Wind power long-time prediction output correction method and system based on covariance weight updating

Abstract

The invention discloses a wind power long-time prediction output correction method and system based on covariance weight updating, and the method comprises the steps: carrying out the collection of historical meteorological data and power data corresponding to different wind power plants, carrying out the preprocessing of the collected data, and determining a training set and a test set based on the preprocessed data; establishing a wind power long-time prediction output correction model, training the model by using the training set, and evaluating the model by using the test set; and obtaining a final prediction result of the long-time output of the typical wind power plant through the wind power long-time prediction output correction model. The wind speed random fluctuation characteristics are described through the joint data and physical driving method, the inter-station space-time correlation of the wind power plants is fully integrated, the long-time output sequence of each wind power plant is further corrected based on the Bayesian optimization covariance matrix, and the overall prediction accuracy and tendency performance of the wind power plants are effectively improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of renewable energy power generation, and in particular relates to a method and system for correcting long-term wind power output forecast based on covariance weight update. Background Art

[0002] Wind power generation has gradually increased its share in the energy structure due to its clean and renewable characteristics. Affected by the inherent characteristics of wind power such as volatility and intermittency, large-scale wind power access and large load demand have significant differences in source-load power characteristics over long time scales, which may lead to insufficient power supply capacity of the system. Under the uncertainty constraints that the power system can bear, it is urgent to further improve the accuracy of long-term wind power output forecasts and reduce the impact of supply guarantee in complex scenarios through dispatch auxiliary strategies.

[0003] According to the different modeling objects, wind power output prediction models can be divided into direct prediction and indirect prediction. The direct wind power output prediction method is based on historical meteorological data and output data for modeling. It usually does not consider the physical process of wind speed change and it is difficult to make full use of meteorological information. The indirect wind power output prediction method uses the nonlinear conversion relationship between wind speed and wind power output. The modeling process is highly interpretable and suitable for long-term wind power output prediction. However, due to the direct influence of wind speed prediction accuracy, traditional wind speed prediction models are mainly based on time series analysis or machine learning methods to capture the time-varying characteristics of wind speed. The spatiotemporal distribution characteristics of wind speed are complex and difficult to accurately describe by relying on a single model. In addition, wind power output is not only affected by meteorological factors, but also by the influence of spatial distribution effects. Different wind farms are coupled with complex spatiotemporal correlations. Geographically adjacent wind farms may show a high output similarity. Deep models are often used to mine the distribution laws of historical data, which improves the spatial consistency and local accuracy of the prediction to a certain extent. However, the correction effect of dynamic changes in correlation between wind farms on long-term predicted output is not considered, which affects the trend performance of wind power output.

[0004] Chinese patent CN114996964A discloses a wind power prediction method, device, equipment and storage medium, which focuses on prediction through data filling and conversion, combined with wind speed Weibull distribution and random difference equation, but rejection sampling and model generation require high computing resources and time costs. Chinese patent CN118017474A discloses a wind power prediction method that considers the spatiotemporal correlation of wind farms and the distribution characteristics of prediction errors. It adopts a hybrid integrated model and an error compensation mechanism, comprehensively considering the spatiotemporal correlation and the distribution characteristics of prediction errors, but the hybrid integrated model used involves multi-model parameter settings and weight calculations, which are prone to error accumulation, and the computational complexity and data volume requirements are high. Summary of the invention

[0005] Aiming at the problems that existing wind power prediction methods are difficult to deeply mine the time-varying characteristics of wind speed relying solely on a single model and ignore the correction effect of the dynamic change of the correlation between wind farms on the long-term predicted output, the present invention provides a method and system for correcting the long-term predicted output of wind power based on covariance weight update, which indirectly obtains and corrects the long-term predicted output of wind power through the combination of data and physical driving technology and the covariance weight dynamic update strategy, thereby effectively improving the prediction accuracy and trend performance.

