Wind speed vertical extrapolation method based on data quality perception and multi-normal form AI

Abstract

The invention discloses a wind speed vertical extrapolation method based on data quality perception and multi-normal form AI, and the method comprises the steps: obtaining and preprocessing multi-source data, carrying out the weighted matching of an optimal meteorological station through information entropy and terrain complexity, and quantifying the data quality to generate a routing flag bit; when the data is scarce, generating an initial estimation value by adopting a physical model and fitting a residual error to generate a correction model; when the data is sufficient, time-varying physical parameters are inverted, and a mapping relation is established through deep learning to generate a dynamic parameterized model; when the performance of a single model is insufficient, starting a meta-learner to integrate the heterogeneous model; and finally, preferentially generating a target extrapolation model on the independent test set and deploying the target extrapolation model. According to the method, the wind speed prediction precision under the complex terrain is remarkably improved, the dependence on wind measurement equipment is reduced, and a customized wind resource evaluation scheme with physical rationality and data adaptability is provided for a wind power plant.

Description

Technical Field

[0001] This invention relates to the fields of wind power resource assessment and meteorological application technology, and in particular to a method for vertical extrapolation of wind speed based on data quality perception and multi-paradigm AI. Background Technology

[0002] As a key clean energy source supporting the "dual carbon" goal, the accuracy of wind energy development heavily relies on precise assessment of wind resources at the hub height (80-150 meters) of wind farms. However, acquiring long-term, continuous hub height observation data is costly, and in practice, vertical extrapolation from ground observation data from nearby meteorological stations is commonly used. Traditional methods mainly rely on power-law models or logarithmic-law wind profile models, whose core parameters (power-law exponent, roughness length) are usually determined by looking up tables based on the underlying surface type. This "one-size-fits-all" parameter fixation method cannot reflect the dynamic changes of parameters with meteorological conditions (stability, wind direction), and can introduce systematic biases in complex terrain, leading to wind resource assessment errors exceeding 15%-20%, seriously affecting the safety of wind farm investment and the accuracy of power generation revenue.

[0003] In recent years, machine learning technology has provided new ideas for wind speed extrapolation, but pure data-driven models have three inherent defects: First, their "black box" nature means they lack physical constraints, which may lead to predictions that violate atmospheric boundary layer dynamics in areas with scarce data or under extreme weather conditions; second, single models are difficult to adapt to the huge differences in wind farm data foundations—insufficient data from newly built farms leads to overfitting, while the data value of long-running farms has not been fully explored; third, general models cannot be tailored to specific situations, ignoring the uniqueness of local topography and meteorological characteristics, resulting in high uncertainty in assessments.

[0004] To address the aforementioned issues, while the industry has attempted to introduce residual learning correction or single deep learning models, significant limitations remain: residual learning relies solely on the physical model and fails to achieve dynamic parameter optimization; deep learning methods often directly map surface wind speed to hub height without explicitly learning the variation patterns of wind profile parameters, resulting in poor model interpretability; and while physical information neural networks have been explored, they lack the ability to model temporal dependencies. Furthermore, existing methods often employ the simple nearest-distance principle in the meteorological station matching process, neglecting the crucial impact of terrain complexity and observational data quality on the representativeness of reference stations.

[0005] In summary, existing technologies have not yet resolved three core contradictions: the contradiction between fixed model parameters and the dynamic nature of meteorology and terrain; the contradiction between single-paradigm modeling and differences in data foundation; and the contradiction between pure data-driven approaches and the constraints of physical laws. These problems result in insufficient accuracy and poor robustness of wind speed extrapolation, limiting its engineering applicability. There is an urgent need for an innovative method that can adapt to data quality, deeply integrate physical mechanisms with multi-paradigm AI, and achieve customized modeling for each event. Summary of the Invention

[0006] To address these issues, this invention provides a wind speed vertical extrapolation method based on data quality perception and multi-paradigm AI, thereby resolving the aforementioned problems in the prior art.

[0007] To achieve the above objectives, this invention provides a method for vertical extrapolation of wind speed based on data quality awareness and multi-paradigm AI, comprising:

[0008] Step S1: Obtain the hub height and wind speed of the target wind farm, historical observation data from surrounding meteorological stations, and geographical information to form a raw dataset. Preprocess the raw dataset to obtain a purified dataset.

[0009] Step S2: Based on the purified dataset, a dynamic weighting strategy of information entropy and terrain complexity is used to calculate the comprehensive score of the candidate stations, determine the optimal meteorological station, and extract its observation data.

[0010] Step S3: Quantify the integrity rate and coverage duration of the hub height wind speed in the purification dataset to generate routing flag bits;

[0011] Step S4: When the routing flag indicates that the data is scarce, based on the optimal meteorological station observation data and the purified dataset, a fixed parameter physical model is used to generate an initial estimate. The systematic deviation between the initial estimate and the measured value is fitted by supervised learning to generate a deviation correction model.

[0012] Step S5: When the routing flag indicates sufficient data, based on the optimal meteorological station observation data and the purified dataset, a time-varying physical parameter sequence is obtained by inversion using a time-series segmentation strategy. A mapping relationship between meteorological characteristics and optimal parameters is established through deep learning to generate a dynamic parameterized model.

[0013] Step S6: Evaluate the extrapolation performance of the residual correction model or the dynamic parameter prediction model on an independent test set, and select the best to generate the target extrapolation model.

[0014] Step S7: Obtain real-time meteorological data as input to the target extrapolation model to obtain the predicted value of hub height wind speed.

[0015] Furthermore, the process of step S2 includes:

[0016] All ground meteorological stations located within a preset radius of the target wind farm are extracted from the purified dataset to form a candidate station set;

[0017] Calculate the distance factor and elevation factor between each candidate station and the wind farm, and use elevation data to quantify the terrain complexity of the area between the candidate station and the wind farm to obtain the terrain complexity factor.

[0018] The information entropy of the historical wind speed sequences of candidate stations is calculated to assess the degree of data uncertainty and normalize it to obtain the data uncertainty factor.

[0019] The terrain complexity factor and the data uncertainty factor are coupled and operated to dynamically generate the altitude similarity weight value, with the distance weight and altitude weight being complementary.

[0020] The product of the distance weight and the proximity factor is added to the product of the altitude weight and the altitude similarity factor to obtain the comprehensive score of each candidate station. The meteorological station with the highest score is selected as the optimal reference meteorological station, and the corresponding observation data is extracted from the purified dataset.

[0021] Furthermore, the process of simultaneously quantifying the terrain complexity of the area between the candidate station and the wind farm using elevation data to obtain a terrain complexity factor includes:

[0022] A buffer zone is delineated along the geographical line connecting the candidate meteorological station and the wind farm;

[0023] Based on the digital elevation model data within the buffer area, the elevation statistical standard deviation is calculated to quantify the degree of terrain undulation.

