A new energy power prediction sample construction and model training method

Abstract

The present application relates to the field of new energy power prediction technology, in particular to a new energy power prediction sample construction and model training method. By synchronously collecting historical actual power data, measured meteorological data and historical numerical weather prediction data, measured meteorological sample set and NWP sample set are constructed respectively; two kinds of sample sets are merged and source identification features are added, and a hybrid prediction model is trained; a hierarchical model is established based on the Bayesian framework, the systematic bias and random error distribution of NWP data are analyzed, diversified virtual NWP samples are generated, and a data enhancement training set is constructed; an enhanced prediction model is trained using the enhanced training set, and it is weighted integrated with the benchmark model, NWP model and hybrid model, and the final prediction model is output. Through the innovative sample construction and training mechanism, the present application effectively solves the problem of inconsistent distribution of training and reasoning data, and significantly improves the accuracy and robustness of power prediction.

Description

Technical Field

[0001] This invention relates to the field of new energy power generation prediction technology, and more specifically, to a method for constructing new energy power prediction samples and training models. Background Technology

[0002] With the large-scale grid connection of new energy sources such as wind power and photovoltaics, the intermittency and volatility of their power output pose challenges to the safe operation of the power grid. High-precision power forecasting is crucial for ensuring the grid's absorption capacity, optimizing dispatch, and participating in market transactions. Existing technologies mainly rely on numerical weather prediction (NWP) data as model input, making predictions by establishing a nonlinear mapping relationship between NWP and power. However, NWP itself has unavoidable forecasting errors, leading to a distribution mismatch between the "ideal" measured meteorological data used in the training phase and the "biased" NWP data used in the inference phase. This results in problems such as decreased model generalization performance and large fluctuations in prediction results.

[0003] Existing methods typically use a single data source (such as only measured meteorological data or only NWP data) for model training, lacking a systematic modeling and compensation mechanism for NWP errors, resulting in limited prediction accuracy, especially under complex meteorological conditions. Therefore, a method for constructing new energy power prediction samples and training models is designed.

[0004] The existing technology has the following technical defects, specifically:

[0005] 1. Existing methods use "perfect" measured meteorological data to train models, but use "imperfect" NWP data for inference. This mismatch in the distribution of input data between the training and inference stages is the root cause of the model's unstable performance in practical applications.

[0006] 2. Existing methods passively accept NWP errors, and the model is very sensitive to disturbances in the input data. This leads to drastic fluctuations or even serious deviations in the prediction results when the accuracy of NWP decreases (such as in complex weather processes).

[0007] 3. Existing methods are mostly simple paradigms based on a single model and a single data source, lacking a systematic strategy to address the NWP error problem, and there is an upper limit to performance improvement. Summary of the Invention

[0008] The purpose of this invention is to provide a method for constructing new energy power prediction samples and training models to solve the problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention aims to provide a method for constructing new energy power prediction samples and training models, including: data collection and sample construction steps: synchronously collecting historical actual power data of the target new energy power station, historical measured meteorological data that is spatiotemporally matched with the power data, and historical numerical weather forecast data of corresponding spatial grid points in the same historical period; constructing a measured meteorological sample set based on the measured meteorological data and the corresponding actual power; and constructing an NWP sample set based on historical NWP data and the actual power of the corresponding time period.

[0010] Hybrid model training steps: merge the measured meteorological sample set and the NWP sample set, add source identification features to each sample, construct a hybrid training sample set, and train the hybrid prediction model based on the sample set.

[0011] Bayesian data augmentation steps: Establish a hierarchical model based on the Bayesian framework, analyze the systematic deviation and random error distribution between historical NWP data and measured meteorological data, generate diverse virtual NWP samples, and merge virtual samples with original samples to construct a data augmentation training set.

[0012] Enhanced model training and integration steps: Train the enhanced prediction model using the data enhancement training set, and then perform a weighted integration of the enhanced prediction model with the baseline model and the NWP model trained based on the measured meteorological sample set and the NWP sample set to output the final prediction model.

[0013] As a further improvement to this technical solution, the specific implementation method for constructing a measured meteorological sample set based on measured meteorological data and corresponding actual power is as follows: each historical measured meteorological data is used as an input feature vector, and the historical actual power data under the corresponding timestamp of the meteorological data is extracted as an output label, and the two are combined to form a measured meteorological data sample set. ,in, Indicates the first Measured meteorological feature vectors Indicates the first Output labels for historical actual power data corresponding to each measured meteorological reading. This represents the total number of samples in the measured meteorological data sample set. This indicates the number of the measured meteorological feature vector.

