Ultra-short-term wind power prediction method and device for extreme weather scene and medium

Abstract

The invention discloses an ultra-short-term wind power prediction method and device for an extreme weather scene and a medium, and relates to the technical field of wind power prediction. According to the method, abnormal value identification is carried out by adopting a progressive fusion mode of coarse screening based on physical constraints and fine screening based on statistical analysis, a multi-strategy processing space is constructed, and weight is adaptively determined by utilizing Bayesian optimization to carry out weighted fusion abnormal value correction; an extreme weather sample is further expanded by adopting a TimeGAN model, weather scene division is firstly carried out in combination with a corrected sample, and then a Transform model and a CNN model are correspondingly trained; and in a prediction stage, adaptively selecting a corresponding trained model for prediction. According to the method, the generation reasons of different abnormal values are more comprehensively considered, a targeted identification and processing method is adopted, and the ultra-short-term wind power prediction of different weather scenes is carried out by adopting a multi-combination model, so that the prediction accuracy is improved.

Description

Technical Field

[0001] This invention relates to the field of wind power prediction technology, and in particular to a method, device and medium for ultra-short-term wind power prediction under extreme weather scenarios. Background Technology

[0002] According to data recently released by the National Energy Administration of China, as of the end of November 2025, the cumulative installed power generation capacity nationwide reached 3.79 billion kilowatts, a year-on-year increase of 17.1%. Among them, wind power installed capacity exceeded 600 million kilowatts, a year-on-year increase of 22.4%. As a clean energy source, wind power output is significantly affected by weather conditions, especially wind speed and direction, exhibiting randomness and volatility. This randomness and volatility of wind power generation pose a severe challenge to power grid dispatching. High-precision wind power forecasting can provide reliable information for the power grid dispatching side, enabling reasonable scheduling of power generation, improving wind power absorption capacity, and ensuring the safe and stable operation of the power system.

[0003] Current research uses wind power datasets to construct and analyze prediction models. However, this method does not fully consider the drastic fluctuations in wind power output caused by extreme weather, which leads to significant nonlinearity and irregularity in weather changes, resulting in numerous outliers in the dataset and further complicating wind power prediction. Therefore, to improve the accuracy of wind power prediction, it is essential to accurately identify and effectively process the large number of outliers occurring in extreme weather scenarios to ensure that the prediction model maintains high accuracy and reliability under complex meteorological conditions. Summary of the Invention

[0004] The purpose of this invention is to provide a method, device and medium for ultra-short-term wind power prediction under extreme weather scenarios, aiming to solve or improve at least one of the above-mentioned technical problems.

[0005] To achieve the above objectives, the present invention provides the following solution: A method for ultra-short-term wind power prediction under extreme weather scenarios includes: An outlier identification method that progressively integrates physical constraint-based coarse screening and statistical analysis-based fine screening is used to identify outliers in the basic training samples. A processing space containing multiple correction strategies is constructed, and the fusion weight of each strategy is determined by adaptive optimization based on the Bayesian optimization framework. The identified outliers are then corrected by weighted fusion. The TimeGAN model is used to augment extreme weather data with small samples. The augmented samples are then combined with the corrected samples. The combined sample data is first divided into weather scenarios and then assigned to CNN and Transformer models for training. The TimeGAN model consists of three parts: a generator, a discriminator, and a supervisor. The samples generated by the generator are passed to the discriminator and compete against real samples in a Minimax game. The discriminator is trained to distinguish between real and generated samples, while the generator is trained to generate samples that are close to the real distribution. Determine the basic data of the wind farm to be measured; the basic data of the wind farm includes the total installed capacity, rated capacity, wind speed at hub height, wind direction at hub height, atmospheric temperature and atmospheric pressure numerical weather forecast; Based on the basic data of the wind farm, NWP data features are extracted to divide the weather scenarios, and the corresponding trained prediction model is adaptively selected to predict the ultra-short-term wind power. The weather scenarios include normal weather and extreme weather.