[0006] The technical solution of the present invention is: A method for correcting the long-term predicted output of wind power based on covariance weight update, including:

[0007] Collect historical meteorological data and power data corresponding to different wind farms, and preprocess the collected data, and determine the training set and test set based on the preprocessed data;

[0008] Establish a model for correcting the long-term predicted output of wind power, train the model using the training set, and evaluate the model using the test set;

[0009] Obtain the final predicted result of the long-term output of a typical wind farm through the model for correcting the long-term predicted output of wind power.

[0010] The preprocessing of the collected data includes data cleaning and feature optimization. Among them, data cleaning is used to remove outliers and fill missing values; feature optimization is carried out using the Pearson correlation coefficient method;

[0011] The preprocessed data is divided after time scale alignment into a training set and a test set.

[0012] When constructing the model for correcting the long-term predicted output of wind power, it includes:

[0013] First, improve the loss function structure of the PINNs network, and then obtain the wind speed time series of each wind farm;

[0014] Secondly, based on the scatter plot of wind speed-wind power output characteristics, express the function mapping relationship between wind speed and wind power output of the wind farm on a long time scale, and obtain the initial predicted result of the long-term output of each wind farm based on the polynomial fitting regression model;

[0015] Then divide the typical wind farms based on the K-means clustering algorithm to obtain the initial data set of the long-term predicted output of the typical wind farms;

[0016] Subsequently, based on Bayes' theorem, the predicted output is used as a new observation condition to dynamically update the historical covariance matrix;

[0017] Finally, based on the magnitude of the covariance weights, the importance of influencing the long-term predicted output of a typical wind farm is characterized, and the initial predicted result of the long-term output of the typical wind farm is corrected by covariance weighting to obtain a corrected model for the long-term predicted output of wind power updated based on covariance weights.

[0018] A PINNs network that incorporates the Weibull distribution constraint is constructed to predict the wind speed time series. After the improvement of its loss function structure, it consists of a data loss, a physical loss, and an additional statistical loss function, specifically as follows:

[0019] L total = λ d L data + λ p L physics + λ w L weibull

[0020] In the formula: L total represents the loss function, L data , L physics , L weibull represent the data loss, the physical loss, and the statistical loss respectively, and λ d , λ p , λ w are the importance hyperparameters of the error terms.

[0021] The polynomial fitting regression model is as follows:

[0022] y = α0 + α1v + α2v 2 + α3v 3

[0023] In the formula: α0, α1, α2, α3 represent the polynomial coefficients; v represents the wind speed; y represents the wind power output time series.

[0024] The K-means clustering algorithm adopted updates the k clustering centers through the distance-based clustering idea until the preset convergence condition is met, and a clustering result reflecting the output characteristics of the wind farm is obtained;

[0025] The output time series of the typical wind farm is preset as different random variables, and the covariance matrix ∑ is used to measure the linear correlation degree between different random variables. Its representation form is as follows:

[0026]

[0027] In the formula: cov(y i , y j j) represents the covariance of y i and y j j, i is from 1 to n, j is from 1 to n, and n represents the number of wind farms.

[0028] Updating the historical covariance matrix based on Bayes' theorem, including:

[0029] Determining the prior distribution P(Σ) of the covariance matrix based on the historical output data of a typical wind farm;

[0030] Taking the polynomial fitting regression result as the evidence of the update model, and calculating the probability distribution P(y|Σ) of observing new data under the condition of a given covariance matrix;

[0031] Using the prior distribution P(Σ) and the probability distribution P(y|Σ), updating the estimation of the covariance matrix through Bayes' theorem, and calculating the posterior distribution P(Σ|y) ∝ P(y|Σ)P(Σ);

[0032] Obtaining the corrected covariance matrix Σ through the calculated posterior probability P(Σ|y) new , reflecting the change in the correlation between wind farms under new observation conditions.

[0033] Using covariance weighting to correct the long-term output sequence of wind farms, and the specific calculation process is as follows:

[0034]

[0035] In the formula: w ij represents the normalized influence weight of the output y of wind farm j j on the output y of wind farm i i ; α represents the adjustment coefficient; represents the corrected predicted output sequence of wind farm i.

