A method and system for multi-source fusion correction of regional wind speed data

By constructing a CNN-LSTM hybrid model and a multi-model integration strategy, the problem of wind speed correction for mesoscale data in complex terrain areas was solved, achieving high-resolution and high-precision wind speed data generation, supporting wind energy resource assessment and wind power project site selection.

CN122654993BActive Publication Date: 2026-11-17POWERCHINA ZHONGNAN ENG
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202611149358.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-07-31
Publication Date
2026-11-17
Estimated Expiration
2046-07-31

AI Technical Summary

Technical Problem

Existing technologies struggle to effectively integrate mesoscale meteorological data, ground-based wind tower measurements, and topographic information, making it impossible to achieve high-resolution and high-precision correction of wind speed data in complex terrain areas. This results in insufficient accuracy in wind energy resource assessment and wind power project site selection.

Method used

A hybrid model combining convolutional neural networks and long short-term memory networks is used for data downscaling, and statistical methods and machine learning algorithms are combined to correct biases. Multi-source data are fused to generate high-resolution wind speed grid data.

Benefits of technology

It improves the spatial resolution and numerical accuracy of wind speed data, reduces the mean absolute error by more than 40%, and increases the correlation coefficient R² from 0.62 to 0.86. It can achieve high-precision wind speed extrapolation in complex terrain areas and support wind energy resource surveys and micro-site selection of wind power projects.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122654993B_ABST
    Figure CN122654993B_ABST
Patent Text Reader

Abstract

This invention provides a multi-source fusion correction method and system for regional wind speed data: It acquires mesoscale meteorological data, ground-measured data, and topographic geographic information data. After quality control, missing value imputation, standardization, and time alignment, a hybrid model of convolutional neural network and long short-term memory network is constructed to downscale the coarse-resolution wind speed field to a high-resolution wind speed field. Then, statistical and machine learning methods are used to jointly correct biases to obtain the corrected wind speed. The downscaling and bias correction model are concatenated and applied to mesoscale meteorological forecast data to generate an hourly high-resolution wind speed grid. The accuracy is verified using untrained ground-measured data. This invention systematically fuses multi-source heterogeneous data, effectively solving the problems of low resolution and sparse meteorological tower sites in mesoscale data through hybrid model downscaling and multi-model integrated correction. It can directly serve wind energy resource surveys, micro-site selection, and power generation assessment, and has significant engineering practical value.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of meteorological data processing technology, specifically to a method and system for multi-source fusion correction of regional wind speed data. Background Technology

[0002] Wind speed data is a crucial foundation for wind energy resource assessment, wind power project site selection, power grid dispatching, and climate research. Currently, the main sources of wind speed data fall into two categories: one is mesoscale meteorological reanalysis data (such as ERA5, CFSR, WRF simulations, etc.) and numerical weather prediction data (such as ECMWF, Pangu meteorological model output); the other is measured observation data from ground-based wind towers or monitoring stations.

[0003] Mesoscale meteorological data has advantages such as wide spatial coverage, strong temporal continuity, and the ability to provide gridded wind field information, and is widely used in regional wind energy resource surveys. However, the spatial resolution of such data is usually low (for example, the spatial resolution of ERA5 reanalysis data is about 0.25°×0.25°, corresponding to a ground grid side length of about 25~30 kilometers), making it difficult to accurately reflect the fine influence of local topography (such as ridges, valleys, and steep slopes) on wind speed, resulting in large errors in wind speed estimation in complex terrain areas.

[0004] Ground-based wind measurement towers or monitoring stations provide highly accurate and reliable data that accurately reflects the wind conditions at their location. However, due to limitations in construction costs and geographical conditions, wind measurement towers are sparsely distributed and cannot achieve full coverage. In high-altitude and complex terrain areas such as autonomous regions, monitoring stations are particularly scarce, and data from a single point cannot represent the overall spatial distribution of wind speed in the region.

[0005] To address the aforementioned issues, researchers have attempted to improve the spatial resolution and correct biases of mesoscale wind speed data using statistical downscaling, dynamic downscaling, or a combination of both. For example, traditional statistical methods such as linear regression and quantile mapping are used for bias correction, while numerical weather prediction models (such as WRF) are employed for dynamic downscaling. However, traditional statistical methods often assume the stationarity of forecast errors, making it difficult to capture the nonlinear variations in wind speed under complex terrain; dynamic downscaling methods are computationally expensive and sensitive to initial fields and boundary conditions.

[0006] In recent years, with the development of artificial intelligence technology, deep learning models have shown potential in the downscaling and bias correction of meteorological elements. Convolutional Neural Networks (CNNs) can effectively extract the spatial features of meteorological fields, Long Short-Term Memory Networks (LSTMs) are adept at capturing the temporal dependence of wind speed, and machine learning models such as XGBoost can be used for nonlinear bias correction. However, existing methods are mostly limited to single correction methods (such as using only CNNs or only statistical correction), lacking a systematic fusion framework from large-scale forecasting → local downscaling → multi-factor bias correction → independent site verification. In particular, how to deeply fuse multi-source heterogeneous data such as mesoscale meteorological data, anemometer tower measured data, and topographic geographic information (elevation, slope, aspect, roughness), and construct a refined wind speed correction model suitable for complex terrain conditions, remains an urgent problem to be solved in the current technical field.