[0006] Optionally, the physical constraint-based coarse screening process specifically includes: Six categories of physical constraint outlier determination rules are defined, as follows: in, v Indicates wind speed. p Indicates power, p max Indicates the maximum power generation capacity. v cut-in Indicates the cut-in wind speed. v cut-out Indicates the cut-out wind speed; If a training sample does not meet any of the rule requirements, it is judged as an outlier.

[0007] Optionally, the statistical analysis-based fine screening process includes: A polynomial extended feature of the wind speed sequence is constructed from the coarsely screened data. The mapping relationship between wind speed and power is fitted using a gradient boosting regression model. The residual between the actual power and the predicted power is calculated. The normal value range is determined based on the interquartile range of the residual. Data with residuals exceeding the range are judged as outliers. The predicted power is determined by the fitted mapping relationship between wind speed and power.

[0008] Optionally, the various correction strategies include: linear interpolation correction, local mean correction, and weighted fusion correction based on gradient boosting regression models.

[0009] Optionally, the Bayesian optimization framework includes: A two-layer optimization problem is constructed to determine the optimal weight vector. For the inner layer optimization, outliers are fused and corrected based on the current weights. For the outer layer optimization, the goal is to minimize the root mean square error of the prediction model on the validation set. A surrogate model of weights and errors is established through Gaussian process regression, and the weights are iteratively updated based on the expectation improvement criterion.

[0010] Optionally, the step of extracting NWP data features based on the wind farm's basic data to segment weather scenarios and adaptively selecting the corresponding trained prediction model for ultra-short-term wind power prediction specifically includes: CEEMDAN was used to decompose and denoise the original power data in the corrected data. Based on the denoised data, NWP data features are extracted, and the current weather scenario is determined according to the NWP data features. If it is normal weather, a trained CNN model is used for ultra-short-term wind power prediction. If it is extreme weather, a trained Transformer model is used for ultra-short-term wind power prediction.

[0011] The present invention also provides an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform the ultra-short-term wind power prediction method for extreme weather scenarios as described above.

[0012] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the ultra-short-term wind power prediction method for extreme weather scenarios as described above.

[0013] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: This invention discloses a method, device, and medium for ultra-short-term wind power prediction under extreme weather scenarios. The method identifies outliers by progressively fusing physical constraint-based coarse screening with statistical analysis-based fine screening, and constructs a multi-strategy processing space. Bayesian optimization is used to adaptively determine weights for weighted fusion outlier correction. Furthermore, the TimeGAN model is used to expand the extreme weather samples. Weather scenarios are first divided based on the corrected samples, and then Transformer and CNN models are trained accordingly. In the prediction phase, the corresponding trained model is adaptively selected for prediction. This invention comprehensively considers the different causes of outliers, adopts targeted identification and processing methods, and uses multiple combined models for ultra-short-term wind power prediction under different weather scenarios, thereby improving prediction accuracy. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a schematic diagram of the overall process of the ultra-short-term wind power prediction method for extreme weather scenarios according to the present invention. Figure 2 This is a diagram illustrating the effect of progressive outlier identification and processing in this embodiment; Figure 3 This is a graph showing the PCA evaluation effect with a small sample size in this embodiment; Figure 4 This is a graph showing the effect of t-SNE evaluation with a small sample size in this embodiment; Figure 5 This is a diagram showing the noise reduction effect of CEEMDAN in this embodiment; Figure 6 This is a comparison chart of the predictions made by the model used in this embodiment and common models; Figure 7 This is a comparison chart of ablation experiment predictions using the method employed in this embodiment. Detailed Implementation

[0016] 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.

[0017] The purpose of this invention is to provide a method, device and medium for ultra-short-term wind power prediction under extreme weather scenarios, aiming to solve or improve at least one of the above-mentioned technical problems.

