, the factors affecting the wind field structure have multi-source and regional differences. Based on systematic error analysis, it is found that the wind speed error shows significant time dependence and underlying surface sensitivity. Specifically, in the time dimension, the stability of the boundary layer increases at night, resulting in a significant increase in wind speed observation error compared to daytime, causing diurnal differences. The error in spring and winter is more significant than that in summer and autumn due to the frequent cold air activities and the coupling of dynamic-thermal processes in the boundary layer. In the underlying surface dimension, high roughness surfaces such as forests, shrubs and urban building areas have significant wind speed simulation errors compared to flat surfaces due to enhanced turbulent dissipation. The error in low mountain and hilly areas and high mountain areas is significantly higher than that in river valley plains and medium-altitude mountain areas due to the influence of topographic waves and mechanical turbulence. Therefore, it is of great significance to construct an intelligent correction method that can fuse multi-source information and take into account the physical background and data characteristics to improve wind speed accuracy.

[0005] Although the current research has made some progress in single-factor modeling, it generally lacks the ability to model multiple factors such as terrain, surface and time period, and cannot effectively realize the fine-grained regional correction of WRF model bias. Therefore, a wind speed correction method that integrates multi-source data and has spatial perception and time period encoding ability is needed, which can not only use the relationship between high-resolution geographic information and meteorological variables, but also use deep neural network model for end-to-end learning to realize higher accuracy and stronger generalization ability of wind speed forecast correction. SUMMARY

[0006] The technical problem to be solved by the present application is to provide a WRF wind speed correction method that integrates surface coverage and time period encoding, which integrates multi-source heterogeneous data such as terrain, surface coverage and time periodicity information, introduces a deep learning method to build a multi-module collaborative bias correction framework, thereby significantly improving the spatial accuracy and regional adaptability of wind speed prediction.

[0007] The technical problem to be solved by the present application is to provide a WRF wind speed correction method that integrates surface coverage and time period encoding, which integrates multi-source heterogeneous data such as terrain, surface coverage and time periodicity information, introduces a deep learning method to build a multi-module collaborative bias correction framework, thereby significantly improving the spatial accuracy and regional adaptability of wind speed prediction. The WRF wind speed correction method integrating surface coverage and time period encoding comprises the following steps: Step 1, collect WRF meteorological data, ERA5-Land data, meteorological station observation data, DEM digital elevation data and surface coverage data of the region to be corrected, construct a data set, and divide it into a training set and a test set; take the 10m wind speed in the WRF meteorological data as the data to be corrected, and take the 10m wind speed in the ERA5-Land data and the meteorological station observation data as the true value data; Step 2, construct a WRF wind speed correction model, which includes four branches, the first branch extracts meteorological features through a meteorological feature extraction module that integrates multi-scale convolution structure and 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 geographical feature extraction module based on multi-scale convolution and adaptive pooling, and the fourth branch extracts surface coverage features through a surface coverage feature extraction module that introduces surface type embedding vectors and combines cross-scale attention; the meteorological features, time period features, terrain features and surface coverage 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; Step 3, train the WRF wind speed correction model constructed in step 2 using the training set, and use an uncertainty weighted loss function to adaptively balance the two true values during the training process to guide the model training, and obtain the trained WRF wind speed correction model; Step 4, input the WRF meteorological data in the test set, DEM digital elevation data and land surface cover data into the trained WRF wind speed correction model, and output the wind speed correction result.

[0008] Compared with the prior art, the present application has the following technical effects: 1. The present application is aimed at the problem that the WRF model has large simulation error for 10m wind speed under complex terrain and time period change conditions. Through spatial and temporal error analysis, the key error area and period are identified, principal component analysis is used for dimension reduction, 6 key correction factors are selected, meteorological factors, time period, terrain and land surface multi-source information are fused, and a unified wind speed error representation vector is formed in the neural network through the multi-module collaborative extraction and feature fusion strategy, which effectively improves the WRF wind speed correction accuracy and has strong generalization ability.

[0009] 2. The present application constructs a meteorological feature extraction module (MSDNet) fusing double attention mechanism and multi-scale structure, which can improve the modeling ability of wind speed error spatial features. The double attention mechanism can capture the spatial correlation and important features of wind speed error, and the multi-scale structure can better process wind speed error features of different scales, so as to more accurately extract meteorological features and provide more rich information for wind speed correction.

[0010] 3. The present application introduces time period sine-cosine coding, represents time period change based on sine-cosine period coding, guides the lightweight convolutional network to model the nonlinear relationship between wind speed and time period, and can effectively capture the influence of time period change on wind speed, solving the problem of insufficient consideration of time change in the prior art and improving the accuracy of wind speed correction in different seasons.