[0007] Therefore, there is an urgent need to provide a wind speed data processing method and system that can integrate multi-source data, consider various terrain factors, and use advanced artificial intelligence algorithms for systematic downscaling and bias correction, so as to improve the spatial resolution and numerical accuracy of wind speed data in complex terrain areas and meet the actual needs of refined wind energy resource assessment and micro-site selection of wind power projects. Summary of the Invention

[0008] The purpose of this invention is to solve the problems mentioned in the background section above, and this invention provides the following technical solution: A multi-source fusion correction method for regional wind speed data includes the following steps: S1. Acquire mesoscale meteorological data, ground-measured data, and topographic geographic information data of the study area, and perform data processing on the acquired data. The data processing includes at least data quality control, missing value imputation, standardization, and time alignment. S2. Construct and train a hybrid model that includes a convolutional neural network and a long short-term memory network. Based on the mesoscale meteorological data and the topographic geographic information data, use the hybrid model to downscale the coarse-resolution mesoscale wind speed data into a high-resolution wind speed field to obtain a downscaled model. S3. During the training phase, based on the downscaled wind speed and the corresponding ground-measured data for the time period, a bias correction model is constructed and trained using statistical methods and machine learning algorithms to correct the bias in the downscaled results. S4. Connect the trained downscaling model and the bias correction model in series and apply them to the mesoscale meteorological forecast data in the study area to generate high-resolution hourly wind speed grid data. S5. Using ground-based measured data that was not used in model training, the accuracy of the generated high-resolution hourly wind speed grid data is verified.

[0009] More preferably, the mesoscale meteorological data includes mesoscale wind speed data and numerical weather prediction data; The ground-based measured data includes measured wind speed data from wind measurement towers or monitoring stations, and observation data from meteorological stations; The topographic geographic information data includes digital elevation model data and land use data, which are used to extract elevation, slope, aspect and surface roughness length.

[0010] More preferably, step S2 specifically includes: S2.1. The spatial features of the mesoscale meteorological element field are extracted using the convolutional neural network, wherein the mesoscale meteorological element field includes at least wind speed, air pressure, temperature and humidity; S2.2. After concatenating the spatial features extracted by the convolutional neural network with the terrain factors, the concatenation network is input into the long short-term memory network to capture the time dependence of wind speed, output the downscaled wind speed value, and establish a downscaled model. The terrain factors include elevation, slope, aspect and roughness length. S2.3 Divide the measured hourly wind speed data at the location of the wind tower or monitoring station into a training set and a validation set. Use the measured hourly wind speed of the training set as the label, and use the mean square error as the loss function to train the hybrid model. Then, select early stop or hyperparameter based on the mean square error of the validation set.

[0011] More preferably, in step S3, a bias correction model is constructed and trained using statistical methods and machine learning algorithms to correct the bias in the downscaling results, specifically including: S3.1. Perform a first-order deviation correction using traditional statistical methods to obtain a first-order correction result. The traditional statistical methods include at least one of mean deviation correction, multiplicative correction, or quantile mapping. S3.2. A machine learning model is used to perform secondary bias correction to obtain the secondary correction result. The machine learning model includes at least one of the XGBoost model or the Long Short-Term Memory network model. S3.3. The first correction result and the second correction result are weighted and integrated to obtain the final corrected wind speed.

[0012] More preferably, the integration weights of the weighted integration in step S3.3 are determined by minimizing the mean square error on the validation set.

[0013] More preferably, in step S4, for areas without ground-based measured data, the spatial proximity method is used to migrate the correction parameters. Specifically, this includes fusing the correction model outputs of multiple surrounding wind measurement towers or monitoring stations using a distance-weighted average method, where the weights are inversely proportional to the square of the distance.

[0014] More preferably, the accuracy verification of the generated high-resolution hourly wind speed grid data specifically includes: The evaluation metrics for accuracy verification include at least one of the following: mean absolute error, root mean square error, coefficient of determination, and mean absolute angular error of wind direction.

[0015] More preferably, the data quality control and the missing value imputation specifically include: For non-continuous missing values, linear interpolation is used for imputation, and the formula is:

[0016] in, The value is the estimated value (meters per second) for the missing time. , The effective values ​​(meters per second) are the values ​​before and after the given time. For data segments that are missing for more than 3 consecutive hours, the entire segment will be removed. Outliers are detected using the 3σ principle, when the following conditions are met: Values ​​that are identified as outliers are either removed or corrected. This is the sample mean (meters per second). Standard deviation (meters per second); The data standardization and time alignment specifically include: Minimum-maximum standardization is performed on features of different dimensions, linearly scaling the features to the [0,1] interval. The standardization formula is as follows:

[0017] in, The standardized dimensionless value. This is the original data. , These are the minimum and maximum values ​​of the data, respectively. For data with different time resolutions, a unified hourly scale is adopted, and linear interpolation is used for time alignment. The interpolation formula is as follows:

[0018] in, For the target time Interpolated wind speed (meters per second). , The wind speed values ​​(in meters per second) are shown before and after the time intervals. The sub-hourly measured data collected by the wind measurement tower or monitoring station are aggregated into an hourly average. The aggregation formula is as follows:

[0019] in, The original number of times within an hour. Let be the average wind speed (meters per second) in the kth hour. For the j-th sampling time within the k-th hour, Let be the measured wind speed value at the j-th sampling time.