[0024] Based on the land cover type data within the buffer zone, the average surface roughness length is estimated to quantify the surface friction effect.

[0025] The standard deviation and roughness length are normalized to obtain the terrain complexity factor.

[0026] Furthermore, the process of step S3 includes:

[0027] The data integrity rate is obtained by calculating the ratio of the number of valid data points to the total number of expected data points.

[0028] Calculate the cumulative duration of effective data coverage to obtain the coverage duration;

[0029] A preset integrity rate threshold and a duration threshold are set, and the integrity rate and the coverage duration are logically compared with the corresponding thresholds to obtain the routing flag bit.

[0030] Further, the process of logically comparing the integrity rate and the coverage duration with corresponding thresholds to obtain the routing flag bit includes:

[0031] When the integrity rate is lower than the integrity rate threshold or the coverage duration is lower than the duration threshold, the routing flag is assigned to the first state to indicate data scarcity; otherwise, it is assigned to the second state to indicate data sufficiency.

[0032] Furthermore, the process of step S4 includes:

[0033] Vertical extrapolation calculations are performed using a power-law wind profile model or a logarithmic-law wind profile model to obtain an initial wind speed estimate sequence for hub height. The exponential parameter of the power-law wind profile model or the roughness length parameter of the logarithmic-law wind profile model adopts a fixed default value obtained by looking up a table based on the underlying surface type.

[0034] The measured wind speed sequence of hub height of the target wind farm is extracted from the cleaned dataset, and the difference between the measured wind speed and the initial wind speed estimate is calculated at each time step to generate a residual sequence.

[0035] Construct a multidimensional feature vector for supervised learning, the feature vector containing ground observation elements, time periodic elements, and the initial wind speed estimate sequence of the optimal meteorological station;

[0036] Using the multidimensional feature vector as input and the residual sequence as output, the training set and test set are divided in chronological order. At least two machine learning algorithms are used for training to generate candidate correction models. The prediction error of each candidate model is evaluated on the test set, and the model that performs best in terms of root mean square error and mean absolute error is automatically selected as the bias correction model.

[0037] Furthermore, the process of constructing a multidimensional feature vector for supervised learning includes:

[0038] Surface wind speed, wind direction, and temperature components are extracted from the optimal meteorological station observation data in the purified dataset.

[0039] The wind direction component is periodically encoded and converted, and the sine and cosine transformation values ​​of the wind direction are calculated using sine and cosine functions, respectively.

[0040] Extract the month and day / hour integer values ​​from the timestamp data;

[0041] The integer values ​​of the month and the integer values ​​of the day and hour are used as periodic features;

[0042] The initial wind speed estimate sequence is used as a physical prior feature.

[0043] All the above components are concatenated into vectors in a preset order to form the multidimensional feature vector.

[0044] Furthermore, the process of step S5 includes:

[0045] The purified dataset is divided into time segments, and time-varying parameters that minimize the error of the physical model are optimized within each segment to form a parameter sequence;

[0046] Meteorological statistical features are extracted for each time period to construct an input feature vector;

[0047] Multiple heterogeneous deep learning architectures are used to learn the mapping relationship between the input feature vector and the parameter sequence;

[0048] The performance of each deep learning architecture is evaluated based on the prediction error, and the optimal architecture is automatically selected as the dynamic parameterized model.

[0049] Furthermore, step S5 also includes:

[0050] When the prediction performance of a single model learned by the various heterogeneous deep learning architectures fails to reach a preset accuracy threshold, a two-layer stacked ensemble learning mechanism is activated:

[0051] The first layer constructs a heterogeneous base model pool, which contains parameter prediction outputs learned from various heterogeneous deep learning architectures;

[0052] The second layer trains a meta-learner, which uses the prediction results of the base model pool as input features and introduces key meteorological features as contextual information to learn the adaptive weight allocation of each base model under different meteorological conditions.

[0053] The meta-learner is trained using gradient boosting tree or ridge regression algorithm to obtain the optimal parameter prediction values;

[0054] The integrated optimal parameters are substituted into the physical wind profile model to calculate the target extrapolated wind speed, thereby constructing the dynamic parameterized model.

[0055] Furthermore, step S6 includes the following process:

[0056] Separate independent test data from the purified dataset;

[0057] The test data is input into the residual correction model and the dynamic parameter prediction model to generate corresponding prediction results;

[0058] Calculate the statistical error index between each prediction result and the measured value;

[0059] Based on the error index, the performance of each candidate model is compared, and the model with the best overall performance is automatically selected as the target extrapolation model.

[0060] Compared with existing technologies, the beneficial effects of this invention are as follows: By deeply coupling terrain complexity, data uncertainty, and modeling strategies through a data quality-aware intelligent routing mechanism, this invention achieves synergistic enhancement of physical mechanisms and multi-paradigm AI: automatically increasing the altitude weight in areas with complex terrain and reducing its impact when data quality is poor, thus optimizing the matching accuracy of meteorological stations from the source; when data is scarce, using a physical wind profile model as a benchmark, supervised learning captures and corrects systematic residual biases to ensure that energy conservation and momentum transfer laws are still followed even with small samples; when data is abundant, time-varying physical parameters are inverted from observation data, and a deep neural network is used to learn the nonlinear mapping between meteorological states and optimal parameters, while a temporal attention mechanism tracks the historical dependence of atmospheric processes, and a physical information neural network embeds power-law or logarithmic-law constraints into the loss function, so that the prediction results naturally satisfy the principles of boundary layer atmospheric dynamics; when the prediction performance of a single architecture is insufficient, the meta-learner adaptively integrates the complementary advantages of heterogeneous models, dynamically weighs the confidence of each architecture under different meteorological conditions, and finally generates a station-specific extrapolation model that combines physical rationality and data adaptability. This mechanism significantly reduces wind speed prediction errors in complex terrain and greatly reduces reliance on expensive wind measurement equipment, providing interpretable and reliable technical support for wind resource assessment throughout the entire life cycle of wind farms. Attached Figure Description

[0061] Figure 1 A flowchart illustrating a wind speed vertical extrapolation method based on data quality awareness and multi-paradigm AI provided by this invention.

[0062] Figure 2 This is a flowchart illustrating step S2 in a wind speed vertical extrapolation method based on data quality awareness and multi-paradigm AI provided by the present invention.

[0063] Figure 3 This is a flowchart illustrating step S3 in a wind speed vertical extrapolation method based on data quality awareness and multi-paradigm AI provided by the present invention.

