A method for correcting incoming wind speed of offshore wind farm based on SAR satellite observation

Abstract

This invention discloses a method for correcting incoming wind speed in offshore wind farms based on SAR satellite observations. First, SAR satellite images are preprocessed and masked to remove interference from bright targets such as wind turbines. Then, a high-resolution near-sea surface wind speed is retrieved from the images using an improved ConvNeXtv2 deep learning network. This is then matched with numerical weather prediction (NMR) wind fields, and the amplitude and spatial structure of the NMR wind fields are corrected using the LightGBM model. Finally, the wind shear index is corrected by combining atmospheric elements at the wind height level with the wind speed extrapolation formula, thereby correcting the extrapolation ratio and obtaining high-precision wind speed extrapolation results at the wind turbine hub height. This invention integrates the high spatial resolution of SAR with the continuity of numerical weather prediction, effectively eliminating wind turbine target interference and numerical weather prediction system bias, significantly improving the accuracy of incoming wind speed at the wind turbine hub height, and providing reliable wind field data for offshore wind farm power prediction and safe operation.

Description

Technical Field

[0001] This invention relates to a method for correcting the incoming wind speed of an offshore wind farm, specifically a method for correcting the incoming wind speed of an offshore wind farm based on SAR satellite observations. Background Technology

[0002] Offshore wind power is an important direction for the development of clean energy, and its construction and operation are gradually expanding to deep water and offshore areas. In the complex environment of the deep sea, the sea surface wind field is a key power input affecting the output characteristics, operational safety, and lifespan assessment of wind turbines. High-precision and high-reliability incoming wind speed information plays a crucial role in wind farm power prediction, performance evaluation, layout optimization, and operation and maintenance decisions.

[0003] Currently, engineering practice mainly relies on Numerical Weather Prediction (NWP) models to provide large-scale, continuous spatiotemporal wind field data. However, NWP wind fields have significant limitations in practical applications: First, wind speed and turbine output power have a cubic relationship, and even small systematic errors in NWP wind speed can lead to significant deviations in power prediction; second, limited by boundary layer parameterization schemes, initial field accuracy, and model resolution, NWP often exhibits persistent systematic biases in densely populated offshore wind farm areas; third, its spatial resolution is relatively coarse (typically on the order of kilometers), making it difficult to finely characterize the spatial distribution and variation characteristics of wind speed within the wind farm scale (such as the wake effect between turbines and the land-sea transition zone effect), thus limiting its application effectiveness in wind farm-level micro-analysis and refined operation.

[0004] Spaceborne Synthetic Aperture Radar (SAR) possesses all-weather, day-and-night observation capabilities and can provide high spatial resolution sea surface backscattered images. Sea surface wind speed can be retrieved by analyzing the physical relationship between sea surface roughness and wind field. This, to some extent, compensates for the shortcomings of NWP in spatial detail description. However, SAR-based wind speed retrieval technology still faces a series of key challenges when practically applied to offshore wind farm inflow assessment:

[0005] 1. Height mismatch: SAR inverted wind speed usually corresponds to a reference value of 10 meters above the sea surface, while the hub height of modern large offshore wind turbines is generally over 100 meters. Directly using SAR wind speed cannot meet the needs of hub height inflow assessment.

[0006] 2. Time discontinuity: SAR observations are constrained by the satellite revisit cycle, resulting in low time resolution. They cannot provide continuous time series of wind fields and are difficult to independently support real-time monitoring and short-term power prediction of wind farm operation.

[0007] 3. Data Contamination and Noise: The inherent speckle noise in SAR images affects the accuracy of inversion. More importantly, wind turbine towers, rotating blades, and auxiliary facilities within the wind farm area generate strong backscattering signals, forming "bright targets" that severely interfere with the extraction of the true wind-generated roughness characteristics of the sea surface, resulting in unreliable wind speed inversion results within the wind farm area.

[0008] 4. Existing fusion methods lack depth: Although the necessity of fusing high-resolution SAR observations with the temporal continuity of NWP is recognized, existing correction methods are mostly simplistic and fail to fully utilize SAR data for systematic and multi-dimensional joint correction of the NWP wind field. In particular, there is a lack of effective mechanisms to simultaneously correct the amplitude deviation and spatial structure error of the NWP wind field, and the filtering of bright target interference in the wind farm area is not adequately considered. As a result, the accuracy and reliability of the fused wind field still have considerable room for improvement.

[0009] Therefore, existing technologies have significant limitations in providing high-precision, high-spatiotemporal resolution hub height and inflow wind speeds suitable for offshore wind farm operation analysis and power assessment. Effectively removing interference from wind turbine structures in SAR imagery, establishing a high-precision sea surface wind speed inversion model, and then deeply integrating the spatial detail advantages of SAR with the spatiotemporal continuity and vertical profile information of NWP (Near-Wave Wind) systems to achieve joint correction of systematic deviations and spatial structures in NWP wind farms, ultimately obtaining accurate and reliable hub height and inflow wind speeds, has become a key technical problem urgently needing to be solved to improve the intelligent operation level of offshore wind farms. Summary of the Invention

[0010] To address the shortcomings of existing technologies, this invention discloses a method for correcting the incoming wind speed of offshore wind farms based on SAR satellite observations. The technical solution is as follows: A method for correcting the incoming wind speed of an offshore wind farm based on SAR satellite observations, characterized by the following steps:

[0011] S1: Acquire multi-temporal synthetic aperture radar (SAR) satellite image data covering the target offshore wind farm area, and perform a series of preprocessing steps on the SAR satellite image data to obtain standardized backscattering coefficient images;

[0012] S2: Based on the backscattering coefficient image, identify and mask the bright target interference area caused by the wind farm turbine array and its auxiliary facilities;

[0013] S3: Based on the masked SAR satellite image data, construct and train a deep learning model for wind speed inversion, and introduce radar line of sight as a geometric auxiliary condition for feature supplementation; use the trained deep learning model for wind speed inversion to process the masked SAR satellite image data and invert the first wind speed field at the near sea level.