[0036] A long-term wind power prediction output correction system updated based on covariance weights, including:

[0037] A preprocessing module, which is used to collect the historical meteorological data and power data corresponding to different wind farms, preprocess the collected data, and determine the training set and the test set based on the preprocessed data;

[0038] A correction module, which is used to establish a long-term wind power prediction output correction model, train the model using the training set, and evaluate the model using the test set;

[0039] A prediction module, which is used to obtain the final long-term output prediction result of a typical wind farm through the long-term wind power prediction output correction model.

[0040] In the operation of the present invention, a wind speed time series is generated by physical and data-driven technical means, and the initial predicted output of a typical wind farm is obtained by combining indirect prediction methods. The spatio-temporal correlation between wind farms is fully integrated to affect the prediction performance. Based on Bayes' theorem, the historical covariance matrix is iteratively optimized. Based on the spatial dependence relationship between typical wind farms, a covariance weight update strategy is proposed to correct the initial predicted output of typical wind farms, thereby enhancing the trend synergy between wind farm output sequences. The corrected long-term power generation prediction results can not only solve the power supply guarantee problem of the new power system in complex scenarios by combining supply-demand assessment and auxiliary decision-making technologies, but also be applied to the long-term planning and operation of the power system, which is of great significance for improving the safe and stable operation of the power grid. Brief Description of the Drawings

[0041] Figure 1 is the flowchart of the method of the present invention,

[0042] Figure 2 is the local comparison curve graph of the predicted output and the actual output of a typical wind farm in the embodiment of the present invention. Detailed Embodiment

[0043] The present invention will be further described below in conjunction with specific examples and drawings.

[0044] As Figure 1 shown, the present invention provides a method for correcting the long-term predicted output of wind power based on covariance weight update,

[0045] Collect the historical meteorological data and power data corresponding to different wind farms, and preprocess the collected data, and determine the training set and the test set based on the preprocessed data;

[0046] Establish a model for correcting the long-term predicted output of wind power, train the model using the training set, and evaluate the model using the test set;

[0047] Obtain the final predicted result of the long-term output of a typical wind farm through the model for correcting the long-term predicted output of wind power.

[0048] In the data preprocessing link of the present invention, first, conventional data cleaning is performed on the historical meteorological data and power data corresponding to different wind farms, mainly including removing outliers and filling missing values; secondly, the Pearson correlation coefficient method (Pearson) is used for feature selection to quantitatively analyze the key meteorological factors affecting wind power output; finally, the cleaned data set is divided into a training set and a test set after being aligned in time scale (i.e., time standard).

[0049] In the model construction stage, the present invention first improves the loss function structure of the PINNs network by adding statistical loss to enhance the mining of the statistical distribution characteristics of wind speed, and then obtains the wind speed time series of each wind farm. Secondly, based on the scatter plot of wind speed-wind power output characteristics, the function mapping relationship between wind speed and wind power output at the long time scale of the wind farm is deeply characterized and expressed, and the initial long-term output prediction results of each wind farm are obtained based on the constructed polynomial fitting regression model. Then, considering the influence of the spatial distribution effect of the wind farm, the typical wind farms are divided based on the K-means clustering algorithm to obtain the initial data set of the long-term predicted output of the typical wind farms. Subsequently, based on Bayes' theorem, the predicted output is used as a new observation condition to dynamically update the historical covariance matrix, thereby reflecting the change in the correlation between typical wind farms. Finally, based on the magnitude of the covariance weight to characterize the importance of influencing the long-term predicted output of the typical wind farms, the initial predicted results of the long-term output of the typical wind farms are corrected by covariance weighting to obtain a corrected model for the long-term prediction of wind power output updated based on covariance weights.

[0050] In the long-term output prediction stage of the present invention, first, the initial long-term output prediction results of each wind farm are indirectly obtained based on the reconstructed wind speed and the non-linear relationship between wind speed and wind power output. Then, the typical wind farm output scenario set is divided through the clustering algorithm, and the covariance weight is dynamically updated through Bayes' theorem to characterize the change in the correlation between typical wind farms. Finally, the final long-term output prediction results of the typical wind farms are obtained through the constructed corrected model for the long-term prediction of wind power output updated based on covariance weights.