[0064] Figure 4 This is a flowchart illustrating step S4 in the wind speed vertical extrapolation method based on data quality perception and multi-paradigm AI provided by the present invention. Detailed Implementation

[0065] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0066] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0067] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0068] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0069] Please see Figure 1 As shown, this invention provides a method for vertical extrapolation of wind speed based on data quality awareness and multi-paradigm AI, comprising:

[0070] Step S1: Obtain the hub height and wind speed of the target wind farm, historical observation data from surrounding meteorological stations, and geographical information to form a raw dataset. Preprocess the raw dataset to obtain a purified dataset.

[0071] Specifically, historical wind speed time series (timestamp, wind speed value) for hub height is extracted from the wind farm's SCADA system; ground observation data (10-meter wind speed, wind direction, temperature) and station geographic coordinates (latitude, longitude, altitude) from surrounding meteorological stations are obtained through the national meteorological data platform or commercial service providers; and elevation information and surface roughness parameters of the target area are obtained using digital elevation models and land cover data. Unified quality control is performed on the three types of data: 1) Physically unreasonable values ​​(such as negative wind speed) and statistical outliers exceeding 3 times the standard deviation are removed; 2) Linear interpolation is used for short-term missing data (<2 hours), and long-term missing data are marked as invalid intervals; 3) The data is uniformly converted to UTC timestamps, and time tolerance alignment is used to eliminate time discrepancies between wind farm and meteorological station data; 4) Wind direction angles are converted to sine / cosine values, and time features such as month and hour are extracted. The processed data are integrated into a structured, cleaned dataset.

[0072] Step S2: Based on the purified dataset, a dynamic weighting strategy of information entropy and terrain complexity is used to calculate the comprehensive score of the candidate stations, determine the optimal meteorological station, and extract its observation data.

[0073] Specifically, step S2 includes the following process:

[0074] All ground meteorological stations located within a preset radius of the target wind farm are extracted from the purified dataset to form a candidate station set;

[0075] Specifically, the geographic coordinates (longitude, latitude, and altitude) of all weather stations are retrieved from the cleaned dataset, and the Euclidean plane distance between each weather station and the center point of the target wind farm is calculated. A preset radius is typically set to 10-20 kilometers. The geodesic function from the GeoPandas library or the Haversine formula is used to accurately calculate the spherical distance, and all weather stations with a distance less than or equal to the preset radius are selected to form a candidate station set. If the set is empty, the radius is expanded for re-selection; if it contains only one station, it is directly determined as the optimal station.

[0076] Calculate the distance factor and elevation factor between each candidate station and the wind farm, and use elevation data to quantify the terrain complexity of the area between the candidate station and the wind farm to obtain the terrain complexity factor.

[0077] Specifically, for each station i in the candidate station set, its straight-line distance to the wind farm is calculated. (Unit: meters), converted to proximity factor: Normalize it to the interval [0,1], with the value increasing as the distance between stations increases. Calculate the absolute value of the elevation difference between the two stations. Find the largest elevation difference in the set. Convert to altitude similarity factor: Similarly, it is normalized to the [0,1] interval and the value increases as the altitude approaches.

[0078] Specifically, the process of simultaneously quantifying the terrain complexity of the area between the candidate station and the wind farm using elevation data to obtain a terrain complexity factor includes:

[0079] A buffer zone is delineated along the geographical line connecting the candidate meteorological station and the wind farm;

[0080] Specifically, a geographic analysis corridor is constructed between candidate meteorological stations and wind farms. The geodesic line between the two points is taken as the central axis, and a horizontal distance of 500 meters is extended to both sides of the axis (this distance can be adjusted according to the complexity of the terrain, usually 300-800 meters), forming a rectangular buffer zone with a length equal to the distance between the two points and a width of 1000 meters.

[0081] Based on the digital elevation model data within the buffer area, the elevation statistical standard deviation is calculated to quantify the degree of terrain undulation.

[0082] Specifically, read the DEM raster data (with a recommended resolution of at least 30 meters) from the cleaned dataset, covering the buffer. Use the `mask` function from the `rasterio` library to clip the DEM raster with buffer polygons, obtaining an array of all elevation cell values ​​within that area. Then, use the `std()` function from the `NumPy` library to calculate the standard deviation of this array. This value directly reflects the degree of dispersion of terrain undulation. The larger the value, the more rugged the terrain and the more drastic the elevation changes within the region. If there are holes in the DEM data (such as lakes) within the buffer zone, the inverse distance weighting method is used to fill the holes before recalculation to avoid standard deviation distortion.

[0083] Based on the land cover type data within the buffer zone, the average surface roughness length is estimated to quantify the surface friction effect.

[0084] Specifically, land cover classification raster data from the same buffer zone is read from the cleaned dataset. An empirical mapping table between land cover type and surface roughness length is established (e.g., water bodies 0.0002 meters, farmland 0.05 meters, forests 1.2 meters, cities 1.8 meters, etc.). For each land cover cell within the buffer zone, a corresponding roughness length value is assigned according to the mapping table, forming a spatial distribution array of roughness lengths. The arithmetic mean of this array is calculated. If the buffer zone spans multiple land cover types, an area-weighted average is required, i.e. This value reflects the average frictional resistance effect of the earth's surface on airflow.

[0085] The standard deviation and roughness length are normalized to obtain the terrain complexity factor.

[0086] Specifically, the calculated and Substitute into the normalization function: This function uses a fractional structure to achieve dynamic normalization: numerator Directly quantify the intensity of terrain undulation; in the denominator By compressing the numerical range of roughness length through logarithmic transformation, we can avoid the denominator being too large due to extremely rough surfaces (such as cities), which would suppress the influence of topography. By adding 1 to the whole, we can ensure that the denominator is not zero and maintain monotonicity. The final TCF value range is (0, +∞). The larger the value, the higher the topographic complexity of the area and the more significant the impact on the spatial heterogeneity of the wind field.

[0087] The information entropy of the historical wind speed sequences of candidate stations is calculated to assess the degree of data uncertainty and normalize it to obtain the data uncertainty factor.

[0088] Specifically, extract the long-term historical wind speed sequence (recommended to be at least 1 year) of candidate station i from the cleanup dataset, and divide the wind speed range into K equal-width intervals (K is usually set to 20-50). Count the number of data points in each interval and calculate the probability distribution. Using the information entropy formula Calculate the uncertainty. Iterate through all candidate stations to find the maximum entropy value. Normalization yields the data uncertainty factor: The larger the value, the stronger the data volatility and the lower the reliability.

[0089] The terrain complexity factor and the data uncertainty factor are coupled and operated to dynamically generate the altitude similarity weight value, with the distance weight and altitude weight being complementary.

[0090] Specifically, for each candidate station i, calculate the altitude similarity weight: This design assigns a higher elevation weight to more complex terrain (TCF↑) and a lower elevation weight to higher data uncertainty (DUF_i↑). Distance weights complement this. .