[0014] As a further improvement to this technical solution, the method for constructing an NWP sample set based on historical NWP data and the actual power of the corresponding time period is as follows: actual power data corresponding to the forecast lead time in the historical NWP data is selected; each historical NWP data is used as an input feature vector; and the synchronous historical actual power data within its forecast period is matched as the output label, thus forming an NWP data training sample set. ,in, Indicates the first NWP feature vectors, Indicates the first The NWP feature vector is the output label of the historical actual power data synchronized within the forecast period. This represents the total number of samples in the NWP data training sample set. This indicates the number of the NWP feature vector.

[0015] As a further improvement to this technical solution, the construction of the hybrid training sample set is specifically implemented by: processing the measured meteorological sample set... With NWP sample set Data format standardization is performed, and samples with missing timestamps or data anomalies are removed. The two sample sets are then directly merged, and a one-dimensional source identifier feature column is added to each merged sample, indicating its origin. The sample with this feature column assigned a value of 0, which comes from The feature column of the sample is assigned a value of 1, thus forming a mixed training sample set. .

[0016] As a further improvement to this technical solution, the specific implementation method for training the hybrid prediction model is as follows:

[0017] Mix training sample sets The dataset is divided into training and validation subsets according to a preset ratio. A gradient boosting tree is used to build the model. The root mean square error and mean absolute error are used as performance evaluation indicators. An early stopping strategy is used to prevent overfitting. The training is iteratively completed until the performance of the validation subset converges.

[0018] As a further improvement to this technical solution, the specific implementation method for analyzing the systematic deviation and random error distribution between historical NWP data and measured meteorological data is as follows: for each core meteorological element, calculate the deviation between historical NWP data and synchronous measured meteorological data, and statistically analyze the mean, standard deviation and distribution characteristics of the deviation.

[0019] As a further improvement to this technical solution, the specific method for generating diverse virtual NWP samples is as follows: Step 1: Establish a Bayesian hierarchical model and define the statistical relationship between NWP data and measured meteorological data.

[0020] Step 2: Based on the data pairs matching historical NWP and measured meteorological data, derive the joint posterior distribution of systematic bias and random error in the model.

[0021] Step 3: For the target measured meteorological data, multiple random NWP simulation data are generated by sampling from the posterior distribution and substituting them into the model.

[0022] As a further improvement to this technical solution, the specific method for constructing the data augmentation training set is as follows: Multiple generated virtual NWP samples are combined with their corresponding actual power labels to form a virtual NWP sample set. After removing abnormal virtual samples that exceed the physically reasonable range, Compared with the original measured meteorological sample set Original NWP sample set The virtual NWP samples are merged, and the source identifier feature column is assigned a value of 1, thus forming the data augmentation training set.

[0023] As a further improvement to this technical solution, the enhanced model training and ensemble are specifically implemented as follows: select a validation set independent of the training data, and initialize the enhanced prediction model. Benchmark Model NWP model Hybrid Model The weights are determined with the goal of maximizing the prediction accuracy of the validation set. By optimizing the weights of each model, the total weights are constrained to be 1 and the value of each individual weight is within the range of 0 to 1. The optimal weights of each model are then obtained. Based on the optimal weights of each model, the prediction results of each model are weighted and summed to obtain the final prediction result of new energy power.

[0024] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0025] 1. It innovatively introduces a training mode based on historical NWP data and integrates the mapping relationship between the measured data and NWP data through hybrid training. This allows the model to become familiar with the characteristics and patterns of NWP data during training and face the input distribution it has "seen" during actual prediction, thereby greatly alleviating the distribution mismatch problem and improving the model's ability to smoothly transition from training to application and its generalization performance.

[0026] 2. A pioneering NWP data augmentation method based on a Bayesian framework is proposed. This method does not avoid NWP errors but actively analyzes their statistical regularities and generates a large number of "virtual" NWP samples simulating real-world biases for training. This is equivalent to providing the model with a "resistant training" environment, allowing it to encounter and adapt to various possible NWP error scenarios during the learning process. Therefore, the trained model no longer provides a fragile "one-to-one" mapping to NWP inputs but a robust "many-to-one" mapping, thus maintaining the stability and reliability of prediction results when faced with NWPs that contain errors in the real world.