[0051] The present invention deeply analyzes the probability distribution characteristics of wind speed, and considers the change in the correlation between wind farms. The long-term output of wind farms is optimized through the covariance weight update strategy. It mainly includes combining data-driven and physics-driven technologies, using PINNs with embedded Weibull distribution constraints to statistically model the wind speed probability distribution, constructing a polynomial fitting regression model based on the deep function mapping relationship between wind speed and wind power output, using the K-means clustering algorithm and covariance matrix to characterize the output correlation between typical wind farms, and using Bayes' theorem to iteratively optimize the covariance matrix, thereby enhancing the accuracy and trend synergy of wind power prediction output through covariance weighting correction.

[0052] The main prediction technical routes in the present invention include:

[0053] (1) Wind speed is the key meteorological element affecting wind power output. Use a PINNs network with embedded Weibull distribution constraints (i.e., Physics-Informed Neural Networks) to statistically model the wind speed probability distribution to generate a wind speed time series that conforms to physical and statistical distribution characteristics;

[0054] (2) Based on the historical data of the wind farm, establish the deep function mapping relationship between wind speed and wind power output, construct a polynomial fitting regression model, and indirectly obtain the long-term predicted wind power output.

[0055] (3) Use the K-means clustering method to conduct a clustering analysis on the output characteristics of the wind farm, and divide the wind farms with similar output characteristics into the same subset to facilitate capturing the correlation between typical wind farms.

[0056] (4) Considering the time-varying characteristics of the output correlation between typical wind farms, iteratively optimize the historical covariance matrix based on the Bayesian model to capture the change in the correlation between wind farms under new observation conditions.

[0057] (5) Considering the output trend synergy and spatial dependence between wind farms, naturally express and process the influence degree of the output between typical wind farms based on the covariance weight update strategy, so as to use covariance weighting to correct the long-term output sequence of the wind farm.

[0058] The present invention takes into account that the output correlation between different wind farms is dynamically changing. In particular, the recent output correlation may have a greater impact than the historical output correlation. By updating the historical covariance matrix, the change in the correlation between current wind farms can be better reflected. In this way, using the correlation for correction will improve the prediction accuracy and trend synergy.

[0059] Specifically, in (1), deeply analyze the wind speed probability distribution, construct a wind speed prediction model that conforms to the physical and statistical distribution characteristics, and improve the accuracy of the wind speed. The method is as follows:

[0060] Wind speed is the key meteorological element affecting the wind power output level. The Weibull distribution is often used to describe the frequency characteristics of wind speed. Combine data-driven and physics-driven technologies to construct a PINNs network that embeds the Weibull distribution constraint to predict the wind speed time series. After the loss function structure is improved, it is composed of data loss, physical loss, and an additional statistical loss function. Train the model until the loss function L total is minimized, and further constrain the prediction result to conform to the differential change law of the meteorological system on the short time scale, and at the same time satisfy the physical law and statistical distribution characteristics. The loss function L total The formula is:

[0061] L total = λ d L data + λ p L physics + λ w L weibull

[0062]

[0063] Lweibull = D KL (P pred ‖P weibull )

[0064] Where: L data , L physics , L weibull represent data loss, physical loss, and statistical loss respectively; v i and represent the actual value and predicted value of the wind speed at the i-th point respectively; R[·] represents the physical residual function; P pred is the wind speed distribution predicted by the model, and P weibull is the Weibull distribution known based on historical wind speed data; D KL (·) represents the KL divergence; n d represents the number of actual observation points; n p represents the number of physical residual sampling points. λ d , λ p , λ w are importance hyperparameters of the error term, flexibly balancing the goals of data fitting, physical constraints, and statistical distribution.