[0091] The product of the distance weight and the proximity factor is added to the product of the altitude weight and the altitude similarity factor to obtain the comprehensive score of each candidate station. The meteorological station with the highest score is selected as the optimal reference meteorological station, and the corresponding observation data is extracted from the purified dataset.

[0092] Specifically, the overall score of candidate station i is calculated: Use Pandas' apply function to iterate through the candidate station set and batch calculate all... The `idxmax` function is used to find the candidate station index corresponding to the highest score; this station is the optimal reference weather station. Finally, all observation data columns (wind speed, wind direction, and temperature) are extracted from the purified dataset based on the station's identifier, serving as input features for subsequent models.

[0093] Step S3: Quantify the integrity rate and coverage duration of the hub height wind speed in the purification dataset to generate routing flag bits;

[0094] Specifically, step S3 includes the following processes:

[0095] The data integrity rate is obtained by calculating the ratio of the number of valid data points to the total number of expected data points.

[0096] Specifically, the hub height and wind speed time series of the target wind farm are extracted from the cleaned dataset. Valid data points are defined as records that simultaneously meet the following conditions: 1) Wind speed value is not NaN or None; 2) Wind speed value is within a physically reasonable range (≥0 m / s and ≤60 m / s); 3) The corresponding timestamp is not marked as an invalid period (e.g., wind turbine shutdown maintenance period). The `dropna()` function of Pandas is used to remove NaN values, Boolean indexing is used to filter records that meet the physical range, and multiple conditions are combined using the `&` logical AND operator to obtain a Boolean mask array. .

[0097] Calculate the cumulative duration of effective data coverage to obtain the coverage duration;

[0098] Specifically, the number of valid data points is counted: The expected total number of data points, N_total, is calculated as: (End timestamp - Start timestamp) / Data sampling interval + 1. The sampling interval is obtained from the time index `freq` attribute of the cleaned dataset; for 10-minute data sets, the interval is 600 seconds. The completeness rate is calculated as: Completeness_Rate = N_valid / N_total. When using Python's `len()` function to obtain the total number of records, it is necessary to first ensure the time series is continuous and without jumps using `resample().asfreq()` to avoid statistical distortion due to missing timestamps.

[0099] A preset integrity rate threshold and a duration threshold are set, and the integrity rate and the coverage duration are logically compared with the corresponding thresholds to obtain the routing flag bit.

[0100] Specifically, the process of logically comparing the integrity rate and the coverage duration with corresponding thresholds to obtain the routing flag includes:

[0101] When the integrity rate is lower than the integrity rate threshold or the coverage duration is lower than the duration threshold, the routing flag is assigned to the first state to indicate data scarcity; otherwise, it is assigned to the second state to indicate data sufficiency.

[0102] In this embodiment, the data integrity rate threshold parameter is set to 80% by default, and the coverage duration threshold parameter is set to 18 months by default. The configuration file is stored in an editable text format, allowing dynamic adjustment of the two threshold parameters according to the wind farm construction stage or data quality requirements. The calculated data integrity rate value is compared with the read integrity rate threshold parameter value. If the integrity rate value is less than the integrity rate threshold parameter value, the first comparison result is recorded as true; otherwise, the first comparison result is recorded as false. The calculated data coverage duration value is compared with the read coverage duration threshold parameter value. If the coverage duration value is less than the coverage duration threshold parameter value, the second comparison result is recorded as true; otherwise, the second comparison result is recorded as false. A logical OR operation is performed on the first comparison result and the second comparison result. That is, the final comparison result is determined to be false if and only if both the first comparison result and the second comparison result are false; otherwise, the final comparison result is true. When the final comparison result is true, the routing flag variable is assigned the first state enumeration value to indicate that the current data condition is a data-scarce mode; when the final comparison result is false, the routing flag variable is assigned the second state enumeration value to indicate that the current data condition is a data-sufficient mode; the routing flag variable is stored in memory in the form of a string or an integer value for subsequent modeling branch judgment calls.

[0103] Step S4: When the routing flag indicates that the data is scarce, based on the optimal meteorological station observation data and the purified dataset, a fixed parameter physical model is used to generate an initial estimate. The systematic deviation between the initial estimate and the measured value is fitted by supervised learning to generate a deviation correction model.

[0104] Specifically, step S4 includes the following process:

[0105] Vertical extrapolation calculations are performed using a power-law wind profile model or a logarithmic-law wind profile model to obtain an initial wind speed estimate sequence for hub height. The exponential parameter of the power-law wind profile model or the roughness length parameter of the logarithmic-law wind profile model adopts a fixed default value obtained by looking up a table based on the underlying surface type.

[0106] Specifically, surface wind speed sequences are extracted from the best meteorological station observation data in the purified dataset. The fixed power law index is looked up from the IEC 61400-1 standard or the land surface classification experience table according to the underlying surface type. (For example, 0.12 for open grasslands and 0.25 for rough terrain) or a fixed roughness length. (For example, 0.1 meters for farmland and 1.0 meter for forest). Extrapolation calculations are performed on the entire time series using vectorized operations: if a power-law model is selected, the initial estimated wind speed sequence at hub height is... ,in This refers to the hub height of the wind farm (e.g., 120 meters). The wind height measured at the weather station is typically 10 meters; if a logarithmic law model is used, then... .

[0107] The measured wind speed sequence of hub height of the target wind farm is extracted from the cleaned dataset, and the difference between the measured wind speed and the initial wind speed estimate is calculated at each time step to generate a residual sequence.

[0108] Specifically, the measured wind speed sequence at the hub height of the target wind farm is extracted from the purified dataset. The Pandas merge_asof function is used to... and Nearest neighbor matching is performed based on timestamps, with a tolerance set to half the sampling period to ensure strict synchronization. Residual values ​​are calculated for the matched data at each time step. Generate a residual sequence. If at a certain time... If a value is missing or invalid (marked as NaN), the residual value at that time is set to NaN and automatically excluded in subsequent steps.

[0109] Construct a multidimensional feature vector for supervised learning, the feature vector containing ground observation elements, time periodic elements, and the initial wind speed estimate sequence of the optimal meteorological station;

[0110] Specifically, the process of constructing a multidimensional feature vector for supervised learning includes:

[0111] Surface wind speed, wind direction, and temperature components are extracted from the optimal meteorological station observation data in the purified dataset.

[0112] Specifically, the observation data table of the optimal reference meteorological station is read from the purified dataset, and three core meteorological element columns are extracted: surface wind speed (in meters per second), wind direction angle (in degrees, ranging from 0 to 360 degrees), and temperature (in degrees Celsius). These three element columns constitute the basic meteorological input layer of the feature vector.

[0113] The wind direction component is periodically encoded and converted, and the sine and cosine transformation values ​​of the wind direction are calculated using sine and cosine functions, respectively.

