A wind speed field correction method and system based on wind field correction operators
By generating an initial residual field through a wind field correction operator and iteratively propagating it along the prevailing wind direction, the problem of wind field continuity being difficult to reflect in existing technologies is solved, achieving efficient wind speed field correction and power prediction, which is suitable for rapid correction and operational applications of large-scale wind farms.
Patent Information
- Application Number
- CN202611282339.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-24
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2046-08-24
AI Technical Summary
In existing technologies, single-point statistical post-processing is difficult to reflect the continuous spatial structure of wind fields, power domain spatiotemporal learning does not explicitly correct the wind field itself, and improved numerical weather prediction models consume a lot of computing power and have high operating costs.
A wind field correction operator-based method is adopted. An initial residual field is generated by a residual initialization operator, and the field anisotropy input operator is used to iteratively propagate along the prevailing wind direction to generate the final residual field. Finally, it is superimposed on the reference wind speed field to achieve wind speed field correction.
It maintains the spatial consistency and physical interpretability of wind speed field correction results, reduces computational overhead, and is suitable for rapid correction and operational applications in large-scale wind farms.
Smart Images

Figure CN122797350B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of wind power data processing technology, and in particular to a wind speed field correction method and system based on wind field correction operators. Background Technology
[0002] As offshore wind power develops towards larger capacity and clustering, the number of turbines within a single wind farm is increasing, wake coupling is strengthening, and the correlation of wind speed errors between turbines is intensifying. Current wind speed forecasting is no longer a problem of single-turbine location errors, but rather a wind farm error problem with a distinct spatial structure. These errors can further propagate into overall power deviations, affecting grid dispatch, reserve arrangements, and system stability. Existing technologies employ two main approaches: First, improving forecast accuracy by refining the physical parameterization scheme within the numerical weather prediction model, such as improving wind farm parameterization, turbulent energy terms, and wake recovery processes. These schemes directly affect the model, resulting in high computational consumption and development and operating costs. Second, single-point statistical post-processing techniques typically treat the wind speed errors at each turbine location or grid point as independent regression problems. Because this technique does not incorporate the physical attribute of a "continuous spatial distribution field" into the model structure, corrections at a certain location cannot naturally propagate to adjacent or downstream areas based on the wind field evolution. The third type of power domain spatiotemporal learning technique typically uses power directly as the prediction target without explicitly modifying the wind field itself. The input information for this type of method usually comes mainly from the SCADA operation data of each wind turbine or measuring point within the station, focusing on mining the short-term temporal and spatial correlations within the station, but lacking an effective characterization of the evolution process of mesoscale weather systems and their impact on future wind field changes.
[0003] The Chinese patent document "High-Precision Correction Method and System for Wind Speed Forecasting in Power System Wind Power Prediction," publication number CN120525133A, published on August 22, 2025, firstly, based on the WRF model and ARW dynamic solver, selects a three-layer nested structure and physical process parameterization scheme with the wind farm location as the regional center, and combines static topography and global meteorological forecast data to achieve high-resolution short-term wind speed forecasts. Then, by modifying boundary layer parameters and introducing data assimilation techniques, multiple differentiated schemes are formed to obtain multiple downscaling forecast results for the same time period. Next, based on the forecast results and wind speed observation data from the corresponding wind farm, various improved artificial intelligence models of different types are used to correct the forecasted wind speed. Finally, an adaptive weighted combination of multiple improved models is used to obtain the final wind speed forecast result. By fully mining the information from topography, observation data, and global meteorological forecast data, the accuracy of wind speed forecast data is effectively improved. This technology improves forecast accuracy by modifying the internal physical parameterization scheme of numerical weather prediction models, but it consumes a lot of computing power and has high development and operation costs. Summary of the Invention
[0004] This invention aims to overcome the problems in existing technologies, such as the difficulty of single-point statistical post-processing in reflecting the continuous spatial structure of wind fields, the lack of explicit correction of wind fields themselves in power domain spatiotemporal learning, and the high computational consumption and operating cost of improved numerical weather prediction models. It provides a wind speed field correction method and system based on wind field correction operators.
[0005] To achieve the above objectives, the present invention adopts the following technical solution: A wind speed field correction method based on a wind field correction operator includes: Acquire baseline forecast data, including the baseline wind speed field, auxiliary meteorological variables from the same source, and historical observation data of wind farm units; Spatiotemporal alignment and preprocessing are performed to construct the input feature tensor and generate a unit mask that identifies the effective observation location; Input the input feature tensor into the residual initialization operator to generate an initial residual field characterizing the magnitude of wind speed deviation; Based on the prevailing wind direction corresponding to the benchmark wind speed field, the initial residual field is iteratively propagated through the field anisotropic input operator, and the spatial redistribution of residual information is carried out along the axis of the prevailing wind direction to obtain the final residual field. The final residual field is superimposed on the reference wind speed field to obtain the corrected wind speed field.
[0006] This invention models day-ahead wind field correction as an operator learning problem. The baseline wind field is the wind speed field output by the WRF-Fitch model, and the actual wind field is monitored within the wind farm area by turbine-level SCADA observations. A wind field correction operator is used to ensure that the corrected wind field equals the baseline wind field plus a residual correction field driven by multi-source meteorological features. This correction field first obtains an initial residual field through a residual initialization operator, and then propagates along the wind direction through a field anisotropy input operator to achieve a unified correction of the wind speed distribution across the entire wind farm, forming a spatially consistent final correction result. This approach maintains the spatial continuity and physical interpretability of the correction results while balancing computational efficiency, operating costs, and engineering deployability, making it suitable for rapid correction and operational applications of day-ahead wind field forecasts.
[0007] As an alternative, the method also includes training the wind field correction operator by inputting the baseline forecast data and the same source auxiliary meteorological variables into the trained wind field correction operator to obtain the corrected wind speed field. The training is optimized using a joint loss function, which adds a smoothing regularization term and an L2 regularization term for the residual field amplitude to the Huber loss with unit masking.
[0008] Preferably, the residual initialization operator is a two-dimensional convolutional encoder; The two-dimensional convolutional encoder contains at least two convolutional layers, each convolutional layer is followed by a nonlinear activation function, and finally a single-channel residual field is output through a two-dimensional convolutional layer. The convolution kernel parameters and bias parameters of the residual output layer are initialized to zero at the start of training and adjusted based on the wind speed error at the effective observation location through backpropagation.
[0009] Preferably, the field anisotropic transport operator is: the difference between the residual field after the (k+1)th propagation and the residual field after the kth propagation is equal to the product of the transport step size and the spatial propagation operator; The spatial propagation operator obtains the residual difference between adjacent grids of the current residual field in each iteration, and assigns higher propagation weights to the directions that are closer to the dominant wind direction in the east-west and north-south directions.