[0065] (2) Based on the deep functional mapping relationship between wind speed and wind power output, a polynomial fitting regression model is constructed. The method is as follows: According to the physical and data-driven technical means, the wind speed time series is predicted, and the functional mapping relationship is established by analyzing the data distribution of the corresponding wind speed and wind power output of the wind farm, so that the predicted wind speed can be indirectly converted into the numerical value of wind power prediction output. The output level of the wind turbine shows a non-linear and positive correlation relationship with the wind speed range. In the stage of changing from the cut-in wind speed to the rated wind speed, the fitting curve of the wind speed and wind power output is close to a cubic polynomial expression, as shown in the following formula:

[0066] y = α0 + α1v + α2v 2 + α3v 3

[0067] Where: α0, α1, α2, α3 represent polynomial coefficients; v represents the wind speed; y represents the wind power output time series.

[0068] Therefore, establishing a polynomial fitting regression model can characterize the wind power output characteristics.

[0069] Based on the scatter plot of the wind speed-wind power output characteristics, the wind power distribution can be clearly observed, and the wind power output can be divided into four stages accordingly:

[0070] The first segment is from 0 to the cut-in wind speed, and the wind power in this segment is 0; the second segment is from the cut-in wind speed to the rated wind speed, and the wind power is positively correlated with the wind speed in this stage; the third segment is from the rated wind speed to the cut-out wind speed, and the wind power is constantly the rated power in this stage; the fourth stage is when the wind speed is greater than the cut-out wind speed, at this time, the wind turbine protection is started and the output is zero. During the actual operation of the wind turbine, different wind farms can refer to their historical data to obtain the wind speed-power characteristic curve during the actual operation process, and the output characteristics of different wind turbines show different output levels following the change of the wind speed range.

[0071] (3) In it, the typical wind farm output scenario set is divided to characterize the output correlation between wind farms. The method is as follows:

[0072] Due to the influence of factors such as weather processes and spatial geographical location information, complex spatio-temporal correlations are coupled between wind farms. Using the K-means clustering algorithm of unsupervised learning to divide wind farms with relatively similar output characteristics into the same subset is conducive to capturing the collaborative and periodic fluctuations of the output trends between typical wind farms by using the correlation between stations. The adopted K-means clustering algorithm updates k clustering centers through the clustering idea based on distance until the preset convergence condition is met, and a clustering result reflecting the output characteristics of the wind farm is obtained.

[0073] To analyze the output correlation between typical wind farms on a long time scale, the output time series of typical wind farms are preset as different random variables, and the covariance matrix ∑ is used to measure the linear correlation degree between different random variables. Its representation form is as follows:

[0074]

[0075] In the formula: cov(y i , y j j) represents the covariance of y i and y j j. i ranges from 1 to n, j ranges from 1 to n, and n represents the number of wind farms. i and j represent different wind farm numbers. Here, the covariance is used to characterize the correlation relationship between wind farm i and wind farm j.

[0076] Among them, cov(y i , y j ) = E[(y i - E[y i )(y j - E[y j )]. E represents the expectation, y i represents the output time series of the i-th wind farm; y j represents the output time series of the j-th wind farm.

[0077] (4) Consider the change of the correlation between wind farms under new observation conditions, and update the historical covariance matrix based on Bayes' theorem. The method is as follows:

[0078] Determine the prior distribution P(Σ) of the covariance matrix based on the historical output data of typical wind farms;

[0079] Take the polynomial fitting regression result as the evidence of the update model, and calculate the probability distribution P(y|Σ) of observing new data under the condition of a given covariance matrix;

[0080] Use the prior distribution P(Σ) and the probability distribution P(y|Σ), and update the estimation of the covariance matrix through Bayes' theorem, and calculate the posterior distribution P(Σ|y) ∝ P(y|Σ)P(Σ);

[0081] Obtain the corrected covariance matrix Σ through the calculated posterior probability P(Σ|y) new , reflecting the change of the correlation between wind farms under new observation conditions.