[0114] Specifically, for the extracted wind direction angle series, trigonometric function transformations are performed to achieve periodic encoding. For each moment's wind direction angle value, it is first multiplied by pi and divided by a conversion factor of 180 to convert it to radians. Then, the sine and cosine functions of this radian value are calculated to generate wind direction sine and cosine transformation series, respectively. This transformation maps the original directional signal, which jumps between 0 degrees and 360 degrees, into two continuous variables that smoothly fluctuate between negative one and positive one, enabling the model to learn the cyclical characteristics of wind direction.

[0115] Extract the month and day / hour integer values ​​from the timestamp data;

[0116] The integer values ​​of the month and the integer values ​​of the day and hour are used as periodic features;

[0117] Specifically, month and hour information are extracted from the timestamp index of the cleaned dataset. For each timestamp, the date portion is parsed to obtain the month integer value (ranging from 1 to 12), and the time portion is parsed to obtain the day and hour integer value (ranging from 0 to 23). These integer values ​​are used directly as time-periodic feature columns without any mathematical transformations to capture the seasonality and intraday variation patterns of wind resources.

[0118] The initial wind speed estimate sequence is used as a physical prior feature.

[0119] Specifically, the initial estimates of hub height and wind speed generated by the aforementioned physical model's initial estimation sub-step are treated as an independent feature column and labeled as physical prior features. This feature column directly reflects the baseline predictive capability of the physical wind profile model, providing a reference benchmark for the machine learning model to be corrected.

[0120] All the above components are concatenated into vectors in a preset order to form the multidimensional feature vector.

[0121] Specifically, all the aforementioned feature columns are horizontally concatenated in a predetermined fixed order: ground wind speed, wind direction sine transformation, wind direction cosine transformation, temperature, month integer values, day / hour integer values, and physical prior features. The concatenated data structure forms a multi-dimensional feature matrix, where each row corresponds to a time sampling point and each column corresponds to a feature dimension. Finally, standardization is performed on each column of this feature matrix, calculating the arithmetic mean and standard deviation of each column. Each element is subtracted from the corresponding column's mean and then divided by the standard deviation, ensuring that the mean and variance of all feature column values ​​are zero, thus guaranteeing numerical stability and convergence speed during subsequent machine learning model training.

[0122] Using the multidimensional feature vector as input and the residual sequence as output, the training set and test set are divided in chronological order. At least two machine learning algorithms are used for training to generate candidate correction models. The prediction error of each candidate model is evaluated on the test set, and the model that performs best in terms of root mean square error and mean absolute error is automatically selected as the bias correction model.

[0123] Specifically, the feature matrix and residual sequence are divided into a 70% training set and a 30% test set in chronological order to ensure that the test set data is not leaked into the training phase. At least two heterogeneous algorithms are used for training: the first is a gradient boosting tree algorithm, with the number of trees n_estimators=200, maximum depth max_depth=5, learning rate learning_rate=0.1, and the objective function set to 'reg:squarederror'; the second is a support vector regression algorithm, with the kernel function chosen as radial basis function 'RBF', penalty coefficient C=10, and kernel coefficient gamma='scale'. Both algorithms are trained on the training set using the fit(X_train, y_train) method, resulting in two candidate correction models. On the test set, the predict() method is called to generate predicted residuals, and the root mean square error (RMSE) between the predicted values ​​and the true residuals is calculated. ) and mean absolute error ( Compare the RMSE and MAE of the two models, and select the model with the smaller values ​​for both indicators as the final deviation correction model; if the indicators have their own advantages and disadvantages, choose the model with the better RMSE, because it has a stronger penalty for large errors, which meets the needs of extreme deviation control in engineering applications.

[0124] Step S5: When the routing flag indicates sufficient data, based on the optimal meteorological station observation data and the purified dataset, a time-varying physical parameter sequence is obtained by inversion using a time-series segmentation strategy. A mapping relationship between meteorological characteristics and optimal parameters is established through deep learning to generate a dynamic parameterized model.

[0125] Specifically, step S5 includes the following process:

[0126] The purified dataset is divided into time segments, and time-varying parameters that minimize the error of the physical model are optimized within each segment to form a parameter sequence;

[0127] Specifically, a sliding time window mechanism is used to segment the high-quality historical wheel height data in the cleaned dataset. The window size is set to thirty days (or adjusted to twenty to sixty days according to the characteristic cycle of atmospheric processes), and the step size is set to one day, achieving an overlap rate of 96.7% between windows. For each window w, the following operation is performed: extract the surface wind speed sequence of the best meteorological station within that window from the cleaned dataset. Measured wind speed sequence at hub height of the target wind farm For the power-law wind profile model, the parameter α to be inverted is used as the optimization variable to construct the objective function. , where i iterates through all time sampling points within the window. The L-BFGS-B algorithm is used for unconstrained optimization, with the maximum number of iterations maxiter=100, gradient tolerance gtol=1e-6, and initial values... The default power law exponent is set to 0.15. After optimization, the optimal power law exponent for this window is obtained. The inversion process for the roughness length z0 in the logarithmic law model is similar, and the objective function is... The optimization boundary is set to [1e-4, 10] meters to conform to physical reality. After traversing all time windows, the optimal parameter sequence with a length equal to the total number of data days minus the window size plus one is obtained. }or{ }

[0128] Meteorological statistical features are extracted for each time period to construct an input feature vector;

[0129] Specifically, for each time window w, five types of meteorological statistical features are calculated based on the best meteorological station observation data in the purified dataset: 1) mean surface wind speed (mean(u_a), which is the arithmetic mean of the u_a sequence within the window; 2) standard deviation of surface wind speed (std(u_a), reflecting the intensity of wind speed fluctuations; 3) predominant wind direction (predominant_wd), calculated by averaging the wind direction angles (converting the wind direction into u and v components, averaging them separately, and then recalculating the angle); 4) mean average temperature (mean(temp); 5) day of year, which is the year number (1-365) of the middle date within the window. The five scalar features are concatenated in sequence to form an input feature vector X_B[w] of shape (5,).

[0130] Multiple heterogeneous deep learning architectures are used to learn the mapping relationship between the input feature vector and the parameter sequence;

[0131] Specifically, three heterogeneous deep learning models are constructed in parallel: 1) Deep Neural Network (DNN): The nn.Sequential container is defined using the PyTorch framework, containing an input layer (5 neurons), four hidden layers (128 neurons per layer, ReLU activation function, using BatchNorm and Dropout layers to prevent overfitting), and an output layer (1 neuron, linear activation). The Adam optimizer is used, with a learning rate of 0.001 and a loss function of mean squared error MSELoss. 2) Temporal Attention Model (Transformer): The input layer expands X_B[w] into a sequence form, adds positional encoding, and then feeds it into a Transformer encoder (2 layers, 4 attention heads, 64 hidden dimensions). The encoder output takes the features from the last time step and feeds them into the fully connected layer for regression parameters. 3) Physical Information Neural Network (PINN): The network outputs two branches, predicting wind speed respectively. and parameters Total loss function λ is an adjustable hyperparameter (default value is 0.1), which implements physical constraints through automatic differentiation. All three models are trained on the training set for 100 epochs, using an early stopping mechanism with a patience of 10.