[0011] 4. The present application introduces a land surface embedded seasonal perception modulator to strengthen land surface heterogeneity modeling. In view of the adjusting effect of land surface type on wind speed, land surface cover features with seasonal specificity and error targeting are extracted. At the same time, combined with feature enhancement, the contribution weight of four seasons features is dynamically generated to improve the local and global structure modeling capability.

[0012] 5. The present application adopts double true value joint supervision and uncertainty weighted loss function. In the supervision mode, ERA5-Land wind speed and station observed wind speed are introduced as double true values, and an uncertainty weighted loss function is introduced to dynamically adjust the weight of the two supervision signals, avoid bias and improve the generalization ability of the model in different regions and time periods. BRIEF DESCRIPTION OF DRAWINGS

[0013] Figure 1 is the overall flowchart of the present application; Figure 2 is the overall framework diagram of the present application; Figure 3is the diurnal wind speed error analysis chart; Figure 4 is the seasonal wind speed error chart; wherein (a) is the spring wind speed error distribution chart, (b) is the summer wind speed error distribution chart, (c) is the autumn wind speed error distribution chart, and (d) is the winter wind speed error distribution chart; Figure 5 is the wind speed simulation error chart of different land cover types; Figure 6 is the terrain wind speed error chart; Figure 7 is the principal component analysis flow chart; Figure 8 is the PCA feature dimension reduction result chart; wherein (a) is the principal component analysis correlation coefficient, (b) is the first four principal component thermal map, and (c) is the principal component cumulative variance explanation; Figure 9 is the meteorological feature extraction module (MSDNet) structure chart; Figure 10 is the cycle information extraction module (CycleNet) structure chart; Figure 11 is the geographic feature extraction module (GeoNet) structure chart; Figure 12 is the land cover feature extraction module (LandNet) structure chart; Figure 13 is the lightweight UNet encoder-decoder structure chart; Figure 14 is the effect comparison chart before and after the wind speed correction in different seasons; wherein (a) is the error comparison chart before and after the spring wind speed correction, (b) is the error comparison chart before and after the summer wind speed correction, (c) is the error comparison chart before and after the autumn wind speed correction, and (d) is the error comparison chart before and after the winter wind speed correction; Figure 15 is the effect comparison chart before and after the wind speed correction in different altitude areas; Figure 16 is the effect comparison chart before and after the wind speed correction in different land cover types; Figure 17 is the effect comparison chart before and after the wind speed correction in different time periods. DETAILED DESCRIPTION

[0014] Embodiments of the present application will be described in detail below, examples of which are shown in the accompanying drawings. The embodiments described below by referring to the accompanying drawings are exemplary and are only used to explain the present application, and cannot be explained as a limitation of the present application.

[0015] In view of the problems of the WRF model in the complex terrain region and the strong surface heterogeneity region, such as insufficient simulation accuracy of 10m wind speed, significant deviation, and lack of targeted correction mechanism, the application provides a multi-module collaborative, feature-guided, lightweight and efficient correction framework, which 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 bias correction framework, thereby significantly improving the spatial accuracy and regional adaptability of wind speed prediction.

[0016] The WRF meteorological data to be corrected, ERA5-Land data and site observation data of the research area (24.6°N-29.3°N, 105.5°E-110.2°E) from 2019 to 2023 are collected, the temporal and spatial error analysis is carried out, and the key error area and time period are identified. The principal component analysis is used for feature dimension reduction, and the meteorological feature extraction module MSDNet, the cycle information extraction module CycleNet, the geographic feature extraction module GeoNet and the land cover feature extraction module LandNet are constructed respectively. The multi-source data features are spliced in the channel, the wind speed map is restored through the decoder, and the high-precision correction result is output combined with ERA5-Land and site observation. The data from 2019 to 2022 are used as the training set, and the data in 2023 are used as the test set, as shown in Figure 1 and Figure 2 The method mainly includes the following steps: Step 1, collect the WRF meteorological data, ERA5-Land data, meteorological site observation data, digital elevation and land cover data of the area to be corrected 24.6-29.3°N, 105.5-110.2°E from 2019 to 2023 with hourly resolution; 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 convective precipitation cumulative cloud, 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, wherein the 10m meridional wind speed component and the 10m latitudinal wind speed component in the WRF data are used as the elements to be corrected, and the 10m meridional wind speed component and the 10m latitudinal wind speed component in the ERA5-Land data are used as the true value data; the temporal and spatial distribution characteristics of wind speed error are analyzed from the time, terrain and surface coverage dimensions, and the significant error area and time period are identified. The method mainly includes the following steps: Step 1.1, the law of wind speed error distribution is analyzed from the time dimension, including classifying the WRF data to be corrected by date according to season and day and night, wherein March to May is spring, June to August is summer, September to November is autumn, and December to February is winter. By comparing with ERA5-Land data and station observation data: calculate the error mean and variance of each season in the region, analyze the error distribution law from the seasonal dimension, and identify the seasons with significant error; calculate the average error of wind speed in the daytime and nighttime. The specific analysis results are shown in Tables 1 and 2.

