A method and system for calibrating the indicated value of wind speed of a power meteorological device based on machine learning

Abstract

This invention relates to a machine learning-based method and system for calibrating wind speed readings on power meteorological devices. The method includes: acquiring raw wind speed measurements and environmental parameters collected synchronously with the raw wind speed measurements; calculating air property parameters based on the environmental parameters, and constructing a physical enhancement feature set based on the air property parameters and the raw wind speed measurements; preprocessing the physical enhancement feature set to obtain a preprocessed feature vector; inputting the raw wind speed measurements into a physical basis model to obtain a basis-calibrated wind speed; using the basis-calibrated wind speed and the preprocessed feature vector as inputs to a residual compensation model to obtain a wind speed residual compensation amount; and obtaining the final calibrated wind speed based on the basis-calibrated wind speed and the wind speed residual compensation amount. Compared with existing technologies, this invention significantly improves the wind speed calibration accuracy and extrapolation robustness of power meteorological devices.

Description

Technical Field

[0001] This invention relates to the field of power meteorological monitoring and wind speed measurement calibration technology, and in particular to a method and system for calibrating wind speed readings of power meteorological devices based on machine learning. Background Technology

[0002] Wind speed is one of the important meteorological elements for power grid operation monitoring and wind disaster early warning. Power meteorological devices, as on-site wind speed monitoring terminals in power grids, are typically installed on transmission towers or in substations to achieve real-time wind speed monitoring. To avoid the problems of high starting wind speed, easy wear and tear, and high maintenance costs associated with traditional mechanical anemometers, power meteorological devices often use sensors based on the ultrasonic time-of-flight method for wind speed measurement.

[0003] However, during the on-site installation of power meteorological devices, the surrounding flow field is easily distorted and exhibits wake effects due to the influence of structures such as columns, Stevenson screens, and supports. This results in complex nonlinear deviations between the local wind speed measured by the sensor and the actual ambient wind speed. Traditional correction methods based on empirical formulas typically possess good global monotonicity, but they are insufficient in describing local nonlinear disturbances such as column wakes and shielding effects. Data-driven machine learning methods can learn nonlinear mapping relationships, but when entering extrapolation intervals with insufficient training sample coverage (such as extreme wind conditions), the model may become unstable or even fail due to a lack of physical constraints.

[0004] Furthermore, in high-reliability applications such as power safety monitoring, simply providing point prediction results is insufficient to support safety decisions. It is necessary to quantify the uncertainty of the wind speed measurement process and output prediction intervals or confidence boundaries that can be used for risk assessment, thereby reflecting potential measurement risks under out-of-distribution (OOD) conditions. For example, patent application CN111024986A discloses a wind speed calibration method based on piecewise linear interpolation of historical data, which is a technique for interpolation calibration by establishing a piecewise linear relationship between measured values ​​and standard values. This method can only approximate nonlinear deviations within the data coverage area piecewise linearly, but cannot accurately compensate for complex local flow field distortions; its calibration relies entirely on historical data intervals, and may fail due to a lack of physical constraints under extrapolation conditions with insufficient training samples; and it can only output point-estimated calibration values, failing to quantify measurement uncertainty to support risk assessment.

[0005] Therefore, there is an urgent need for a wind speed calibration method that can balance calibration accuracy, extrapolation stability, and output risk characterization. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art by providing a machine learning-based method and system for calibrating wind speed readings of power meteorological devices, which significantly improves the accuracy of wind speed calibration and the robustness of extrapolation of power meteorological devices.

[0007] The objective of this invention can be achieved through the following technical solutions: A machine learning-based method for calibrating wind speed readings on power meteorological devices includes the following steps: Obtain the original wind speed measurement value, as well as the environmental parameters collected synchronously with the original wind speed measurement value; Air physical property parameters are calculated based on the environmental parameters, and a set of physical enhancement features is constructed based on the air physical property parameters and the original wind speed measurement values. The physical enhancement feature set is preprocessed to obtain a preprocessed feature vector; The original wind speed measurement value is input into the physical base model to obtain the base-calibrated wind speed; The base-calibrated wind speed and the preprocessed feature vector are used as inputs to the residual compensation model to obtain the wind speed residual compensation amount. The final calibrated wind speed is obtained based on the base calibration wind speed and the wind speed residual compensation amount.

