Wind resource assessment method based on millimeter-wave wind radar

Abstract

This invention relates to the field of wind energy resource assessment technology and discloses a wind resource assessment method based on millimeter-wave wind-measuring radar. The method includes: acquiring wind field data of a target area using millimeter-wave wind-measuring radar and preprocessing it to generate a standardized dataset; inputting the data into a prediction model to obtain a wind energy distribution feature map; generating a parameter set based on the feature map using threshold segmentation and dynamically constructing an evaluation function for simulation assessment; simultaneously collecting actual wind energy data and calculating the deviation matrix between the predicted and actual data to quantify the assessment error; using this error to correct the parameters of the prediction model to obtain an optimized model; and finally, regenerating the feature map using the optimized model to complete the resource assessment. This method, through online model calibration and dynamic construction of the evaluation function, achieves closed-loop optimization and data-driven assessment of the assessment process, improving the accuracy of the assessment results and its adaptability to specific sites.

Description

Technical Field

[0001] This invention relates to the field of wind energy resource assessment technology, specifically a wind resource assessment method based on millimeter-wave wind measuring radar. Background Technology

[0002] Traditional wind resource assessment heavily relies on the establishment of anemometer towers for fixed-point, limited-height wind speed and direction observations. The acquired data needs to be spatially extrapolated using wind field models based on computational fluid dynamics to estimate the wind energy distribution across the entire target area. This method is highly dependent on the accuracy and universality of the model; once the model parameters are set, they typically remain fixed throughout the assessment period. The assessment process is unidirectional, from inputting measurement data into the model to the final output of the assessment report, lacking a step for optimizing the model based on actual wind condition feedback. The criteria and functions used in the assessment are often pre-defined based on general experience, failing to adapt to the complex wind resource distribution patterns of specific sites.

[0003] Existing technical solutions have shortcomings. Wind measurement towers have limited representative range and are costly and time-consuming to construct. Static models struggle to accurately reflect the fine structure of wind fields in complex terrain, leading to unquantifiable and uncorrectable systematic biases between assessment results and actual wind farm output. Fixed assessment functions are insufficient to automatically identify the core areas with the highest development value in regions with uneven wind energy distribution, resulting in inadequate accuracy and site adaptability of the assessment results. The precision and reliability of wind resource assessment are limited by inherent processes, necessitating a dynamic assessment method capable of self-verification and self-correction. Summary of the Invention

[0004] The purpose of this invention is to provide a wind resource assessment method based on millimeter-wave wind radar to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, this invention provides a wind resource assessment method based on millimeter-wave wind-measuring radar, the method comprising:

[0006] A millimeter-wave wind radar system is deployed in the target area to collect raw wind field data streams. The raw wind field data streams are then time-series aligned and outlier filtered to generate a standardized wind dataset.

[0007] A standardized wind dataset is input into a wind energy potential prediction model, and the model calculates and outputs a wind energy distribution feature map, which includes wind speed distribution and energy density information.

[0008] Based on the wind energy distribution feature map, a threshold segmentation algorithm is used to generate an initial evaluation parameter set, and an initial evaluation function is constructed based on the initial evaluation parameter set;

[0009] The initial evaluation function is executed to simulate and evaluate wind resources, while the actual wind energy output data of the target area is collected to form an actual wind energy dataset.

[0010] Calculate the deviation matrix between the actual wind energy dataset and the wind energy distribution characteristic map, and evaluate the error through deviation matrix analysis;

[0011] The parameters of the wind energy potential prediction model are corrected using the evaluation error to generate an optimized wind energy potential prediction model.

[0012] The wind energy distribution feature map was recalculated using an optimized wind energy potential prediction model, and a final wind resource assessment was performed based on the new feature map.

[0013] Preferably, the millimeter-wave wind radar system deployed in the target area collects raw wind field data streams, performs time-series alignment and outlier filtering on the raw wind field data streams, and generates a standardized wind dataset, including:

[0014] Configure the scanning parameters of the millimeter-wave wind radar to capture raw wind speed and direction data of the target area at fixed time intervals, forming a raw wind field data stream;

[0015] The original wind field data stream is segmented using a sliding window algorithm. Within each data segment, median filtering is used to remove impulse noise, resulting in a filtered data stream.

[0016] The filtered data stream is timestamped and aligned, and interpolation algorithms are used to compensate for missing data points to ensure data continuity.

[0017] The aligned data stream is normalized to a unified dimension to generate a standardized wind dataset.

[0018] Preferably, the step of inputting the standardized wind dataset into the wind energy potential prediction model and calculating and outputting a wind energy distribution feature map through the model includes:

[0019] The wind energy potential prediction model is constructed as a multilayer perceptron network. The input layer receives a standardized wind dataset, the hidden layer performs feature transformation, and the output layer generates a wind energy distribution feature map.

[0020] The wind energy potential prediction model is trained using historical wind energy data, and the network weights are adjusted through the backpropagation algorithm to minimize the prediction error.

[0021] The standardized wind dataset is input into the trained wind energy potential prediction model, and the model outputs a grid map of wind speed distribution and energy density values ​​to form a wind energy distribution feature map.

[0022] Preferably, the step of generating an initial evaluation parameter set based on a threshold segmentation algorithm using a wind energy distribution feature map, and constructing an initial evaluation function based on the initial evaluation parameter set, includes:

[0023] An adaptive threshold segmentation algorithm is applied to the wind energy distribution feature map to divide the feature map into high, medium and low wind energy regions based on the energy density value;

[0024] Statistical characteristics of each region are extracted, including average wind speed and energy variance, to form an initial set of evaluation parameters;

[0025] Based on the initial evaluation parameter set, the evaluation function is designed as a linear weighted combination, with weights allocated based on regional importance, thus generating the initial evaluation function.

