NWP wind speed frequency division correction method, device, equipment and storage medium

Abstract

This application provides a method, apparatus, device, and storage medium for day-ahead NWP wind speed frequency division correction considering spatiotemporal correlation and feature reconstruction. It relates to the field of numerical weather prediction correction technology. The method includes: selecting reference wind farms associated with the target wind farm based on spatial distance constraints and similarity calculation formulas; correcting the historical actual wind speeds of these wind farms using an outlier correction method based on sliding multi-window and fitted average wind speed-power curves; decomposing the wind speed data of the wind farms; reconstructing low-frequency and high-frequency components using the permutation entropy corresponding to each sub-component; and correcting the high-frequency and low-frequency components respectively using an artificial intelligence model and a similar high-frequency component error matching method to obtain the final NWP correction result. Experimental verification shows that the proposed method reduces RMSE by 0.4268 m / s and MAE by 0.2256 m / s compared to the original NWP correction.

Description

Technical Field

[0001] This application relates to the field of numerical weather prediction correction technology, and in particular to a day-ahead NWP wind speed frequency division correction method, apparatus, equipment and storage medium. Background Technology

[0002] Wind power is widely used due to its green, clean, and renewable characteristics. However, wind power is random, intermittent, and fluctuating. Directly connecting wind power to the grid may cause short-term power surges, impacting the power system and even damaging electrical equipment. Wind power forecasting can predict the output of wind farms in advance, allowing grid dispatchers to develop scheduling plans ahead of time. It also helps wind farms to rationally arrange operation and maintenance plans, thus reducing operating costs. Wind speed information from numerical weather prediction (NWP) is a crucial feature for day-ahead wind power forecasting. Accurate NWP wind speed information allows forecasted wind speeds to more closely approximate actual wind speeds, enabling wind power forecasting models to more accurately establish the mapping relationship between NWP wind speed and wind power, thereby improving the accuracy of day-ahead wind power forecasting. However, in reality, the difference between the NWP wind speed on a daytime scale and the actual wind speed is significant. This is mainly reflected in the fact that the NWP wind speed may not be able to keep up with the trend of the actual wind speed in a timely manner, the numerical difference between the two wind speeds is large, and the NWP wind speed is relatively smooth, lacking detailed fluctuation information.

[0003] Therefore, the technical problem that urgently needs to be solved in the existing technology is: (1) Data-driven artificial intelligence prediction methods can predict NWP wind speed in future periods by establishing mapping relationships between data, and can achieve higher accuracy NWP correction by further improving the prediction performance of artificial intelligence models. However, this method only focuses on the data characteristics of its own wind farm and does not further explore the associated time series characteristics of wind farms with spatial correlation within a large-scale wind farm group; (2) The error between NWP wind speed and actual wind speed has the characteristics of high frequency and low amplitude. It is difficult to predict the error using only artificial intelligence methods. However, complex error prediction models have long running time and low efficiency.

[0004] (3) In order to further explore the hidden information of time series features, data decomposition models are widely used in the NWP correction process. If all decomposed components are analyzed and processed by artificial intelligence models, it will result in low efficiency and long running time. (4) If the original decomposed components are reconstructed by calculating the permutation entropy of each component, the classification algorithm alone may cause classification errors due to outliers in the permutation entropy. Summary of the Invention

[0005] This application provides a day-ahead NWP wind speed frequency division correction method, apparatus, equipment, and storage medium. The aim is to provide a scientifically sound, physically meaningful, fast, efficient, simple, practical, and highly accurate day-ahead NWP wind speed frequency division correction scheme that reflects the dynamic characteristics of the system and considers the spatiotemporal correlation between wind farms. This scheme considers reference wind farm information associated with the target wind farm in a large-scale wind farm cluster, corrects abnormal data of actual wind speed, and establishes a reasonable threshold selection mathematical model for reconstructing time series data after decomposition. This yields reasonable low-frequency and high-frequency components, and finally, frequency division correction is applied to the high-frequency and low-frequency components to obtain the final NWP wind speed correction result.