[0132] The performance of each deep learning architecture is evaluated based on the prediction error, and the optimal architecture is automatically selected as the dynamic parameterized model.

[0133] Specifically, on the test set, predictions are performed for each model: the output predictions of both DNN and Transformer. The sequence is substituted into the power-law formula to calculate the extrapolated wind speed; PINN outputs directly. Calculate the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination for each model. The comparison process is as follows: first, select the one with the smallest RMSE; if the RMSE difference is less than 0.05 m / s, then compare... The model with the largest value is selected; if a distinction still cannot be made, the model with the shortest training time is chosen. The final selected model is saved as an ONNX format or a PyTorch state_dict, and its configuration parameters and normalizer object are recorded, thus completing the generation of the dynamically parameterized model.

[0134] Specifically, step S5 further includes:

[0135] When the prediction performance of a single model learned by the various heterogeneous deep learning architectures fails to reach a preset accuracy threshold, a two-layer stacked ensemble learning mechanism is activated:

[0136] Specifically, after the model evaluation and selection sub-step is completed, it is checked whether the root mean square error of the best single model exceeds a preset accuracy threshold (the default value is 0.75 m / s). If the threshold is exceeded, or if the test set determination coefficients of the three heterogeneous deep learning models are all below 0.85, the ensemble learning mechanism is activated, and the output mode of the base model pool is switched from single model to ensemble mode.

[0137] The first layer constructs a heterogeneous base model pool, which contains parameter prediction outputs learned from various heterogeneous deep learning architectures;

[0138] Specifically, a pre-trained deep neural network, a temporal attention model, and a physical information neural network are loaded simultaneously. For each time window in the test set, the input feature vector X_B[w] is sequentially input into the three models to obtain the predicted values ​​of the three parameters: , and .

[0139] The second layer trains a meta-learner, which uses the prediction results of the base model pool as input features and introduces key meteorological features as contextual information to learn the adaptive weight allocation of each base model under different meteorological conditions.

[0140] Specifically, the three predicted values ​​will be compared with key meteorological features (surface average wind speed) in the original input features. and average temperature The features are concatenated to form the input feature vector of the meta-learner. The first three are the base model output layers, and the last two are the context information layers.

[0141] The meta-learner is trained using gradient boosting tree or ridge regression algorithm to obtain the optimal parameter prediction values;

[0142] Specifically, N samples are extracted from the validation time period of the cleaned dataset to construct the training dataset for the meta-learner. For each validation sample i, the vector V[i] is calculated, and the optimal parameters obtained from the inversion within the corresponding time window are used. The first three components (base model output) of V[i] were standardized using StandardScaler, and the last two components (meteorological features) were processed using RobustScaler to eliminate the influence of outliers. The processed feature matrix has a shape of (N, 5), and the target vector has a shape of (N, 1). The vectors were divided into a meta-training set and a meta-validation set in an 80:20 ratio.

[0143] When choosing the gradient boosting tree algorithm, set the hyperparameters as follows: n_estimators=100 (number of trees), max_depth=3 (tree depth limit to prevent overfitting), learning_rate=0.05 (learning rate), subsample=0.8 (sample sampling ratio), colsample_bytree=0.8 (feature sampling ratio), and objective='reg:squarederror' (regression task objective function). Train the model on the meta-training set using the `fit()` method. During training, monitor performance on the meta-validation set using the `early_stopping_rounds=10` parameter. Terminate training when the validation error does not decrease for 10 consecutive rounds. After training, obtain the feature importance score of the meta-model to analyze the contribution of each base model.

[0144] When choosing the Ridge Regression algorithm, set the hyperparameters as follows: alpha = 1.0 (L2 regularization strength), fit_intercept = True (fit intercept term), normalize = False (due to pre-standardization). Use 5-fold cross-validation combined with RidgeCV to automatically search for the optimal alpha value, with the search range set to [0.1, 1.0, 10.0]. Retrain the final model with the determined alpha on the complete meta-training set. The output of Ridge Regression is a linear combination of weights, which can be directly interpreted as the confidence coefficients of each base model.

[0145] The trained meta-learner receives a new feature vector V[w] and outputs the ensembled optimal parameter predictions. For gradient boosting trees, the predict() method outputs the nonlinear fusion result; for ridge regression, the parameter calculation formula is... The weight coefficient wi is automatically learned by the meta-learner and satisfies the dynamic change of weight with meteorological conditions (e.g., increase the PINN weight during low wind speed periods and increase the Trans weight during periods of drastic wind direction changes).

[0146] The integrated optimal parameters are substituted into the physical wind profile model to calculate the target extrapolated wind speed, thereby constructing the dynamic parameterized model.

[0147] Specifically, Substitute the values ​​into the power-law wind profile model to calculate the final extrapolated wind speed: , in, The optimal power-law exponent after ensemble learning of the meta-learner is defined. This parameter integrates the prediction results of deep neural networks, temporal attention models, and physical information neural networks, enabling adaptive modeling of wind shear characteristics under different meteorological conditions. The ensemble model (containing three heterogeneous architectures, a meta-learner, a normalizer, and a physical model) is saved via pickle serialization or exported in ONNX format. The metadata records the ensemble strategy, architecture list, and performance improvement, forming a complete dynamic parameterized model.

[0148] Step S6: Evaluate the extrapolation performance of the residual correction model or the dynamic parameter prediction model on an independent test set, and select the best to generate the target extrapolation model.

[0149] Specifically, step S6 includes the following process:

[0150] Separate independent test data from the purified dataset;

[0151] Specifically, the last 30% of the time segment is extracted from the purified dataset in chronological order as independent test data, ensuring that the optimal meteorological station observation data and the measured hub height data of the target wind farm within this segment are not involved in any model training process. The time span of the test data is no less than six months to cover the complete seasonal cycle. If the total data duration is less than two years, a five-fold time series cross-validation method is used to construct the test set, that is, each time a continuous 20% of the data is taken as the test set, and the rest is used as the training set, and the average performance is taken after five iterations.

[0152] The test data is input into the residual correction model and the dynamic parameter prediction model to generate corresponding prediction results;

[0153] Specifically, the test data is divided into an input side and a ground truth side: the input side includes the surface wind speed, wind direction, temperature sequence, and timestamp from the optimal weather station; the ground truth side is the measured wind speed sequence at the hub height of the target wind farm. For the residual correction model, the input data is input into the model to obtain the predicted residual sequence, which is then superimposed with the initial estimate from the physical model to obtain the final wind speed prediction result. For the dynamic parameter prediction model, the input side data is input into the model to obtain the prediction parameter sequence, which is then substituted into the power-law wind profile formula to calculate the extrapolated wind speed. Both models use the same input data in their prediction processes to ensure fairness in the evaluation.