[0026] Preferably, the step of executing the initial evaluation function to perform wind resource simulation evaluation, and simultaneously collecting actual wind energy output data of the target area to form an actual wind energy dataset, includes:

[0027] Deploy wind energy monitoring equipment in the target area to record actual wind speed and power output data, with the time synchronized with the simulation evaluation cycle;

[0028] Run the initial evaluation function, output the simulated wind energy distribution results, and record the simulated values;

[0029] Collect actual wind energy output data and match it with simulated values ​​at specific times to form an actual wind energy dataset.

[0030] Preferably, the step of calculating the deviation matrix between the actual wind energy dataset and the wind energy distribution characteristic map, and evaluating the error through deviation matrix analysis, includes:

[0031] Align the actual wind energy dataset with the predicted values ​​of the wind energy distribution feature map by time series and calculate the absolute deviation value at each time point;

[0032] The organizational deviation values ​​are formed into a matrix structure, where rows represent time points and columns represent spatial locations, thus generating a deviation matrix.

[0033] Principal component analysis was applied to the deviation matrix to extract the main error components and quantify the error.

[0034] Preferably, the step of using the evaluation error to correct the parameters of the wind energy potential prediction model and generate an optimized wind energy potential prediction model includes:

[0035] The evaluation error is used as the loss function, and the gradient descent algorithm is used to adjust the network weights of the wind energy potential prediction model.

[0036] The model parameters are updated through an iterative optimization process until the loss function converges.

[0037] Verify the performance of the optimized model on the validation dataset and generate an optimized wind energy potential prediction model.

[0038] Preferably, the step of recalculating the wind energy distribution characteristic map using an optimized wind energy potential prediction model and performing a final wind resource assessment based on the new characteristic map includes:

[0039] Input the latest standardized wind dataset into the optimized wind energy potential prediction model and output the updated wind energy distribution feature map.

[0040] Based on the updated wind energy distribution feature map, a new set of evaluation parameters is generated by reapplying the threshold segmentation algorithm.

[0041] The final assessment function is constructed based on the new assessment parameter set, and the final wind resource assessment is performed.

[0042] Preferably, the step of applying the sliding window algorithm to segment the raw wind field data stream includes:

[0043] Set the size and step size of the sliding window; the window size is determined based on the typical cycle of wind speed changes.

[0044] Wavelet transform is applied to the data within each window to extract time-frequency features;

[0045] Calculate the statistical characteristics of the data within the window, including mean, variance, and skewness;

[0046] Feature vectors are constructed based on statistical features and used for subsequent anomaly detection.

[0047] Preferably, updating the model parameters through the iterative optimization process includes:

[0048] Set an early stopping mechanism to terminate training when the validation set loss function fails to improve for several consecutive iterations.

[0049] An adaptive learning rate adjustment strategy is adopted to dynamically adjust the learning rate based on the direction of gradient change.

[0050] A weight decay term is added during the optimization process to control model complexity;

[0051] Record the loss function value for each iteration and plot the convergence curve to monitor the training process.

[0052] Compared with the prior art, the beneficial effects of the present invention are:

[0053] By synchronously collecting power output data from actual wind farm operations, a real-world wind energy dataset is constructed that is comparable in time and space to the wind energy distribution feature map predicted by the model. The deviation matrix between the two is then precisely calculated. This deviation matrix quantifies the systematic error between the model prediction and actual wind conditions at different locations and speed ranges. This matrix is ​​used to perform targeted correction of the internal parameters of the wind energy potential prediction model, rather than relying on preset general parameters or offline, intermittent calibration. A continuous feedback loop is established, transforming the model from a static prediction tool into an intelligent system capable of sensing its own prediction deviations and dynamically adjusting. By continuously absorbing real-world operational feedback from specific sites, the model gradually learns and adapts to the unique wind field patterns caused by complex factors such as local topography, atmospheric boundary layer characteristics, and wake effects. This improves the reliability and accuracy of future wind energy potential predictions for that site, effectively overcoming the inherent defect of accuracy degradation in traditional general models applied to complex terrain.

[0054] Based on high-resolution wind energy distribution feature maps generated by millimeter-wave radar, an adaptive threshold segmentation algorithm is used to automatically identify key intervals and spatial distribution patterns of wind speed and energy density in the maps. A set of evaluation parameters is extracted from these identified key features, and a dedicated evaluation function is dynamically constructed based on these parameters. This eliminates the reliance on fixed empirical formulas or uniform thresholds for evaluation criteria, instead allowing them to be directly driven by the inherent statistical characteristics and spatial distribution patterns of the current dataset. The logical structure and weight allocation of the evaluation function are fully adapted to the actual wind resource situation in the target area. For complex sites with multiple wind energy enrichment centers or extremely uneven resource distribution, the function can automatically focus on the areas with the highest commercial development value, achieving differentiated and accurate evaluation of wind resource potential. This avoids resource misjudgment or underestimation that may result from fixed evaluation criteria, enhancing the relevance and decision support value of the evaluation conclusions. Attached Figure Description

[0055] Figure 1 This is a schematic diagram illustrating the working principle of the wind resource assessment method based on millimeter-wave wind radar described in this invention.