[0017] Table 1 WRF model 10m wind speed simulation error in different seasons

[0018] Table 2 WRF model 10m wind speed simulation error in day and night

[0019] Step 1.2, the wind speed error distribution is analyzed from the terrain dimension, including classifying the WRF area to be corrected according to the altitude, wherein the ground to 500m is the valley plain, 500m to 1000m is the low mountain and hill, 1000m to 1500m is the medium altitude mountain, and 1500m to 3000m is the high altitude mountain. The specific analysis results are shown in Table 3.

[0020] Table 3 WRF model 10m wind speed simulation error in different terrains

[0021] Step 1.3, the wind speed error distribution law is analyzed from the surface cover dimension, the WRF area to be corrected is classified according to the surface cover type, and the hourly average error of 10m wind speed under various surface cover types is calculated, and the results are shown in Table 4.

[0022] Table 4 WRF model 10m wind speed simulation error in different surface cover types

[0023] Through the above analysis of time period, terrain and surface cover dimension, the WRF model 10m wind speed error has significant distribution law in space and time, mainly affected by three factors of time period, terrain and surface cover. The main conclusions are as follows: The 10m wind speed error in time scale is shown in Figure 3 , 4 , Figure 4The error distribution of the wind speed in spring, summer, autumn, and winter is shown in (a)-(d) in FIG. 6. The conclusion is that the performance in spring is the worst, followed by winter, with an average error of about 1.09 m / s and 0.97 m / s, respectively. The performance in summer and autumn is relatively good, with an average error of 0.67 m / s and 0.64 m / s, respectively. The error in daytime is higher than that in nighttime, with an average error of 1.18 m / s, which is 28.3% higher than that in nighttime. The peak value appears at 10 o'clock (1.82 m / s), which is close to twice the average value in nighttime. The error in nighttime is lower, with an average error of 0.92 m / s, and the minimum value appears at 0 o'clock (0.52 m / s).

[0024] The error performance of the 10m wind speed in the spatial scale is shown in FIG. 7. Figure 5 Figure 6 The error of the 10m wind speed in the high-altitude mountainous area and the low mountainous area is significant, reaching 1.29 m / s and 1.23 m / s, respectively. The error in the valley plain area is 1.03 m / s, and the error in the medium-altitude mountainous area is relatively small, with an error of 0.78 m / s. The areas covered by forests and shrubs have the worst performance, with an average error of 1.9 m / s and 1.25 m / s, respectively. The areas covered by grasslands and buildings have a relatively poor performance, with an average error of 1.2 m / s. The average wind speed error of the remaining coverage types is less than 1 m / s.

[0025] Based on the above error analysis, the error of the 10m wind speed simulated by WRF shows significant differences under different terrain, land cover, season, and day-night conditions. To improve the correction accuracy, the synergistic effect of multiple data sources needs to be considered. By integrating the above multi-source heterogeneous information through feature extraction modules such as MSDNet, GeoNet, CycleNet, and LandNet, the mapping relationship between error and geography, time, and surface features can be established, and high-precision post-correction of WRF wind speed simulation can be realized.

[0026] Step 2: Use principal component analysis method to reduce the dimensionality of WRF meteorological elements, as shown in FIG. 8. Figure 7 The first four elements with high contribution are selected as auxiliary meteorological elements, together with the 10m meridional wind speed component and the 10m zonal wind speed component, to form the key features for wind speed correction. Meanwhile, the key features are projected and converted from Lambert projection to equal latitude and longitude projection, and the coordinates are consistent with ERA5-Land grid data. The unified spatial and temporal resolution is hourly, 0.1°x0.1°, and the data size is 48x48.

[0027] ​Step 2.1, the hourly WRF data is E=8760 samples, each sample contains G=18 meteorological feature variables (downward shortwave radiation SWDOWN, 2m temperature T2, cumulative grid snow and ice amount SNOWNC, latent heat flux LH, upward heat flux HFX, upward longwave radiation OLR, terrain height HGT, specific humidity Q2, surface pressure PSFC, snow water equivalent SNOW, land mask LANDMASK, surface temperature TSK, total convective precipitation cumulative RAINC, total grid precipitation cumulative RAINNC, instantaneous downward shortwave flux SWDNB, instantaneous upward shortwave flux SWUPB, instantaneous downward longwave flux LWDNB, and instantaneous upward longwave flux LWUPB) to form a data matrix For each feature, the mean is 0 and the variance is 1, and the standardized matrix Z is obtained, and the specific formula is as follows: , Wherein, is a matrix representing the original WRF meteorological data, E=8760 corresponds to 1 year of hourly data, and G=18 corresponds to the number of meteorological elements of each sample, represents the th feature value of the e th sample, represents the th feature sample mean, represents the th feature sample standard deviation, represents the standardized feature value, the mean is 0, the variance is 1, and the dimensional difference is eliminated.