[0014] S4: Obtain numerical weather prediction (NWP) wind field data that is spatiotemporally matched with the SAR satellite image data, align the first wind speed field with the near-sea wind field in the NWP wind field data spatiotemporally, establish a correction model based on the aligned data, and use the correction model to correct the NWP near-sea wind field to obtain the corrected second wind speed field.

[0015] S5: The vertical wind speed shear index is corrected by combining upper-level meteorological elements. Based on the corrected extrapolation model, the second wind speed field is extrapolated to the hub height of the wind turbine to obtain the incoming wind speed field at the hub height.

[0016] The present invention also discloses an incoming wind speed correction system for offshore wind farms based on SAR satellite observation. The system is used to execute the aforementioned method for correcting incoming wind speeds for offshore wind farms based on SAR satellite observation. The system is characterized by comprising: a data acquisition and preprocessing module for executing step S1; a bright target identification and masking module for executing step S2; a high-resolution wind speed inversion module for executing step S3; a wind field fusion correction module for executing step S4; and a wind speed vertical extrapolation module for executing step S5.

[0017] The present invention also discloses an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above method. Attached Figure Description

[0018] Figure 1 This is a flowchart of the method of the present invention.

[0019] Figure 2 Scatter density map of inversion results based on ResNet-18 network;

[0020] Figure 3 Scatter density plot of evaluation results obtained from ConvNeXtv2 backbone network inversion;

[0021] Figure 4 This is a masking result of a bright target in a VV polarimetric SAR image on a certain day.

[0022] Figure 5 The graph shows the extrapolated wind speed from the hub height after wind shear index correction. Detailed Implementation

[0023] Example 1

[0024] A method for correcting incoming wind speed in offshore wind farms based on SAR satellite observations includes the following steps:

[0025] S1: Acquire multi-temporal synthetic aperture radar (SAR) satellite image data covering the target offshore wind farm area, and perform a series of preprocessing steps on the SAR satellite image data to obtain standardized backscattering coefficient images;

[0026] This step aims to acquire high-quality SAR satellite imagery data and perform a series of preprocessing operations to eliminate systematic errors, suppress noise, and standardize the data format and coordinate system, laying the data foundation for subsequent wind speed inversion. The specific implementation includes the following two main parts:

[0027] 1. Acquisition and preliminary screening of SAR image data

[0028] Data Source and Mode Selection: The study area is the target offshore wind farm and its surrounding waters. Multi-temporal spaceborne synthetic aperture radar (SAR) satellite imagery is acquired. The preferred method is to use dual-polarization (VV and VH polarization) ground-range detected (GRD) products acquired by satellites such as Sentinel-1 in interferometric wide (IW) imaging mode. Dual-polarization data helps to comprehensively utilize the sensitivity of different polarization modes to sea surface scattering mechanisms, obtaining richer information on sea surface conditions. The GRD product has undergone multi-view processing based on single-view complex (SLC) data, improving the stability of intensity information and effectively reducing speckle noise.

[0029] Temporal and Quality Control: SAR images covering typical wind conditions, different seasons, and various operating states of wind turbines are collected, typically spanning years. Before data download, preliminary visual retrieval and screening are performed to remove images whose backscattering signals from the sea surface are distorted due to large-scale meteorological phenomena (such as heavy rain) or marine phenomena (such as oil spills), ensuring the quality of data used for subsequent analysis.

[0030] 2. Systematic preprocessing of SAR image data

[0031] After acquiring the raw SAR image data of the target sea area, the SAR image is subjected to standardized preprocessing to obtain a backscattering coefficient image for wind speed inversion. The standardized preprocessing includes one or more of the following: orbit correction, noise removal, radiometric calibration, speckle noise suppression, geometric correction, target region cropping, and decibel conversion.

[0032] Specifically, firstly, orbit correction is performed on the SAR image based on precise orbit data to improve the spatial positioning accuracy of the image; then, boundary noise and thermal noise are removed from the SAR image to reduce the impact of non-sea surface scattering signals and system noise on subsequent feature extraction; subsequently, radiometric calibration is performed on the image according to sensor calibration parameters, converting the original digital quantization values ​​into physically meaningful backscattering coefficients. The radiometric calibration process can be expressed as:

[0033]

[0034] Wherein, DN represents the image grayscale value. This represents the noise parameter, and A represents the backscattering correction parameter, taken from the corresponding file of the downloaded SAR image. This is the result for the backscattering coefficient.

[0035] To suppress speckle noise caused by SAR coherent imaging, adaptive speckle noise filtering is applied to the radiometrically calibrated backscattering coefficient image. Then, the image is geometrically corrected by combining orbital information and geographic reference information, and cropped according to the target wind farm or target sea area to obtain a subset of the target area image.