[0006] In a first aspect, this application provides a day-ahead NWP wind speed frequency division correction method that considers spatiotemporal correlation and feature reconstruction, including: Calculate the distance between the target wind farm and other wind farms, filter out wind farms whose distance from the target wind farm is less than a set distance, and calculate the score of each screened wind farm based on the historical actual wind speed of the screened wind farm and the historical actual wind speed of the target wind farm. Select the wind farms with the highest scores as reference wind farms. Outliers in the actual wind speed-power curve are removed using the quartile method under a multi-scale time window. The wind speed data are arranged in descending order of quantity, and an average wind speed-power curve is fitted to obtain the missing wind speed-power data. Based on the missing wind speed-power data, the actual wind speed data of the target wind farm is corrected to obtain the corrected actual wind speed data. The actual wind speed-power curve is determined based on the historical actual wind speeds of the target wind farm and the reference wind farm. The wind speed sequence and wind power sequence of the target wind farm are decomposed to obtain their respective wind speed sub-components and power sub-components. The wind speed sequence includes corrected historical actual wind speed data, historical NWP wind speed data, and future NWP wind speed data. Each wind speed sub-component and power sub-component is reconstructed into a trend component, an oscillation component, and a random component, respectively. The trend component and the oscillation component are added to obtain the low-frequency component of the first decomposition. The random component is decomposed a second time to obtain the secondary decomposition sub-components. The trend component and the oscillation component in the secondary decomposition sub-components are added to the low-frequency component of the first decomposition to obtain the final low-frequency component. The remaining random component after the secondary decomposition is taken as the high-frequency component. A TCN-LSTM-Transformer model is constructed to correct the final low-frequency component, resulting in a low-frequency corrected component. The high-frequency components are corrected by a high-frequency component correction method that matches errors in historical and future periods, resulting in high-frequency corrected components. The low-frequency correction component and the high-frequency correction component are added together to obtain the NWP correction result for the future time period.

[0007] In one possible design, the distances between the target wind farm and other wind farms are calculated. Wind farms whose distances from the target wind farm are less than a set distance are selected. A score is calculated for each selected wind farm based on its historical actual wind speed and the historical actual wind speed of the target wind farm. The wind farms with the highest scores are selected as reference wind farms, including: Obtain the set of straight-line distances between all wind farms and the target wind farm, denoted as: ,in, d 1. d 2 and d n Indicates the target wind farm and the first, second, and third n The straight-line distance between the target wind farm and the first wind farm i Straight-line distance between wind farms d i The calculation formula is: (1) In the formula, It is the Earth's radius, and the geographical coordinates of the target wind farm are... , No. i The geographical coordinates of the reference wind farms are as follows , Based on a set distance, several wind farms that meet the conditions are selected from the set of straight-line distances; The actual wind speeds of the selected wind farms are scored using the following formula: (2) In the formula, Represents normalized mutual information. This represents the HuberLoss loss value. Indicates the weight value. s Indicates the score; The formula for calculating normalized mutual information is: (5) In the formula, This represents the normalized mutual information between variables X and Y. This represents the mutual information between two time series. and Let X and Y represent the entropy of variables X and Y respectively, and calculate them using the following formulas: (6) (7) In the formula, p ( x i )and p ( y i They are respectively X and Y The marginal probability distribution function; The formula for calculating HuberLoss is as follows: (8) In the formula, This represents the HuberLoss loss value between two time series. y This represents the wind speed sequence of the target wind farm. f ( x ) represents the wind speed sequence of other wind farms. This represents the hyperparameters in the HuberLoss loss function.