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] like Figure 1 As shown, this invention provides an ultra-short-term wind power prediction method for extreme weather scenarios, including: S1. An outlier identification method that progressively integrates physical constraint-based coarse screening and statistical analysis-based fine screening is used to identify outliers in the basic training samples.

[0020] As a specific implementation method, the progressive outlier identification method based on physical constraints and statistical analysis in this step specifically includes the following processing steps: ① Use physical constraints to coarsely screen outliers: The operating characteristics of wind turbine generators need to strictly adhere to certain physical constraints, and samples that violate any physical constraint rule are marked as outliers. Six categories of physical constraint outlier determination rules are defined for this purpose, as shown in formula (1):

[0021] in, v Indicates wind speed. p Indicates power, p max Indicates the maximum power generation capacity. v cut-in Indicates the cut-in wind speed. v cut-out This indicates the cut-out wind speed.

[0022] ② Use statistical analysis methods to screen outliers: A polynomial extended feature of the wind speed sequence is constructed from the coarsely screened data. A gradient boosting regression model is used to fit the mapping relationship between wind speed and power. The residual between the actual power and the predicted power is calculated. Based on the interquartile range of the residual, the normal value interval is determined, and data with residuals exceeding the interval are identified as outliers. Specifically: Outlier screening was achieved using a statistical analysis method combining GBR fitting data regularity and residual IQR threshold analysis. First, feature construction was performed. The wind speed sequence was extended using a cubic polynomial feature expansion according to formula (2) to adapt the nonlinear mapping relationship between wind speed and power, ultimately forming an extended feature matrix. As shown in formula (3):

[0023] in, Indicates the first i The extended feature vector of each sample, Indicates the first i Wind speed values ​​for each training sample. , They represent the first i The quadratic and cubic terms of wind speed for each training sample.

[0024] Next, an initialization model is established, and the error is corrected by iteratively training weak learners. All weak learners are accumulated, and finally, a GBR model that can fit the wind speed-power mapping relationship is obtained as shown in formula (4):

[0025] in, Indicates the processM The final gradient boosting regression model after rounds of iterations, This indicates the initialization of the model. Indicates the learning rate. M Indicates the total number of iterations. Indicates the first m Step size of the wheel weak learner Indicates the first m A weak learner trained in rounds.

[0026] Finally, the residual *r* between the actual power and the GBR predicted power is calculated, and the interquartile range (IQR) of the residual is calculated. The range of normal values ​​is determined by the IQR, and samples whose residuals do not meet the range are judged as outliers. The calculation of the IQR and the determination of the normal value range are shown in formulas (5) and (6), respectively:

[0027] Where Q1 is the 25th percentile of the residuals and Q3 is the 75th percentile of the residuals.

[0028] S2. Construct a processing space containing multiple correction strategies, and adaptively optimize based on the Bayesian optimization framework to determine the fusion weight of each strategy, and perform weighted fusion correction on the identified outliers.

[0029] As a specific implementation method, this step constructs a multi-strategy processing space and a Bayesian optimization framework, dynamically selects the optimal outlier correction strategy, and restores the reasonable values ​​of outliers to the maximum extent while preserving effective temporal characteristics.

[0030] Let the set of outliers be... The corresponding original power sequence position index is Define the candidate processing strategy set as follows: The corresponding methods are linear interpolation correction based on linear interpolation of adjacent normal points to replace outliers, local mean correction based on the mean of normal points within the sliding window to replace outliers, and prediction correction based on the GBR model for outliers, as shown in formulas (7), (8), and (9):

[0031] in, This represents the characteristic of the rate of change of wind speed.

[0032] Define the strategy selection vector ,in Representation strategy The fusion weights satisfy The final corrected value for outliers is a weighted fusion of the values ​​from all strategies:

[0033] The optimal weight vector is determined by constructing an inner and outer two-layer optimization problem. For inner layer optimization: based on the current weights Outliers are fused and corrected to obtain a cleaned dataset. For outer layer optimization: with the goal of minimizing the RMSE of the prediction model on the validation set, the optimal weights are searched through Bayesian optimization.