[0008] Furthermore, the environmental parameters include ambient temperature and ambient air pressure, and the air physical properties include air density and aerodynamic viscosity. The formula for calculating air density is: In the formula, air density, For ambient air pressure, air is the gas constant. The ambient temperature; The formula for calculating the aerodynamic viscosity is: In the formula, Aerodynamic viscosity, For reference dynamic viscosity, For reference temperature, is Sutherland's constant.

[0009] Furthermore, the set of physical enhancement features includes Reynolds number features, dynamic pressure features, and turbulence intensity features. The formula for calculating the Reynolds number characteristic is as follows: In the formula, Let Reynolds number be 1. air density, This is the original wind speed measurement value. The diameter of the support column for the power meteorological device. Aerodynamic viscosity, This is a low-wind-speed numerical stability disturbance term; The formula for calculating the dynamic pressure characteristic is: In the formula, For dynamic pressure; The turbulence intensity characteristics are obtained using sliding window statistics, and the calculation formula is as follows: In the formula, for Turbulence intensity at time t, The standard deviation of wind speed within the sliding window. This represents the average wind speed within the sliding window. This is a small constant correction term.

[0010] Furthermore, the physical enhancement feature set is preprocessed by using the Z-Score normalization method to standardize the physical enhancement feature vectors in order to eliminate the dimensional differences between different features.

[0011] Furthermore, the physical basis model is a monotonically increasing model in the form of a power-law function, used to characterize the global trend relationship between the original wind speed measurement and the basis-calibrated wind speed. Specifically, the physical basis model is expressed as follows: In the formula, For base calibration of wind speed, This is the original wind speed measurement value. The goal is to obtain model parameters by fitting the training data.

[0012] Furthermore, the residual compensation model is an ensemble model based on the Bagging strategy, which outputs the wind speed residual compensation amount through the prediction ensemble of multiple base learners. The calculation formula for the wind speed residual compensation amount is as follows: In the formula, This is the wind speed residual compensation amount. The total number of base learners, For the first Adaptive weights for each base learner To preprocess the feature vectors, For base calibration of wind speed, For the first The residual compensation amount predicted independently by each base learner Among them, the first The adaptive weights of the base learners are: In the formula, For weighted adjustment of hyperparameters, For the first Variance estimation of cognitive uncertainty for each base learner given an input. For the first Estimation of the variance of cognitive uncertainty of each base learner given an input.

[0013] Furthermore, the final calibrated wind speed is: In the formula, For final wind speed calibration, For base calibration of wind speed, This is the wind speed residual compensation amount.

[0014] Furthermore, based on the final calibrated wind speed, the prediction interval of the final calibrated wind speed at a preset confidence level is estimated by synthesizing the total prediction standard deviation and the scaling factor calibrated from out-of-package samples. The prediction interval is: In the formula, For the prediction interval, For final wind speed calibration, The scaling factor is calibrated using out-of-package samples. This represents the composite total forecast standard deviation; The scaling factor calibrated from the out-of-package sample is: In the formula, It is a 0.95 quantile function. This is the true reference wind speed value for a single sample outside the package. For the predicted wind speed values ​​of the samples outside the package, For scale parameters; The standard deviation of the composite overall prediction is: In the formula, The standard deviation of cognitive uncertainty, The standard deviation of random uncertainty, Standard deviation For the first The residual compensation amount predicted independently by each base learner For a fixed offset coefficient, This is the proportionality coefficient.

[0015] Furthermore, the size of the prediction interval is matched with the reliability of the final calibrated wind speed result.