[0020] More preferably, the weighted integration of the first correction result and the second correction result to obtain the final corrected wind speed specifically includes: Minimize the mean square error between the weighted ensemble wind speed and the measured wind speed on the validation set, while satisfying the constraint that the sum of the weights is 1; the calculation formula for the weighted ensemble is as follows:

[0021] in, For the final corrected wind speed, The corrected wind speed values ​​are output by three bias correction methods: quantile mapping, XGBoost model, and long short-term memory network model. They are respectively The corresponding weighting coefficients.

[0022] A regional wind speed data multi-source fusion correction system, applied to the aforementioned regional wind speed data multi-source fusion correction method, includes: The data acquisition and preprocessing module acquires and preprocesses mesoscale meteorological data, ground-measured data, and topographic geographic information data of the study area. The AI ​​downscaling module has a built-in and trained hybrid model that includes convolutional neural networks and long short-term memory networks, which is used to downscale coarse-resolution mesoscale wind speed data into high-resolution wind speed fields. The deviation correction module employs statistical and machine learning methods to perform multi-model ensemble correction on the downscaled wind speed. The deviation correction module includes: a traditional statistical correction unit for performing quantile mapping correction; a machine learning correction unit for performing XGBoost and / or LSTM model correction; and an ensemble learning unit for weighted fusion of at least two correction results. The grid generation module connects the trained downscaling model and the bias correction model in series and applies them to the mesoscale meteorological forecast data within the study area to generate high-resolution hourly wind speed grid data. The verification module uses ground-based measured data that was not involved in model training to verify the accuracy of the generated high-resolution hourly wind speed grid data.

[0023] Compared with existing technologies, the multi-source fusion correction method and system for regional wind speed data provided by this invention can bring the following significant benefits: (1) This invention breaks through the limitations of traditional methods that rely on a single data source by systematically integrating mesoscale reanalysis data, ground-based wind tower measured data, meteorological station observation data, and topographic geographic information data (elevation, slope, aspect, roughness). By constructing a CNN-LSTM hybrid model for spatiotemporal downscaling and employing a bias correction strategy that combines statistical and machine learning methods, it effectively solves the problems of low spatial resolution of mesoscale data and the difficulty in reflecting the influence of complex terrain due to the sparseness of wind tower stations. Implementation examples demonstrate that the mean absolute error (MAE) of wind speed compared to independent meteorological stations is reduced by more than 40% after correction, and the coefficient of determination R² increases from 0.62 to 0.86.

[0024] (2) This invention overcomes the limitations of single correction methods by employing a weighted ensemble strategy combining quantile mapping, XGBoost, and LSTM. The weights are adaptively determined by minimizing the mean square error on the validation set, thus integrating the advantages of statistical and machine learning methods. This method not only corrects the mean deviation and distribution pattern of wind speed but also captures temporal dependence and nonlinear characteristics. Validation results under extreme weather conditions (cold wave processes) show that the correlation coefficient between the corrected wind speed and the measured value reaches 0.87–0.89, demonstrating good generalization ability and stability.

[0025] (3) This invention connects the trained downscaling model with the bias correction model to generate hourly wind speed grid data with a resolution of 1km×1km. For areas without meteorological tower coverage, the spatial proximity method (distance-weighted average) is used to integrate the correction parameters of surrounding stations, achieving high-precision wind speed extrapolation in areas without stations. This method effectively solves the problem of insufficient representativeness of traditional methods relying on a single meteorological tower, and can directly serve wind energy resource surveys, micro-site selection of wind power projects, and power generation assessment, with significant engineering practical value. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the overall process of multi-source fusion correction of regional wind speed data in this invention; Figure 2 This is a standardized raster image of wind speed data after preprocessing ERA5 mesoscale data in January according to an embodiment of the present invention; in the figure, (a) is the uncalibrated data and (b) is the calibrated data; Figure 3 This is a standardized raster image of wind speed data after preprocessing ERA5 mesoscale data in February according to an embodiment of the present invention; in the figure, (a) is the uncalibrated data and (b) is the calibrated data; Figure 4 This is a standardized raster image of wind speed data after preprocessing ERA5 mesoscale data in March according to an embodiment of the present invention; in the figure, (a) is the uncalibrated data and (b) is the calibrated data; Figure 5This is a standardized raster image of wind speed data after preprocessing ERA5 mesoscale data in April according to an embodiment of the present invention; in the figure, (a) is the uncalibrated data and (b) is the calibrated data; Figure 6 This is a standardized raster image of wind speed data after preprocessing ERA5 mesoscale data in May according to an embodiment of the present invention; in the figure, (a) is the uncalibrated data and (b) is the calibrated data; Figure 7 This is a standardized raster image of wind speed data after preprocessing ERA5 mesoscale data in June according to an embodiment of the present invention; in the figure, (a) is the uncalibrated data and (b) is the calibrated data; Figure 8 This is a standardized raster image of wind speed data after preprocessing ERA5 mesoscale data in July according to an embodiment of the present invention; in the figure, (a) is the uncalibrated data and (b) is the calibrated data; Figure 9 This is a standardized raster image of wind speed data after preprocessing ERA5 mesoscale data in August according to an embodiment of the present invention; in the figure, (a) is the uncalibrated data and (b) is the calibrated data; Figure 10 This is a standardized raster image of wind speed data after preprocessing ERA5 mesoscale data in September according to an embodiment of the present invention; in the figure, (a) is the uncalibrated data and (b) is the calibrated data; Figure 11 This is a standardized raster image of wind speed data after preprocessing ERA5 mesoscale data in October according to an embodiment of the present invention; in the figure, (a) is the uncalibrated data and (b) is the calibrated data; Figure 12 This is a standardized raster image of wind speed data after preprocessing ERA5 mesoscale data in November according to Embodiment 1 of the present invention; in the figure, (a) is the uncalibrated data and (b) is the calibrated data; Figure 13 This is a standardized raster image of wind speed data after preprocessing ERA5 mesoscale data in December according to Embodiment 1 of the present invention; in the figure, (a) is the uncalibrated data and (b) is the calibrated data; Figure 14 This is a standardized raster map showing the effect of improving regional wind speed data in the embodiments of the present invention. Detailed Implementation

[0027] 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.