[0034] in This is a lightweight proxy prediction model trained on corrected data, used for rapid evaluation of processing performance.

[0035] The optimization process uses Gaussian process regression to establish a surrogate model between the weight vector and RMSE, and iteratively updates the weight parameters based on the expected improvement criterion.

[0036] in, To achieve the current optimal RMSE, For weight The corresponding validation set error. This method achieves adaptive fusion of outlier handling strategies, automatically adjusting the contribution weight of the correction strategy based on the contextual features of different outliers. This avoids the limitations of a single processing method and ensures that the quality of the processed data is maximized through a data-driven optimization process, which is beneficial to downstream prediction tasks. This significantly improves the scientific and systematic nature of outlier handling.

[0037] S3. The TimeGAN model is used to augment extreme weather data with small samples. The augmented samples are combined with the corrected samples. The combined sample data is first divided into weather scenarios and then assigned to CNN and Transformer models for training. The TimeGAN model consists of three parts: generator, discriminator and supervisor. The generated samples are passed to the discriminator to compete against real samples in a Minimax game. The discriminator is trained to distinguish between real samples and generated samples, while the generator is trained to generate samples that are close to the real distribution.

[0038] As a specific implementation method, this step is used to build a scene recognition module in the multi-combination model ultra-short-term wind power prediction method based on weather scene identification, which can determine the weather scene corresponding to the input wind speed in real time. Extreme weather scenes are typical small-sample events, and the small number of sample points will lead to insufficient model learning rate, which will affect the prediction accuracy. TimeGAN is used to expand the extreme weather samples to small samples. The TimeGAN model consists of three parts: generator (G), discriminator (D) and supervisor (S). The generated samples of the generator will be passed through the discriminator to compete with real samples through Minimax game, and the discriminator will be trained to distinguish between real samples and generated samples. At the same time, the generator will be trained to generate samples that are close to the real distribution. The mathematical expression of Minimax game is shown in formula (13):

[0039] in Represents generator G Attempting to minimize the value function Discriminator D Attempting to maximize the antagonistic relationship of the value function, Represents the expectation of real data. This represents the expectation of generating data. This indicates that the discriminator is effective against real latent samples. x The probability of judgment, This indicates that the generator will generate noise. z The mapping yields potential samples.

[0040] After expanding the small sample of extreme wind speeds, the ultra-short-term wind power prediction method based on the multi-combination model of weather scene identification calls the corresponding prediction model through the built-in scene recognition module. Its prediction logic is defined by a piecewise function, as shown in formula (14):

[0041] S4. Determine the basic data of the wind farm to be measured; the basic data of the wind farm includes the total installed capacity, rated capacity, wind speed at hub height, wind direction at hub height, atmospheric temperature, and numerical weather forecast of atmospheric pressure.

[0042] S5. Based on the basic data of the wind farm, extract NWP data features to divide the weather scenarios, and adaptively select the corresponding trained prediction model to predict the ultra-short-term wind power; the weather scenarios include normal weather and extreme weather.

[0043] As a specific implementation method, this method also includes simulation calculation and error analysis.

[0044] First, determine the simulation inputs: Analyze the measured data of the wind farm to determine the total installed capacity of the wind farm; the rated capacity of the wind farm; the wind speed at hub height; the wind direction, atmospheric temperature, and atmospheric pressure at hub height; and the numerical weather forecast; the data sampling interval is 15 minutes; and obtain the real-time prediction results of the wind power of the entire farm according to the above steps.

[0045] Then, error analysis is performed.