[0016] According to another aspect of the present invention, a machine learning-based wind speed indication calibration system for power meteorological devices is provided, comprising: The data acquisition module is used to acquire the raw wind speed measurement value, as well as environmental parameters acquired synchronously with the raw wind speed measurement value; The physical enhancement feature set construction module is used to calculate air physical property parameters based on the environmental parameters, and construct a physical enhancement feature set based on the air physical property parameters and the original wind speed measurement value; The preprocessing module is used to preprocess the physical enhancement feature set to obtain a preprocessed feature vector; The base calibration wind speed acquisition module is used to input the original wind speed measurement value into the physical base model to obtain the base calibration wind speed. The wind speed residual compensation quantity acquisition module is used to take the base calibration wind speed and the preprocessed feature vector as inputs to the residual compensation model to obtain the wind speed residual compensation quantity. The final calibration wind speed acquisition module is used to obtain the final calibration wind speed based on the base calibration wind speed and the wind speed residual compensation amount.

[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention integrates physical enhancement features such as Reynolds number, dynamic pressure, and turbulence intensity, and uses a physical basis model to obtain the basis calibration wind speed. Based on the basis calibration wind speed and preprocessed feature vectors, the wind speed residual compensation amount is obtained through a residual compensation model, thus obtaining the final calibration wind speed. This overcomes the limitations of traditional empirical formulas in accurately describing complex nonlinear turbulence deviations and the instability of pure data-driven models in extrapolation under extreme wind conditions. It also solves the problems of large local errors in wind speed calibration in complex flow field environments and prediction failure under unknown strong wind conditions, thus improving the accuracy of wind speed calibration and the robustness of extrapolation.

[0018] 2. Based on the final calibrated wind speed, this invention estimates the prediction range of the final calibrated wind speed at a preset confidence level by synthesizing the total prediction standard deviation and the scaling factor calibrated by the out-of-package samples. This overcomes the limitation of most existing technologies that lack the quantification of the uncertainty of the prediction results, and solves the problem that point prediction values ​​alone cannot assess risks and support safety decisions in high-reliability scenarios such as power safety monitoring. This improves the model's risk perception capability, prediction credibility, and decision support reliability. Attached Figure Description

[0019] Figure 1This is a flowchart illustrating a machine learning-based method for calibrating wind speed readings on power meteorological devices, as proposed in this invention. Figure 2 This is a schematic diagram of the structure of a machine learning-based wind speed calibration system for power meteorological devices proposed in this invention. Detailed Implementation

[0020] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0021] Example 1 This embodiment provides a machine learning-based method for calibrating wind speed readings on power meteorological devices, such as... Figure 1 As shown, it includes the following steps: S1. Obtain the original wind speed measurement value, as well as the environmental parameters collected synchronously with the original wind speed measurement value.

[0022] Obtain the raw wind speed measurements from the power meteorological device. Environmental parameters include ambient temperature and atmospheric pressure, while air properties include air density and aerodynamic viscosity. Standard wind speeds should also be obtained as calibration values ​​during model training and parameter identification.

[0023] The formula for calculating air density is: In the formula, air density, For ambient air pressure, Let be the gas constant for air, and take . , The ambient temperature.

[0024] The formula for calculating aerodynamic viscosity is: In the formula, Aerodynamic viscosity, For reference dynamic viscosity, For reference temperature, Let be Sutherland's constant, where is taken as . , , .

[0025] This embodiment uses the FWS700 power meteorological device as the object of measurement. Its wind speed measurement range is 0–45 m / s, with a resolution of 0.1 m / s and an accuracy of ±0.3 m / s or 3%. The device under test is installed at the center of the closed-pipe wind tunnel test section, and the ultrasonic measurement path is strictly orthogonal to the incoming flow direction in order to simulate the measurement deviation when the device is subjected to the wake interference of complex mechanical structures during actual operation.

[0026] To ensure the reliability of the reference true value, this embodiment adopts a segmented hybrid reference strategy: when the reference wind speed is less than 4.5 m / s, a high-sensitivity thermal anemometer is used as the reference; when the reference wind speed is greater than or equal to 4.5 m / s, a calibrated standard Pitot hydrostatic tube and a digital differential pressure gauge are used as the reference measurement chain.

[0027] S2. Calculate the air physical property parameters based on the environmental parameters, and construct a set of physical enhancement features based on the air physical property parameters and the original wind speed measurement values.