[0027] 3. This invention is not a single improvement, but rather a complete technical system encompassing comparative analysis, hybrid training, advanced data augmentation, and ultimately model integration. This systematic approach ensures that we can fully leverage the advantages of different data sources (real-world testing vs. NWP) and different training strategies (standard training vs. augmented training). Ultimately, through integration, the wisdom of each sub-model is combined, resulting in a prediction system that not only surpasses traditional methods in accuracy but also achieves a qualitative leap in stability and robustness, providing a more reliable decision-making basis for power grid dispatch. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art 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.

[0029] Figure 1 This is a schematic diagram of the implementation steps of the method of the present invention. Detailed Implementation

[0030] 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.

[0031] Example: Please refer to Figure 1 As shown, a method for constructing new energy power prediction samples and training models is provided, including: data collection and sample construction steps: synchronously collecting historical actual power data of the target new energy power station, historical measured meteorological data that is spatiotemporally matched with the power data, and historical numerical weather forecast data of corresponding spatial grid points in the same historical period; constructing a measured meteorological sample set based on the measured meteorological data and the corresponding actual power; and constructing an NWP sample set based on historical NWP data and the actual power of the corresponding time period.

[0032] Collect high-resolution (typically at 15-minute or 1-hour intervals) actual power output data from the target wind farm or photovoltaic power plant for the same historical period. This data serves as the ground truth (label) benchmark for all model training. Collect historical measured meteorological data, such as wind speed, located at the site and fully synchronized with the power data in time. ,wind direction Irradiance ,temperature Measured meteorological data can typically be obtained from environmental monitoring instruments, irradiance meters (photovoltaic stations), and wind measurement equipment (wind farms) installed within renewable energy power plants. This data represents the closest approximation of actual local meteorological conditions. Historical NWP data, collected from the same historical period as the power data and corresponding spatial grid locations, is also included. The key point is that this collection focuses on historically published NWP data that forecasts future periods. For example, it collects meteorological element data forecasted daily from 00:00 last year for the next 24-48 hours, aligning it with the actual power and timestamps of the measured meteorological data for that period. The NWP field must cover the measured meteorological element field information and be able to correspond one-to-one. This forms the basis for evaluating historical NWP bias and conducting subsequent training.

[0033] In one specific embodiment, the method for constructing a measured meteorological sample set based on measured meteorological data and corresponding actual power is as follows: Each historical measured meteorological data point is used as an input feature vector; the historical actual power data under the corresponding timestamp of that meteorological data point is extracted as an output label; and these are combined to form a measured meteorological data sample set. ,in, Indicates the first Measured meteorological feature vectors Indicates the first Output labels for historical actual power data corresponding to each measured meteorological reading. This represents the total number of samples in the measured meteorological data sample set. This indicates the number of the measured meteorological feature vector.

[0034] In one specific embodiment, the method for constructing an NWP sample set based on historical NWP data and the actual power of the corresponding time period is as follows: actual power data corresponding to the forecast lead time in the historical NWP data is selected; each historical NWP data is used as an input feature vector; and the synchronous historical actual power data within its forecast period is matched as an output label, thus forming an NWP data training sample set. ,in, Indicates the first NWP feature vectors, Indicates the first The NWP feature vector is the output label of the historical actual power data synchronized within the forecast period. This represents the total number of samples in the NWP data training sample set. This indicates the number of the NWP feature vector.

[0035] Hybrid model training steps: merge the measured meteorological sample set and the NWP sample set, add source identification features to each sample, construct a hybrid training sample set, and train the hybrid prediction model based on the sample set.

[0036] In one specific embodiment, the construction of the hybrid training sample set is specifically implemented by: processing the measured meteorological sample set... With NWP sample set Data format standardization is performed, and samples with missing timestamps or data anomalies are removed. The two sample sets are then directly merged, and a one-dimensional source identifier feature column is added to each merged sample, indicating its origin. The sample with this feature column assigned a value of 0, which comes from The feature column of the sample is assigned a value of 1, thus forming a mixed training sample set. = .

[0037] Mixed sampling enables the model to learn power mapping patterns under both "real weather conditions" and "forecast weather conditions" simultaneously. By using source identification features, the model can distinguish input types and adaptively adjust its mapping strategy.

[0038] In one specific embodiment, the training of the hybrid prediction model is specifically implemented by: using a hybrid training sample set... The dataset is divided into training and validation subsets according to a preset ratio. A gradient boosting tree is used to build the model. The root mean square error and mean absolute error are used as performance evaluation indicators. An early stopping strategy is used to prevent overfitting. The training is iteratively completed until the performance of the validation subset converges.