[0010] Preferably, the method for determining the iteration number and transport step size of the field anisotropic transport operator is as follows: Set the candidate value ranges for the number of iterations and the transport step size respectively; Based on a pre-set validation sample set, wind speed error and spatial continuity index are compared under different parameter combinations. Among the parameter combinations with progressively increasing iterations, the first parameter combination is selected where the wind speed error decrease is less than the preset decrease threshold, or the spatial continuity index is lower than the preset index threshold.
[0011] Preferably, the spatiotemporal alignment and preprocessing to construct the input feature tensor includes: Unify the time resolution of all data; For data with different spatial resolutions, the data is mapped to a two-dimensional wind field grid consistent with the reference wind speed field based on interpolation or grid resampling, and the observation data of multiple units within the same grid are averaged. After periodically encoding or standardizing the spatiotemporally aligned data, they are concatenated into an input feature tensor.
[0012] Preferably, the auxiliary meteorological variable of the same source is a meteorological variable field from the same source as the baseline forecast data, which identifies the spatial distribution characteristics of wind speed error under different weather backgrounds; The acquired data also includes external auxiliary meteorological forecast data, which, after spatiotemporal alignment and preprocessing, is used as additional feature input to the residual initialization operator.
[0013] Preferably, a power compliance verification step is also included: Wind field features are extracted based on the corrected wind speed field, and feature vectors are constructed by combining the baseline power and meteorological background factors. Build and train a power correction model between the baseline power and the actual power, and output the corrected power without applying a physical upper limit; Determine the upper limit of the effective full-power generation of the wind farm; apply non-negative constraints and upper limit constraints of the effective full-power generation to the corrected power, and generate the final day-ahead power prediction results for the entire farm.
[0014] Preferably, the method for determining the effective full-power upper limit includes: Using the first preset high quantile of the reference power as a standard, the corresponding high power candidate time period is identified; Filter out time periods where the high-power state lasts for a preset length; Within the selected time period, the second preset high quantile of the actual power is extracted as the effective upper limit of the full power.
[0015] A wind speed field correction system based on a wind field correction operator includes: The data acquisition module acquires baseline forecast data, source auxiliary meteorological variables, and historical observation data of wind farm units, including the baseline wind speed field. The preprocessing module performs spatiotemporal alignment and preprocessing on the data, constructs the input feature tensor, and generates the unit mask; The residual initialization module receives the input feature tensor and generates an initial residual field; The anisotropic transport module iteratively propagates the initial residual field and spatially redistributes the residual information along the axis of the prevailing wind direction to obtain the final residual field. The field synthesis module superimposes the final residual field onto the reference wind speed field to obtain a spatially continuous corrected wind speed field.
[0016] Compared to existing technologies, this invention has the following advantages: Treating wind speed correction as an operator operation applied to the entire wind farm naturally maintains spatial consistency, making it suitable for large-scale wind farms or wind power clusters; by combining residual initialization with anisotropic transport along the wind direction, it separates the modeling of "where the deviation occurs" from "how the deviation propagates," resulting in strong physical interpretability; unlike data assimilation, it does not require repeated integration of dynamic models, and correction can be completed in the inference stage by inputting only baseline forecast data and similar auxiliary meteorological variables, resulting in lower computational overhead and easy integration into existing operational links; it can improve overall prediction accuracy while maintaining the dynamic characteristics of the wind farm, making it suitable for offshore wind farm scenarios with significant dominant wind direction, wake propagation, and spatial correlation errors. Attached Figure Description
[0017] Figure 1 This is a flowchart of a wind speed field correction method based on a wind field correction operator in this invention.
[0018] Figure 2 This is a schematic diagram of the turbine distribution within a wind farm in an embodiment of the present invention.
[0019] Figure 3 This is the correction and improvement effect and wind rose diagram under directional dispersion wind conditions in the embodiments of the present invention. Among them, (a) is the spatial RMSE diagram of the test set in May 2024; (b) is the wind rose diagram (ZKTJ u / v@100m) in May 2024, with data from the 26th and later as the test set, and Adir=22.5°.
[0020] Figure 4 This is the correction and improvement effect and wind rose diagram under continuous directional wind conditions in the embodiments of the present invention. Among them, (a) is the test set spatial RMSE diagram in July 2024; (b) is the wind rose diagram in July 2024 (ZKTJ u / v@100m), with data from the 26th and later as the test set, and Adir=22.5°.
[0021] Figure 5 This is a comparison chart of wind speed time series for the test months in this embodiment of the invention.
[0022] Figure 6 This is a comparison chart of the overall power time series in an embodiment of the present invention. Detailed Implementation
[0023] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be noted that all data in the present invention were collected under the premise of legality and compliance.
[0024] A wind speed field correction method based on a wind field correction operator includes: Acquire baseline forecast data, including the baseline wind speed field, auxiliary meteorological variables from the same source, and historical observation data of wind farm units; Spatiotemporal alignment and preprocessing are performed to construct the input feature tensor and generate a unit mask that identifies the effective observation location; Input the input feature tensor into the residual initialization operator to generate an initial residual field characterizing the magnitude of wind speed deviation; Based on the prevailing wind direction corresponding to the benchmark wind speed field, the initial residual field is iteratively propagated through the field anisotropic input operator, and the spatial redistribution of residual information is carried out along the axis of the prevailing wind direction to obtain the final residual field. The final residual field is superimposed on the reference wind speed field to obtain the corrected wind speed field.
[0025] This invention models day-ahead wind field correction as an operator learning problem. The baseline wind field is the wind speed field output by the WRF-Fitch model, and the actual wind field is monitored within the wind farm area by turbine-level SCADA observations. A wind field correction operator is used to ensure that the corrected wind field equals the baseline wind field plus a residual correction field driven by multi-source meteorological features. This correction field first obtains an initial residual field through a residual initialization operator, and then propagates along the wind direction through a field anisotropy input operator to achieve a unified correction of the wind speed distribution across the entire wind farm, forming a spatially consistent final correction result. This approach maintains the spatial continuity and physical interpretability of the correction results while balancing computational efficiency, operating costs, and engineering deployability, making it suitable for rapid correction and operational applications of day-ahead wind field forecasts.
[0026] Optional, such as Figure 1 As shown, a training phase is also included. After obtaining the corrected wind speed field through the wind field correction operator, the wind field correction operator is trained by optimizing and fitting the parameters of the operator model to obtain the trained wind field correction operator. After training is completed, in the subsequent deployment and inference phases, it is only necessary to input the baseline forecast data and the same source auxiliary meteorological variables into the trained wind field correction operator to obtain the corrected wind speed field.