[0056] Figure 2 Flowchart for generating wind energy distribution characteristic map;

[0057] Figure 3 A flowchart for constructing the initial evaluation function;

[0058] Figure 4 This is a wind energy region partitioning map based on adaptive threshold segmentation;

[0059] Figure 5 Heat map of the wind resource assessment and prediction deviation matrix. Detailed Implementation

[0060] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] Please see Figure 1 This invention provides a wind resource assessment method based on millimeter-wave wind-measuring radar. The method includes: deploying a millimeter-wave wind-measuring radar system in a target area, collecting raw wind field data streams, and performing time-series alignment and outlier filtering on the raw wind field data streams to generate a standardized wind dataset. The standardized wind dataset is input into a wind energy potential prediction model, which calculates and outputs a wind energy distribution feature map, which includes wind speed distribution and energy density information. Based on the wind energy distribution feature map, a threshold segmentation algorithm is used to generate an initial assessment parameter set, and an initial assessment function is constructed based on the initial assessment parameter set. The initial assessment function is executed to perform wind resource simulation assessment, while simultaneously collecting actual wind energy output data from the target area to form an actual wind energy dataset. The deviation matrix between the actual wind energy dataset and the wind energy distribution feature map is calculated, and the assessment error is analyzed using the deviation matrix. The assessment error is used to correct the parameters of the wind energy potential prediction model, generating an optimized wind energy potential prediction model. The optimized wind energy potential prediction model is used to recalculate the wind energy distribution feature map, and a final wind resource assessment is performed based on the new feature map.

[0062] Example 1: In practical implementation, deploying a millimeter-wave wind measurement radar system in the target area requires selecting a representative wind field area, such as open plains or coastal areas. The installation of the millimeter-wave wind measurement radar system must ensure that the radar scanning sector is unobstructed, and the radar antenna base remains horizontal and stable. Precise positioning and timing are achieved through GPS and BeiDou modules. The power supply for the millimeter-wave wind measurement radar system uses a combination of solar panels and batteries, and is equipped with redundant power modules to ensure continuous data acquisition. A data transmission link is established between the millimeter-wave wind measurement radar system and the data receiving station via a wireless data transmission radio or mobile communication network. Configuring the scanning parameters of the millimeter-wave wind measurement radar is a prerequisite for data acquisition. Scanning parameters include scanning mode, elevation angle sequence, range length, and pulse repetition frequency. The fixed time interval is set to a configurable value within the range of 10 minutes to 1 hour. The millimeter-wave wind measurement radar transmits millimeter-wave signals at this fixed time interval and receives backscattered signals from atmospheric particles. Radial wind speed and wind direction are calculated through Doppler frequency shift, forming a raw wind field data stream containing timestamps, latitude and longitude coordinates, altitude layers, wind speed values, and wind direction angles. In practice, the raw wind farm data stream is stored in real time on a local storage device in binary or custom format, and simultaneously sent to a remote data center via a data transmission link. During storage and transmission, the raw wind farm data stream employs data encryption and verification mechanisms to prevent data tampering and loss.

[0063] A sliding window algorithm is applied to segment the raw wind field data stream. This algorithm uses a fixed-length window that slides across the time series. The window size is determined based on the typical period of wind speed variation, which is referenced to the pulsation characteristics of wind speed in the atmospheric boundary layer and is typically set as an integer multiple of 10 minutes to 1 hour. The sliding step size is set to half or one-third of the window size. In practice, each sliding window covers a continuous time series data segment containing wind speed and direction observations from multiple scan cycles. The window size can be configured to 30 minutes, and the sliding step size to 15 minutes, enabling continuous segmentation of the data stream. Wavelet transform is applied to the data within each window, using the Mallat algorithm for multi-resolution analysis. The Daubechies or Symlets wavelet systems are selected as basis functions. Convolution operations are used to extract time-frequency features at different scales. These features include high-frequency detail coefficients and low-frequency approximation coefficients, reflecting the local fluctuation characteristics and trend components of the wind speed signal. The statistical characteristics of the data within the window are calculated, including mean, variance, and skewness. The mean reflects the average level of wind speed, the variance characterizes the intensity of wind speed fluctuations, and the skewness describes the asymmetry of wind speed distribution. A multi-dimensional feature vector is constructed based on these statistical characteristics, and this feature vector is used for subsequent anomaly detection and data quality control. In some embodiments, median filtering is used to remove impulse noise within each data segment. Median filtering selects a 3-point or 5-point sliding window. After sorting the data within the window by numerical value, the median value is used to replace the original value, effectively suppressing isolated abnormal impulse interference and generating a smooth filtered data stream.

[0064] The filtered data stream undergoes timestamp alignment, based on standard UTC time. The continuity of timestamps for each data point is checked. For missing timestamps due to data transmission delays or device sleep, linear interpolation or spline interpolation algorithms are used to compensate for the missing data points. Linear interpolation calculates the missing position based on the preceding and following valid data points, while spline interpolation uses a cubic spline function to fit the data curve, ensuring data continuity in the time dimension. In practice, the timestamp alignment process also includes removing redundant data corresponding to duplicate timestamps, retaining the earliest received or highest quality data record. The timestamp-aligned data stream forms a regular sequence with equal time intervals. The aligned data stream is then normalized to a unified dimension. Normalization uses either min-max scaling or Z-score standardization. Min-max scaling linearly transforms the data to the [0,1] interval, while Z-score standardization converts the data into a distribution with a mean of 0 and a standard deviation of 1. Unified dimension eliminates the scale effect of unit differences in wind speed and direction values. Understandably, normalized data facilitates the training and convergence of wind energy potential prediction models, generating standardized wind datasets that can be used as model input. Optionally, the standardized wind dataset is stored in NetCDF or HDF5 format. The dataset header contains metadata information, such as data source, processing time, and coordinate reference system. The dataset body is a multi-dimensional array structure, organizing wind speed and direction values ​​by time, altitude, and latitude / longitude dimensions.