[0039] The preset ratio is usually set based on the data scale and model validation requirements, commonly using a 7:3 or 8:2 division. It can be flexibly adjusted according to the actual amount of data and training stability to ensure that the training subset is sufficient to support model learning and the validation subset can effectively evaluate generalization performance. The gradient boosting tree and early stopping strategy mentioned are existing technologies and will not be described in detail here.

[0040] The model is trained using a mixed sample set. During training, the source identifier feature is used as an additional input feature. During actual prediction, since the input is always NWP data, this identifier feature is fixed at 1 (representing NWP input). By fusing the two types of information, this model achieves more stable and accurate predictive performance than the pure NWP model (Mnwp) because it internally learns the relationship between the two data distributions and is insensitive to NWP bias.

[0041] Bayesian data augmentation steps: Establish a hierarchical model based on the Bayesian framework, analyze the systematic deviation and random error distribution between historical NWP data and measured meteorological data, generate diverse virtual NWP samples, and merge virtual samples with original samples to construct a data augmentation training set.

[0042] In one specific embodiment, the analysis of the systematic deviation and random error distribution between historical NWP data and measured meteorological data is specifically implemented by: for each core meteorological element, calculating the deviation between historical NWP data and synchronous measured meteorological data, and statistically analyzing the mean, standard deviation, and distribution characteristics of the deviation.

[0043] For example, regarding wind speed: = and statistics The distribution characteristics. For example, most commonly, it is considered that... Follows a normal distribution , and The mean and standard deviation of the error distribution can be statistically derived from observations of historical data. This step prepares for advanced data augmentation by precisely characterizing the statistical regularity of NWP errors.

[0044] In one specific embodiment, the method for generating diverse virtual NWP samples is as follows: Step 1: Establish a Bayesian hierarchical model and define the statistical relationship between NWP data and measured meteorological data.

[0045] Define the statistical relationship between NWP forecast meteorological element x and actual observed element y as follows: .in, The systematic bias of the NWP model is an unknown parameter to be estimated. This represents random error, assumed to have a mean of zero and a variance of . The normal distribution, i.e. ( .

[0046] Based on this, given hour, The likelihood function (i.e., the data distribution) is: .in, The mean is variance is It follows a normal distribution.

[0047] For Bayesian inference, the parameters are unknown. and Define the prior distribution:

[0048] deviation Priors:

[0049] variance Priors: Where IG represents the Inverse-Gamma distribution, and the hyperparameters are... These hyperparameters can be initialized based on statistical analysis of historical error data, or optimized and determined through model selection techniques such as grid search and cross-validation, for example, by setting them based on prior meteorological knowledge (e.g., assuming average deviation). and give it greater uncertainty. ).

[0050] Step 2: Based on the data pairs matching historical NWP and measured meteorological data, derive the joint posterior distribution of systematic bias and random error in the model.

[0051] Given a set of historical matching data pairs Calculate parameters using Bayes' theorem and Joint posterior distribution: .in, It is the likelihood function.

[0052] Given the chosen conjugate prior distribution, the posterior distribution has an analytical solution form:

[0053] deviation The marginal posterior distribution is still a normal distribution:

[0054] variance The marginal posterior distribution remains an inverse-Gamma distribution: In the formula, the parameter Historical data The parameters are obtained from the analytical calculation of prior hyperparameters.

[0055] Step 3: For the target measured meteorological data, multiple random NWP simulation data are generated by sampling from the posterior distribution and substituting them into the model.

[0056] For a new real wind speed observation value The goal is to generate One possible NWP simulation value This is achieved by sampling from the posterior prediction distribution. The posterior prediction distribution is obtained through the following steps:

[0057] (1) From the learned posterior distribution A set of parameters was sampled in the middle. .

[0058] (2) According to the model ,in A simulated value was obtained through calculation. .

[0059] (3) Repeat the above process. Next, get A random simulation of data.

[0060] The invention will be further illustrated below with a simplified numerical example:

[0061] (1) Data preparation: Collect the wind speed NWP forecast values ​​of a certain station over the past several days. and corresponding actual observation values (Unit: m / s), a total of 30 pairs of data.

[0062] (2) Model specification: The prior hyperparameters are set as follows: .