[0046] set up for i The actual average power over the period for i Predicted average power for the time period The total time period for daily assessment Let be the operating capacity of the wind farm. Then, the normalized root mean square error is defined as equation (15):

[0047] The normalized mean absolute error is defined as Equation (16):

[0048] The coefficient of determination is defined as Equation (17):

[0049] Following step 4), input the simulation input quantity, and perform error calculation between the predicted power calculated by the model and the measured power using the error evaluation criteria (15), (16) and (17) in step 5) to obtain the prediction accuracy.

[0050] Based on the above technical solution, the following embodiments are provided.

[0051] In this embodiment, the main steps include: 1) Constructing a progressive outlier identification method that integrates physical constraint coarse screening and statistical learning fine screening. First, explicit outlier data is removed through physical constraints, and then implicit outlier data is captured through statistical learning, so as to achieve comprehensive and accurate removal of outliers; 2) Adaptive optimization and fusion of outlier processing methods. By constructing a multi-strategy processing space and a Bayesian optimization framework, the optimal outlier correction strategy is dynamically selected; 3) TimeGAN small sample expansion is used to solve the problem of insufficient learning degree of small sample models for extreme wind speeds; 4) CEEMDAN is used to denoise the original power, reducing the impact of noise on the power prediction accuracy while retaining the effective temporal features of power; 5) Key NWP data features such as wind speed and wind direction are extracted to classify weather scenarios; 6) Transformer and CNN models are adaptively selected based on the differences in characteristics of different weather scenarios, which can ensure the prediction accuracy of wind power and maximize the prediction efficiency of wind power.

[0052] To maximize the accuracy of wind power prediction, this embodiment primarily selects multiple combined models to perform ultra-short-term predictions of wind power on the test set. The actual operational data from the "Renewable Energy Generation Forecasting Competition" held by the State Grid Corporation of China, specifically the data from wind farm site 1 (2019-01-01 00:00:00—2020-12-31 23:45:00), was used, with a sampling interval of 15 minutes. The dataset includes parameters such as hub height, wind speed, and power, and is divided into training and test sets in an 8:2 ratio.

[0053] Based on the above dataset, a more specific processing procedure is provided: Step 1: Progressive Outlier Identification Combining Physical Constraint Coarse Screening and Statistical Learning Fine Screening. This progressive outlier identification method, combining physical constraint coarse screening and statistical learning fine screening, first uses physical constraints based on the actual operating characteristics of the wind turbine for coarse screening, and then employs GBR fitting data pattern-residual IQR threshold analysis for statistical learning fine screening. This method accurately identifies outliers, significantly reducing the number of outliers and improving the wind speed-power curve fitting accuracy. The outlier identification performance evaluation is shown in Table 1.

[0054] Table 1 Outlier Identification Results

[0055] Step 2: Adaptive Optimization and Outlier Processing Integration. A multi-strategy processing space is constructed, defining three outlier correction strategies: linear interpolation, local mean, and GBR prediction. The outlier correction values ​​are then integrated using a weighted weight vector. Bayesian optimization is then used to determine the optimal weights for both inner and outer layers, ultimately achieving adaptive optimization and outlier processing integration. Through outlier identification in Step 1 and outlier processing in Step 2, the wind speed-power curve is significantly improved, as shown below. Figure 2 As shown.

[0056] Step 3: TimeGAN Extreme Weather Small Sample Augmentation Method. TimeGAN is used to augment extreme weather data with small samples, addressing the issue of insufficient model learning in extreme weather conditions. After TimeGAN small sample augmentation, the proportion of extreme wind speed samples increased from 5.32% to 9.32%. Two-component dimensionality reduction is then performed using PCA (Principal Component Analysis) and t-SNE (T-Distributed Random Nearest Neighbor Embedding), and the data distribution is evaluated using a visual scatter plot to verify the authenticity of the synthesized samples. Figure 3 and Figure 4The visualization evaluation results show that the synthetic data is highly correlated with the real data, thus verifying the effectiveness of the TimeGAN extreme weather small sample augmentation method.