[0028] According to the standardized matrix Z, the covariance matrix C and the eigenvalues and eigenvectors are calculated, the eigenvalues are sorted from large to small, and the corresponding first four eigenvectors form the projection matrix P. The standardized data Z is projected onto the first four principal components, as shown in (a)-(c) in the following figure, and finally four wind speed high impact factors are obtained. Joint to be corrected elements as meteorological features for subsequent correction tasks, the calculation formula is as follows: Figure 8 , Wherein, C represents the covariance matrix, which represents the covariance relationship between features, Z and represent the standardized data matrix and the transpose matrix respectively, represents the degree of freedom correction, which ensures the unbiased estimation of the covariance matrix, represents the variance contribution degree corresponding to the principal component, is the eigenvector, which represents the projection direction of the data in the principal axis direction, ​denotes the projection matrix, the eigenvectors corresponding to the first four largest eigenvalues are selected, and the projection matrix is formed by arranging them in columns to retain most of the variance information. denotes the data matrix after dimensionality reduction.

[0029] Step 2.2, projection conversion of the data to be corrected WRF model, WRF model default output Lambert projection, and ERA5-Land data is equal to the latitude projection. The inconsistency of the two projections will cause spatial misplacement and affect data matching. Resampling the uniform spatial resolution, through the bilinear interpolation method, the WRF data after projection conversion is resampled to the equal latitude grid of 0.1°x0.1°, so that it is accurately aligned with the ERA5-Land grid. All data are cropped to the target area range (24.6°N-29.3°N, 105.5°E-110.2°E), and the corresponding grid size is 48x48. The coordinate conversion formula is as follows: , wherein, denotes the coordinates of the WRF output data in the Lambert projection plane coordinate system, denotes the projection center point coordinates of the Lambert projection, denotes the Euclidean distance from the point to the projection center, denotes the polar angle of the point relative to the projection center, denotes the radian latitude of the point , denotes the scale factor of the Lambert projection, denotes the radian longitude of the point , denotes the central meridian of the Lambert projection, corresponding to STAND_LON in WRF parameters. Through this conversion, the WRF data can be accurately aligned with the ERA5-Land on the same grid, avoiding the wind speed correction deviation caused by the difference in projection.

[0030] Step 3, construct a meteorological feature extraction module (MSDNet) that integrates multi-scale perception and double attention mechanism, as shown in Figure 9 The module receives meteorological data with size [B,6,48,48], and through multi-scale convolution feature extraction and channel-space double attention mechanism, effectively enhances the model's response ability to meteorological factors under different scales and regional conditions, and provides discriminative deep meteorological representation.

[0031] Step 3.1, the input data is passed through three parallel convolution branches to extract local, small-scale and mesoscale meteorological response features, respectively. The convolution kernel sizes are 3x3, 5x5 and 7x7, the padding step is 1, and the output channel number of each branch is set to 16. The corresponding output size is [B, 16, 48, 48]. Then, the channel splicing operation is used to form an intermediate feature map with a dimension of [B, 48, 48, 48]. To reduce the computational complexity and maintain semantic richness, a 1x1 convolution is used to compress the 48-channel feature to obtain a feature map with 16 channels and a spatial size of 48x48 . The multi-scale perception convolution calculation formula is as follows: , wherein, represents the input data, represents the neural network training batch size, represents the rectified linear unit activation function, represents the convolution and splicing operation, respectively, and represent local, small-scale and mesoscale meteorological features, respectively, represents the feature map after channel axis splicing, with a size of [B, 48, 48, 48], represents the compressed fusion feature.

[0032] Step 3.2, the compressed fusion feature is subjected to global average pooling in the spatial dimension to obtain a global description vector for each channel. The description vector is mapped through two linear layers to generate query vector Q, key vector K and value vector V, respectively. Each vector retains the original channel dimension of 16, and the channel attention matrix is calculated using the attention mechanism. Specifically, let the number of attention heads be h, then the input channel of each head is 16 / h, and the attention output of all heads is spliced and mapped back to the original channel dimension through a linear layer. Finally, the channel weight vector is generated by Sigmoid function activation . The weight is multiplied with channel by channel to complete the weighted adjustment of the channel dimension, generating the channel enhanced feature . The channel attention calculation formula is as follows:

[0033] Step 3.3, the channel enhanced feature is subjected to average pooling and max pooling along the channel dimension to obtain two single-channel spatial feature maps and , both of which are [B, 1, 48, 48] in size. The two are concatenated along the channel dimension to obtain a joint spatial map, which is input to a padded 7x7 convolutional layer with an output dimension of [B, 1, 48, 48]. The result is activated by a sigmoid function to obtain a spatial attention map for use in conjunction with pixel-wise multiplication to generate a spatially enhanced feature map , which has a dimension of [B, 16, 48, 48] and represents high-order meteorological feature representations enhanced by multi-scale semantics and dual attention. The spatial attention calculation formula is as follows:

[0034] Step 4: Construct a cycle information extraction module (CycleNet) as shown in Figure 10 , which obtains time series information, extracts the season, month, and day-night information corresponding to each time point from the time variables in the WRF model and ERA5-Land reanalysis data, and further converts them to obtain season numbers , which correspond to the four seasons of spring, summer, autumn, and winter, respectively, and are used to express macro seasonal information; the month information is an integer value ranging from 1 to 12; the day-night information is an integer value , which is used to express fine-grained time periods.