[0028] Example 1: This invention provides a multi-source fusion correction method for regional wind speed data, comprising the following steps: S1. Acquire mesoscale meteorological data, ground-measured data, and topographic geographic information data of the study area, and process the acquired data. The data processing includes at least data quality control, missing value imputation, standardization, and time alignment. In step S1: mesoscale meteorological data includes mesoscale wind speed data and numerical weather prediction data; ground-measured data includes measured wind speed data from wind towers or monitoring stations and observation data from meteorological stations; topographic geographic information data includes digital elevation model data and land use data, used to extract elevation, slope, aspect, and surface roughness length. Data quality control and missing value imputation specifically include: For non-continuous missing values, linear interpolation is used for imputation, and the formula is:

[0029] in, The value is the estimated value (meters per second) for the missing time. , The effective values ​​(meters per second) are the values ​​before and after the given time. For data segments that are missing for more than 3 consecutive hours, the entire segment will be removed. Outliers are detected using the 3σ principle, when the following conditions are met: Values ​​that are identified as outliers are either removed or corrected. This is the sample mean (meters per second). Standard deviation (meters per second); The data standardization and time alignment specifically include: Minimum-maximum standardization is performed on features of different dimensions, linearly scaling the features to the [0,1] interval. The standardization formula is as follows:

[0030] in, The standardized dimensionless value. This is the original data. , These are the minimum and maximum values ​​of the data, respectively. For data with different time resolutions, a unified hourly scale is adopted, and linear interpolation is used for time alignment. The interpolation formula is as follows:

[0031] in, For the target time Interpolated wind speed (meters per second). , The wind speed values ​​(in meters per second) are shown before and after the time intervals. The sub-hourly measured data collected by the wind measurement tower or monitoring station are aggregated into an hourly average. The aggregation formula is as follows:

[0032] in, The original number of times within an hour. Let be the average wind speed (meters per second) in the kth hour. For the j-th sampling time within the k-th hour, Let be the measured wind speed value at the j-th sampling time.

[0033] S2. Construct and train a hybrid model containing a convolutional neural network and a long short-term memory network. Based on the mesoscale meteorological data and the topographic geographic information data, the hybrid model is used to downscale the coarse-resolution mesoscale wind speed data to a high-resolution wind speed field, resulting in a downscaled model. Step S2 specifically includes: S2.1. Using the convolutional neural network to extract the spatial features of the mesoscale meteorological element field, which includes at least wind speed, air pressure, temperature, and humidity. S2.2. After concatenating the spatial features extracted by the convolutional neural network with topographic factors, input them into the long short-term memory network to capture the time dependence of wind speed, output the downscaled wind speed value, and establish a downscaled model. The topographic factors include elevation, slope, aspect, and roughness length. S2.3. Divide the measured hourly wind speed data at the location of the wind tower or monitoring station into a training set and a validation set. Using the measured hourly wind speed of the training set as the label, the hybrid model is trained using mean squared error as the loss function, and early stopping or hyperparameter selection is performed based on the mean squared error of the validation set.

[0034] S3. During the training phase, based on the downscaled wind speed and the corresponding ground-measured data for the time period, a bias correction model is constructed and trained using statistical methods and machine learning algorithms to correct the bias in the downscaling results. Step S3, which uses statistical methods and machine learning algorithms to construct and train the bias correction model to correct the bias in the downscaling results, specifically includes: S3.1. Performing a first-order bias correction using traditional statistical methods to obtain a first-order correction result. The traditional statistical methods include at least one of average bias correction, multiplicative correction, or quantile mapping. S3.2. Performing a second-order bias correction using a machine learning model to obtain a second-order correction result. The machine learning model includes at least one of an XGBoost model or a Long Short-Term Memory network model. S3.3. Weightedly integrating the first-order correction result and the second-order correction result to obtain the final corrected wind speed, specifically including: Minimize the mean square error between the weighted ensemble wind speed and the measured wind speed on the validation set, while satisfying the constraint that the sum of the weights is 1; the calculation formula for the weighted ensemble is as follows:

[0035] in, For the final corrected wind speed, The corrected wind speed values ​​are output by three bias correction methods: quantile mapping, XGBoost model, and long short-term memory network model. They are respectively The corresponding weighting coefficients.

[0036] The weights of the weighted ensemble satisfy the non-negativity constraint and are determined by minimizing the mean square error on the validation set.

[0037] S4. The trained downscaling model and the bias correction model are concatenated and applied to the mesoscale meteorological forecast data within the study area to generate high-resolution hourly wind speed grid data. In step S4, for areas without ground-based measured data, the spatial proximity method is used to migrate the correction parameters. Specifically, this includes fusing the correction model outputs from multiple surrounding wind towers or monitoring stations using a distance-weighted average method, where the weights are inversely proportional to the square of the distance.