[0036] Furthermore, the backscattering coefficients in the linear domain are converted to decibels to obtain a decibel-domain backscattering coefficient image:

[0037]

[0038] In the formula, The backscattering coefficient is... The backscattering coefficient value represents the decibel level.

[0039] Through the above systematic data acquisition and preprocessing process, a batch of SAR backscattering coefficient images with accurate geographic coordinates, standardized radiometric measurements, controlled noise, and focused on the target sea area were finally obtained, providing high-quality data input for the next step of identifying wind farm interference targets and retrieving sea surface wind speed.

[0040] S2: Based on the backscattering coefficient image, identify and mask the bright target interference area caused by the wind farm turbine array and its auxiliary facilities;

[0041] This step aims to accurately identify strong scattering signals (bright targets) generated by wind turbines and their auxiliary facilities in SAR images, thereby isolating such non-sea surface wind-induced signals in subsequent processing and preventing them from contaminating sea surface wind speed inversion.

[0042] S2.1: Extracting candidate bright target pixels based on statistical thresholds

[0043] Calculation and Threshold Setting: The preprocessed VV polarization (vertical transmission, vertical reception) decibel-modulated backscattering coefficient image is used as the primary data source. First, the mean and standard deviation of the backscattering coefficients of all pixels in the entire image or target area are calculated. Then, the standardized statistical value is calculated for each pixel using the following formula:

[0044]

[0045] in, This represents the VV polarization backscattering coefficient of a pixel. This represents the mean of the VV polarization backscattering coefficients across all samples. Let be the standard deviation, and z be the standardized statistic.

[0046] Candidate point extraction: A threshold k is set, typically ranging from 3.0 to 5.0; in this embodiment, k=4.5 is preferred. Pixels with z>k are identified as "extremely bright candidate points." These pixels have significantly higher backscattering intensity than the background sea surface, and are highly likely to correspond to strong scattering objects such as wind turbine towers, rotating blades, ships, or platforms. All candidate points are marked as 1 (foreground), and the remaining pixels are marked as 0 (background), generating an initial binary detection map.

[0047] S2.2: Morphological dilation and connected component analysis

[0048] Morphological dilation: Due to the characteristics of SAR imaging and the fact that a single wind turbine target may appear as a scattering cluster composed of main lobes and side lobes in the image, the initial binary image may be fragmented. To connect multiple pixels belonging to the same physical target into a continuous region, a morphological dilation operation is employed. Specifically, a 3×3 square structuring element is used to traverse the image: if at least one foreground pixel (value 1) exists in the neighborhood covered by the structuring element, the center pixel of the output image is set to 1. This operation can moderately expand the candidate bright target region, ensuring that the pixel set corresponding to a single wind turbine target is spatially connected.

[0049] Connected component labeling: Perform connected component analysis on the dilated binary image, assign a unique label to each interconnected foreground pixel group, and count the total number of pixels contained in each connected component, i.e., the area of ​​the connected component.

[0050] S2.3: Area Filtering and Mask Generation

[0051] Setting an area filtering threshold: Not all bright connected regions are wind turbine targets of interest. Connected regions that are too small may be noise, while those that are too large may be other large structures or anomalous areas. Therefore, based on the typical pixel size range of wind turbine targets at a specific image resolution, a preset area filtering interval [area_min, area_max] is established, where area_min is the minimum area and area_max is the maximum area. For example, based on experimental analysis, area_min = 3 and area_max = 400 can be set.

[0052] Generate the final mask: Traverse all connected components, retaining only those with areas within the preset interval [area_min, area_max]. These retained connected components are considered to correspond to artificial bright targets that need to be avoided, such as wind turbine arrays, ships, or fixed platforms. The union of all connected components that pass the area filtering is used to generate the final bright target mask file. In this mask file, the target area is marked with a specific value (e.g., 1), while the background sea surface area is marked with another value (e.g., 0).

[0053] The above steps enable a systematic and adaptive detection of strong interference targets introduced by wind farm facilities from SAR imagery, and generate accurate masks. This effectively eliminates the adverse effects of artificial structures on the accuracy of sea surface wind field inversion, and significantly improves the reliability and consistency of wind speed inversion results within the wind farm area.

[0054] S3: Based on the masked SAR satellite image data, construct and train a deep learning model for wind speed inversion, and introduce radar line of sight as a geometric auxiliary condition for feature supplementation; use the trained deep learning model for wind speed inversion to process the masked SAR satellite image data and invert the first wind speed field at the near sea level (i.e., the near sea level wind speed field based on SAR image, defined as the first wind speed field).

[0055] This step aims to train a deep learning model for wind speed inversion based on masked SAR image data, and then use the trained model to retrieve high spatial resolution near-shore wind fields. The process includes two stages: model building and training, and wind speed inversion application.

[0056] (I) Model Construction and Training

[0057] The wind speed inversion deep learning model described in this invention employs an improved ConvNeXtv2 network architecture. Its construction and training process is as follows:

[0058] 1) Constructing the network input tensor. The preprocessed sub-image dataset contains VV polarization backscattering coefficients, VH polarization backscattering coefficients, incident angle, and multi-channel raster data of bright target mask files, forming a tensor with a size of [size missing]. The input tensor is defined as follows: B is the batch size, C is the number of channels, and H and W are the height and width of the image. The mean and standard deviation of the VV-polarized backscattering coefficients, VH-polarized backscattering coefficients, and incident angle are calculated on the training set to standardize the input data and improve the network's stability.