[0035] The cycle information extraction module includes season numbers , day-night encodings , and month numbers , which are respectively encoded using sine and cosine to capture the periodic characteristics of time, including the four seasons, day and night, and monthly cycles. Perform sine and cosine transformations, concatenate the six encoding results in order to form a periodic time feature vector with a length of 6 and a size of [6, 48, 48], which is used to express the time semantic information of wind speed correction. The specific calculation formula is as follows:

[0036] Step 5: Construct a geographical feature extraction module (GeoNet) as shown in Figure 11 , which uses the Sobel operator to calculate terrain slope and slope direction, uses the ArcGIS window statistical function to calculate terrain roughness R, and uses the quadratic gradient solving method to calculate slope variability Sr. The altitude, slope, slope direction, terrain roughness, and slope variability are concatenated along the channel and input to the module. Through multi-source terrain parameter fusion and terrain complexity adaptive dynamic extraction, precise geographical feature capture is achieved, and a 48x48 scale DEM feature map is output. The module includes the following steps: Step 5.1, terrain information extraction and feature construction, using ArcGIS spatial analysis tools to process digital elevation data, directly obtaining elevation A, using Sobel operator to perform convolution operation on DEM in horizontal and vertical directions respectively, obtaining gradient components dx and dy of each pixel, calculating slope Slope and aspect, and normalizing elevation A, slope Slope and aspect to [0, 1] interval, calculating terrain roughness R through ArcGIS window statistics function, calculating slope rate Sr through quadratic gradient solving method, completing full extraction of multi-source terrain parameters (A, Slope, Aspect, R, Sr). Concatenate the above 5 types of terrain parameters along the channel dimension to obtain multi-source terrain input features of [B, 5, 256, 256], the specific calculation formula is as follows: , , , Among them, represents the average value of 3x3 window elevation, and S represents the slope.

[0037] Step 5.2, basic feature encoding, using a double-branch parallel encoding structure to extract initial terrain semantic representation F0, which specifically includes: branch one performs 3x3 convolution on the concatenated [B, 5, 256, 256] feature, followed by batch normalization and ReLU activation function, to strengthen local terrain spatial correlation feature extraction; branch two performs 1x1 convolution on the same concatenated feature, with 32 convolution kernel numbers, step 1, zero padding, and realizes semantic fusion of terrain parameters through channel dimension linear transformation; element-wise sum of the output features of branch one and branch two, to obtain the initial terrain semantic representation F0 with dimension [B, 32, 256, 256], which retains local spatial details and strengthens semantic correlation between channels.

[0038] Step 5.3, terrain complexity classification and mask generation, based on terrain roughness R and slope rate Sr to construct terrain complexity index Tc, the calculation formula is as follows: , Among them, the weight coefficients 0.4, 0.6 are set according to physical mechanism, the slope rate Sr reflects the degree of terrain fluctuation, and the terrain roughness R reflects the surface friction coefficient. According to the range of Tc, three terrain grades are divided, and the corresponding binary masks are generated. Tc<0.2 is low complexity terrain, mask is Tc_low, 0.2≤Tc≤0.6 is medium complexity terrain, mask is Tc_mid, Tc>0.6 is high complexity terrain, mask is Tc_high, three types of masks maintain [B, 1, 256, 256] dimension, which is completely aligned with the spatial scale of F0.

[0039] Step 5.4, terrain adaptive dynamic extraction, based on terrain complexity level to dynamically match different receptive field of dilated convolution, to achieve accurate capture of multi-scale geographic features, including: The low complexity branch inputs F0 into a 3x3 dilated convolution with a dilation rate of 1 and padding of 1, followed by batch normalization and a GELU activation function. The small receptive field focuses on the local microscopic details of low relief terrain. The medium complexity branch inputs F0 into a 3x3 dilated convolution with a dilation rate of 3 and padding of 3, followed by batch normalization and a GELU activation function. The medium receptive field balances the local details of hills and low-altitude mountains. The high complexity branch inputs F0 into a 3x3 dilated convolution with a dilation rate of 5 and padding of 5, followed by batch normalization and a GELU activation function. The large receptive field captures the global terrain structure of medium and high-altitude mountains and steep slopes. The output features of the three branches are multiplied element-wise with the corresponding complexity mask to retain only the effective features of the corresponding complexity region. The filtered three types of features are added element-wise to obtain adaptive dynamic extraction features of [B, 32, 256, 256]. The calculation formula of the above-mentioned dilated convolution is as follows: , wherein, represents the basic feature encoding output feature, represents the dilated convolution kernel, which is obtained by model initialization and training learning, t represents the dilation rate, which can be 1, 3, and 5, corresponding to Tc_low, Tc_mid, and Tc_high, represents the output feature of the corresponding branch.