[0028] The set of physical enhancement features includes Reynolds number features, dynamic pressure features, and turbulence intensity features.

[0029] The Reynolds number characteristic reflects the essential properties of a structure under wind load, helping the model understand the impact of different flow states on the structure's wind pressure distribution and vibration characteristics. The formula for calculating the Reynolds number characteristic is: In the formula, Let Reynolds number be 1. air density, This is the original wind speed measurement value. The diameter of the support column for the power meteorological device. Aerodynamic viscosity, This is a numerical stability disturbance term for low wind speeds.

[0030] Dynamic pressure is directly related to the wind load on the structure. Dynamic pressure characteristics provide the model with a physical quantity proportional to the structural forces, enhancing the model's direct perception of the wind load magnitude. The formula for calculating dynamic pressure characteristics is: In the formula, It is dynamic pressure.

[0031] Turbulence intensity is a key indicator for evaluating wind field fluctuations and structural fatigue loads. Through sliding window calculations, the turbulence intensity characteristic becomes a dynamic, time-varying feature, capable of capturing real-time changes in the wind field's turbulent state, which is crucial for predicting the dynamic response of structures. The turbulence intensity characteristic is obtained using sliding window statistics, and its calculation formula is as follows: In the formula, for Turbulence intensity at time t, The standard deviation of wind speed within the sliding window. The average wind speed within the sliding window is measured at a frequency of 10Hz. This is a small constant correction term.

[0032] These features can reveal the physical mechanisms of wind-structure interaction more deeply, providing key physical prior knowledge for subsequent machine learning models, thereby improving the accuracy, interpretability and generalization ability of the models.

[0033] S3. Preprocess the physical enhancement feature set to obtain the preprocessed feature vector.

[0034] After constructing the physical augmentation feature set, it needs to be preprocessed to form standardized input data suitable for subsequent model training. Preprocessing the physical augmentation feature set involves using the Z-Score standardization method to standardize the physical augmentation feature vectors, eliminating dimensional differences between different features.

[0035] Z-Score normalization transforms physical enhancement features with different physical meanings and dimensions into a purely mathematical preprocessed feature vector with a mean of 0 and a standard deviation of 1. This vector eliminates dimensional differences, making the data distribution more consistent with the input assumptions of typical machine learning algorithms, thus laying a reliable data foundation for subsequent model building.

[0036] S4. Input the original wind speed measurement value into the physical base model to obtain the base-calibrated wind speed.

[0037] The physical basis model is a monotonically increasing model in the form of a power-law function, used to characterize the global trend relationship between the original wind speed measurement and the basis-calibrated wind speed. The physical basis model is specifically expressed as follows: In the formula, For base calibration of wind speed, This is the original wind speed measurement value. The goal is to obtain model parameters by fitting the training data.

[0038] This step aims to eliminate or mitigate the global impact of systematic errors in the measuring device itself (such as instrument constant deviations, flow field distortions caused by installation location, etc.).

[0039] S5. Using the base calibration wind speed and preprocessed feature vector as inputs to the residual compensation model, the wind speed residual compensation amount is obtained.

[0040] The residual compensation model is an ensemble model based on the Bagging strategy. Its core idea is as follows: First, multiple distinct sub-training sets are generated from the training data using a bootstrap sampling method. Then, multiple homogeneous or heterogeneous base learners (e.g., decision trees, neural networks, etc.) are independently trained on these sub-training sets. Each base learner learns the mapping relationship from the input to the wind speed residual. In this embodiment, the experimental data acquisition frequency is 10Hz, with a moving average recorded every 5 seconds. The training dataset contains 5264 valid samples, covering wind speeds from 0 to 42 m / s. The Train subset contains 3950 samples, serving as the core source for model parameter learning; the Blind_Interp subset contains 800 samples, used to evaluate interpolation accuracy; and the Blind_Extrap subset contains 514 samples, used to evaluate extrapolation stability near the training boundary. An additional independent test set of 250 samples, with an upper wind speed limit extended to 47.4 m / s, is used to test the model's generalization ability under extreme strong wind conditions outside the distribution range.