[0057] Step 4: CEEMDAN decomposes and denoises the raw power data. CEEMDAN decomposes the raw power data into four components of different frequencies, IMF1-IMF4. The highest-order IMF4 is high-frequency noise. After filtering out IMF4, the remaining three components are reconstructed to minimize the impact of noise on power prediction accuracy while preserving the original sequence trend characteristics, thereby improving the reliability of wind power prediction. The raw power signal is decomposed into four different frequencies using CEEMDAN, and the highest-frequency noise is filtered out before reconstructing the remaining components. Figure 5 It can be clearly seen that the power signal after denoising is smoother. The variance of the power signal before and after denoising is reduced from 240.3377 to 216.4324, and the signal-to-noise ratio is increased to 12.6462, with a higher proportion of effective signal.

[0058] Step 5: Classify weather scenarios based on NWP data characteristics. In actual wind farm operation, various weather scenarios exist. This embodiment extracts key NWP data features such as wind speed and direction, classifying weather scenarios into two main categories: normal weather and abnormal weather, and assigning corresponding labels. This facilitates the use of different prediction models for ultra-short-term wind power prediction in Step 6.

[0059] Step 6: Adaptively select a model for prediction based on the differences in characteristics of different weather scenarios. Different weather scenarios exhibit significant differences in characteristics. For extreme weather scenarios, it is crucial to consider the nonlinearity and irregularity of weather changes, which make it difficult to capture temporal relationships. Therefore, models capable of capturing complex temporal relationships should be selected for extreme weather scenarios. For normal weather scenarios, the temporal relationships are simpler, requiring a balance between prediction accuracy and efficiency. This embodiment compares three models: Model 1 uses a Transformer model for extreme weather scenarios, relying on a self-attention mechanism to capture long-range complex dependencies, while a CNN model is used for normal weather scenarios, leveraging local feature extraction capabilities and parallel computing advantages to improve prediction accuracy and efficiency. Model 2 uses a single Transformer model for ultra-short-term wind power prediction regardless of weather scenario. Model 3 uses a single CNN model for ultra-short-term wind power prediction regardless of weather scenario. The comparison of these three models verifies the effectiveness of the ultra-short-term wind power prediction scheme adopted in this embodiment, which adaptively selects a model based on the differences in characteristics of different weather scenarios.

[0060] According to specific experiments, the prediction model 1 used in this embodiment outperforms prediction models 2 and 3. The accuracy of prediction model 1 is 88.94%, with a mean absolute error of 0.1106 and a correlation coefficient of 0.7362, representing a 2.69% improvement in accuracy compared to prediction model 2 and a 0.58% improvement compared to prediction model 3. Ablation experiments were conducted to further verify the reliability of the method in this embodiment. The method was compared with methods without CEEMDAN for original power sequence denoising and without the progressive outlier handling proposed in this embodiment. Specific experimental results show that without CEEMDAN for original power denoising, the wind power prediction accuracy drops to 85.29%, and the RMSE increases to 0.1471; without outlier handling, the wind power prediction accuracy drops to 86.03%, and the RMSE increases to 0.1397. Based on the above comparative experiments and ablation experiments, and as shown above… Figure 6 and Figure 7 As shown in the comparison of predicted waveforms, the reliability of the ultra-short-term wind power prediction method for progressive identification and processing of outliers in extreme weather scenarios in this embodiment is demonstrated. The evaluation of prediction performance of different models is shown in Table 2, and the comparative evaluation of ablation experiments is shown in Table 3.