[0040] Step 5.5, feature refining and scale adaptation, input the adaptive dynamic extraction features into the feature refining unit, which consists of a 1x1 convolution with 32 kernel numbers, a step of 1, zero padding, batch normalization, and a GELU activation function. The feature refining strengthens the core geographic features. Finally, the 256x256 spatial scale is compressed to 48x48 by adaptive average pooling, and the final output is a DEM feature map of [B, 32, 48, 48].

[0041] Step 6, construct the land cover feature extraction module (LandNet), as shown in Figure 12 , input the spring, summer, autumn, and winter ESRIs discrete classification land cover data and the actual seasonal identifier corresponding to the sample into the dual input. The core consists of a seasonal perception modulator and a feature compression part. The land cover features are extracted with seasonal specificity and error targeting. The output land cover feature map includes the following steps: Step 6.1, Season-aware Modulator, implements seasonal dynamic adaptation and feature enhancement, the ESRI classified four-season land cover data is spliced in order into a tensor of [B, 4, 256, 256] dimensions, while a sample season identification tensor of [B, 1] dimensions is constructed to mark the actual meteorological season corresponding to each input sample, with discrete values 1-4 corresponding to spring, summer, autumn, and winter, respectively. The discrete classification values of each season are converted into continuous semantic vectors of [B, 4, 256, 256, 16] through a learnable embedding layer, and then passed through a 1x1 convolution and rearrangement to obtain a tensor of [B, 128, 256, 256]. The features are split into 4 [B, 32, 256, 256] sub-features through layer normalization, corresponding to the four seasons of spring, summer, autumn, and winter. Each sub-feature is configured with a dedicated extraction branch, and each branch structure is uniformly 3x3 convolution, batch normalization, and GELU activation function. The convolution kernel weights of each branch are independently initialized and updated. Finally, four independent seasonal features F_spr, F_sum, F_aut, and F_win are output, each with a size of [B, 32, 256, 256] dimensions. The following is the calculation formula of layer normalization: , where represents the feature mean, represents the feature variance, are learnable scaling and bias parameters, is a small value to avoid division by zero.

[0042] Step 6.2, dynamically generate the contribution weight of four-season features based on the season identification of the input sample. The season identification of [B, 1] is converted into a [B, 4] vector through one-hot encoding, and then projected into a low-dimensional seasonal semantic vector of [B, 4] through a 1-layer fully connected network . A 4x4 learnable weight matrix is introduced, and the seasonal semantic vector is multiplied by the weight matrix to obtain dynamic weights and through the Softmax function normalization. Broadcast to [B, 4, 32, 256, 256], and multiply element-wise with the four independent seasonal features to obtain a seasonal-aware fusion feature with a size of [B, 128, 256, 256] along the channel dimension.

[0043] ​Step 6.3, feature compression, realizes dimension adaptation and core feature refinement, performs 1x1 convolution on the seasonal perception fusion feature, the number of convolution kernels is 64, the step is 1, zero padding, and is followed by batch normalization and GELU activation, and outputs refined features of [B, 64, 256, 256]. Further, an adaptive average pooling technique and a 3x3 convolution layer are used to compress high-resolution features into mesoscale features of [B, 32, 48, 48] that retain regional-level surface structure information. The module takes four seasonal high-resolution land cover classification data and meteorological data as input.

[0044] Step 7, build a multi-source data fusion module to integrate multi-source heterogeneous feature information, including meteorological features, time period features, geographic features, and land cover features, and model the spatial distribution of wind speed bias uniformly, which includes the following steps: Step 7.1, multi-source feature standardization and size alignment. The meteorological features, time period features, geographic features, and land cover features are spliced in the channel dimension to form a fusion input feature tensor. Each branch output feature is normalized before input to ensure numerical consistency and avoid gradient drift during fusion:

[0045] Step 7.2, channel compression and semantic fusion. The 86-channel high-dimensional feature after splicing is compressed and semantically fused through two consecutive ConvBlock blocks. Each module contains a 3x3 convolution layer, a batch normalization layer, and a ReLU activation function to extract joint expression relationships between cross-modal features. Finally, it is compressed to 64 channels to obtain a unified wind speed bias representation vector :