[0041] For a new input, the predictions from all base learners are integrated using an adaptive weighting mechanism to obtain the final residual compensation. This effectively reduces the model's variance and improves its generalization ability and robustness.

[0042] The wind speed residual compensation is output by ensemble predictions from multiple base learners. The formula for calculating the wind speed residual compensation is as follows: In the formula, This is the wind speed residual compensation amount. The total number of base learners, For the first Adaptive weights for each base learner To preprocess the feature vectors, For base calibration of wind speed, For the first The residual compensation amount predicted independently by each base learner.

[0043] Among them, the The adaptive weights of the base learners are: In the formula, For weighted adjustment of hyperparameters, For the first Variance estimation of cognitive uncertainty for each base learner given an input. For the first Estimation of the variance of cognitive uncertainty of each base learner given an input.

[0044] S6. Obtain the final calibrated wind speed based on the base calibration wind speed and the wind speed residual compensation amount.

[0045] This approach combines global trend calibration with local fine-tuning compensation. The base-calibrated wind speed corrects for systematic biases, while the residual compensation further corrects for nonlinear and transient errors that the physical base model fails to describe. The sum of these two methods yields a final wind speed estimate that is optimized in both accuracy and dynamic response.

[0046] The final calibrated wind speed is: In the formula, For final wind speed calibration, For base calibration of wind speed, This is the wind speed residual compensation amount.

[0047] To assess the reliability of the final calibrated wind speed, this embodiment further constructs its prediction interval, which represents the possible range of true wind speed values ​​at a given confidence level (e.g., 95%).

[0048] Based on the final calibrated wind speed, the prediction interval for the final calibrated wind speed at a preset confidence level is estimated by combining the overall prediction standard deviation and the scaling factor calibrated from out-of-package samples. The prediction interval is: In the formula, For the prediction interval, For final wind speed calibration, The scaling factor is calibrated using out-of-package samples. This represents the composite total prediction standard deviation.

[0049] The prediction interval width is proportional to the total standard deviation, while the scaling factor controls the confidence level of the interval. It is not simply a quantile of the standard normal distribution (e.g., 1.96 corresponds to 95%), but rather post-calibrated using out-of-package samples to ensure the prediction interval reaches a nominal confidence level on unknown data. The scaling factor calibrated using out-of-package samples is: In the formula, It is a 0.95 quantile function. This is the true reference wind speed value for a single sample outside the package. For the predicted wind speed values ​​of the samples outside the package, This is the scale parameter.

[0050] If the model error perfectly follows a normal distribution It will be close to 1.96. If the model error distribution is wider or narrower, It will automatically zoom in or out, so that the final prediction interval can accurately cover 95% of the true value in the actual (outside the package) data.

[0051] The composite total prediction standard deviation is: In the formula, The standard deviation of cognitive uncertainty, The standard deviation of random uncertainty, Standard deviation For the first The residual compensation amount predicted independently by each base learner For a fixed offset coefficient, This is the proportionality coefficient.

[0052] The size of the prediction interval should be matched to the reliability of the final calibrated wind speed result. A wide prediction interval indicates that the model believes there is high uncertainty regarding the final calibrated wind speed under current conditions, and therefore the reliability of the result is relatively low. Users should use this single-point prediction value with caution and may need to refer to the upper and lower bounds of the interval for risk assessment, or make a comprehensive judgment in conjunction with other information.

[0053] A narrow prediction interval indicates that the model assumes low uncertainty regarding the final calibrated wind speed under current conditions, thus resulting in relatively high reliability. Users can place high confidence in the calibrated wind speed at this point, making it suitable for applications requiring high precision.

[0054] To verify the effectiveness of the method proposed in this embodiment, it was compared with other methods such as Power-law Only, Cubic Polynomial, SVR-RBF, RandomForest, and QuantileGBDT on independent test sets. Test metrics included root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²). 2 ), Predicted Interval Coverage (PICP) and Average Predicted Interval Width (MPIW).

[0055] The performance comparison results of each algorithm on the full range of the independent test set are shown in the table below.