[0082] (5) In this method, consider the collaborative trend and spatial dependence of the output between wind farms, and correct the long-term predicted output based on the covariance weight update strategy. The method is as follows:

[0083] Measure the long-term output correlation relationship between typical wind farms through the covariance matrix, use Bayes' theorem to capture the change of the output correlation of typical wind farms, naturally express and process the influence degree of the output between typical wind farms based on the covariance weight update strategy, and further use covariance weighting to correct the long-term output sequence of wind farms, so that the predicted output can better reflect the actual situation. The specific calculation process is as follows:

[0084]

[0085] In the formula: w ij represents the normalized influence weight of the output y j of wind farm j on the output y i of wind farm i; α represents the adjustment coefficient, which controls the balance between the initial predicted value and the covariance weighted correction value; represents the corrected predicted output sequence of wind farm i.

[0086] The present invention also provides a wind power long-term predicted output correction system based on covariance weight update, including:

[0087] A preprocessing module, which is used to collect the historical meteorological data and power data corresponding to different wind farms, preprocess the collected data, and determine the training set and the test set based on the preprocessed data;

[0088] A correction module, which is used to establish a wind power long-term predicted output correction model, train the model using the training set, and evaluate the model using the test set;

[0089] A prediction module for obtaining the final long-term output prediction result of a typical wind farm through the long-term wind power prediction output correction model.

[0090] The present invention constructs a PINNs network model with Weibull distribution constraints embedded, combines data and physics-driven technologies to deeply explore the time-varying characteristics of wind speed, and thus effectively characterizes the stochastic fluctuation characteristics of wind speed. By Bayesian optimizing the historical covariance matrix between typical wind farms and combining the covariance weight update strategy to correct the long-term prediction output, the synergy of the output trends between typical wind farms is effectively enhanced. In application, the corrected long-term power generation prediction result can be combined with supply-demand assessment and auxiliary decision-making technologies to effectively solve the power supply guarantee problem of the new power system in complex scenarios.

[0091] To fully demonstrate the improvement effect of the wind power long-term prediction output correction method based on covariance weight update on prediction accuracy and trend performance, the root mean square error (RMSE) and the similarity (S) based on Euclidean distance are used as evaluation indicators, and the error analysis before and after the prediction output correction of typical wind farms is compared, as shown in Table 1 specifically:

[0092]

[0093] Where: D represents the Euclidean distance after flattening two correlation matrices.

[0094] Table 1 Error analysis before and after the prediction output correction of typical wind farms

[0095]

[0096] From the comparative analysis of the two evaluation indicators of RMSE and similarity S in Table 1, after considering the correlation correction in the long-term wind power prediction output, the prediction accuracy is effectively improved. As Figure 2 shown in the local comparison curve of the predicted output and the actual output of the typical wind farm, the prediction model considering the correlation correction has a higher degree of coincidence with the actual value. Especially in the period when the power fluctuates greatly or reaches the peak, the correlation-based prediction can more accurately reflect the change of the actual power. In addition, when the power fluctuates violently, especially in the climbing stage, the correlation-based prediction method effectively reduces the time-delay phenomenon, objectively indicating the promoting effect of considering the correlation correction prediction output on enhancing the trend performance.

[0097] In summary, the present invention considers the dynamic change of correlation to correct the long-term prediction output of the wind farm, which not only effectively improves the prediction accuracy, but also has more stable performance in capturing the trend change of the long-term output.

[0098] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.

Claims

1. A method for correcting wind power long-term forecast output based on covariance weight update, characterized in that: include: Collect historical meteorological data and power data corresponding to different wind farms, pre-process the collected data, and determine the training set and test set based on the pre-processed data; Establish a wind power long-term forecast output correction model, train the model using the training set, and evaluate the model using the test set; The final prediction result of the long-term output of a typical wind farm is obtained through the wind power long-term prediction output correction model.

2. The method for correcting wind power long-term forecast output based on covariance weight update according to claim 1, characterized in that: The preprocessing of collected data includes data cleaning and feature optimization, where data cleaning is used to remove outliers and fill missing values; feature optimization is performed using the Pearson correlation coefficient method; The preprocessed data are divided into training and testing sets after time-scale alignment.