[0046] Step 7.3, introduce residual connection and attention reinforcement. The high-dimensional spliced feature is compressed to 64 channels through a 1x1 convolution, and is added to to form a residual fusion output unified multi-source heterogeneous fusion feature:

[0047] Step 8, build a lightweight UNet encoding-decoding structure as shown in Figure 13 , which takes multi-source heterogeneous fusion features as input, uses an encoder-decoder structure to realize nonlinear modeling and spatial information recovery of fusion features, outputs revised high-precision 10m wind speed revision results, and uses a double-truth joint supervision mechanism and an uncertainty weighted loss function for training, which includes the following steps: Step 8.1, the encoder adopts a 3-layer convolution + ReLU block for feature compression and abstract representation, with channel numbers of 64→128→256, convolution kernel size of 3×3, and stride of 2 to realize spatial down-sampling. The bottleneck module uses two lightweight convolution residual blocks Residual Block, which consists of two 3×3 convolution layers and a short circuit connection, effectively enhancing the nonlinear modeling capability and keeping the feature size and channel number unchanged. The decoder uses 2 layers of deconvolution for up-sampling, with channel numbers gradually decreasing from 256→128→64, and sizes recovering layer by layer from 6×6→12×12→24×24, and using jump connection with the corresponding layer of the encoder for feature fusion to preserve edge and detail information. Finally, up-sampling and compressed channels are used to restore the original resolution, and the final wind speed correction map [B, 48, 48] is obtained.

[0048] Step 8.2, joint supervision mechanism and uncertainty weighted loss function design. To utilize ERA5-Land reanalysis data and ground station observation information simultaneously, a dual-truth 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. The learnable uncertainty parameters and are set for the ERA5-Land and station observation error branches respectively, and the total loss function is derived by maximum likelihood estimation to make the model adaptively adjust the weights of the two types of loss during training. This mechanism effectively alleviates the training imbalance problem caused by data heterogeneity and improves the overall performance of wind speed correction results in large-scale structure and local accuracy. The final total loss function is as follows:

[0049] Step 8.3, model optimization to realize WRF wind speed correction, using the backpropagation algorithm based on Adam optimizer to train the whole process model end-to-end. The weighted loss function constructed by the joint supervision mechanism is used as the optimization objective to guide the model to learn the correction relationship between large-scale wind field structure and local fine error. According to the principle of minimum comprehensive loss on the validation set, the model parameter file with the best performance is automatically saved. The high-precision post-processing correction of WRF simulated 10m wind speed in any region is realized.

[0050] The application proposes a WRF wind speed revision method fused with surface cover and time period coding, aiming to improve the precision of WRF mode simulation of 10m wind speed. The method collects nearly 5 years of meteorological data to be revised in the study area, ERA5-Land data and site observation data, performs spatial and temporal error analysis, and identifies key error areas and time periods. Principal component analysis is used for feature dimension reduction, and meteorological feature extraction module MSDNet, cycle information extraction module CycleNet, geographic feature extraction module GeoNet and surface cover feature extraction module LandNet are constructed respectively. The channel splicing of multi-source data features is output through the lightweight UNet network, and the high-precision revision result is output combined with ERA5-Land and site observation.

[0051] To verify the actual role of each module in the revision process. Different module combinations are used to process the WRF data to be revised, and ablation experiments are carried out. A-G is set as 7 groups of control experiments, epoch is set as 20, and the training loss and validation loss of each group of experiments are observed. The specific test settings and performances are shown in Table 5.

[0052] Table 5 Control test settings

[0053] The experiment system evaluates the influence of multi-module combination on the wind speed revision model. The benchmark model MSDNet (A) has a verification loss of 0.3284. Adding CycleNet (B), GeoNet (C) or LandNet (D) module alone can improve the performance, among which GeoNet has the best effect, and the verification loss is 0.2964. In the double module combination, MSDNet+CycleNet+GeoNet (E) performs best, and the verification loss is 0.2850. The complete revision model (G) fuses all modules and achieves the best effect, with a verification loss of 0.2817, which verifies the effectiveness of multi-feature fusion.

[0054] To verify the revision effect of the application in the dimensions of season, day and night, terrain and surface cover, the complete revision model (G) is used to revise the 10m wind speed simulation results of the WRF mode in 2023. The experiment uses all-year data, and analyzes according to season, day and night, terrain and surface cover. The revision effect of the model in each dimension is systematically evaluated. The results are shown in Tables 6-9 and related figures, and the model has achieved significant improvement in all dimensions. Tables 6 and Figure 17 show the revision performance of day and night; Tables 7 and Figure 14 show the revision performance of the revision model in the seasonal dimension, Figure 14 (a)-(d) in Table 8 and Figure 15 show the revision performance of different terrains, and Tables 9 and Figure 16The performance before and after the wind speed revision for different land surface coverage types.