[0038] S5. Using ground-based measured data that was not used in model training, verify the accuracy of the generated high-resolution hourly wind speed grid data. Specifically, this includes: The evaluation metrics for accuracy verification include at least one of the following: mean absolute error, root mean square error, coefficient of determination, and mean absolute angular error of wind direction.

[0039] This embodiment also provides a regional wind speed data multi-source fusion correction system, applied to the aforementioned regional wind speed data multi-source fusion correction method, including: The data acquisition and preprocessing module acquires and preprocesses mesoscale meteorological data, ground-measured data, and topographic geographic information data of the study area. The AI ​​downscaling module has a built-in and trained hybrid model that includes convolutional neural networks and long short-term memory networks, which is used to downscale coarse-resolution mesoscale wind speed data into high-resolution wind speed fields. The deviation correction module employs statistical and machine learning methods to perform multi-model ensemble correction on the downscaled wind speed. The deviation correction module includes: a traditional statistical correction unit for performing quantile mapping correction; a machine learning correction unit for performing XGBoost and / or LSTM model correction; and an ensemble learning unit for weighted fusion of at least two correction results. The grid generation module connects the trained downscaling model and the bias correction model in series and applies them to the mesoscale meteorological forecast data within the study area to generate high-resolution hourly wind speed grid data. The verification module uses ground-based measured data that was not involved in model training to verify the accuracy of the generated high-resolution hourly wind speed grid data.

[0040] This embodiment uses a certain autonomous region as the research area, selecting a typical complex terrain area within the latitude and longitude range (78°E~100°E, 28°N~36°N), and employs the method of this invention to achieve multi-source fusion correction of regional wind speed data. This region has an average altitude of over 4000m, sparsely distributed monitoring stations, and strong spatial variability in wind speed, making it an ideal experimental area for verifying the effectiveness of this invention.

[0041] I. Multi-source data acquisition and preprocessing 1.1 Multi-source data collection This embodiment collects four types of basic data: (1) Mesoscale data: The ECMWF ERA5 reanalysis dataset was used, with a time resolution of 1 hour and a spatial resolution of 0.25×0.25. The variables included wind speed at 100 meters above the ground, wind direction, air temperature, air pressure, and relative humidity. The time span was from January 1, 2020 to December 31, 2024.

[0042] (2) Monitoring station data: Measured wind speed data from 25 monitoring stations within the autonomous region were collected with a time resolution of 10 minutes and an installation height of 120 meters. After quality control, the data were aggregated into hourly average values. The monitoring stations covered different terrain types (mountains, valleys, plateaus, and plains) at altitudes ranging from 3000m to 5200m.

[0043] (3) Meteorological station data: Measured wind speed data (hourly average) from 30 national meteorological stations in the region were collected, of which 25 were used for model training and calibration, and 5 were used for independent verification.

[0044] (4) Topographic geographic information data: acquire SRTM 90m resolution digital elevation model (DEM) data, and extract elevation, slope, aspect and topographic relief from it; at the same time, acquire GlobCover land use data and calculate the surface roughness length.

[0045] 1.2 Data Quality Control The raw data undergoes integrity checks and outlier identification. For non-continuous missing data (single point or less than 3 hours of continuous data) at monitoring stations, linear interpolation is used to fill in the gaps, using the following formula:

[0046] in, The value is the estimated value (meters per second) for the missing time. , These are the valid values ​​(meters per second) for the preceding and following times.

[0047] Data segments that are missing for more than 3 consecutive hours will be removed entirely.

[0048] The three-standard-deviation criterion is used to detect outliers, as shown in the following formula:

[0049] in, This is the sample mean (m / s). The standard deviation is denoted as m / s. Points exceeding the threshold are considered outliers and are removed or corrected. Statistically, the average wind speed at the monitoring stations is 5.8 m / s, with a standard deviation of 2.6 m / s. The removal threshold is set to ≥20 m / s or ≤0 m / s.

[0050] 1.3 Data Standardization and Time Alignment Mini-maximum standardization is applied to features of different dimensions to linearly scale the data to the [0,1] interval, as shown in the following formula:

[0051] in, The standardized dimensionless value. This is the original data. , These are the minimum and maximum values ​​of the data, respectively. For data with different time resolutions, a unified hourly scale is adopted, and linear interpolation is used for time alignment. The interpolation formula is as follows:

[0052] in, For the target time Interpolated wind speed (meters per second). , These are the wind speed values ​​(in meters per second) at the previous and next times, respectively.

[0053] The sub-hourly measured data collected by the wind measurement tower or monitoring station are aggregated into an hourly average. The aggregation formula is as follows:

[0054] in, The original number of times within an hour. Let be the average wind speed (meters per second) in the kth hour. For the j-th sampling time within the k-th hour, Let be the measured wind speed value at the j-th sampling time.

[0055] II. Spatiotemporal scaling based on artificial intelligence 2.1 Model Construction A downscaling model combining a convolutional neural network (CNN) and a long short-term memory network (LSTM) is constructed. The input is an ERA5 grid meteorological field (5×5 grid, covering approximately 125km×125km around the target location), including four channels: wind speed, air pressure, temperature, and humidity. The CNN part consists of two convolutional layers (3×3 kernel, stride 1, and 64 and 128 filters), employing the ReLU activation function and max pooling operation to extract the spatial features of the meteorological field and flatten them into a one-dimensional vector.