[0065] The scanning parameter configuration of a millimeter-wave wind-measuring radar system needs to consider atmospheric environmental adaptability. In practice, scanning parameters are dynamically adjusted for different weather conditions. For example, signal transmission power is increased to compensate for attenuation during rainfall, and the scanning sequence is optimized to improve spatiotemporal resolution during clear weather. The fixed time interval of the millimeter-wave wind-measuring radar can be dynamically adjusted according to assessment needs. A longer interval is used in the preliminary wind resource assessment stage to reduce data volume, while a shorter interval is used in the detailed assessment stage to capture rapid wind speed changes. This flexible configuration of scanning parameters allows the millimeter-wave wind-measuring radar system to adapt to diverse wind resource assessment scenarios. The window size in the sliding window algorithm is determined based on the typical period of wind speed change. The typical period is obtained through spectral analysis of historical wind speed data. The power spectral density of the wind speed time series is calculated, and significant periodic components are identified as a reference for setting the window size. In practice, the window size can be set as an integer multiple of the main periodic components in the power spectrum to cover the complete wind speed fluctuation cycle. The sliding step size is smaller than the window size to achieve data segment overlap and avoid the loss of effective information. Optionally, the sliding window algorithm uses an overlapping window approach, where there is partial data overlap between adjacent windows. The data in the overlapping areas are then fused through weighted averaging in subsequent processing to improve the smoothness of data segment boundaries.

[0066] The time-frequency features extracted by wavelet transform are used to construct feature vectors. In addition to mean, variance, and skewness, these feature vectors can also incorporate statistics such as kurtosis and turbulence intensity to form a more comprehensive description of wind conditions. In specific implementations, the feature vectors are fed into an anomaly detection model. This model uses either the isolated forest algorithm or the local outlier factor algorithm to identify and mark outliers in the data segments. Outliers may be caused by equipment failure or extreme weather. Marked outliers are then removed or corrected in subsequent processing. In some embodiments, the feature vectors are also used for data quality identification. Data quality levels are classified based on the numerical range of the feature vectors. High-quality data segments are directly used for subsequent analysis, while low-quality data segments trigger re-acquisition or manual review. The median filtering method removes impulse noise while preserving the edge features of the wind speed signal. The window size for median filtering is selected based on the noise characteristics; a smaller window is used for high-frequency impulse noise, and a larger window is used for broadband noise. In practice, median filtering can be applied iteratively multiple times. After each filtering, the smoothness of the data is checked until the derivative of the data curve tends to stabilize. The filtered data stream needs to be validated. The validation methods include range checking and consistency checking. Range checking removes data points that exceed the physically reasonable value. Consistency checking compares whether the differences between adjacent data points are within the threshold.

[0067] The interpolation algorithm selection during timestamp alignment depends on the data missing pattern. Linear interpolation is used for random single-point missing data, while spline interpolation is used for continuous multi-point missing data. The interpolation algorithm is only applied to segments with a missing rate below a preset threshold; when the missing rate is too high, the segment is directly marked as invalid data. In practice, after timestamp alignment, a complete time series is generated. The time series uses a standard time axis as a reference, arranging data points at equal intervals. The continuity of the time series is verified by calculating the time difference between adjacent points; if the time difference exceeds the tolerance range, an alarm is triggered. The normalization process uses a unified unit to facilitate the fusion and analysis of data from different sources. Normalization parameters such as minimum, maximum, mean, and standard deviation are calculated from the training dataset and applied throughout the data processing flow to ensure the consistency of data transformation. The standardized wind dataset is ultimately stored in a distributed file system, establishing a data index and fast retrieval mechanism to support efficient access to the wind energy potential prediction model.

[0068] Example 2: See Figure 2In practical implementation, the wind energy potential prediction model is constructed as a multilayer perceptron network. The multilayer perceptron network structure includes an input layer, at least two hidden layers, and an output layer. The input layer is designed to receive a standardized wind dataset. Each sample in the standardized wind dataset contains wind speed, wind direction, and spatial coordinate information at a specific time point. The number of neurons in the input layer matches the feature dimension of the standardized wind dataset; for example, when each sample contains 10 feature values, the input layer has 10 neurons. The hidden layers perform feature transformation. Each hidden layer consists of a fully connected layer and an activation function. The first hidden layer receives the feature vector output from the input layer, performs a linear transformation using the weight matrix and bias vector, and then applies the ReLU activation function for nonlinear mapping. The second hidden layer processes the output of the first hidden layer in a similar manner, progressively extracting high-level feature representations. The output layer generates a wind energy distribution feature map. The number of neurons in the output layer is determined by the resolution of the target wind energy distribution feature map. For example, when the feature map is divided into a 100x100 grid, the output layer has 10,000 neurons, with each neuron corresponding to the predicted energy density value of a grid point.

[0069] The wind energy potential prediction model is trained using historical wind energy data, which includes standardized wind datasets collected in the target area over the past few years and corresponding measured wind energy distribution data. During training, the historical data is divided into training and validation sets. The training set is used to adjust the parameters of the multilayer perceptron network, while the validation set is used to monitor the training process and prevent overfitting. The network weights are adjusted using the backpropagation algorithm, which calculates the error gradient between the predicted output and the true value. This error gradient propagates backward from the output layer to the input layer, updating the weights and bias parameters based on the gradient descent principle.

[0070] The mean squared error is used as the loss function to minimize the prediction error. This loss function measures the difference between the output value of the wind energy potential prediction model and the actual wind energy distribution. The optimization process iteratively reduces the loss function value. The trained model can extract features from a new standardized wind dataset and generate accurate prediction results. The standardized wind dataset is input into the trained wind energy potential prediction model. The model performs forward propagation calculations. After transformation through each layer, the output layer generates the wind speed distribution and energy density values ​​for each grid point. These values ​​are organized into a two-dimensional grid map, forming the final wind energy distribution feature map. In practice, the number of hidden layers in the multilayer perceptron network can be adjusted according to the data complexity. For wind field evaluation with complex terrain, three or four hidden layers can be added to improve the model's expressive power. The number of neurons in each hidden layer is usually decreased layer by layer to achieve feature compression. In addition to ReLU, LeakyReLU or ELU functions can be used for the activation functions of the hidden layers to alleviate the gradient vanishing problem. The output layer selects a linear activation function or a sigmoid function based on the range of the predicted values. The careful design of the network structure directly affects the model's performance.