[0059] 2) Constructing a SAR image feature extraction branch. The backbone network includes a feature embedding module, multiple hierarchical feature extraction stages, and a multi-scale feature aggregation module. The feature embedding module maps the multi-channel SAR sub-image input to a high-dimensional feature space and performs initial downsampling to obtain a basic feature map. Subsequently, it sequentially performs multiple stages of hierarchical convolutional feature extraction. Each stage consists of several stacked convolutional blocks used to progressively extract multi-scale spatial features characterized by sea surface backscattering intensity, polarization differences, and incident angle modulation effects. Downsampling layers are set between adjacent stages to gradually reduce the spatial resolution of the feature map and improve the channel-dimensional representation capability, thereby enhancing the modeling ability for a wider range of spatially dependent and scale-invariant features.

[0060] In this embodiment, the ConvNeXtv2 network adopts a four-stage hierarchical structure. The stage depth parameter is set to depth=[2,2,6,2], the stage channel width is set to hidden_sizes=[96,192,384,768], and the activation function is GELU. The feature embedding module maps the input to a 96-dimensional feature space and performs initial downsampling. Subsequent stages use downsampling layers to halve the feature map space size and double the number of channels.

[0061] 3) ConvNeXtv2 basic unit structure. First, k×k depthwise separable convolutions are used to expand the effective receptive field while maintaining a low number of parameters. Then, Global Response Normalization (GRN) is introduced to replace the traditional normalization strategy to enhance the inter-channel response modeling capability. In the channel dimension, nonlinear transformations of dimensionality increase / decrease are achieved through point convolutions or equivalent mappings, and optional activation functions are introduced to improve expressive power. Finally, the input and output of this unit are fused through residual connections, thereby simultaneously characterizing the influence of bipolar scattering differences, incident angle modulation effects, and bright target interference on the backscattered field within the convolutional framework, providing discriminative feature representations for wind speed regression.

[0062] 4) Construct the radar observation geometry coding branch. Obtain the satellite platform heading angle from the annotation file of the original SAR satellite product, and calculate the radar line-of-sight angle based on the satellite platform heading angle. This angle is used to characterize the observation direction information during SAR imaging. The calculation formula is as follows:

[0063]

[0064] in For the satellite platform's heading angle, This represents the radar line-of-sight angle.

[0065] Simultaneously, considering the periodicity of the angle variable, the radar line-of-sight angle is encoded using sine and cosine periods:

[0066]

[0067] in, For the satellite platform's heading angle, Radar line-of-sight angle, π is the constant value of pi, and g is the encoding result.

[0068] The encoded angle vector is then input into the geometric encoder. After passing through a fully connected layer and a nonlinear activation function, the low-dimensional angle information is mapped into high-dimensional geometric embedding features. These features are then concatenated with the multi-scale image features extracted by the SAR backbone network to obtain a fused feature vector.

[0069] 4) Construct a wind speed regression output module. At the end of the network, global aggregation of high-rise features is performed to obtain a global representation vector, which is then normalized and output with a linear regression head to produce the predicted wind speed value at a height of 10m for the corresponding sub-map.

[0070] 5) Loss Function Design. This invention uses the Smooth L1 loss function to measure the deviation between the predicted wind speed and the monitored wind speed, for a batch of samples. loss function Defined as:

[0071]

[0072] Where N is the batch size. For the estimated wind speed value, To monitor the true wind speed values ​​of the samples. The Smooth L1 loss function can be expressed as:

[0073]

[0074]

[0075] In the formula, The threshold parameter is used for smoothing. This loss function employs an approximate squared penalty when the prediction error is small, which helps improve the accuracy of wind speed prediction; when the prediction error is large, it employs an approximate linear penalty, which can reduce the impact of abnormal samples, local strong scattering interference, or label errors on model parameter updates, thereby improving training stability and robustness.

[0076] 6) Evaluation index design. For model performance evaluation, the coefficient of determination is used. The correlation coefficient (Corr), root mean square error (RMSE), bias (Bias), and mean absolute error (MAE) were used as evaluation indicators to perform statistical calculations on the test samples. and and represent the predicted wind speed and monitored wind speed values ​​for the i-th sample, respectively, and M is the total number of samples.

[0077] Coefficient of determination This represents the proportion of the actual wind field that can be explained by the model inversion results:

[0078]

[0079] in, This is the mean of the true values ​​for all samples.

[0080] The correlation coefficient is used to measure the consistency between the predicted wind speed and the monitored wind speed in terms of their trends.

[0081]

[0082] in, and These are the sample mean values ​​for predicted wind speed and monitored wind speed, respectively.

[0083] Root mean square error is used to measure the average error between predicted and supervised values.

[0084]

[0085] Bias is used to measure the degree to which prediction results are systematically overestimated or underestimated:

[0086]

[0087] Mean absolute error is used to measure the average deviation between model predictions and actual values.

[0088]

[0089] The supervisory labels used for model training and validation are derived from multi-source wind speed data, including measured data from buoys and anemometer towers, as well as reanalysis wind field data from ERA5 and CCMP. They are matched to the pixel center positions of SAR images using bilinear interpolation to ensure spatiotemporal alignment.