[0061] Table 2 Evaluation of Prediction Performance of Different Models

[0062] Table 3 Comparative Evaluation of Ablation Experiments

[0063] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0064] This embodiment uses specific examples to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only for the purpose of helping to understand the core ideas of the present invention; at the same time, for those skilled in the art, there will be changes in specific implementation methods and application scope based on the ideas of the present invention. In summary, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for ultra-short-term wind power prediction under extreme weather scenarios, characterized in that, include: An outlier identification method that progressively integrates physical constraint-based coarse screening and statistical analysis-based fine screening is used to identify outliers in the basic training samples. A processing space containing multiple correction strategies is constructed, and the fusion weight of each strategy is determined by adaptive optimization based on the Bayesian optimization framework. The identified outliers are then corrected by weighted fusion. The TimeGAN model is used to augment extreme weather data with small samples. The augmented samples are then combined with the corrected samples. The combined sample data is first divided into weather scenarios and then assigned to CNN and Transformer models for training. The TimeGAN model consists of three parts: a generator, a discriminator, and a supervisor. The samples generated by the generator are passed to the discriminator and compete against real samples in a Minimax game. The discriminator is trained to distinguish between real and generated samples, while the generator is trained to generate samples that are close to the real distribution. Determine the basic data of the wind farm to be measured; the basic data of the wind farm includes the total installed capacity, rated capacity, wind speed at hub height, wind direction at hub height, atmospheric temperature and atmospheric pressure numerical weather forecast; Based on the basic data of the wind farm, NWP data features are extracted to divide the weather scenarios, and the corresponding trained prediction model is adaptively selected to predict the ultra-short-term wind power. The weather scenarios include normal weather and extreme weather.

2. The ultra-short-term wind power prediction method for extreme weather scenarios according to claim 1, characterized in that, The physical constraint-based coarse screening process specifically includes: Six categories of physical constraint outlier determination rules are defined, as follows: in, v Indicates wind speed. p Indicates power, p max Indicates the maximum power generation capacity. v cut-in Indicates the cut-in wind speed. v cut-out Indicates the cut-out wind speed; If a training sample does not meet any of the rule requirements, it is judged as an outlier.

3. The ultra-short-term wind power prediction method for extreme weather scenarios according to claim 1, characterized in that, The specific process of the statistical analysis-based fine screening includes: A polynomial extended feature of the wind speed sequence is constructed from the coarsely screened data. The mapping relationship between wind speed and power is fitted using a gradient boosting regression model. The residual between the actual power and the predicted power is calculated. The normal value range is determined based on the interquartile range of the residual. Data with residuals exceeding the range are judged as outliers. The predicted power is determined by the fitted mapping relationship between wind speed and power.

4. The ultra-short-term wind power prediction method for extreme weather scenarios according to claim 1, characterized in that, The various correction strategies include: linear interpolation correction, local mean correction, and weighted fusion correction based on gradient boosting regression models.

5. The ultra-short-term wind power prediction method for extreme weather scenarios according to claim 1, characterized in that, The Bayesian optimization framework includes: A two-layer optimization problem is constructed to determine the optimal weight vector. For the inner layer optimization, outliers are fused and corrected based on the current weights. For the outer layer optimization, the goal is to minimize the root mean square error of the prediction model on the validation set. A surrogate model of weights and errors is established through Gaussian process regression, and the weights are iteratively updated based on the expectation improvement criterion.

6. The ultra-short-term wind power prediction method for extreme weather scenarios according to claim 1, characterized in that, The step of extracting NWP data features based on the wind farm's basic data to segment weather scenarios and adaptively selecting the corresponding trained prediction model for ultra-short-term wind power prediction specifically includes: CEEMDAN was used to decompose and denoise the original power data in the corrected data. Based on the denoised data, NWP data features are extracted, and the current weather scenario is determined according to the NWP data features. If it is normal weather, a trained CNN model is used for ultra-short-term wind power prediction. If it is extreme weather, a trained Transformer model is used for ultra-short-term wind power prediction.

7. An electronic device, characterized in that, The device includes a memory and a processor, the memory being used to store a computer program, and the processor running the computer program to enable the electronic device to perform the ultra-short-term wind power prediction method for extreme weather scenarios according to any one of claims 1-6.

8. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the ultra-short-term wind power prediction method for extreme weather scenarios as described in any one of claims 1-6.

Images