[0056] The convolutional layer operation is defined as follows:

[0057] in, For the first Layer feature map, For convolution kernel, For convolution operations, For bias, For ReLU activation function: (x) = max(0,x).

[0058] The vector is concatenated with terrain factors (elevation difference, slope, aspect, roughness length) and then input into a two-layer LSTM (128 hidden units) to capture the temporal dependency of wind speed. The LSTM unit update rules are as follows:

[0059]

[0060]

[0061]

[0062]

[0063]

[0064] in, , , These are the forget gate, input gate, and output gate, respectively. For memory units, Candidate memory units, In hidden state, This indicates element-wise multiplication. Input the feature vector at the current time. The hidden state from the previous moment.

[0065] Take the hidden state at the last time step As a summary of temporal features, it is input into a fully connected output layer to obtain downscaled wind speed. (meters per second), the formula is as follows:

[0066] in, For output layer weights, This is a bias term.

[0067] 2.2 Model Training The measured hourly wind speed at the location of the wind measurement tower (or monitoring station) (meters per second) is the label. Let the total number of samples participating in training be (number of samples), and the loss function be the mean squared error:

[0068] The Adam optimizer was used, with a learning rate of 0.001, a batch size of 32, and a training epoch. ( ≥100), and adopt the early stopping method (stop if the loss does not decrease after 10 rounds to prevent overfitting).

[0069] III. Wind Speed ​​Deviation Correction 3.1 Traditional Bias Correction (Quantile Mapping QM) A cumulative distribution function mapping relationship between predicted and measured wind speeds is constructed based on historical samples. For any predicted value, its quantile in the predicted distribution is first calculated, and then the wind speed value corresponding to the quantile in the measured distribution is used as the correction result.

[0070] Traditional deviation correction methods fall into three categories: mean deviation correction, multiplicative correction, and quantile mapping. Mean deviation correction calculates the average deviation *b* (meters per second) between predicted and measured values ​​in historical data sets.

[0071] Corrected wind speed (meters per second), the calculation formula is as follows:

[0072] The multiplicative correction method is suitable for cases where the relative error is stable, and the calculation formula is as follows:

[0073]

[0074] in Dimensionless.

[0075] Quantile mapping methods can correct not only the mean but also the entire distribution, and are used to construct wind speed forecasts based on historical samples. Empirical cumulative distribution function and measured wind speed cumulative distribution function For any forecast value Its correction value is:

[0076] 3.2 Machine Learning Bias Correction (XGBoost) An XGBoost regression model was constructed, with input features including: downscaled wind speed, elevation difference, slope, aspect, hour (0-23), month (1-12), and wind direction. The target variable was the measured wind speed at the monitoring station.

[0077] The objective function of XGBoost is as follows:

[0078] The meanings of each parameter are as follows: The actual output of the i-th sample, i.e., the corrected wind speed. .

[0079] The predicted output of the i-th sample, i.e., the model's prediction of the i-th sample. The estimated value .

[0080] Loss function: measures the difference between predicted and actual values. For regression tasks like wind speed correction, squared loss is typically used: .

[0081] : No. Each decision tree (or regression tree) corresponds to a base learner.

[0082] Regularization term: Used to control model complexity and prevent overfitting.

[0083] The function takes the following form:

[0084] in: : The number of leaf nodes in the tree.

[0085] : No. The weight of each leaf node (i.e., the output value of that leaf node).

[0086] : Regularization hyperparameters, which control the number of leaf nodes and the weight of each leaf node.

[0087] The total number of training samples (corresponding to different time series) ).

[0088] The total number of decision trees in XGBoost (number of iterations).

[0089] The input features are used to construct the decision rules for each tree during model training, but they are not explicitly included in the expression of the objective function.

[0090] 3.3 Timing Deviation Correction (LSTM) Temporal bias correction is performed using Long Short-Term Memory (LSTM) network temporal modeling, as shown in the following formula:

[0091]

[0092]

[0093]

[0094]

[0095]

[0096] in, , , These are the forget gate, input gate, and output gate, respectively. For memory units, Candidate memory units, In hidden state, This indicates element-wise multiplication. Input the feature vector at the current time. This represents the hidden state from the previous time step. The input feature vector is... , Output the correction value at the corresponding time.

[0097] The input sequence length is 12 hours, and the features are the same as XGBoost. The LSTM unit update rule is the same as in step 2.1. The network structure is a two-layer LSTM (64 hidden units), and the output layer is a fully connected layer. Training is performed for 100 epochs, using early stopping to prevent overfitting.

[0098] 3.4 Multi-model integrated correction The correction results of the three methods (quantile mapping, XGBoost, and LSTM) in steps 3.1-3.3 are then weighted and integrated:

[0099] Weight Determined by minimizing the mean square error on the validation set, satisfying =1 in, For the final corrected wind speed, The corrected wind speed values ​​are output by three bias correction methods: quantile mapping, XGBoost model, and long short-term memory network model. They are respectively The corresponding weighting coefficients.

[0100] This invention employs a weighted ensemble strategy combining quantile mapping, XGBoost, and LSTM methods. The weights are adaptively determined by minimizing the mean squared error on the validation set, thus integrating the advantages of statistical and machine learning methods. Extreme weather validation results demonstrate that the calibration model exhibits good generalization ability and stability.