[0063] (3) Posterior calculation: 30 pairs of data Substituting into the analytical formula, the posterior parameters are calculated. .

[0064] (4) Stochastic simulation: Assuming today's observations .

[0065] First, from Sampling yields a .

[0066] Next, from Sampling yields a .

[0067] Then, from Sample a random error .

[0068] Finally, calculate the first simulated value: .

[0069] (5) Repeat step (4) 1000 times to obtain 1000 random NWP simulation data around 5.0 m / s. These data as a whole reflect the possible biases and uncertainties of the NWP model under this weather condition.

[0070] This is a key innovative step. The core idea is to decompose the difference between the predicted and actual NWP values ​​into two parts: systematic bias and random error. Then, using Bayes' theorem, the probability distributions of these two components are learned from historical data. After learning, for any new actual observation (or baseline value), multiple corresponding possible simulated NWP values ​​can be generated by sampling from the learned posterior distribution. By sampling from the distribution, K virtual samples with random perturbations can be generated for each historical NWP sample. This method systematically expands the diversity of NWP training samples, especially covering various possible error scenarios, allowing the model to "experience" various possible input biases during the training phase, greatly improving robustness.

[0071] Enhanced model training and integration steps: Train the enhanced prediction model using the data enhancement training set, and then perform a weighted integration of the enhanced prediction model with the baseline model and the NWP model trained based on the measured meteorological sample set and the NWP sample set to output the final prediction model.

[0072] In one specific embodiment, the method for constructing the data augmentation training set is as follows: multiple generated virtual NWP samples are combined with their corresponding actual power labels to form a virtual NWP sample set. After removing abnormal virtual samples that exceed the physically reasonable range, Compared with the original measured meteorological sample set Original NWP sample set By merging the data and assigning a value of 1 to the source identifier feature column of the virtual NWP samples, a data augmentation training set is finally formed. .

[0073] In one specific embodiment, the enhanced model training and ensemble is implemented by: selecting a validation set independent of the training data and initializing the enhanced prediction model. Benchmark Model NWP model Hybrid Model The weights are determined with the goal of maximizing the prediction accuracy of the validation set. By optimizing the weights of each model, the total weights are constrained to be 1 and the value of each individual weight is within the range of 0 to 1. The optimal weights of each model are then obtained. Based on the optimal weights of each model, the prediction results of each model are weighted and summed to obtain the final prediction result of new energy power.

[0074] The training of the enhanced prediction model, the baseline model, the NWP model, and the hybrid model all follow the following general process: the corresponding training set is divided into a training subset and a validation subset according to a preset ratio; the model is constructed using gradient boosting tree or artificial neural network algorithms; the root mean square error (RMSE) and the mean absolute error (MAE) are used as performance evaluation indicators; an early stopping strategy is used to prevent overfitting; and the training is iteratively continued until the performance of the validation subset converges.

[0075] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A method for constructing samples and training a model for predicting new energy power, characterized in that, include: Data collection and sample construction steps: Simultaneously collect historical actual power data of the target new energy power station, historical measured meteorological data that is spatiotemporally matched with the historical actual power data, and historical numerical weather forecast data of corresponding spatial grid points in the same historical period. Construct a measured meteorological sample set based on the historical measured meteorological data and the corresponding historical actual power, and construct a numerical weather forecast sample set based on the historical numerical weather forecast data and the historical actual power of the corresponding time period. Hybrid model training steps: merge the measured meteorological sample set and the numerical weather prediction sample set, add source identification features to each sample, construct a hybrid training sample set, and train the hybrid prediction model based on the hybrid training sample set; Bayesian data augmentation steps: Based on the Bayesian framework, a hierarchical model is established to analyze the systematic deviation and random error distribution between historical numerical weather forecast data and historical measured meteorological data, generate diversified virtual numerical weather forecast samples, and fuse the virtual numerical weather forecast samples with the measured meteorological sample set and the numerical weather forecast sample set to construct a data augmentation training set. Enhanced model training and integration steps: Train the enhanced prediction model using the data enhancement training set, and then perform a weighted integration of the enhanced prediction model, the baseline model trained based on the measured meteorological sample set, the numerical weather prediction model trained based on the numerical weather prediction sample set, and the hybrid prediction model to output the final prediction model.