[0027] Under this optional scheme, the trained wind field correction operator can serve as a relationship model between the baseline wind speed field and the corrected wind speed field. Only the baseline forecast data containing the baseline wind speed field and the same source auxiliary meteorological variables need to be input to obtain the corrected wind speed field, which is more convenient for engineering deployment and rapid correction and operational application of wind field forecast results.
[0028] During training, all historical data can be divided into training, validation, and testing sets. These sets are arranged chronologically, and within each month, a "first N days for training, second N days for validation, and third N days and beyond for testing" approach is used. For example, the 1st to the 20th of each month can be used as training samples, the 21st to the 25th as validation samples, and the 26th and beyond as testing samples. Model parameter fitting is performed solely based on the training samples, hyperparameter selection and early stopping are determined using the validation samples, and the testing samples are used only for final performance evaluation, thus preventing future information from being incorporated into the model training process.
[0029] Training is optimized using a joint loss function. Training is considered complete when the number of training epochs reaches a preset upper limit or the change in the joint loss function value is less than a preset threshold. The joint loss function, based on the Huber loss with a unit mask, incorporates a smoothing regularization term and an L2 regularization term for the residual field amplitude. Specifically, it can be expressed as L = L Huber +λs·L smooth +λr·LL2 , where λs and λr are the spatial consistency loss weight and residual regularization term weight, respectively, and their values can be selected based on actual needs and application scenarios.
[0030] Huber lost L Huber Equals Huber(WScorr-WSreal)⊙M turbine WScorr represents the corrected wind speed field, WSreal represents the true wind speed field, and M turbine For unit masking, Huber loss with unit masking is used to monitor the corrected wind field, balancing the sensitivity of mean square error and the robustness of absolute error. Data consistency error is calculated only at grid locations where valid unit observations exist. Smoothing regularization term L smooth The sum of squares of the differences between the residuals of all adjacent grids in the final residual field is used to suppress high-frequency noise in the sparse supervision region and maintain the spatial continuity of the correction field; the L2 regularization term of the residual field amplitude constrains L L2 To prevent global drift and overfitting and limit the overall amplitude of the residual field, it can be calculated as the sum of the squares of the residuals of each grid in the residual field (or it can be averaged by dividing by the total number of grids).
[0031] The spatial consistency loss weight λs is used to constrain the magnitude of residual variation between adjacent grids or related grids. If λs is too small, the model may introduce local noise; if λs is too large, it may weaken the true spatial differences. Therefore, λs can be gradually increased starting from zero or a small weight, and a value that maintains both low error and good spatial continuity on the validation set can be selected.
[0032] The residual regularization term weight λr is used to limit the magnitude of outlier residuals, preventing the model from overfitting to individual outlier observations or local noise. λr can be adjusted based on the difference between the training and validation set errors. When the training error is significantly lower than the validation error, λr can be increased appropriately; when the model is underfitting and both the training and validation errors are high, λr can be decreased appropriately. λs and λr can be set to 0, 10, or 10 respectively. -4 10 -3 10 -2 10 -1 The search is conducted among candidate values of equal magnitude to determine the answer.
[0033] As a specific implementation, the acquired data mainly includes baseline forecast data, auxiliary meteorological variables from the same source, optional external auxiliary meteorological forecast data, and historical observation data of wind farm units. Before entering the wind farm correction operator, data from different sources need to undergo spatiotemporal alignment and standardization processing to construct a unified wind farm correction input.
[0034] The baseline forecast data was used to generate high-resolution numerical forecast results for the target wind farm area using the WRF-Fitch model. The WRF-Fitch model is a forecasting model that incorporates a wind farm parameterization scheme into the WRF numerical weather prediction model, enabling it to characterize the impact of wind turbines on the local flow field to a certain extent. The WRF-Fitch model employs a multi-layered nested grid, with the innermost grid covering the target wind farm area and outputting the hub height wind speed field, wind direction field, or horizontal wind speed component for the next 24 hours. This wind speed field serves as the baseline wind speed field to be corrected in this invention.
[0035] The auxiliary meteorological variables are meteorological variable fields from the same source as the baseline forecast data, synchronously output by the WRF-Fitch model or the same numerical weather prediction system. These variables include, but are not limited to, boundary layer height, temperature, air pressure, water vapor mixing ratio, precipitation, near-surface wind speed and direction, and horizontal wind speed components at different altitudes. These auxiliary meteorological variables are not the direct correction targets of this invention, but rather serve as auxiliary inputs to explain wind speed forecast biases, helping the wind field correction operator identify the spatial distribution characteristics of wind speed errors under different weather conditions. For example, when the boundary layer height is low, near-surface stability is strong, or precipitation has a significant impact, WRF-Fitch forecast wind speeds may exhibit different types of systematic biases; by introducing these auxiliary variables, the residual initialization operator can more accurately determine which areas in the wind field need correction and the magnitude of the correction.
[0036] External auxiliary meteorological forecast data is optional; any external auxiliary meteorological forecast data other than the baseline WRF-Fitch can be selected as supplementary input to the regional weather background. External auxiliary meteorological forecast data typically includes regional high-resolution meteorological forecast data, artificial intelligence meteorological forecast data, or other similar meteorological forecast products. In specific implementations, data such as the Zhongke Tianji China Regional High-Precision Meteorological Forecast Data and GRAPES regional forecast data can be used. External auxiliary meteorological forecast data is not a necessary limitation for the implementation of this invention. Even without external auxiliary meteorological forecast data, this invention can still complete wind field correction based on the WRF-Fitch baseline wind speed field, source-related auxiliary meteorological variables, and SCADA observation data. When external auxiliary meteorological forecast data is available, it can be time-synchronized, spatially interpolated, and standardized as variable inputs to the residual initialization operator, thereby enhancing the wind field correction operator's ability to characterize large-scale weather backgrounds, regional wind field changes, and complex weather processes.
[0037] Historical observation data for wind farm turbines utilizes turbine-level observation data collected by the wind turbine SCADA system, serving as supervisory information for training and validating wind farm correction operators. Turbine observation data includes nacelle wind speed, nacelle wind direction, actual power, and turbine operating status. During preprocessing, samples showing faults, outages, power curtailment, maintenance, communication anomalies, and obvious abnormalities are removed, retaining only valid observation data under normal operating conditions. Subsequently, based on the wind turbine's geographical location, the turbine-level observation data is mapped to a corresponding two-dimensional wind farm grid, forming a gridded wind speed field. For cases where multiple turbines exist within the same grid, the average, median, or weighted average can be used as the representative observation value for that grid.