[0071] The preparation of historical wind energy data must ensure both time span and data quality. The time span should cover different seasons and weather patterns to provide comprehensive learning samples. Data quality is improved by cleaning outliers and noise. The division of the training and validation sets uses random sampling or time-series cross-validation to ensure consistent distribution between the two sets. The implementation of the backpropagation algorithm relies on automatic differentiation techniques, calculating the partial derivatives of the loss function with respect to each network parameter during training iterations.

[0072] The optimization process to minimize prediction error can be achieved by updating the weights using the following formula:

[0073] ,

[0074] in: This represents the connection weights from neuron i to neuron j in the l-th layer. The learning rate hyperparameter controls the step size for parameter updates. For loss function, This represents the gradient of the loss function with respect to the weights. Besides the basic gradient descent, optimization algorithms such as the momentum method or the Adam algorithm can be used to accelerate convergence.

[0075] The model training process sets a maximum number of iterations and an early stopping condition. Training terminates when the validation set loss no longer decreases for several consecutive cycles, and the model parameters with the best performance on the validation set are saved. The trained wind energy potential prediction model needs to be evaluated offline, using an independent test set to calculate prediction accuracy metrics such as mean absolute error or coefficient of determination. Optionally, the model can be deployed as an online learning system, periodically receiving new wind field data and fine-tuning parameters to adapt to long-term changes in the wind field environment. When inputting the standardized wind dataset into the model, the data needs to undergo the same preprocessing operations as in the training phase, including feature scaling and dimension alignment. The model's forward propagation is performed layer by layer using matrix multiplication and activation functions. The grid map values ​​generated by the output layer represent the predicted energy density, typically in kilowatts per square meter. The spatial extent and resolution of the grid map are determined during the model design phase, and it must cover the target evaluation area and meet accuracy requirements. In some embodiments, the wind energy distribution feature map can be further converted into a grayscale image or pseudo-color image for easier visualization and analysis. Optionally, the wind energy potential prediction model can integrate an uncertainty estimation module to generate a confidence interval map while outputting predicted values. The confidence intervals are implemented using Dropout technology or a Bayesian neural network, providing a reliable reference for wind resource assessment. The generation frequency of the wind energy distribution feature map can be synchronized with millimeter-wave radar data acquisition, for example, outputting an updated feature map every hour to achieve near real-time wind energy monitoring. The final generated wind energy distribution feature map is stored in a geospatial data format, embedding metadata information such as coordinate system and timestamps for subsequent evaluation modules to access.

[0076] Example 3: See Figure 3 In practical implementation, an adaptive threshold segmentation algorithm is applied based on the wind energy distribution feature map. This algorithm dynamically calculates the segmentation threshold based on the energy density value of each pixel in the wind energy distribution feature map. For example, the Otsu method is used to traverse all possible thresholds, maximizing the inter-class variance of high, medium, and low wind energy regions. The feature map is divided into high, medium, and low wind energy regions based on energy density values. High wind energy regions correspond to the set of pixels with energy density values ​​greater than the upper threshold, medium wind energy regions to the set of pixels with energy density values ​​between the upper and lower thresholds, and low wind energy regions to the set of pixels with energy density values ​​less than the lower threshold. Statistical features of each region are extracted, including the average wind speed and energy variance within each connected region. The average wind speed is calculated as the arithmetic mean of the wind speed values ​​of all pixels within that region, and the energy variance is calculated as the sum of the squared deviations of the energy density values ​​of all pixels within that region from their mean, divided by the number of pixels minus one. An initial evaluation parameter set is then formed. This initial evaluation parameter set is a structured dataset containing the identifier, geographical extent, average wind speed, energy variance, and percentage area information for each region.

[0077] Based on the initial evaluation parameter set, an evaluation function is designed as a linear weighted combination, mapping multiple statistical features to a comprehensive evaluation score. Weights are assigned based on regional importance, which is determined by both regional area share and average energy density. The weight coefficient for high-wind-energy regions is typically set to the maximum, followed by medium-wind-energy regions, and then the minimum for low-wind-energy regions. An initial evaluation function is generated, which can be expressed in the following form:

[0078] ,

[0079] in: This is a dimensionless comprehensive evaluation score. The average wind speeds for high, medium, and low wind energy areas. For reference wind speed, For energy variance, For reference energy density, This refers to the dimensionless weighting coefficients. It can be understood that by introducing a reference physical quantity, all terms are transformed into dimensionless numbers, ensuring consistency of dimensions on both sides of the formula. In the actual wind energy dataset collection process, measured wind speed data also needs to be normalized by dividing by the same reference wind speed to ensure comparability between simulated and measured values.

[0080] Wind energy monitoring equipment is deployed in the target area, including cup anemometers and ultrasonic anemometers on weather towers, as well as SCADA systems on wind turbine generators, to record actual wind speed and power output data. Time is synchronized with the simulation evaluation cycle, which is consistent with the generation cycle of the wind energy distribution feature map, for example, both being 1 hour. All device clocks are synchronized via Network Time Protocol (NTP) to ensure that the timestamp alignment accuracy between actual and simulated data is within seconds. An initial evaluation function is run, outputting simulated wind energy distribution results, including a comprehensive evaluation score S and detailed statistics for each region, and the simulated values ​​are recorded in the database. Actual wind energy output data is collected in real time from the wind energy monitoring equipment, aggregated along the time dimension after data verification and quality control, with the aggregation cycle matching the simulation evaluation cycle. Time point matching is performed with simulated values, based on a unified timestamp index, associating actual measured values ​​with model predicted values ​​within the same time window to form an actual wind energy dataset. The actual wind energy dataset is stored in tabular form, with each row containing a timestamp, geographical location, measured average wind speed, measured power output, and corresponding simulated prediction value.