2. The method for constructing new energy power prediction samples and training models according to claim 1, characterized in that, The specific implementation method for constructing the measured meteorological sample set based on the historical measured meteorological data and the corresponding historical actual power is as follows: Each historical measured meteorological data point is used as the input feature vector, and the historical actual power data corresponding to the timestamp of that historical measured meteorological data point is extracted as the output label. These are combined to form a sample set of measured meteorological data. ,in, Indicates the first Historical measured meteorological feature vectors, Indicates the first Output labels for historical actual power data corresponding to each historical measured meteorological data point. This represents the total number of samples in the historical measured meteorological data sample set. This indicates the number of the historical measured meteorological feature vector.

3. The method for constructing new energy power prediction samples and training models according to claim 2, characterized in that, The specific implementation method for constructing a numerical weather forecast sample set based on the historical numerical weather forecast data and the corresponding historical actual power for the same period is as follows: The actual power data corresponding to the forecast lead time in historical numerical weather prediction data are selected. Each historical numerical weather prediction data is used as an input feature vector, and the synchronous historical actual power data within its forecast period is matched as the output label. This combination forms a training sample set of numerical weather prediction data. ,in, Indicates the first Historical numerical weather forecast feature vectors Indicates the first Each historical numerical weather prediction feature vector is output with labels from the historical actual power data synchronized within the forecast period. This represents the total number of samples in the training sample set of historical numerical weather prediction data. This indicates the number of the historical numerical weather forecast feature vector.

4. The method for constructing new energy power prediction samples and training models according to claim 3, characterized in that, The specific method for constructing the hybrid training sample set is as follows: For the measured meteorological sample set Numerical Weather Prediction Sample Set Data format standardization is performed, and samples with missing timestamps or data anomalies are removed. The two sample sets are then directly merged, and a one-dimensional source identifier feature column is added to each merged sample, indicating its origin. The sample with this feature column assigned a value of 0, which comes from The feature column of the sample is assigned a value of 1, thus forming a mixed training sample set. .

5. The method for constructing new energy power prediction samples and training models according to claim 4, characterized in that, The specific implementation method of training the hybrid prediction model is as follows: Mix training sample sets The dataset is divided into training and validation subsets according to a preset ratio. A gradient boosting tree is used to build the model. The root mean square error and mean absolute error are used as performance evaluation indicators. An early stopping strategy is used to prevent overfitting. The training is iteratively completed until the performance of the validation subset converges.

6. The method for constructing new energy power prediction samples and training models according to claim 5, characterized in that, The analysis of the systematic deviations and random error distributions between historical numerical weather forecast data and historical measured meteorological data is specifically implemented as follows: For each core meteorological element, the deviation between historical numerical weather forecast data and synchronous historical measured meteorological data is calculated, and the mean, standard deviation, and distribution characteristics of the deviation are statistically analyzed.

7. The method for constructing new energy power prediction samples and training models according to claim 6, characterized in that, The specific method for generating diverse virtual numerical weather forecast samples is as follows: Step 1: Establish a Bayesian hierarchical model and define the statistical relationship between historical numerical weather forecast data and historical measured meteorological data; Step 2: Based on the data pairs matched with historical numerical weather forecast data and historical measured meteorological data, derive the joint posterior distribution of systematic bias and random error in the Bayesian hierarchy; Step 3: For the target measured meteorological data, multiple random numerical weather prediction simulation data are generated by sampling from the posterior distribution and substituting it into the Bayesian hierarchical model.

8. The method for constructing new energy power prediction samples and training models according to claim 7, characterized in that, The specific implementation method for constructing the data augmentation training set is as follows: The generated virtual numerical weather forecast samples are combined with their corresponding historical actual power labels to form a virtual numerical weather forecast sample set. After removing abnormal virtual samples that exceed the physically reasonable range, the virtual numerical weather forecast sample set is... With measured meteorological sample set Numerical weather forecast sample set The data are merged, and the source identifier feature column of the virtual numerical weather forecast samples is assigned a value of 1, thus forming the data augmentation training set.

9. The method for constructing new energy power prediction samples and training models according to claim 8, characterized in that, The specific implementation method for the enhanced model training and ensemble is as follows: Select a validation set independent of the training data to initialize the augmented prediction model. Benchmark Model Numerical weather prediction models Hybrid Model The weights are determined with the goal of maximizing the prediction accuracy of the validation set. By optimizing the weights of each model, the total weights are constrained to be 1 and the value of each individual weight is within the range of 0 to 1. The optimal weights of each model are then obtained. Based on the optimal weights of each model, the prediction results of each model are weighted and summed to obtain the final prediction result of new energy power.

Images