[0038] Furthermore, the baseline forecast data, source-related auxiliary meteorological variables, optional external auxiliary meteorological forecast data, and observational data are unified to the same or corresponding time scale, such as a 15-minute resolution. For data with different spatial resolutions, they are unified to the two-dimensional wind field grid of the target wind farm (consistent with the baseline wind speed field) through bilinear interpolation, nearest neighbor interpolation, or grid resampling. After completing the spatiotemporal alignment, the baseline wind speed field, source-related auxiliary meteorological variables, and optional external auxiliary meteorological forecast data are encoded or preprocessed and then used to construct the input feature tensor.
[0039] The encoding and preprocessing methods for input features include: for periodic angular variables such as wind direction, a sine-cosine encoding method is used to reduce discontinuities caused by angle jumps near 0° and 360°. For continuous physical variables such as wind speed, power, boundary layer height, temperature, air pressure, air density, humidity, precipitation, sensible heat flux, and latent heat flux, their original physical quantity values are used as input, or normalization, standardization, or quantile scaling is performed according to actual needs. For example, for tree model-type residual correctors, continuous numerical features can be used directly; for neural network models, the mean and variance of the training set are used to standardize continuous variables.
[0040] For time variables, hours and months are used as numerical features, or periodic encoding is used to represent daily and seasonal variations. For spatial location variables, numerical features such as latitude and longitude, relative coordinates, grid index, station zoning identifiers, terrain height, roughness, or land-sea masking are used. For missing or anomalous features, time interpolation, forward padding, backward padding, neighboring grid padding, or missing marker variables are used for processing. The above encoding methods do not change the core structure of wind field residual initialization, anisotropic transport, and supervised correction in this invention; they are only used to improve the continuity and numerical stability of input features.
[0041] Furthermore, for generating the turbine mask, since SCADA observations only exist at the turbine locations and do not cover the entire two-dimensional wind field grid, a turbine mask needs to be constructed to identify grid cells with valid observations. During wind field correction operator training, supervised loss is calculated only at locations with valid turbine observations; data errors are not directly calculated for grid cells without observations, but they still participate in overall wind field correction and spatial constraints. Through the above processing, multi-source meteorological forecast data and turbine observation data are unified into the same temporal-spatial framework, providing structured input for subsequent residual initialization, field anisotropic transport, and correction wind field generation.
[0042] As a specific implementation, the wind field correction operator for wind speed field correction mainly includes three parts: residual initialization, field anisotropic transport, and corrected wind field synthesis. Its core idea is to first determine which regions of the WRF-Fitch wind field may have systematic deviations, then propagate this deviation information spatially according to wind direction constraints, and finally add the obtained residual field back to the original WRF wind speed field to form a continuous corrected wind field.
[0043] The residual initialization operator is a two-dimensional convolutional encoder; the two-dimensional convolutional encoder contains at least two convolutional layers, each of which is followed by a nonlinear activation function, and finally outputs a single-channel residual field through a two-dimensional convolutional layer; the convolutional kernel parameters and bias parameters of the residual output layer are zero-initialized at the beginning of training, and the parameters are adjusted based on the wind speed error of the effective observation location through backpropagation.
[0044] It should be noted that the residual initialization operator generates a preliminary wind speed residual field. The residual here can be understood as an addition or subtraction correction to the WRF-Fitch wind speed: if the WRF wind speed in a certain area is too high, the residual should tend to be negative; if the WRF wind speed in a certain area is too low, the residual should tend to be positive.
[0045] In practice, the WRF-Fitch wind speed field, generator mask, wind direction information, and multi-source meteorological features are first concatenated into a multi-channel input tensor. For external meteorological features that only have a time dimension and no spatial grid, they are first standardized and then copied onto the entire two-dimensional wind field grid so that they can be used as input to the convolutional network along with the WRF wind speed field.
[0046] The residual initialization operator employs a two-dimensional convolutional encoder. In this embodiment, two 3×3 convolutional layers are used, each followed by GELU nonlinear activation, and then a single-channel residual field is output through another 3×3 convolution. The purpose of the 3×3 convolution is to ensure that the output of each grid point depends not only on its own WRF wind speed, but also on the wind speed distribution in its surrounding neighborhood, the generator mask, wind direction characteristics, and multi-source meteorological background information. Therefore, the convolutional encoder can extract local wind speed gradients, regionally high or low wind speeds, boundary anomalies, and local deviation structures formed under the influence of wake or topography.
[0047] It's worth noting that the initial residual field isn't based on SCADA observations from the outset, nor is it directly calculated from the current SCADA data. In this embodiment, the residual output layer uses zero initialization; the output residual is close to 0 at the beginning of training, essentially meaning the WRF wind field isn't corrected temporarily. Subsequently, the wind field correction operator adds the residual to the WRF wind speed to obtain the corrected wind speed, and calculates the error between the corrected wind speed and the SCADA-observed wind speed at the effective wind turbine locations. This error is propagated back to the convolutional encoder, allowing the operator to gradually learn under what wind field structures and meteorological conditions should positive or negative residuals be output. Therefore, the initial residual field output by the residual initialization operator can represent "where correction is needed and how much initial correction is required."
[0048] The initial residual field obtained by the residual initialization operator mainly reflects the location of local deviations and the initial correction magnitude. However, this result is still biased towards a local representation and does not fully reflect the characteristics of wind speed deviations, wake deficits, or incoming flow errors being continuously distributed along the prevailing wind direction. In actual wind farms, wind field errors usually do not exist in isolation at a single grid point, but often form a continuous structure in space along the incoming flow direction. Therefore, it is necessary to further introduce a field anisotropic transfer operator to propagate the initial residual field with directional constraints, so that the residual information can be selectively updated and redistributed in the two-dimensional wind field according to the prevailing wind direction.
[0049] The field anisotropic transport operator is: the difference between the residual field after the (k+1)th propagation and the residual field after the kth propagation is equal to the product of the transport step size and the spatial propagation operator. In each iteration, the spatial propagation operator obtains the residual difference between adjacent grids of the current residual field and assigns higher propagation weights to the directions that are closer to the dominant wind direction in the east-west and north-south directions.