[0081] In some embodiments, the adaptive threshold segmentation algorithm can employ a local adaptive strategy. When the spatial heterogeneity of the wind energy distribution feature map is high, the feature map is divided into several sub-windows, and the segmentation threshold is calculated independently within each sub-window before region merging. When extracting statistical features, in addition to average wind speed and energy variance, optional features such as turbulence intensity and prevailing wind direction frequency in the wind rose diagram can be added to enrich the information dimensions of the initial evaluation parameter set. It is understood that more comprehensive statistical features help improve the discriminative power of the evaluation function.

[0082] The weighting process can be quantified using the Analytic Hierarchy Process (AHP). By constructing a judgment matrix, the relative importance weights of each region and feature can be calculated, improving the scientific rigor of the weighting allocation. The initial evaluation function is not limited to linear weighting; in some embodiments, nonlinear or interaction terms can be introduced to capture complex relationships between features, such as adding a product term of average wind speed and energy variance. After generating the initial evaluation function, sensitivity analysis needs to be performed on a historical dataset to verify whether the function output's response to changes in input parameters meets expectations.

[0083] The deployment plan for wind energy monitoring equipment needs to consider spatial representativeness. Equipment locations should cover high, medium, and low wind energy areas, avoiding concentration in the center of wind farms. Data acquisition frequency should be higher than the evaluation cycle, for example, collecting raw data every minute, and then averaging or integrating according to the evaluation cycle. A regular calibration process should be established for the time synchronization mechanism to prevent data misalignment caused by clock drift. It is understandable that high-quality actual wind energy datasets are the foundation for model bias analysis. Before being stored, actual wind energy datasets need to undergo integrity checks and outlier labeling. Missing data should be repaired using interpolation methods or marked as invalid to ensure the continuity and reliability of the dataset.

[0084] See Figure 4This image is a core visualization result of the regional division and assessment parameter extraction stages in the wind resource assessment method based on millimeter-wave wind radar. The image constructs a 10×10km spatial grid of the target area using X and Y coordinates. Through an adaptive threshold segmentation algorithm, based on the energy density information of the wind energy distribution feature map, the area is divided into three categories: low-wind-energy, medium-wind-energy, and high-wind-energy regions. First, the adaptive threshold segmentation algorithm is applied to the wind energy distribution feature map acquired by the millimeter-wave radar and output by the model to maximize the inter-class variance of high, medium, and low-wind-energy regions, thereby accurately identifying spatial blocks with different wind energy potential. This type of regional division intuitively presents the spatial heterogeneity of wind energy resources in the target area, providing a spatial basis for subsequent extraction of statistical characteristics such as average wind speed and energy variance of each region, and for constructing targeted assessment functions. Compared to the limitations of traditional methods that rely on fixed empirical thresholds, the dynamic regional division shown in this image can adapt to the wind energy distribution patterns under complex terrain, effectively avoiding resource misjudgment. It is a key step in achieving differentiated and accurate evaluation of wind resources, supporting subsequent model correction and final assessment, and providing an intuitive and scientific spatial basis for prioritizing wind farm development areas.

[0085] Example 4: In specific implementation, calculating the deviation matrix between the actual wind energy dataset and the wind energy distribution feature map first requires aligning the predicted values ​​of the actual wind energy dataset and the wind energy distribution feature map according to time series. Time series alignment is based on Coordinated Universal Time (UTC). By matching timestamps, a correspondence is established between the actual wind speed value and the model's predicted value at each measurement point. The absolute deviation value at each time point is calculated using the method of subtracting the predicted value from the actual value and taking the absolute value. The deviation values ​​are organized into a matrix structure, where the row indices of the matrix correspond to consecutive time point sequences, and the column indices correspond to different spatial location codes. This generates the deviation matrix, which is mathematically represented as a two-dimensional array, where each element stores the deviation value at a specific time and spatial point. Refer to Table 1 to show the basic structure of the deviation matrix.

[0086] Table 1: Data Organization Structure of the Deviation Matrix

[0087] Time point number Latitude and longitude coordinates A deviation value Latitude and longitude coordinate B deviation value Latitude and longitude coordinate C deviation value T001 0.25 0.31 0.19 T002 0.18 0.27 0.22 T003 0.33 0.24 0.28

[0088] Principal component analysis (PCA) is applied to the deviation matrix. PCA extracts eigenvalues ​​and eigenvectors by calculating the covariance matrix and performing eigenvalue decomposition. Eigenvalues ​​reflect the magnitude of the variance of the error components, while eigenvectors indicate the main direction of error change. The main error components are extracted by selecting the first few eigenvectors with larger eigenvalues. The assessment error is quantified by calculating the cumulative variance contribution rate of the principal components, which represents the proportion of the total variance explained by the first k principal components. In essence, PCA effectively identifies the dominant error patterns in the deviation matrix. The assessment error is then used to correct the parameters of the wind energy potential prediction model. The assessment error is used as the loss function, defined as the Frobenius norm or other matrix norm of the deviation matrix. Gradient descent is employed to adjust the network weights of the wind energy potential prediction model. The gradient descent algorithm determines the update direction by calculating the partial derivative of the loss function with respect to the network weights. The model parameters are updated through an iterative optimization process. This process includes setting a maximum number of iterations and a loss convergence threshold, and implementing an early stopping mechanism. Training terminates when the loss function on the validation set fails to improve over several consecutive iterations; the number of consecutive iterations for early stopping can be configured to be 5 to 10. An adaptive learning rate adjustment strategy is employed, dynamically adjusting the learning rate based on the gradient direction. For example, the learning rate is increased when the gradient direction is consistent and decreased when the gradient direction oscillates. A weight decay term is added during optimization to control model complexity; this term adds a penalty term based on the parameter norm to the loss function. The loss function value for each iteration is recorded, and a convergence curve is plotted to monitor the training process, with the number of iterations on the x-axis and the loss value on the y-axis. The optimized model is then validated on the validation dataset, with performance metrics including mean absolute error and coefficient of determination, resulting in an optimized wind energy potential prediction model.