[0056] Table 1. Performance comparison results of each algorithm on the full range of the independent test set. As shown in Table 1, the method proposed in this embodiment achieves an RMSE of 0.7256 m / s and a MAPE of 2.56% on the full-range independent test set, with overall accuracy superior to pure data-driven models such as SVR-RBF, RandomForest, and QuantileGBDT. Meanwhile, the constructed 95% prediction interval coverage reaches 95.2%, indicating that the method can not only improve the accuracy of point prediction but also has good probabilistic reliability.

[0057] The performance comparison between the interpolation region and the extrapolation region is shown in the table below.

[0058] Table 2. Performance Comparison Results of Interpolation and Extrapolation Regions As shown in Table 2, in the interpolation region with sufficient training sample coverage (≤40m / s), the RMSE of the method in this embodiment is only 0.1127m / s. 2 The value reaches 0.9999, indicating its ability to accurately fit the nonlinear deviation caused by structural turbulence in the device; in the extrapolation region (>40 m / s), the R value of SVR-RBF is similar to that of RandomForest. 2 The values ​​dropped to -6.8279 and -4.6959 respectively, showing significant failure, while the method in this embodiment still maintained relatively stable extrapolation performance (RMSE of 2.5368 m / s, R...). 2 The value of -0.3154 indicates that the physical basis constraint significantly improves the robustness under unknown strong wind conditions.

[0059] Example 2 This embodiment provides a machine learning-based system for calibrating wind speed readings on power meteorological devices, such as... Figure 2 As shown, it includes: The data acquisition module is used to acquire raw wind speed measurements, as well as environmental parameters acquired synchronously with the raw wind speed measurements. The physical enhancement feature set construction module is used to calculate air physical property parameters based on environmental parameters, and to construct a physical enhancement feature set based on air physical property parameters and original wind speed measurements. The preprocessing module is used to preprocess the physical enhancement feature set to obtain a preprocessed feature vector; The base calibration wind speed acquisition module is used to input the original wind speed measurement value into the physical base model to obtain the base calibration wind speed; The wind speed residual compensation acquisition module is used to take the base calibration wind speed and the preprocessed feature vector as inputs to the residual compensation model to obtain the wind speed residual compensation amount. The final calibration wind speed acquisition module is used to obtain the final calibration wind speed based on the base calibration wind speed and the wind speed residual compensation amount.

[0060] The rest is the same as in Example 1.

[0061] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for calibrating wind speed readings on power meteorological devices based on machine learning, characterized in that, Includes the following steps: Obtain the original wind speed measurement value, as well as the environmental parameters collected synchronously with the original wind speed measurement value; Air physical property parameters are calculated based on the environmental parameters, and a set of physical enhancement features is constructed based on the air physical property parameters and the original wind speed measurement values. The physical enhancement feature set is preprocessed to obtain a preprocessed feature vector; The original wind speed measurement value is input into the physical base model to obtain the base-calibrated wind speed; The base-calibrated wind speed and the preprocessed feature vector are used as inputs to the residual compensation model to obtain the wind speed residual compensation amount. The final calibrated wind speed is obtained based on the base calibration wind speed and the wind speed residual compensation amount.

2. The method for calibrating wind speed readings of power meteorological devices based on machine learning according to claim 1, characterized in that, The environmental parameters include ambient temperature and ambient air pressure, and the air physical properties include air density and air dynamic viscosity. The formula for calculating air density is: In the formula, air density, For ambient air pressure, air is the gas constant. Ambient temperature; The formula for calculating the aerodynamic viscosity is: In the formula, Aerodynamic viscosity, For reference dynamic viscosity, For reference temperature, is Sutherland's constant.

3. The method for calibrating wind speed readings of power meteorological devices based on machine learning according to claim 1, characterized in that, The set of physical enhancement features includes Reynolds number features, dynamic pressure features, and turbulence intensity features. The formula for calculating the Reynolds number characteristic is as follows: In the formula, The Reynolds number is... air density, This is the original wind speed measurement value. The diameter of the support column for the power meteorological device. Aerodynamic viscosity, This is a low-wind-speed numerical stability disturbance term; The formula for calculating the dynamic pressure characteristic is: In the formula, For dynamic pressure; The turbulence intensity characteristics are obtained using sliding window statistics, and the calculation formula is as follows: In the formula, for Turbulence intensity at time t, The standard deviation of wind speed within the sliding window. This represents the average wind speed within the sliding window. This is a small constant correction term.