[0050] The prevailing wind direction θ is inferred from the wind direction variable in meteorological characteristics, representing the prevailing wind direction for the current sample across the entire field. 0 degrees represents true north, 90 degrees clockwise represents true east, 180 degrees represents true south, and 270 degrees represents true west. The propagation and update of the residual field adopts a multi-step iterative form, which can be expressed as: u k+1 =u k +α·L θ (u k), where k = 0, 1, ..., K-1, u k Let L represent the residual field after the k-th propagation; α is the learnable transport step size, used to control the propagation strength at each step; K is the number of propagation iterations; L θ (u k ) is a wind-constrained spatial propagation operator, which can be expressed as |sinθ| multiplied by u k The second derivative with respect to x, plus |cosθ| multiplied by u k The second derivative of y is used, where the x-axis represents the east-west direction and the y-axis represents the north-south direction, without distinguishing between positive and negative directions. In this embodiment, α is converted to a non-negative step size using the softplus function before being multiplied by a scaling factor to improve training stability.
[0051] In practical propagation, the spatial propagation operator reads the current residual field in each iteration, calculates the residual difference between each grid point and its surrounding neighbors (up, down, left, and right), and assigns different weights to the differences in different directions based on the prevailing wind direction. The east-west propagation weight is represented by |sinθ|, and the north-south propagation weight is represented by |cosθ|. When the prevailing wind direction is closer to the east-west direction, the east-west residual update is stronger; when the prevailing wind direction is closer to the north-south direction, the north-south residual update is stronger.
[0052] It should be noted that this operator does not uniformly smooth the residual field in all directions. Instead, it enhances the propagation along the corresponding spatial axis according to the prevailing wind direction, causing the local deviation information to expand along a direction that better matches the wind field structure, forming a more continuous and directionally consistent residual field. Furthermore, since |cosθ| and |sinθ| are used as directional weights, it primarily achieves anisotropic propagation along the axis containing the prevailing wind direction, without strictly distinguishing between downwind and upwind directions. After K propagations, the initial residual field u... 0 Become u K The process in each iteration can be understood as follows: the residual initialization operator first gives a local judgment on "where correction is needed and how much initial correction is needed"; the field anisotropy output operator then spatially redistributes the residual information according to the prevailing wind direction, transforming it from a local residual template into a final residual field with directional continuity.
[0053] Optionally, the field anisotropic output operator also includes a group mask, residual amplitude limits, and a learnable gating mechanism. The group mask is used to set the residuals of invalid meshes to zero; the residual amplitude limits are used to prevent excessive and unreasonable corrections to the output; and the learnable gating is used to control the residual correction strength, so that the operator maintains a weak correction in the early stages of training and gradually increases its effective correction capability as training progresses.
[0054] The method for determining the number of iterations and the transport step size of the field anisotropic transport operator is as follows: Set the candidate value ranges for the number of iterations and the transport step size respectively; Based on a pre-set validation sample set, wind speed error and spatial continuity index are compared under different parameter combinations. Among the parameter combinations with progressively increasing iterations, the first parameter combination is selected where the wind speed error decrease is less than the preset decrease threshold, or the spatial continuity index is lower than the preset index threshold.
[0055] The number of iterations, K, controls the spatial range of residual information propagation along the prevailing wind direction. When K is small, the correction is mainly limited to the local grid; when K is large, the residual information can propagate over a larger spatial range, but an excessively large K may lead to over-smoothing of the residual field. Therefore, K can be selected from several discrete candidate values, such as 1 to 6 propagations, and the stopping criterion is when the validation set error no longer decreases significantly or spatial details begin to be over-smoothed.
[0056] The transport step size α is used to control the update amplitude of neighborhood residual information in a single propagation. If α is too small, spatial propagation will be insufficient; if α is too large, it may lead to residual field oscillation or excessive diffusion. Therefore, α and K can be adjusted together to select a value with lower validation set error while ensuring stable updates of the residual field. Candidate values can be set in the range of 0.05 to 0.5, and the specific value is determined according to the grid resolution, wind farm scale, and prevailing wind direction stability.
[0057] After anisotropic transport is completed, the final residual field is superimposed on the original WRF-Fitch reference wind speed field to obtain the corrected wind speed field: WScorr = WS WRF +ΔWS, while preserving the WRF physical background field, performs constrained addition and subtraction corrections on regions with systematic biases. In actual calculations, non-negative constraints can also be applied to the corrected wind speed field, and the residuals of invalid meshes can be set to zero using a unit mask to ensure that the output results have physical meaning.
[0058] Unlike traditional point-by-point regression, this invention does not build an independent correction model for each wind turbine or each grid point. Instead, it first identifies deviations across the entire two-dimensional wind field network and then propagates them spatially based on wind direction constraints. Therefore, the output is a continuous corrected wind field with consistent wind direction, which is more suitable for describing spatially correlated wind speed errors and power prediction deviations in large wind farms.
[0059] As a specific implementation, to verify whether the wind speed correction results can be stably transmitted to the power generation side, a power consistency verification is performed after the wind farm correction. This involves further converting the corrected wind speed field into a total power prediction result and comparing it with the actual generated power, thereby verifying the consistency and usability of the wind farm correction results in power generation operations. This step is not a necessary component of the wind farm correction operator, but rather an optional verification and extension part.
[0060] Power conformance verification can be performed using the following steps.
[0061] Acquire and clean the power samples. Read the actual generated power, the overall power prediction from numerical weather forecasts or WRF-Fitch output, multi-source meteorological characteristics, and wind speed field related characteristics obtained by the wind field correction operator at the same time resolution. Sort and align the timestamps uniformly; for a small number of missing actual generated power values, use time interpolation to fill in the gaps; for negative power values, prune them to zero to meet the physical condition that the actual power of the wind farm is non-negative.
[0062] Determine the baseline power. The total day-ahead power output from the WRF-Fitch model is used as the baseline power. This baseline power is the original forecast result on the generation side, and in this embodiment, it is cleaned and denoted as Ptotal.
[0063] Estimating the effective maximum power output. To avoid issues such as theoretical installed capacity failing to reflect power curtailment, availability, equipment status, and actual operational constraints, the effective maximum power output of the power station is automatically estimated based on historical power samples. Specifically, high-power candidate periods are identified using the first preset high quantile of the baseline power Ptotal, and the duration of this high-power state must reach a preset length. Within the period that meets the sustained high-power condition, the second preset high quantile of the actual generated power is extracted as the effective maximum power output (CAP). If no high-power period meets the duration requirement, the process can degenerate to extracting the third preset high quantile of the actual generated power from all high-power candidate periods as the CAP. This CAP represents the effective maximum power output that the power station can achieve under actual operating conditions.