[0090] A training dataset is constructed using multi-channel SAR image patches as input, radar line-of-sight angle as geometric aid, and spatiotemporal matching with supervised labels. This dataset is then used to supervise the training of the network, optimizing parameters by minimizing the RMSE loss function until the model converges. After training, the model performance is validated using the aforementioned evaluation metrics.

[0091] (II) Application of wind speed inversion

[0092] Once the trained and validated model is applied to new masked SAR image data, the image to be inverted is constructed as an input tensor in the same manner as described above. This input tensor is then fed into the model for forward computation, which outputs the high spatial resolution near-sea surface wind speed field at the corresponding location.

[0093] Using the aforementioned deep learning model based on ConvNeXtv2, end-to-end, high-precision inversion of sea surface wind speed from SAR image features was achieved. The model can automatically learn and fuse dual-polarization scattering differences, incident angle modulation effects, and avoid interference from bright targets, effectively suppressing the influence of inherent SAR speckle noise. Ultimately, it obtains a fine wind field with spatial details far superior to traditional numerical weather prediction, providing high-quality observational constraints for subsequent fusion correction.

[0094] S4: Obtain the numerical weather prediction (NWP) wind field data that is spatiotemporally matched with the SAR image, align the SAR inverted wind field (i.e., the first wind speed field) with the near-shore wind field in the NWP wind field spatiotemporally, establish a correction model based on the aligned data, and use the correction model to correct the near-shore wind field of the NWP to obtain the corrected near-shore wind speed field (i.e., the second wind speed field).

[0095] 1) Construct a matching dataset. Download background wind field data of NWPs within the target sea area, selecting forecast products consistent with operational forecasts, including information such as wind speed at 10m height, temperature at 2m height, humidity at 2m height, and wind direction at 10m height, with coordinates uniformly in the WGS-84 latitude and longitude coordinate system. Select the nearest neighbor time from the SAR satellite data, extract the latitude and longitude coordinates of the corresponding wind speed center from the SAR data, and perform bilinear interpolation on the NWP grid to obtain the corresponding 10m height features of the NWPs.

[0096] 2) Establish a bias correction model. Construct the SAR data observation constraint residuals as the supervision signal:

[0097]

[0098] in, The wind speed at a height of 10m above the sea surface is obtained from SAR data inversion. This represents the wind speed corresponding to NWP data at the same latitude and longitude. This represents the result deviation.

[0099] The LightGBM regressor is used to fit the nonlinear regression mapping relationship between the residuals and the input features to obtain the residual prediction function:

[0100]

[0101] Where x is the input feature vector, which contains the features extracted from the NWP background wind field at the matching point and its neighborhood, including at least: wind speed at 10 meters height, wind direction at 10 meters height, temperature at 2 meters height, humidity at 2 meters height, and spatial gradients or local statistical features (such as the local standard deviation of wind speed and the spatial rate of change of wind direction) calculated based on these elements. For LightGBM regressors, The estimated observation constraint residuals.

[0102] The final corrected wind speed is:

[0103]

[0104] in, The wind speed at a height of 10 meters for the NWP background wind field. The wind speed deviation value is obtained by fitting the LightGBM regressor.

[0105] S5: Based on the wind shear index after correction of the meteorological element variables, the corrected near-sea surface wind speed field (second wind speed field) is extrapolated to the hub height of the wind turbine to obtain the incoming wind speed field at the hub height.

[0106] 1) Calculate the basic wind shear index based on the empirical formula.

[0107] First, the average wind speed and local disturbance terms of the scenario are calculated from the near-sea surface wind speed:

[0108]

[0109]

[0110] in, The wind speed at a height of 10m is the true value during training and the predicted value during testing, in order to calculate the average wind speed of the scene. The wind speed at a height of 10m at the current pixel location (i.e., the value of the second wind speed field obtained after correction in step S4 at this point). This represents the average wind speed measured within the same scene.

[0111] Secondly, a polynomial function is constructed based on the scene's average wind speed:

[0112]

[0113] The coefficients A, B, C, D, and E are obtained by least-squares fitting of historical wind speed observation data (such as buoy, anemometer, or reanalysis data). In this embodiment, the typical coefficients obtained by fitting wind speed data based on ERA5 reanalysis are: A = -0.0003, B = 0.0082, C = -0.072, D = 0.41, and E = 0.58. In practical applications, the coefficients can be refitted based on local measured data.

[0114] Further, the equivalent roughness is obtained:

[0115]

[0116] In the formula, g is the acceleration due to gravity. The parameter is 0.011.

[0117] Next, based on the idea of ​​logarithmic wind profile, a vertical structure factor is constructed:

[0118]

[0119] Where z is the target height.

[0120] Combining the above vertical structure factors and local perturbation terms, the basic extrapolation ratio is obtained:

[0121]

[0122] Where k is an empirical correction coefficient, which can be approximated as 0.0019 by linear fitting.

[0123] Calculate the basic wind shear index using the power-law extrapolation formula. ,in, This is a reference height (usually taken as 10m). If the target extrapolated height (e.g., wind turbine hub height, typically 100m) is given, then:

[0124]

[0125] in, for , Extrapolation ratio of wind speed at altitude.

[0126] 2) Based on key meteorological element variables, derived variables are constructed, and the wind shear index correction term is obtained through a multilayer perceptron coding network. The wind speed at the hub height is calculated using the second wind speed field as the reference wind speed.