[0154] Calculate the statistical error index between each prediction result and the measured value;

[0155] Specifically, the two types of prediction results are compared with the measured wind speed series from the true value side. Perform time-by-time alignment matching. Calculate three core statistical indicators: 1) Root mean square error 1) N is the number of test samples; 2) Mean absolute error 3) Coefficient of determination ,in, , representing the sample mean of the measured wind speed. If the prediction error at a certain moment exceeds three times the RMSE, that point is not included in the MAE calculation but is retained in the RMSE to ensure sensitivity to large errors.

[0156] Based on the error index, the performance of each candidate model is compared, and the model with the best overall performance is automatically selected as the target extrapolation model.

[0157] Specifically, a three-layer decision-making logic is established: the first layer compares the RMSE and selects the model with the smaller RMSE as the candidate optimal model; if the difference in RMSE between the two models is less than 0.05 m / s, then the second layer of comparison is initiated. ,choose The larger one; if If the difference is still less than 0.02, the process proceeds to the third level, comparing model complexity. The residual correction model with fewer parameters is prioritized to ensure computational efficiency. The decision logic is implemented using nested if-elif structures in Python, and the final selected model is stored in a dictionary variable. The model's type identifier ('residual' or 'dynamic'), evaluation metric value, and training timestamp are written to a JSON-formatted model metadata file. If the performance of both models fails to meet the standard (RMSE > 1.5 m / s or R² < 0.7), an alarm is triggered and the model reverts to a pure physics model with fixed parameters as the baseline.

[0158] Step S7: Obtain real-time meteorological data as input to the target extrapolation model to obtain the predicted value of hub height wind speed.

[0159] Specifically, a real-time data connection channel is established through the application programming interface (API) provided by the meteorological data platform (such as a RESTful API or MQTT message queue) to periodically pull or subscribe to the latest ground observation data from the optimal reference meteorological station. The request frequency is set to once every ten minutes to match the data resolution during model training. The JSON data packet returned by the API must contain at least four fields: timestamp, wind speed, wind direction, and temperature. If the API call fails or the returned data is delayed by more than five minutes, the system switches to a backup data source and reads the most recent valid data from the local cache as temporary filler. The acquired real-time observation data undergoes preprocessing operations that are completely consistent with those in the training phase: 1) Filter outliers (negative wind speed, temperature exceeding -40 to 50°C) using the same physical plausibility check rules; 2) Use the last observation value forward fill strategy for occasional missing values; 3) Substitute the wind direction angle value into the sine and cosine transform coefficients saved during training for the same transformation; 4) Extract the current month and hour values ​​from the real-time timestamp; 5) Call the StandardScaler object persisted during the training phase to standardize the wind speed and temperature, and scale the sine and cosine values ​​of the wind direction in the same way. The preprocessed data is organized into a real-time feature vector of shape (1, 7), with the feature order strictly aligned with the training features. The serialized file of the target extrapolation model is read from the storage medium (hard disk or cloud storage). If it is a residual correction model, the Pickle package containing the physical model initial estimation module, a gradient boosting tree model (XGBRegressor) instance, and a feature normalizer is loaded; if it is a dynamic parameter prediction model, the TorchScript file containing the DNN / Transformer / PINN network structure, training weights, and a parameter normalizer is loaded. After loading, the model is placed in evaluation mode and a forward propagation is performed once to initialize the cache. The real-time feature vector is input into the target extrapolation model to perform prediction: for the residual correction model, the initial physical model estimate u_p is calculated first, then input into the XGBoost model to obtain the residual correction Δu, and finally the wind speed prediction value. For dynamic parameter prediction models, feature vectors are input into a neural network to obtain predicted parameters. Substitute into the power law formula Calculate the extrapolated wind speed. The inference process uses a prediction mode with a batch size of 1, and the calculation time is controlled within 100 milliseconds to meet real-time requirements. Negative predictions and abnormal results exceeding 60 meters per second are eliminated. An alarm is triggered when the predicted value exceeds 1.5 times the historical maximum wind speed. The verified predicted values ​​are retained to two decimal places and returned in JSON format, containing the following fields: wind farm ID, prediction timestamp, hub height wind speed prediction, model version number, and confidence score (calculated based on the model's error distribution on the test set). Simultaneously, the prediction results are written to a time-series database (InfluxDB) and pushed to the monitoring dashboard, completing the prediction loop.

[0160] Specifically, this invention achieves synergistic enhancement of physical mechanisms and multi-paradigm AI by deeply coupling terrain complexity, data uncertainty, and modeling strategies through a data quality-aware intelligent routing mechanism: automatically increasing altitude weight in areas with complex terrain and reducing its impact when data quality is poor, thus optimizing the matching accuracy of meteorological stations from the source; when data is scarce, using a physical wind profile model as a benchmark, supervised learning captures and corrects systematic residual biases to ensure that energy conservation and momentum transfer laws are still followed even with small samples; when data is abundant, time-varying physical parameters are inverted from observational data, and a deep neural network is used to learn the nonlinear mapping between meteorological states and optimal parameters, while a temporal attention mechanism tracks the historical dependence of atmospheric processes, and a physical information neural network embeds power-law or logarithmic-law constraints into the loss function, so that the prediction results naturally satisfy the principles of boundary layer atmospheric dynamics; when the prediction performance of a single architecture is insufficient, the meta-learner adaptively integrates the complementary advantages of heterogeneous models, dynamically weighs the confidence of each architecture under different meteorological conditions, and finally generates a station-specific extrapolation model that combines physical rationality and data adaptability. This mechanism significantly reduces wind speed prediction errors in complex terrain and greatly reduces reliance on expensive wind measurement equipment, providing interpretable and reliable technical support for wind resource assessment throughout the entire life cycle of wind farms.