[0064] Power correction features are constructed. Using the baseline power Ptotal as the core input, time features, wind speed features, wind direction features, corrected wind field-related features, and meteorological background factors such as boundary layer height, precipitation, temperature, air pressure, humidity, air density, and sensible or latent heat flux are incorporated to form a feature vector for power correction. These features are used to describe the systematic bias of the original power forecast under different inflow, boundary layer states, and seasonal diurnal variations.
[0065] A power correction model is trained. A supervised learning regression model is used to establish the correction relationship between the baseline power and the actual power. The regression model can be a gradient boosting tree, random forest, neural network, or other nonlinear regression model. In this embodiment, an XGBoost regression model is used, with the feature vectors during the training period as input and the actual power as the supervision label for training. A validation period is used to control model complexity and overfitting. The model output can be represented as the corrected power without physical upper bound constraints. In another alternative, the model objective can be set as the power residual between the actual power and the baseline power. The model first outputs the power residual, and then adds the residual to the baseline power to obtain the corrected power. For the implementation of directly outputting the corrected power, the difference between the model output and the baseline power can also be understood as the power residual obtained by implicit learning. Therefore, the residual power correction of this invention includes both explicit residual output and an equivalent method of directly outputting the corrected power with the baseline power as the core input.
[0066] Physical boundary constraints are applied to the power generation side. For the power output of the power correction model, a non-negativity constraint is first applied, clipping predicted power values less than zero to zero; then, an effective full-power constraint is applied, clipping predicted power values exceeding the CAP to the CAP. This yields the corrected power result after adding physical boundaries. This constraint prevents the machine learning model from generating power predictions exceeding the actual operating capacity of the power plant under high wind speeds or abnormal weather conditions. Parameters such as cut-in wind speed, rated wind speed, cut-out wind speed, and rated power are physical consistency constraints in this embodiment, not the source of the reference power in this embodiment. When the reference power is already directly provided by the existing WRF-Fitch business system, there is no need to re-execute the unit power curve mapping; the aforementioned wind speed boundary is mainly used to check or limit whether the wind speed-power relationship conforms to the unit operating rules. For example, when the corrected wind speed is lower than the cut-in wind speed, no positive single-unit power generation contribution should be generated; when the corrected wind speed reaches or exceeds the rated wind speed and the cut-out protection is not triggered, the single-unit power should not exceed the rated power; when the corrected wind speed exceeds the cut-out wind speed or meets the shutdown protection conditions, it should be handled as a shutdown or restricted operation state. For the site-level implementation, the above single-unit constraints can be equivalently reflected as the total site power not being lower than zero and not exceeding the effective full-power CAP.
[0067] Generate day-ahead power verification results for the entire field. Compare the original reference power, the corrected power without CAP constraints, and the corrected power with CAP constraints to the actual generated power, and calculate the normalized error and prediction accuracy. Using CAP as the normalization denominator, first calculate the relative error between the predicted power and the actual generated power at each time point, and limit abnormal relative errors to a preset range. Then calculate the normalized root mean square error to obtain the prediction accuracy. If the power prediction results after incorporating wind field correction features and physical boundary constraints show lower error and higher accuracy compared to the reference power during the test period, it indicates that the wind speed correction results can be effectively transmitted to the generation side.
[0068] In addition to a wind speed field correction system based on a wind field correction operator, this invention also provides a wind speed field correction system based on a wind field correction operator, comprising: The data acquisition module acquires baseline forecast data, source auxiliary meteorological variables, and historical observation data of wind farm units, including the baseline wind speed field. The preprocessing module performs spatiotemporal alignment and preprocessing on the data, constructs the input feature tensor, and generates the unit mask; The residual initialization module receives the input feature tensor and generates an initial residual field; The anisotropic transport module iteratively propagates the initial residual field and spatially redistributes the residual information along the axis of the prevailing wind direction to obtain the final residual field. The field synthesis module superimposes the final residual field onto the reference wind speed field to obtain a spatially continuous corrected wind speed field.
[0069] The parameter selection in this invention is not based solely on the error of a single moment or a single unit, but rather comprehensively considers the wind speed error at the effective unit location, the overall spatial continuity, the rationality of the prevailing wind direction propagation, and the power consistency verification results. When the errors of multiple parameter combinations on the validation set are similar, the parameter combination with fewer iterations, smaller transport step size, lower loss weight, and a smoother and more stable spatial residual field is preferentially selected to reduce the complexity of the operator model and improve generalization ability.
[0070] This invention can be applied to offshore wind farms, onshore wind farms, large-scale wind power base clusters, and other scenarios requiring spatially consistent meteorological field correction. For offshore wind farms, the transport direction and neighborhood weights are constructed primarily using prevailing wind direction, land-sea thermal differences, boundary layer state, and wake propagation relationships. For onshore wind farms, features such as terrain height, slope, roughness, underlying surface type, and terrain shading relationships are further introduced. For large-scale wind power base clusters, site-level or regional-level topological relationships are constructed based on the distance between sites, prevailing flow direction, and electrical zoning. The above extensions only involve adjustments to input features, neighborhood relationships, and parameter values, and do not change the basic process of this invention for achieving spatially consistent wind field correction based on residual initialization and directional transport.
[0071] Example 1: To further illustrate the feasibility and technical effectiveness of the method of the present invention (Wind Farm Correction Operator WFCO), a day-ahead wind farm correction task in a certain region is used as an example. The location of the wind farm and the distribution of the turbines are as follows: Figure 2 As shown, the study includes multiple large-capacity offshore wind turbines, covering 11 consecutive months from January 2024 to November 2024, which can reflect the wind field variation characteristics under different seasons, different prevailing wind directions, and different boundary layer conditions.
[0072] The input data used in this embodiment includes: wind field from the WRF-Fitch numerical weather prediction driven by ECMWF IFS, high-resolution regional forecast products from CAS Tianji, and SCADA operation data of wind turbines. The WRF-Fitch model employs a multi-layered nested region, with the innermost grid covering the target wind farm, outputting hub height wind field data at a 15-minute time resolution, along with auxiliary meteorological variables such as boundary layer height, temperature, air pressure, humidity, and precipitation. CAS Tianji's regional forecast products serve as a supplementary meteorological forecast source to enhance the representation of regional wind field characteristics. SCADA data includes variables such as turbine hub height wind speed, wind direction, and active power. SCADA wind speed serves as the primary monitoring signal for wind field correction, while active power is used for subsequent power consistency verification.
[0073] WRF-Fitch, CAS Tianji, and SCADA data were unified to a 15-minute time resolution and aligned according to timestamps. For SCADA data, abnormal operating conditions such as power curtailment, faults, maintenance, abnormal pitch angles, and obviously unreasonable wind speeds were removed, retaining only valid observation data under normal power generation conditions.