[0101] IV. Generation of Regional Wind Speed ​​Grids The trained downscaling model is concatenated with the bias correction model and applied to regional mesoscale meteorological forecast data (such as ERA5 reanalysis data or numerical weather prediction data) to generate a 1km×1km resolution hourly wind speed grid. For areas without wind towers (or monitoring stations), the spatial proximity method is used to migrate correction parameters: the correction model outputs from multiple surrounding wind towers (or monitoring stations) are fused using a distance-weighted average method, with the weights being the reciprocal of the square of the distance, as shown in the following formula:

[0102] in, denoted as , where is the distance (km) between the grid point and the j-th wind measurement tower (or monitoring station). The corrected wind speed (m / s) is calculated using the tower correction model. To represent the total number of wind measurement towers or monitoring stations participating in the weighted fusion, this embodiment uses the three nearest surrounding stations, i.e. .

[0103] Therefore, this invention can generate hourly wind speed grid data with a resolution of 1km×1km, and use the spatial proximity method (distance-weighted average) to extrapolate wind speed in areas without wind tower coverage, thus achieving high-precision wind speed extrapolation in areas without stations, which can directly serve wind energy resource surveys, micro-site selection and power generation assessment.

[0104] V. Verification by Independent Meteorological Stations 5.1 Overall Result Verification Measured wind speed data from meteorological stations not involved in the modeling within the region were compared and validated with the corrected wind speed grid data. The following evaluation indicators were calculated: mean absolute error (MAE), root mean square error (RMSE), and correlation coefficient (R²), as shown in the following formulas:

[0105]

[0106]

[0107]

[0108] in, It refers to the number of points used in the comparative verification. This is the corrected wind speed grid data. These are measured wind speed data from meteorological stations that were not involved in the modeling. It is the mean absolute error. It is the root mean square error. It is the correlation coefficient. The wind direction error is calculated using the average absolute angle error. The verification results are shown in Table 1 below.

[0109] Table 1 Verification Results

[0110] Therefore, this invention effectively solves the problems of low resolution and sparse meteorological tower stations in mesoscale data by using multi-source data fusion, CNN-LSTM downscaling, and statistical-machine learning joint correction. After correction, the mean absolute error between wind speed and independent meteorological stations is reduced by more than 40%, and the correlation coefficient R² is increased from 0.62 to 0.86.

[0111] 5.2 Spatial Distribution of Accuracy Improvement Accuracy Improvement Spatial Distribution. Based on the accuracy improvement effect calculated for each point in the entire area according to this embodiment, the results for some points are shown in Table 2. The improvement range is defined as (reduction in corrected average absolute error / original average absolute error) × 100%.

[0112] Table 2. Improvement in wind speed correction accuracy at some locations in the autonomous region

[0113] Statistics show that the average improvement across the region is 12%, while the improvement can reach more than 14% in the vicinity of monitoring stations and in areas with complex terrain. The closer to the monitoring station, the more significant the improvement in accuracy.

[0114] 5.3 Extreme Weather Verification Two typical cold wave events in winter were selected for testing. The correlation coefficients between the corrected wind speed and the measured value were 0.87 and 0.89, respectively, and the mean absolute errors were 1.35 m / s and 1.22 m / s, respectively. This shows that the correction model of the present invention still has good generalization ability under extreme weather conditions.

[0115] This embodiment successfully improved the resolution of the autonomous region's ERA5 mesoscale reanalysis data from 0.25°×0.25° to 1km×1km. After correction, the mean absolute error of wind speed compared to independent meteorological stations was reduced by more than 40%, and R² was improved from 0.62 to 0.86. This method effectively solves the problems of sparse monitoring stations and large mesoscale data errors in high-altitude and complex terrain areas, providing a highly reliable data foundation for the refined assessment of wind energy resources.

[0116] In summary, this invention has beneficial effects such as improving accuracy through multi-source fusion, enhancing generalization ability through multi-model integration, and supporting engineering applications through high-resolution mesh.

[0117] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for multi-source fusion correction of regional wind speed data, characterized in that, Includes the following steps: S1. Acquire mesoscale meteorological data, ground-measured data, and topographic geographic information data of the study area, and perform data processing on the acquired data. The data processing includes at least data quality control, missing value imputation, standardization, and time alignment. S2. Construct and train a hybrid model that includes a convolutional neural network and a long short-term memory network. Based on the mesoscale meteorological data and the topographic geographic information data, use the hybrid model to downscale the coarse-resolution mesoscale wind speed data into a high-resolution wind speed field to obtain a downscaled model. Step S2 specifically includes: S2.

1. The spatial features of the mesoscale meteorological element field are extracted using the convolutional neural network, wherein the mesoscale meteorological element field includes at least wind speed, air pressure, temperature and humidity; S2.

2. After concatenating the spatial features extracted by the convolutional neural network with the terrain factors, the concatenation network is input into the long short-term memory network to capture the time dependence of wind speed, output the downscaled wind speed value, and establish a downscaled model. The terrain factors include elevation, slope, aspect and roughness length. S2.3 Divide the measured hourly wind speed data at the location of the wind tower or monitoring station into a training set and a validation set. Use the measured hourly wind speed of the training set as the label, and use the mean square error as the loss function to train the hybrid model. Then, select early stop or hyperparameter based on the mean square error of the validation set. S3. During the training phase, based on the downscaled wind speed and the corresponding ground-measured data for the time period, a bias correction model is constructed and trained using statistical methods and machine learning algorithms to correct the bias in the downscaled results. In step S3, a bias correction model is constructed and trained using statistical methods and machine learning algorithms to correct the bias in the downscaling results. Specifically, this includes: S3.