[0055] Table 6 Day and night revision performance

[0056] Table 7 Revision performance in seasonal dimension

[0057] Table 8 Revision performance for different terrains

[0058] Table 9 Performance before and after the wind speed revision for different land surface coverage types

[0059] The present application significantly improves the accuracy and reliability of wind speed prediction through comprehensive and multi-level data processing and innovative modeling methods, and has wide application prospects and practical value.

[0060] Based on the same inventive concept, an embodiment of the present application provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the steps of the aforementioned WRF wind speed revision method fusing land surface coverage and time period coding when executing the computer program.

[0061] Based on the same inventive concept, an embodiment of the present application provides a computer readable storage medium, which stores a computer program, and the computer program implements the steps of the aforementioned WRF wind speed revision method fusing land surface coverage and time period coding when executed by a processor.

[0062] Those skilled in the art will understand that embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage, etc.) containing computer-usable program code.

[0063] The present application is described with reference to flowcharts according to the method, device (system), and computer program product of the embodiments of the present application. It should be understood that each flow in the flowcharts, and the combination of the flows in the flowcharts, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a machine that implements the functions described in the flowcharts.Figure 1 an apparatus that performs the function specified in the flow or flows.

[0064] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the flow Figure 1 an apparatus that performs the function specified in the flow or flows.

[0065] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the flow Figure 1 an apparatus that performs the function specified in the flow or flows.

[0066] The above embodiments are merely intended for describing the technical idea of the present application, and cannot be used to limit the protection scope of the present application. Any modification made according to the technical idea of the present application, on the basis of the technical scheme, falls within the protection scope of the present application.

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 surface 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; Step 3, train the WRF wind speed correction model constructed in step 2 using the training set, and use an uncertainty weighted loss function to adaptively balance the two true values during the training process to guide the model training, thereby obtaining a 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 to output wind speed correction results.

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 are collected, wherein the WRF meteorological data includes 10m meridional wind speed component, 10m zonal 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 accumulation, total grid precipitation accumulation, 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; and 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 the step 2, before the meteorological feature and time period feature extraction, the principal component analysis method is used to reduce the dimension of the remaining WRF meteorological data except the 10m meridional wind speed component and the 10m zonal wind speed component, and the first four meteorological factors with the greatest influence on the wind speed are selected from the 18 meteorological factors affecting the wind speed as auxiliary meteorological factors, which together with the 10m meridional wind speed component and the 10m zonal wind speed component form the key features of the wind speed correction; and the key features are converted from Lambert projection to latitude and longitude projection, which is consistent with the ERA5-Land data coordinates, and the spatial and temporal resolutions are 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. 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.

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 the step 2, the meteorological feature extraction module includes three parallel convolution branches, which extract local, small-scale and mesoscale meteorological features respectively, and the convolution kernel sizes of the three parallel convolution branches are 3x3, 5x5 and 7x7 respectively; the input of each convolution branch is a 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 , multiply the channel weight with the intermediate feature spliced along the channel axis to obtain the channel enhanced feature ; The specific formula is as follows: ， wherein, denotes a channel dimension component of the global description vector z, denotes the height and width of the compressed feature map respectively, denotes the spatial pixel value of the feature map at the spatial position , denotes a Sigmoid activation function, denotes the output linear layer weight of the multi-head attention mechanism, and denotes three learnable weight matrices in the attention mechanism respectively, denotes the dimension of the query vector Q, denotes the Hadamard product, T denotes 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 , The spatial attention map Pixel by pixel multiplication 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 day-night information corresponding to each time in the time sequence, and converts to obtain the season number , day-night code and month number ; , corresponding to spring, summer, autumn and winter respectively, March to May is spring, June to August is summer, September to November is autumn, and December to February is winter; , corresponding to January to December respectively; , 1 for day and 0 for night, 8-20 for day and 20-8 the next day for night; Based on season number , day and night coding , 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 as the output of the periodic information extraction module, and 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. In the 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, and 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 variability; the above multi-source terrain parameters are spliced ​​along the channel dimension to obtain multi-source terrain input features; An initial terrain semantic representation F0 is extracted by using a double-branch parallel coding 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 levels 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 by using a low complexity branch, a medium complexity branch and a high complexity branch respectively, 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 branch, the medium complexity branch and the high complexity branch 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, medium and 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 geographical feature extraction module, i.e., the terrain features.

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 the 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 1*1 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 3*3 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 and are obtained by normalizing through the Softmax function , and the seasonal perception fusion features are obtained by element-wise multiplication of the four independent seasonal features and channel dimension splicing The season-aware 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, 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 the step 3, the uncertainty weighted loss function is as follows: ， wherein, represents an uncertainty weighted loss function, and and and represent the contribution of 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, The processor executes the computer program to realize the steps of the WRF wind speed revision method for fusing land cover and time period coding according to any one of claims 1 to 9.

Images