[0089] The parameter update process uses the following formula:

[0090] ,

[0091] in: Let be the model parameters for the τth iteration. For learning rate, For loss function, For gradient, To prevent division by zero by small constants, the early stopping mechanism monitors both the loss function value and the weight decay term as dimensionless values ​​during optimization. The adaptive learning rate adjustment strategy is based on changes in the dimensionless gradient direction.

[0092] See Figure 5This graph is a key visualization result in the deviation analysis and model calibration stages of wind resource assessment. With time points as the vertical axis and monitoring locations as the horizontal axis, the graph uses a red-blue gradient heatmap to visually present the distribution of deviations between model predictions and actual wind energy data at different time points and spatial locations. First, the actual wind energy dataset and the predicted values ​​from the wind energy distribution feature map are precisely aligned according to the time series. The absolute deviation value for each time point and spatial location is calculated, and these deviation values ​​are then organized into a matrix structure, ultimately visualized as a heatmap. This deviation matrix quantifies the model's prediction errors in different spatiotemporal dimensions, serving as the core basis for subsequent principal component analysis to extract major error components, quantify and assess errors, and ultimately calibrate the parameters of the wind energy potential prediction model. Compared to the limitations of traditional methods that lack spatial dimension analysis of errors, this graph clearly identifies which regions and time periods have insufficient prediction accuracy, providing intuitive and precise guidance for targeted model optimization. It is a key supporting tool for achieving dynamic iteration of wind resource assessment models and improving prediction reliability.

[0093] Example 5: In specific implementation, recalculating the wind energy distribution feature map using the optimized wind energy potential prediction model first requires inputting the latest standardized wind dataset into the optimized wind energy potential prediction model. The latest standardized wind dataset contains preprocessed and standardized wind field observation data of the target area within the current assessment period. The optimized wind energy potential prediction model is a multilayer perceptron network with improved prediction accuracy after parameter correction. After receiving the standardized input data, the model calculates through forward propagation and outputs an updated wind energy distribution feature map. The updated wind energy distribution feature map contains optimized wind speed distribution and energy density information, and its spatial resolution and coverage remain consistent with the original feature map. Based on the updated wind energy distribution feature map, a threshold segmentation algorithm is reapplied to generate a new set of assessment parameters. The threshold segmentation algorithm uses the same adaptive threshold calculation method as the initial assessment stage, but the segmentation threshold is recalculated based on the updated feature map data, dividing the feature map into high, medium, and low wind energy regions. Statistical features, including the region's average wind speed, energy density variance, and region area percentage, are extracted from each region to form a new set of assessment parameters. It can be understood that the new set of assessment parameters reflects the more accurate wind energy distribution characteristics after model optimization.

[0094] The final evaluation function is constructed based on the new evaluation parameter set. This final evaluation function adopts a linear weighted combination form, but the weight coefficients are recalibrated according to the importance of the optimized regional characteristics. The weight calibration considers the relative contribution of each region in the updated wind energy distribution characteristic map; the weight of high-wind-energy regions is typically increased, while the weight of low-wind-energy regions is correspondingly decreased. The final wind resource assessment is performed by calculating a comprehensive evaluation index from the final evaluation function. This comprehensive evaluation index quantitatively characterizes the wind energy development potential level of the target region. In some embodiments, the final evaluation function can be expressed as:

[0095] ,

[0096] in: The final evaluation score is dimensionless. These are dimensionless weighting coefficients. For the characteristic transformation function, Let the feature parameters be those of the k-th class region. These are reference values ​​for the corresponding features.

[0097] In practical implementation, the latest standardized wind dataset needs to undergo integrity verification before input, checking whether the data's time span and spatial coverage meet the model's input requirements, and filling in missing data using interpolation methods. The optimal parameter configuration of the wind energy potential prediction model after training is loaded and optimized. During forward propagation, each hidden layer uses the same activation function and normalization processing as in the training phase. The updated wind energy distribution feature map is stored in a grid data format, with each grid cell containing latitude and longitude coordinates, predicted wind speed, and energy density values. When implementing the threshold segmentation algorithm, an optional multi-scale segmentation strategy can be adopted, applying threshold segmentation at different spatial scales and then fusing the results to improve the accuracy of regional division. In addition to basic statistical features, the new evaluation parameter set can also include derived parameters such as wind energy availability hours and effective wind frequency distribution to enrich the evaluation dimensions.

[0098] The final evaluation function's weight calibration process can incorporate the entropy weighting method for objective weighting, calculating weights based on the variability of each evaluation parameter to avoid the influence of subjective factors. During the final wind resource assessment, the comprehensive evaluation indicators can be compared with industry standard levels to determine the feasibility level of wind farm development. Optionally, the final evaluation results can generate a wind resource assessment report, including a wind energy distribution map, regional division results, and development recommendations. In some embodiments, the final wind resource assessment can also perform uncertainty analysis, evaluating the confidence interval of the assessment results using Monte Carlo simulation. All intermediate data and final results from the entire evaluation process are stored in a wind resource database, providing data support for subsequent wind farm planning and design. It is understood that recalculating the feature map based on the optimized model can significantly improve the reliability and practicality of the evaluation results.