[0127] First, the original meteorological variables ,

[0128] in, The air temperature is 2m. The dew point temperature is 2m. Sea surface temperature, Sea level pressure The boundary layer height, For latent heat flux, For sensible heat flux.

[0129] Based on the above variables, derived variables are constructed as follows:

[0130] Nearshore air temperature - sea surface temperature difference:

[0131] Dew point difference:

[0132] Sensible heat to latent heat flux ratio: ,in To prevent extremely small positive numbers with a denominator of zero, we take 10. -6 .

[0133] The final input meteorological variables are:

[0134] The final input formula for the prior feature variables is:

[0135] Furthermore, the formula characteristics and meteorological characteristics are standardized: ; ,

[0136] in, and Let be the mean vector of the prior features of the formula and each dimension of the meteorological features on the training set. and Let be the standard deviation vector of each dimension of the formula prior features and meteorological features on the training set.

[0137] The network input is obtained by splicing: ;

[0138] Next, the meteorological feature vectors are input into a multilayer perceptron coding network to learn the wind shear index correction term:

[0139] ,

[0140] in, This indicates the meteorological coding correction network. This represents the wind shear index correction amount obtained from learning. This is the maximum permissible correction amount, used to limit the correction magnitude to avoid instability in the extrapolation results. In this embodiment, we take... =0.15, which can be adjusted between 0.1 and 0.3 in practical applications depending on the wind speed distribution and hub height.

[0141] Finally, the corrected wind shear index is: ;

[0142] Height extrapolation ratio: ;

[0143] Finally, the predicted wind speed at a height of 100m was obtained: ;

[0144] The final output includes the corrected 10-meter wind speed field and the hub height wind speed field.

[0145] Application Examples

[0146] A study area was selected, and 147 Sentinel-1 dual-polarization images from 2021 to 2025 were collected. Taking one image from a specific day as an example, a threshold k of 4.5 was set, and the bright target masking result after threshold segmentation is shown below. Figure 4 As shown, automated batch preprocessing is performed using professional SAR processing software (such as ESA's SNAP).

[0147] The supervision label used was the 10-meter height wind speed (hourly, 0.25°×0.25° grid) from the ECMWF ERA5 reanalysis data. Bilinear interpolation was used to match the ERA5 wind speed to the pixel center of the SAR image, ensuring spatiotemporal alignment.

[0148] The network skeleton used is a four-stage hierarchical structure, with stage depth parameters set to depths=[2, 2, 6, 2] and stage channel widths set to hidden_sizes=[96, 192, 384, 768]. The activation function is GELU. The convolutional feature embedding (Stem) module maps the input X to a 96-dimensional feature space and performs initial downsampling to obtain a low-resolution, high-channel basic feature map. The four-stage hierarchical convolutional module then enters Stage 1, which stacks two ConvNeXtv2 Blocks on the 96 channels to extract basic representations such as sea surface stripes, texture, and local intensity statistics. Between stages, downsampling layers halve the feature map space size and double the number of channels. Stage 2 stacks two blocks on 192 channels to enhance the joint modeling of mesoscale texture and incident angle effects. Stage 3 stacks six blocks on 384 channels as the core computational segment of the backbone and introduces an SE channel attention mechanism to learn stronger scale-invariant features and a wider range of spatial dependencies. Stage 4 stacks two blocks on 768 channels to finally integrate high-level semantic features and obtain a globally discriminative representation sensitive to wind speed changes. Finally, global average pooling (GAP) is performed on the output feature map of Stage 4 to obtain a 768-dimensional patch representation vector. In the regression stage, a multi-scale feature fusion regression head is used to perform global average pooling (GAP) and linear mapping on the output features of each stage and then fuse them to output the predicted wind speed value. The resulting high-precision incoming wind speed field at the hub height can be directly input into the wind farm power prediction model or the unit load assessment system to improve the accuracy of power generation prediction and the level of safe operation control of the unit.

[0149] Example 2

[0150] This embodiment provides a software system for implementing the aforementioned method. This system is organized with a modular architecture, and through the collaborative work of its various functional modules, it automatically completes the entire correction process from SAR data to hub height and incoming wind speed. The overall system architecture is as follows: Figure 1 As shown (ensuring the flowchart includes all modules and data flows), it mainly includes the following five core functional modules:

[0151] 1. Data Acquisition and Preprocessing Module

[0152] Module composition and working process:

[0153] SAR Data Acquisition Unit: Automatically schedules and downloads multi-temporal SAR satellite imagery data (preferably Sentinel-1 IW mode GRD dual-polarization products) covering the target sea area via a network interface. This unit integrates data retrieval and filtering logic, and can automatically complete the initial data selection based on preset spatiotemporal range, cloud cover, and data quality indicators (such as removing large-area rain clusters and oil spill images).

[0154] Preprocessing pipeline unit: Integrates specialized SAR processing algorithms to form a standardized batch processing workflow. This unit sequentially executes: 1) Orbit correction, applying precise ephemeris data; 2) Noise removal, including eliminating GRD boundary noise and system thermal noise; 3) Radiometric calibration, converting the original DN values ​​into backscattering coefficients. ; 4) Speckle noise filtering (using Refined Lee and other filters); 5) Geometric correction and subset clipping: based on DEM and orbit data, the image is reprojected to a standard geographic coordinate system and the target area is clipped out; 6) Decibel conversion: to generate a standardized backscattering coefficient image.