[0074] In terms of spatial processing, the WRF-Fitch wind field and SCADA observed wind speeds are mapped onto the same two-dimensional wind field grid. For grids containing valid unit observations, the average SCADA wind speed of normally operating units within that grid is used as the observed wind speed. For grids without valid unit observations, they are not included in the supervised loss calculation, but are still retained in the complete spatial wind field to ensure that the correction results are output as a continuous wind field of the entire field.
[0075] The samples are divided according to the time sequence within the month: the 1st to the 20th of each month are used as training samples, the 21st to the 25th as validation samples, and the 26th and thereafter as test samples. Model parameter fitting uses only the training samples, hyperparameter selection and early termination judgment use only the validation samples, and the test samples are only used for the final independent evaluation, thereby avoiding future information leakage and simulating the actual application scenarios in day-ahead forecasting operations.
[0076] like Figure 5 The table shows a time series comparison of actual observations, the original WRF wind field, and the wind speed corrected by the WFCO method of this invention. On independent test samples, the root mean square error (RMSE) of the hub height wind speed at the effective unit grid was calculated for both the original WRF-Fitch wind field and the wind field corrected by this invention. As shown in Table 1, the test results indicate that in the 11-month test sample from January to November 2024, the method of this invention achieved positive improvement in every month, demonstrating that the method is not only effective for individual months or weather events, but also has a stable improvement effect across multiple months and wind conditions. Overall, the RMSE of the hub height wind speed decreased from 2.359 m / s in the original WRF-Fitch result to 1.109 m / s, an absolute reduction of 1.249 m / s and a relative reduction of 53.0%.
[0077] Table 1 Monthly Performance Summary (Test samples are from the 26th of each month onwards)
[0078] From a spatial distribution perspective, the method of this invention shows an overall positive error reduction across the effective unit grid, with no systematic degradation regions observed. The error improvement is not concentrated in a few isolated grids, but rather exhibits a relatively wide spatial distribution within the wind farm area, indicating that the correction process of this invention has farm-level consistency, rather than being a simple local location fitting.
[0079] The monthly wind direction concentration coefficient in Table 2 is used to characterize the concentration of the dominant wind direction within the test month, while the spatial concentration is used to determine whether the error improvement is excessively concentrated in a few grids. The results show that the persistence of wind direction varies greatly in different months, but the spatial concentration remains within a narrow range, indicating that the spatial improvement of this invention has good field-level distribution characteristics.
[0080] Table 2 Monthly Spatial Consistency Diagnostic Indicators
[0081] To further characterize the incoming wind direction organization and spatial distribution characteristics of error improvement in different test months, this embodiment introduces the intra-month wind direction concentration factor R. dir Spatial concentration S of error improvement 10 .
[0082] Monthly wind direction concentration factor R dir This is used to measure the concentration of incoming wind direction within a test month. Let θ be the representative incoming wind direction at the t-th valid time point. t If the total number of valid moments is N, then the monthly wind direction concentration factor R is... dir for When the incoming wind direction is relatively consistent at various times, the superposition of unit direction vectors results in a large composite length, R. dir Approaching 1; when the wind direction is distributed over a wide angular range, vectors of different directions cancel each other out, R dir Close to 0.
[0083] Spatial Concentration S 10 This represents the proportion of the positive error reduction contributed by the approximately 10% of effective unit grids with the largest error improvement, out of the total positive error reduction. ,in, S represents the reduction in positive error across each effective unit grid after adopting the revised method of this invention. When the error improvement is approximately uniformly distributed across the grids, S... 10 Approaching 0.1; when the error improvement is mainly concentrated in a few grids, S 10 Increase. Therefore, the lower S 10 This indicates that the error improvement generated by the present invention has better field-level spatial distribution characteristics, rather than being dominated by significant improvements in a few local grids.
[0084] To further illustrate how this invention adapts to different wind conditions and organizational patterns, this embodiment selects two typical test months, May 2024 and July 2024, for comparative analysis. Among them, as... Figure 3 The May figure shows a month with more dispersed wind directions, such as... Figure 4 The figure for July represents a month with highly concentrated wind direction. By comparing the wind rose diagrams and spatial error improvement distributions for two typical months, it can be further verified that the present invention does not simply apply a uniform translation across the entire field, but is able to form different spatial correction morphologies based on the incoming flow structure.
[0085] The low wind direction persistence coefficient of the test samples in May 2024 indicates that the distribution of incoming flow directions was relatively dispersed that month, with no single, long-term stable dominant incoming flow channel. Under these wind conditions, the spatial improvement corrected by this invention was not concentrated on a single narrow downstream path, but rather showed a positive error reduction in multiple areas of the wind farm. This suggests that when wind direction changes frequently, the residual initialization operator is mainly used to identify the overall field background deviation, while the anisotropic input operator distributes the residual information under different incoming flow direction constraints, thus forming a more uniform field-level correction effect.
[0086] The test sample from July 2024 exhibited a high wind direction persistence coefficient, and the wind rose diagram showed that the incoming flow direction was concentrated within a narrow sector, representing a typical stable dominant incoming flow condition. Under this type of wind condition, the spatial improvement field of this invention exhibits more pronounced directional organization characteristics, with a larger RMSE reduction region expanding downstream, consistent with the physical laws governing wake propagation and wind field error propagation under stable incoming flow conditions. This result demonstrates that the anisotropic transport operator can enhance residual propagation along the wind direction when the dominant wind direction is clear, rather than learning arbitrary spatial correlations.
[0087] As can be seen from the two typical months mentioned above, the present invention can perform background deviation correction across the entire field in months with dispersed wind direction, and can perform structured correction along the dominant incoming flow direction in months with continuous wind direction. In both cases, there was no improvement contributed only by a few isolated grids, indicating that the present invention has good spatial consistency and wind condition adaptability.
[0088] To verify whether the wind speed correction results can be stably transmitted to the power generation side, this embodiment further constructs a wind speed-power consistency verification module. This module uses the total power output by the WRF-Fitch as the reference power and establishes a residual power correction model based on it to learn the deviation between the actual generated power and the reference power. The inputs to the power correction model include the wind speed correction magnitude, the effective wind speed characteristics constrained by the cut-in wind speed and the rated wind speed, the cubic growth characteristics within the cut-in to rated range, and physically interpretable characteristics such as the WRF-Fitch reference power.