3. The method for correcting wind power long-term forecast output based on covariance weight update according to claim 1, characterized in that: When constructing the wind power long-term forecast output correction model, the following steps are included: Firstly, the loss function structure of PINNs network is improved to obtain the wind speed time series of each wind farm; Secondly, based on the wind speed-wind power output characteristic scatter plot, the wind speed-wind power output function mapping relationship in the long-term scale of the wind farm is expressed, and the initial prediction results of the long-term output of each wind farm are obtained based on the polynomial fitting regression model; Then, the typical wind farms are divided based on the K-means clustering algorithm to obtain the initial data set of long-term predicted output of typical wind farms; Then, based on Bayesian theorem, the predicted output is used as the new observation condition to dynamically update the historical covariance matrix; Finally, based on the size of the covariance weight, the importance of the long-term predicted output of a typical wind farm is represented. The initial prediction result of the long-term output of a typical wind farm is corrected by covariance weighting, and a wind power long-term predicted output correction model based on covariance weight update is obtained.

4. The method for correcting wind power long-term forecast output based on covariance weight update according to claim 3 is characterized in that: A PINNs network with embedded Weibull distribution constraints is constructed to predict wind speed time series. The loss function structure is improved to consist of data loss, physical loss and additional statistical loss functions, specifically: L total =λ d L data +λ p L physics +λ w L weibull Where: L total represents the loss function, L data , L physics , L weibull Represent data loss, physical loss and statistical loss respectively, λ d , p , w is the error term importance hyperparameter.

5. The method for correcting wind power long-term forecast output based on covariance weight update according to claim 3, characterized in that: The polynomial fitting regression model is: y=α0+α1v+α2v 2 +α3v 3 Where: α0, α1, α2, α3 represent polynomial coefficients; v represents wind speed; y represents wind power output time series.

6. The method for correcting wind power long-term forecast output based on covariance weight update according to claim 5, characterized in that: The K-means clustering algorithm used updates k cluster centers through the distance-based clustering idea until the preset convergence conditions are met, and obtains the clustering results that reflect the output characteristics of the wind farm; The typical wind farm output time series is preset as different random variables, and the covariance matrix ∑ is used to measure the linear correlation between different random variables. Its representation form is as follows: Where: cov(y i ,y j ) represents y i and j The covariance of , i is 1~n, j is 1~n, and n represents the number of wind farms.

7. The method for correcting wind power long-term forecast output based on covariance weight update according to claim 6, characterized in that: Update the historical covariance matrix based on Bayes' theorem, including: Determine the prior distribution P(Σ) of the covariance matrix based on the historical output data of typical wind farms; Use the polynomial fitting regression results as evidence to update the model and calculate the probability distribution P(y|Σ) of observing new data under the condition of a given covariance matrix; Using the prior distribution P(Σ) and the probability distribution P(y|Σ), the estimate of the covariance matrix is ​​updated through Bayes’ theorem to calculate the posterior distribution P(Σ|y)∝P(y|Σ)P(Σ); The modified covariance matrix Σ is obtained by calculating the posterior probability P(Σ|y) new , reflecting the change in correlation between wind farms under new observation conditions.

8. The method for correcting wind power long-term forecast output based on covariance weight update according to claim 7, characterized in that: The covariance weighted correction of the long-term output sequence of the wind farm is carried out. The specific calculation process is as follows: Where: w ij Expressed as wind farm j output y j Output y for wind farm i i The normalized influence weight of; α is represented by the adjustment coefficient; It represents the corrected predicted output sequence of wind farm i.

9. A wind power long-term forecast output correction system based on covariance weight update, characterized in that: include: A preprocessing module is used to collect historical meteorological data and power data corresponding to different wind farms, preprocess the collected data, and determine the training set and test set based on the preprocessed data; The correction module is used to establish a correction model for long-term wind power output forecast, train the model using a training set, and evaluate the model using a test set; The prediction module is used to obtain the final prediction result of the long-term output of a typical wind farm through the wind power long-term prediction output correction model.

Images