[0155] Data quality control unit: performs quality checks and logs on the outputs of each preprocessing stage to ensure that the data input to the next module meets the specifications.

[0156] Output: Standardized, geographically consistent dual-polarization (VV / VH) backscattering coefficient imagery.

[0157] 2. Bright Target Recognition and Masking Module

[0158] Statistical threshold detection unit: Reads the preprocessed VV polarization image. First, it calculates the mean μ and standard deviation σ of the backscattering coefficients for the entire image. Then, it calculates the z-score value for each pixel. ,in, z: Standardized statistical value (z-score), used to identify bright target pixels; : VV polarization backscattering coefficient of a pixel; μ: The mean value of the backscattering coefficients of all pixels in the entire image (or target area); σ: Standard deviation of the backscattering coefficients of all pixels in the entire image (or target area).

[0159] Pixels with z values ​​greater than a preset threshold (e.g., 4.5) are marked as candidate bright targets, and an initial binary image is generated.

[0160] Morphological processing and connected component analysis unit: A morphological dilation operation (e.g., using a 3×3 structuring element) is performed on the initial binary graph to connect discrete scattering points of the same physical target (e.g., a wind turbine). The dilation results are then labeled with connected components, and the area of ​​each connected component is calculated.

[0161] Mask generation unit: Based on the typical pixel size range of the wind turbine target in the image, a connected component area filtering threshold (e.g., [3,400]) is set. This unit filters out connected components that are too large or too small, merges the remaining connected components, and generates the final bright target mask file. In this mask, bright target areas are marked with specific values ​​(e.g., 1), and sea surface areas are marked with background values ​​(e.g., 0).

[0162] Output: Bright target mask file spatially aligned with SAR imagery.

[0163] 3. High-resolution wind speed inversion module

[0164] Model Deployment and Inference Unit: The core of this module is a pre-trained deep learning model for wind speed inversion (based on the ConvNeXtv2 architecture). The model takes the aforementioned four-channel data blocks (VV backscattering coefficient, VH backscattering coefficient, incident angle, and bright target mask) as input. This unit is responsible for loading the model and organizing the masked SAR image data, cropping it into sub-image blocks according to the same size (e.g., HxW) as during training, and constructing the input batch.

[0165] Input data standardization unit: Before inference, the mean (μ) and standard deviation (σ) parameters stored in the model configuration are automatically invoked to standardize the VV, VH and incident angle channels of the input data block ((x-μ) / σ), while the mask channel remains unchanged.

[0166] Working Process: Standardized four-channel data blocks are fed into a deep learning model for forward propagation. The model extracts features step-by-step through its multi-layer convolutional structure (ConvNeXt Blocks), adds a radar line-of-sight geometric auxiliary branch for feature stitching, and finally outputs the predicted wind speed at a near-sea surface height of 10 meters corresponding to the center position of each sub-image block by the regression head. After processing all sub-blocks, a high spatial resolution SAR inversion wind field covering the entire target area is generated.

[0167] Output: High spatial resolution near-sea surface wind speed field (i.e., first wind speed field).

[0168] 4. Wind Field Fusion Correction Module

[0169] Module composition and working process:

[0170] Data Spatiotemporal Matching Unit: This module simultaneously connects to external numerical weather prediction (NWP) data sources. First, it acquires the NWP grid data (including 10m wind speed, wind direction, 2m temperature, humidity, etc.) closest to the SAR imaging time. Then, it performs spatiotemporal alignment between each wind speed point (latitude and longitude) of the SAR-retrieved wind field and the NWP grid: spatially, bilinear interpolation is used to interpolate the NWP meteorological elements to the precise location of the SAR wind speed points; temporally, the closest forecast time is selected.

[0171] The calibration model application unit incorporates a pre-trained LightGBM regressor as the calibration model. Its workflow is as follows: 1) Calculate the residual between the SAR wind speed and the NWP wind speed at the matching point; 2) Input the NWP features (wind speed, wind direction, temperature, humidity) at that location into the LightGBM model; 3) The model predicts the wind speed deviation at that point; 4) Correct the original NWP wind speed. This process is repeated iteratively for all grid points or locations within the target area that require calibration.

[0172] Output: Spatially continuous high-precision near-sea surface wind speed field (i.e., the second wind speed field) after SAR observation constraint correction.

[0173] 5. Vertical wind speed pushout module

[0174] Module composition and working process:

[0175] Wind shear index correction calculation unit: Based on empirical formulas, the basic wind shear index is first obtained, and then the extrapolation scaling factor is corrected by combining key meteorological elements.

[0176] Vertical extrapolation calculation unit: This unit uses the corrected near-sea surface (10 meters) wind speed field as a benchmark, and combines the corrected wind shear index α and the target wind turbine hub height to perform vertical extrapolation calculations using a power law formula. Through this calculation, the sea surface wind speed is vertically mapped to the wind turbine hub height.

[0177] Output: The final product is high-precision, high-resolution information on the incoming wind speed field and its spatiotemporal distribution at the hub height of the wind turbine.