[0089] Regarding power-side constraints, operating boundaries are applied to the corrected power results: power is zero below the cut-in wind speed, power does not exceed the rated power after reaching the rated wind speed, and a monotonically increasing relationship is maintained between the cut-in wind speed and the rated wind speed. The final power results are then trimmed to upper and lower limits to ensure they remain within a reasonable operating range. Figure 6 The diagram shows a comparison of the actual observed wind speed, the original WRF wind field, and the wind speed corrected by the WFCO method of this invention to obtain the overall power. Test results show that the accuracy of the overall day-ahead power prediction improved from 77.0% of the WRF-Fitch baseline result to 86.6% after correction, an improvement of 9.6 percentage points.
[0090] The above embodiments are further elaborations and descriptions of the present invention to facilitate understanding, and are not intended to limit the present invention in any way. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A wind speed field correction method based on a wind field correction operator, characterized in that, The wind field correction operator includes the residual initialization operator, the field anisotropy input operator, and the corrected wind field synthesis; Acquire baseline forecast data, including the baseline wind speed field, auxiliary meteorological variables from the same source, and historical observation data of wind farm units; Spatiotemporal alignment and preprocessing are performed to construct the input feature tensor and generate a unit mask that identifies the effective observation location; Input the input feature tensor into the residual initialization operator to generate an initial residual field characterizing the magnitude of wind speed deviation; Based on the prevailing wind direction θ corresponding to the reference wind speed field, the initial residual field is iteratively propagated through the field anisotropic input operator to obtain the final residual field; The field anisotropic transport operator is: the residual field u after the (k+1)th propagation. k+1 The residual field u after the kth propagation k The difference is equal to the difference between the transport step size and the spatial propagation operator L. θ (u k The product of ) L θ (u k ) is |sinθ| multiplied by u k The second derivative with respect to x, plus |cosθ| multiplied by u k The second derivative with respect to y, with the x-axis in the east-west direction and the y-axis in the north-south direction; The final residual field is superimposed on the reference wind speed field to obtain the corrected wind speed field.
2. The wind speed field correction method based on a wind field correction operator according to claim 1, characterized in that, It also includes training the wind field correction operator, inputting the baseline forecast data and the same source auxiliary meteorological variables into the trained wind field correction operator to obtain the corrected wind speed field; The training is optimized using a joint loss function, which adds a smoothing regularization term and an L2 regularization term for the residual field amplitude to the Huber loss with unit masking.
3. A wind speed field correction method based on a wind field correction operator according to claim 1 or 2, characterized in that, The residual initialization operator is a two-dimensional convolutional encoder; The two-dimensional convolutional encoder contains at least two convolutional layers, each convolutional layer is followed by a nonlinear activation function, and finally a single-channel residual field is output through a two-dimensional convolutional layer. The convolution kernel parameters and bias parameters of the residual output layer are initialized to zero at the start of training and adjusted based on the wind speed error at the effective observation location through backpropagation.
4. The wind speed field correction method based on a wind field correction operator according to claim 1, characterized in that, The method for determining the iteration number and transport step size of the field anisotropic transport operator is as follows: Set the candidate value ranges for the number of iterations and the transport step size respectively; Based on a pre-set validation sample set, wind speed error and spatial continuity index are compared under different parameter combinations. Among the parameter combinations with progressively increasing iterations, the first parameter combination is selected where the wind speed error decrease is less than the preset decrease threshold, or the spatial continuity index is lower than the preset index threshold.
5. A wind speed field correction method based on a wind field correction operator according to claim 1 or 2, characterized in that, The process of performing spatiotemporal alignment and preprocessing to construct the input feature tensor includes: Unify the time resolution of all data; For data with different spatial resolutions, the data is mapped to a two-dimensional wind field grid consistent with the reference wind speed field based on interpolation or grid resampling, and the observation data of multiple units within the same grid are averaged. After periodically encoding or standardizing the spatiotemporally aligned data, they are concatenated into an input feature tensor.
6. The wind speed field correction method based on a wind field correction operator according to claim 5, characterized in that, The aforementioned auxiliary meteorological variables are meteorological variable fields from the same source as the baseline forecast data, which identify the spatial distribution characteristics of wind speed errors under different weather backgrounds; The acquired data also includes external auxiliary meteorological forecast data, which, after spatiotemporal alignment and preprocessing, is used as additional feature input to the residual initialization operator.
7. A wind speed field correction method based on a wind field correction operator according to claim 1 or 2, characterized in that, It also includes a power compliance verification step: Wind field features are extracted based on the corrected wind speed field, and feature vectors are constructed by combining the baseline power and meteorological background factors. Build and train a power correction model between the baseline power and the actual power, and output the corrected power without applying a physical upper limit; Determine the upper limit of the effective full-capacity power of the wind farm; Apply nonnegative constraints and effective full-power upper limit constraints to the corrected power to generate the final day-ahead power prediction results for the entire field. The predicted daytime power output is compared with the actual power output, and the normalized error is calculated.
8. The wind speed field correction method based on a wind field correction operator according to claim 7, characterized in that, The method for determining the effective full-power upper limit includes: Using the first preset high quantile of the reference power as a standard, the corresponding high power candidate time period is identified; Filter out time periods where the high-power state lasts for a preset length; Within the selected time period, the second preset high quantile of the actual power is extracted as the effective upper limit of the full power.
9. A wind speed field correction system based on a wind field correction operator, applicable to the wind speed field correction method as described in any one of claims 1-8, characterized in that, include: The data acquisition module acquires baseline forecast data, source auxiliary meteorological variables, and historical observation data of wind farm units, including the baseline wind speed field. The preprocessing module performs spatiotemporal alignment and preprocessing on the data, constructs the input feature tensor, and generates the unit mask; The residual initialization module receives the input feature tensor and generates an initial residual field; The anisotropic transport module, based on the prevailing wind direction θ corresponding to the reference wind speed field, iteratively propagates the initial residual field through the field anisotropic transport operator to obtain the final residual field; the field anisotropic transport operator is: the residual field u after the (k+1)th propagation. k+1 The residual field u after the kth propagation k The difference is equal to the difference between the transport step size and the spatial propagation operator L. θ (u k The product of ) L θ (u k ) is |sinθ| multiplied by u k The second derivative with respect to x, plus |cosθ| multiplied by u k The second derivative with respect to y, with the x-axis in the east-west direction and the y-axis in the north-south direction; The field synthesis module superimposes the final residual field onto the reference wind speed field to obtain a spatially continuous corrected wind speed field.
Citation Information
Patent Citations
Forecasting wind speed high-precision correction method and system oriented to wind power prediction of electric power system
CN120525133A
Wind power plant wind speed correction method and system based on dynamic space-time modeling
CN120687794A
CNN-Transform-based near-surface meteorological element forecast deviation correction method and system
CN122388397A