[0161] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0162] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A wind speed vertical extrapolation method based on data quality-aware and multi-paradigm AI, characterized in that, The method comprises the following steps: Step S1, obtaining the hub height wind speed of the target wind farm, the historical observation data of the surrounding meteorological stations and the geographic information to form an original data set, and preprocessing the original data set to obtain a purified data set; Step S2, according to the purified data set, a candidate station comprehensive score is calculated by using an information entropy and terrain complexity dynamic weighting strategy, the optimal meteorological station is determined, and the observation data thereof is extracted; Step S3, the completeness rate and coverage duration of the hub height wind speed in the purified data set are quantified to generate a routing flag; Step S4, when the routing flag indicates data scarcity, based on the observation data of the optimal meteorological station and the purified data set, an initial estimate value is generated by using a fixed parameter physical model, and a deviation correction model is generated by fitting the systematic deviation between the initial estimate value and the measured value through supervised learning; Step S5, when the routing flag indicates data sufficiency, based on the observation data of the optimal meteorological station and the purified data set, a time series segmentation strategy is used to obtain a time-varying physical parameter sequence, and a dynamic parameterization model is generated by establishing a mapping relationship between meteorological features and optimal parameters through deep learning; Step S6, according to the residual correction model or the dynamic parameter prediction model, the extrapolation performance thereof is evaluated on an independent test set, and a target extrapolation model is generated by selection; Step S7, real-time meteorological data is obtained as the input of the target extrapolation model to obtain a hub height wind speed prediction value.

2. The method of claim 1, wherein, The process of step S2 comprises: extracting all ground meteorological stations within a preset radius range of the target wind farm from the purified data set to form a candidate station set; calculating the distance factor and altitude factor of each candidate station from the wind farm, and quantifying the terrain complexity between the candidate station and the wind farm using elevation data to obtain a terrain complexity factor; calculating the information entropy of the historical wind speed sequence of the candidate station to evaluate the data uncertainty degree and normalize it to obtain a data uncertainty factor; coupling the terrain complexity factor and the data uncertainty factor to dynamically generate the weight value of the altitude similarity, and the distance weight and the altitude weight are complementary; adding the product of the distance weight and the distance proximity factor to the product of the altitude weight and the altitude similarity factor to obtain the comprehensive score of each candidate station, selecting the meteorological station with the highest score as the optimal reference meteorological station, and extracting the corresponding observation data from the purified data set.

3. The method of claim 2, wherein, The process of simultaneously quantifying the terrain complexity between the candidate station and the wind farm using elevation data to obtain the terrain complexity factor comprises: drawing a buffer area on the geographical line between the candidate meteorological station and the wind farm; based on the digital elevation model data in the buffer area, calculating the standard deviation of the elevation statistics to quantify the terrain fluctuation degree; based on the land cover type data in the buffer area, estimating the average roughness length of the ground surface to quantify the surface friction effect; normalizing the standard deviation and roughness length to obtain the terrain complexity factor.

4. The wind speed vertical extrapolation method based on data quality-aware and multi-paradigm AI of claim 3, wherein, The process of step S3 comprises: calculating the ratio of the number of valid data points to the total number of expected data points to obtain the data completeness rate; calculating the cumulative duration of valid data coverage to obtain the coverage duration; A preset completeness rate threshold and a duration threshold are provided, and the completeness rate and the duration are compared with the corresponding thresholds respectively to obtain the routing flag.

5. The wind speed vertical extrapolation method based on data quality-aware and multi-paradigm AI of claim 4, wherein, The process of comparing the completeness rate and the duration with the corresponding thresholds respectively to obtain the routing flag comprises: When the completeness rate is lower than the completeness rate threshold or the duration is lower than the duration threshold, the routing flag is assigned a first state to indicate data scarcity, otherwise, the routing flag is assigned a second state to indicate data abundance.

6. The wind speed vertical extrapolation method based on data quality-aware and multi-paradigm AI of claim 5, wherein, The process of step S4 comprises: A power-law wind profile model or a logarithmic-law wind profile model is used for vertical extrapolation calculation to obtain a wind speed initial estimation sequence at the hub height, wherein an index parameter of the power-law wind profile model or a roughness length parameter of the logarithmic-law wind profile model adopts a fixed default value obtained based on a lookup table of underlying surface types; A measured wind speed sequence at the hub height of a target wind farm is extracted from the purified dataset, and a difference between the measured wind speed and the wind speed initial estimation is calculated at each time instant to generate a residual sequence; A multi-dimensional feature vector for supervised learning is constructed, which includes ground observation elements of the optimal meteorological station, time periodicity elements, and the wind speed initial estimation sequence; The multi-dimensional feature vector is taken as input and the residual sequence is taken as output, and a training set and a test set are divided in time sequence, at least two machine learning algorithms are used for training to generate candidate correction models, and the prediction error of each candidate model is evaluated on the test set to automatically select a model with optimal performance in root mean square error and mean absolute error indicators as the bias correction model.

7. The wind speed vertical extrapolation method based on data quality-aware and multi-paradigm AI of claim 6, wherein, The process of constructing the multi-dimensional feature vector for supervised learning comprises: A ground wind speed component, a wind direction component, and a temperature component are extracted from the observation data of the optimal meteorological station in the purified dataset; The wind direction component is periodically encoded and converted, and the sine and cosine transformed values of the wind direction are calculated by sine and cosine functions respectively; Month integer values and day-hour integer values are extracted from timestamp data; The month integer values and day-hour integer values are taken as periodic features; The wind speed initial estimation sequence is taken as a physical prior feature; All the components are vector spliced in a preset order to form the multi-dimensional feature vector.

8. The wind speed vertical extrapolation method based on data quality-aware and multi-paradigm AI of claim 7, wherein, The process of step S5 comprises: The purified dataset is time segmented, and time-varying parameters that minimize the physical model error are optimized and solved in each segment to form a parameter sequence; For each time segment, meteorological statistical features are extracted to construct an input feature vector; A mapping relationship between the input feature vector and the parameter sequence is learned using multiple heterogeneous deep learning architectures; The performance of each deep learning architecture is evaluated based on the prediction error, and the optimal architecture is automatically selected as the dynamic parameterization model.

9. The wind speed vertical extrapolation method based on data quality-aware and multi-paradigm AI of claim 8, wherein, The step S5 further comprises: When the single model prediction performance learned by the multiple heterogeneous deep learning architectures does not reach a preset precision threshold, a two-layer stacked ensemble learning mechanism is started: The first layer constructs a pool of heterogeneous base models, including parameter prediction outputs of the multiple heterogeneous deep learning architectures; The second layer training meta-learner takes the prediction results of the base model pool as input features and introduces key meteorological features as context information to learn adaptive weight distribution of each base model under different meteorological conditions; The meta-learner is trained by using gradient boosting tree or ridge regression algorithm to obtain optimal parameter prediction value; The integrated optimal parameters are substituted into the physical wind profile model to obtain the target extrapolated wind speed to construct the dynamic parameterized model.

10. The wind speed vertical extrapolation method based on data quality-aware and multi-paradigm AI of claim 9, wherein, The process of the step S6 includes: Divide independent test data from the purified data set; Input the test data into the residual correction model and the dynamic parameter prediction model to generate corresponding prediction results; Calculate the statistical error index between each prediction result and the measured value; Based on the error index, compare the performance of each candidate model, and automatically select the one with the best comprehensive performance as the target extrapolation model.

Images