[0178] The five modules described above are connected in series to form an automated processing chain. Raw SAR data flows into the data acquisition and preprocessing module. The processed standardized image, along with the mask generated by the bright target identification and masking module, is sent to the high-resolution wind speed inversion module to obtain a refined wind field. This wind field is then fused and corrected with externally input NWP data in the wind field fusion and correction module. Finally, the wind speed vertical extrapolation module performs the height conversion, outputting the final wind field product that can directly serve wind farm operation. Intermediate results are transmitted between modules via a standard data interface, enabling the system to operate in batches and automatically.

[0179] Example 3

[0180] Based on Embodiment 1, this embodiment discloses an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method described in Embodiment 1.

[0181] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A method for correcting the incoming wind speed of an offshore wind farm based on SAR satellite observations, characterized in that, Includes the following steps: S1: Acquire multi-temporal synthetic aperture radar (SAR) satellite image data covering the target offshore wind farm area, and perform a series of preprocessing steps on the SAR satellite image data to obtain standardized backscattering coefficient images; S2: Based on the backscattering coefficient image, identify and mask the bright target interference area caused by the wind farm turbine array and its auxiliary facilities; S3: Based on the masked SAR satellite image data, construct and train a deep learning model for wind speed inversion, and introduce radar line of sight as a geometric auxiliary condition for feature supplementation; use the trained deep learning model for wind speed inversion to process the masked SAR satellite image data and invert the first wind speed field at the near sea level. S4: Obtain numerical weather prediction (NWP) wind field data that is spatiotemporally matched with the SAR satellite image data, align the first wind speed field with the near-sea wind field in the NWP wind field data spatiotemporally, establish a correction model based on the aligned data, and use the correction model to correct the NWP near-sea wind field to obtain the corrected second wind speed field. S5: The vertical wind speed shear index is corrected by combining upper-level meteorological elements. Based on the corrected extrapolation model, the second wind speed field is extrapolated to the hub height of the wind turbine to obtain the incoming wind speed field at the hub height.

2. The method of claim 1, wherein, The series of preprocessing steps in step S1 also includes: using a speckle filter to filter the SAR satellite image data to suppress speckle noise, and converting the backscattering coefficient to decibels.

3. The method of claim 1, wherein, Step S2 specifically includes: S2.1: Based on the statistical values ​​of the backscattering coefficients of the pixels in the backscattering coefficient image, candidate bright target pixels are extracted by setting a threshold; S2.2: Morphological dilation is performed on the candidate bright target pixels using a square structuring element of a preset size, and connected component analysis is performed on the dilation result; S2.3: Filter the connected components according to the preset area range, and merge the filtered connected components to generate a bright target mask.

4. The method of claim 1, wherein, In step S3, the input to the wind speed inversion deep learning model is an image patch containing multi-channel feature information extracted from SAR satellite image data after mask processing. The method further includes: obtaining the satellite platform heading angle from the annotation file of the SAR satellite image, calculating the radar line-of-sight angle based on the satellite platform heading angle, and spatially matching the radar line-of-sight angle with the corresponding image patch. The multi-channel feature information includes at least: backscattering coefficients of different polarization modes, incident angle, and bright target mask information generated in step S2. The radar line-of-sight angle is used as a geometric auxiliary condition, and after angle encoding, it is fused with the features of the image patch to supplement the geometric information of SAR imaging observation.

5. The method of claim 4, wherein, The wind speed inversion deep learning model is trained through the following steps: Constructing a training dataset with multi-channel SAR satellite image data blocks as input and reanalysis wind speeds at corresponding spatiotemporal locations as supervision labels; using the training dataset to train an initial model based on a convolutional neural network, calculating the predicted wind speed using forward propagation, calculating the error between the predicted wind speed and the supervision labels using a loss function, and updating the parameters of the initial model through backpropagation until the initial model converges, and using the converged model as the wind speed inversion deep learning model.

6. The method of claim 1, wherein, In step S4, a correction model is established based on the aligned data, specifically as follows: Calculate the wind speed residual between the first wind speed field and the near-shore wind field of the NWP at the matching spatial location; use the meteorological element features of the NWP wind field at the matching spatial location and its spatially derived features as input, and use the wind speed residual as a supervision signal to train a machine learning regression model to obtain the correction model; the meteorological element features include at least one of wind speed, wind direction, temperature and humidity; the spatially derived features include spatial gradients or local statistical features calculated based on the meteorological element features.

7. The method of claim 6, wherein, The machine learning regression model is a gradient boosting decision tree model, and the model is the LightGBM model.

8. The method according to claim 1, characterized in that, Step S5 specifically includes: S5.1: Calculate the basic wind shear index; S5.2: Based on key meteorological element variables, construct derived variables, obtain the wind shear index correction term through a multilayer perceptron coding network, and calculate the wind speed at the hub height using the second wind speed field as the reference wind speed; wherein, the derived variables include at least one of nearshore air temperature-sea temperature difference, dew point difference, or sensible heat-latent heat flux ratio.

9. A system for correcting the incoming wind speed of an offshore wind farm based on SAR satellite observations, the system being used to execute the method for correcting the incoming wind speed of an offshore wind farm based on SAR satellite observations as described in claim 1, characterized in that, include: The data acquisition and preprocessing module is used to perform step S1; Bright target recognition and masking module is used to perform step S2; A high-resolution wind speed inversion module is used to perform step S3; The wind field fusion correction module is used to perform step S4; The wind speed vertical outward push module is used to execute step S5.

10. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the offshore wind speed correction method based on SAR satellite observation as described in claim 1.

Images