4. The method for calibrating wind speed readings of power meteorological devices based on machine learning according to claim 1, characterized in that, The physical enhancement feature set is preprocessed by using the Z-Score normalization method to standardize the physical enhancement feature vectors to eliminate the dimensional differences between different features.

5. The method for calibrating wind speed readings of power meteorological devices based on machine learning according to claim 1, characterized in that, The physical basis model is a monotonically increasing model in the form of a power-law function, used to characterize the global trend relationship between the original wind speed measurement and the basis-calibrated wind speed. Specifically, the physical basis model is expressed as follows: In the formula, For base calibration of wind speed, This is the original wind speed measurement value. The goal is to obtain model parameters by fitting the training data.

6. The method for calibrating wind speed readings of power meteorological devices based on machine learning according to claim 1, characterized in that, The residual compensation model is an ensemble model based on the Bagging strategy. It outputs the wind speed residual compensation amount by integrating the predictions of multiple base learners. The calculation formula for the wind speed residual compensation amount is as follows: In the formula, This is the wind speed residual compensation amount. The total number of base learners, For the first Adaptive weights for each base learner To preprocess the feature vectors, For base calibration of wind speed, For the first The residual compensation amount predicted independently by each base learner Among them, the first The adaptive weights of the base learners are: In the formula, For weighted adjustment of hyperparameters, For the first Variance estimation of cognitive uncertainty for each base learner given an input. For the first Estimation of the variance of cognitive uncertainty of each base learner given an input.

7. The method for calibrating wind speed readings of power meteorological devices based on machine learning according to claim 1, characterized in that, The final calibrated wind speed is: In the formula, For final wind speed calibration, For base calibration of wind speed, This is the wind speed residual compensation amount.

8. The method for calibrating wind speed readings of power meteorological devices based on machine learning according to claim 1, characterized in that, Based on the final calibrated wind speed, the prediction interval of the final calibrated wind speed at a preset confidence level is estimated by combining the overall prediction standard deviation and the scaling factor calibrated from out-of-package samples. The prediction interval is: In the formula, For the prediction interval, For final wind speed calibration, The scaling factor is calibrated using out-of-package samples. This represents the composite total forecast standard deviation; The scaling factor calibrated from the out-of-package sample is: In the formula, It is a 0.95 quantile function. This is the true reference wind speed value for a single sample outside the package. For the predicted wind speed values ​​of the samples outside the package, For scale parameters; The standard deviation of the composite overall prediction is: In the formula, The standard deviation of cognitive uncertainty, The standard deviation of random uncertainty, Standard deviation For the first The residual compensation amount predicted independently by each base learner For a fixed offset coefficient, This is the proportionality coefficient.

9. The method for calibrating wind speed readings of power meteorological devices based on machine learning according to claim 8, characterized in that, The size of the prediction interval is matched to the reliability of the final calibrated wind speed result.

10. A machine learning-based calibration system for wind speed indications in power meteorological devices, characterized in that, include: The data acquisition module is used to acquire the raw wind speed measurement value, as well as environmental parameters acquired synchronously with the raw wind speed measurement value; The physical enhancement feature set construction module is used to calculate air physical property parameters based on the environmental parameters, and construct a physical enhancement feature set based on the air physical property parameters and the original wind speed measurement value; The preprocessing module is used to preprocess the physical enhancement feature set to obtain a preprocessed feature vector; The base calibration wind speed acquisition module is used to input the original wind speed measurement value into the physical base model to obtain the base calibration wind speed. The wind speed residual compensation quantity acquisition module is used to take the base calibration wind speed and the preprocessed feature vector as inputs to the residual compensation model to obtain the wind speed residual compensation quantity. The final calibration wind speed acquisition module is used to obtain the final calibration wind speed based on the base calibration wind speed and the wind speed residual compensation amount.

Images