[0099] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0100] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A wind resource assessment method based on millimeter-wave wind-measuring radar, characterized in that, The method includes: A millimeter-wave wind radar system is deployed in the target area to collect raw wind field data streams. The raw wind field data streams are then time-series aligned and outlier filtered to generate a standardized wind dataset. A standardized wind dataset is input into a wind energy potential prediction model, and the model calculates and outputs a wind energy distribution feature map, which includes wind speed distribution and energy density information. Based on the wind energy distribution feature map, an initial evaluation parameter set is generated using a threshold segmentation algorithm, and an initial evaluation function is constructed based on the initial evaluation parameter set. This includes: applying an adaptive threshold segmentation algorithm to the wind energy distribution feature map and dividing the feature map into high, medium, and low wind energy regions based on the energy density value. Statistical characteristics of each region are extracted, including average wind speed and energy variance, to form an initial set of evaluation parameters; Based on the initial evaluation parameter set, the evaluation function is designed as a linear weighted combination, with weights allocated based on regional importance, thus generating the initial evaluation function; the initial evaluation function has the following form: ， in: This is a dimensionless comprehensive evaluation score. The average wind speeds for high, medium, and low wind energy areas. For reference wind speed, For energy variance, For reference energy density, These are dimensionless weighting coefficients; The initial evaluation function is executed to simulate and evaluate wind resources. At the same time, the actual wind energy output data of the target area is collected to form an actual wind energy dataset, including: deploying wind energy monitoring equipment in the target area to record actual wind speed and power output data, with the time synchronized with the simulation evaluation cycle; Run the initial evaluation function, output the simulated wind energy distribution results, and record the simulated values; Collect actual wind energy output data and match it with simulated values ​​at specific times to form an actual wind energy dataset. Calculate the deviation matrix between the actual wind energy dataset and the wind energy distribution feature map, and evaluate the error through deviation matrix analysis, including: aligning the predicted values ​​of the actual wind energy dataset and the wind energy distribution feature map by time series, and calculating the absolute deviation value at each time point; The organizational deviation values ​​are formed into a matrix structure, where rows represent time points and columns represent spatial locations, thus generating a deviation matrix. Principal component analysis was applied to the deviation matrix to extract the main error components and quantify the error. The parameters of the wind energy potential prediction model are corrected using the evaluation error to generate an optimized wind energy potential prediction model. The wind energy distribution feature map was recalculated using an optimized wind energy potential prediction model, and a final wind resource assessment was performed based on the new feature map.

2. The wind resource assessment method based on millimeter-wave wind radar according to claim 1, characterized in that, The millimeter-wave wind radar system is deployed in the target area to collect raw wind field data streams. The raw wind field data streams are then time-aligned and outlier filtered to generate a standardized wind dataset, including: Configure the scanning parameters of the millimeter-wave wind radar to capture raw wind speed and direction data of the target area at fixed time intervals, forming a raw wind field data stream; The original wind field data stream is segmented using a sliding window algorithm. Within each data segment, median filtering is used to remove impulse noise, resulting in a filtered data stream. The filtered data stream is timestamped and aligned, and interpolation algorithms are used to compensate for missing data points to ensure data continuity. The aligned data stream is normalized to a unified dimension to generate a standardized wind dataset.

3. The wind resource assessment method based on millimeter-wave wind radar according to claim 2, characterized in that, The step of inputting a standardized wind dataset into a wind energy potential prediction model and calculating and outputting a wind energy distribution feature map through the model includes: The wind energy potential prediction model is constructed as a multilayer perceptron network. The input layer receives a standardized wind dataset, the hidden layer performs feature transformation, and the output layer generates a wind energy distribution feature map. The wind energy potential prediction model is trained using historical wind energy data, and the network weights are adjusted through the backpropagation algorithm to minimize the prediction error. The standardized wind dataset is input into the trained wind energy potential prediction model, and the model outputs a grid map of wind speed distribution and energy density values ​​to form a wind energy distribution feature map.

4. The wind resource assessment method based on millimeter-wave wind-measuring radar according to claim 1, characterized in that, The step of using the evaluation error to correct the parameters of the wind energy potential prediction model and generate an optimized wind energy potential prediction model includes: The evaluation error is used as the loss function, and the gradient descent algorithm is used to adjust the network weights of the wind energy potential prediction model. The model parameters are updated through an iterative optimization process until the loss function converges. Verify the performance of the optimized model on the validation dataset and generate an optimized wind energy potential prediction model.

5. The wind resource assessment method based on millimeter-wave wind radar according to claim 4, characterized in that, The process of recalculating the wind energy distribution characteristic map using an optimized wind energy potential prediction model and performing a final wind resource assessment based on the new characteristic map includes: Input the latest standardized wind dataset into the optimized wind energy potential prediction model and output the updated wind energy distribution feature map. Based on the updated wind energy distribution feature map, a new set of evaluation parameters is generated by reapplying the threshold segmentation algorithm. The final assessment function is constructed based on the new assessment parameter set, and the final wind resource assessment is performed.

6. The wind resource assessment method based on millimeter-wave wind radar according to claim 2, characterized in that, The step of applying the sliding window algorithm to segment the raw wind field data stream includes: Set the size and step size of the sliding window; the window size is determined based on the typical cycle of wind speed changes. Wavelet transform is applied to the data within each window to extract time-frequency features; Calculate the statistical characteristics of the data within the window, including mean, variance, and skewness; Feature vectors are constructed based on statistical features and used for subsequent anomaly detection.

7. The wind resource assessment method based on millimeter-wave wind radar according to claim 4, characterized in that, The step of updating model parameters through iterative optimization includes: Set an early stopping mechanism to terminate training when the validation set loss function fails to improve for several consecutive iterations. An adaptive learning rate adjustment strategy is adopted to dynamically adjust the learning rate based on the direction of gradient change. A weight decay term is added during the optimization process to control model complexity; Record the loss function value for each iteration and plot the convergence curve to monitor the training process.

Images