1. Perform a first-order deviation correction using traditional statistical methods to obtain a first-order correction result. The traditional statistical methods include at least one of mean deviation correction, multiplicative correction, or quantile mapping. S3.

2. A machine learning model is used to perform secondary bias correction to obtain the secondary correction result. The machine learning model includes at least one of the XGBoost model or the Long Short-Term Memory network model. S3.

3. The first correction result and the second correction result are weighted and integrated to obtain the final corrected wind speed; In step S3.3, the integration weights of the weighted integration satisfy the non-negativity constraint and are determined by minimizing the mean square error on the validation set. S4. Connect the trained downscaling model and the bias correction model in series and apply them to the mesoscale meteorological forecast data in the study area to generate high-resolution hourly wind speed grid data. In step S4, for areas without ground-based measured data, the spatial proximity method is used to migrate the correction parameters. Specifically, this includes fusing the correction model outputs of multiple surrounding wind towers or monitoring stations using a distance-weighted average method, where the weights are inversely proportional to the square of the distance. S5. Using ground-based measured data that was not used in model training, the accuracy of the generated high-resolution hourly wind speed grid data is verified.

2. The method for multi-source fusion correction of regional wind speed data according to claim 1, characterized in that, In step S1: The mesoscale meteorological data includes mesoscale wind speed data and numerical weather prediction data; The ground-based measured data includes measured wind speed data from wind measurement towers or monitoring stations, and observation data from meteorological stations; The topographic geographic information data includes digital elevation model data and land use data, which are used to extract elevation, slope, aspect and surface roughness length.

3. The method for multi-source fusion correction of regional wind speed data according to claim 1, characterized in that, The accuracy verification of the generated high-resolution hourly wind speed grid data specifically includes: The evaluation metrics for accuracy verification include at least one of the following: mean absolute error, root mean square error, coefficient of determination, and mean absolute angular error of wind direction.

4. The method for multi-source fusion correction of regional wind speed data according to claim 1, characterized in that, The data quality control and the missing value imputation specifically include: For non-continuous missing values, linear interpolation is used for imputation, and the formula is: in, For the missing time, , The values ​​are valid values ​​before and after the given time. For data segments that are missing for more than 3 consecutive hours, the entire segment will be removed. Outliers are detected using the 3σ principle, when the following conditions are met: Values ​​that are identified as outliers are either removed or corrected. The sample mean. Standard deviation; The data standardization and time alignment specifically include: Minimum-maximum standardization is performed on features of different dimensions, linearly scaling the features to the [0,1] interval. The standardization formula is as follows: in, The standardized dimensionless value. The original data, , These are the minimum and maximum values ​​of the data, respectively. For data with different time resolutions, a unified hourly scale is adopted, and linear interpolation is used for time alignment. The interpolation formula is as follows: in, For the target time interpolated wind speed, , These are the wind speed values ​​at the previous and next times, respectively; The sub-hourly measured data collected by the wind measurement tower or monitoring station are aggregated into an hourly average. The aggregation formula is as follows: in, The original number of times within an hour. Let be the average wind speed in the kth hour. For the j-th sampling time within the k-th hour, Let be the measured wind speed value at the j-th sampling time.

5. The method for multi-source fusion correction of regional wind speed data according to claim 1, characterized in that, The step of weightedly integrating the first correction result and the second correction result to obtain the final corrected wind speed specifically includes: Minimize the mean square error between the weighted ensemble wind speed and the measured wind speed on the validation set, while satisfying the constraint that the sum of the weights is 1; the calculation formula for the weighted ensemble is as follows: in, For the final corrected wind speed, The corrected wind speed values ​​are output by three bias correction methods: quantile mapping, XGBoost model, and long short-term memory network model. They are respectively The corresponding weighting coefficients.

6. A regional wind speed data multi-source fusion correction system, applied to the regional wind speed data multi-source fusion correction method described in any one of claims 1 to 5, characterized in that, include: The data acquisition and preprocessing module acquires and preprocesses mesoscale meteorological data, ground-measured data, and topographic geographic information data of the study area. The AI ​​downscaling module has a built-in and trained hybrid model that includes convolutional neural networks and long short-term memory networks, which is used to downscale coarse-resolution mesoscale wind speed data into high-resolution wind speed fields. The deviation correction module employs statistical and machine learning methods to perform multi-model ensemble correction on the downscaled wind speed. The deviation correction module includes: a traditional statistical correction unit for performing quantile mapping correction; a machine learning correction unit for performing XGBoost and / or LSTM model correction; and an ensemble learning unit for weighted fusion of at least two correction results. The grid generation module connects the trained downscaling model and the bias correction model in series and applies them to the mesoscale meteorological forecast data within the study area to generate high-resolution hourly wind speed grid data. The verification module uses ground-based measured data that was not involved in model training to verify the accuracy of the generated high-resolution hourly wind speed grid data.

Citation Information

Patent Citations

  • Long-term wind speed forecast correction method based on network fusion

    CN120974409A

  • Sub-kilometer scale near-earth forecast wind field real-time correction and downscaling method and model

    CN121682231A

  • Photovoltaic power prediction method and system based on multi-source data fusion and deep learning

    CN121705643A