[0008] In one possible design, outliers in the actual wind speed-power curve are removed using the quartile method under a multi-scale time window. The wind speed data are arranged in descending order of quantity, and an average wind speed-power curve is fitted to obtain the missing wind speed-power data. Based on this missing data, the actual wind speed data of the target wind farm is corrected to obtain the corrected actual wind speed data, including: The data in the historical actual wind speed-power curve are subjected to the quartile method using a sliding window with a first predetermined wind speed range width to remove outliers within that window scale. The interquartile method was applied to all data in the historical actual wind speed-power curve to remove the remaining outliers; A sliding window with a width greater than the width of the first predetermined wind speed interval is used to apply the quartile method to the data in the historical actual wind speed-power curve to further remove outliers; The wind speed data after removing outliers is arranged in descending order, and the power data is adjusted synchronously with the corresponding wind speed data position; a data window of a predetermined length is set, and data is removed so that the remaining data length is an integer multiple of the predetermined length; The wind speed-power data is sampled by dividing the data window according to the predetermined length. The average wind speed and average power are calculated in each window. Based on the average wind speed and average power calculated from all windows, the average wind speed-power curve is fitted. Using the average wind speed power curve described in the steps, the corresponding wind speed value is derived from the historical actual power data to correct missing or abnormal historical actual wind speed data.

[0009] In one possible design, the wind speed sequence and wind power sequence of the target wind farm are decomposed separately to obtain their respective wind speed sub-components and power sub-components, or the random components are decomposed a second time, including: Empirical mode decomposition is performed on the input sequence to obtain sub-components; wherein, the input sequence includes the wind speed sequence, wind power sequence, or random components obtained from a single decomposition of the target wind farm; By reconstructing the individual components using permutation entropy, we obtain random components, oscillatory components, and trend components.

[0010] In one possible design, the individual components are reconstructed using permutation entropy to obtain random components, oscillatory components, and trend components, including: By removing outliers from the permutation entropy set using the quartile method, the permutation entropy value without outliers is obtained. ;in, E 1 represents the permutation entropy of the first component; E 2 represents the permutation entropy of the second component; E i Indicates the first i The permutation entropy of each component E i+1 Indicates the first i +1 component permutation entropy, E j Indicates the first j The permutation entropy of each component E j+1 Indicates the first j +1 component permutation entropy; E n Indicates the first n The permutation entropy of each component; The range-average value of the permutation entropy is calculated using the following formula. : (11) In the formula, The maximum value of the permutation entropy. This represents the minimum value of the permutation entropy. Let the critical threshold for the random component and the oscillating component be... The critical thresholds for oscillating and trend components are: ,when hour, The calculation formula is: (12) when hour, for and The average between; When the calculated permutation entropy does not satisfy Then, take the average entropy of the two permutations with the smallest number of permutations. The calculation formula is: (13) Passing the critical threshold and The permutation entropy is divided into three categories, and the sub-components of each category are reconstructed into random components, oscillating components, and trend components.

[0011] In one possible design, the TCN-LSTM-Transformer model includes a temporal convolutional network (TCN), an LSTM model, and a Transformer model. The LSTM model is placed after the multi-head attention mechanism module of the Transformer model to form the LSTM-Transformer model. The final low-frequency component is corrected through the TCN-LSTM-Transformer model, and the low-frequency corrected component is obtained by means of: The final low-frequency component is input into the temporal convolutional network (TCN), and the data processing procedure of the TCN includes: Temporal features are extracted layer by layer using causal convolution and dilated convolution. The calculation process is represented as follows: (15) In the formula, This represents the expansion factor, used to control the sampling interval. This represents the weight of the k-th convolutional kernel. Indicates the convolution kernel weights, This represents the historical index of the convolution kernel. x t-d·k Indicates the first Time series at each point in time y t This represents the output of the convolutional layer at time t; The input and output of the convolutional layer are changed by N A stacked structure consisting of residual blocks, where the output of a single residual block is: (16) In the formula, Indicates the first n The output of each residual block This represents the activation function. Indicates the first nThe input of each residual block, Represents the residual function; based on N The output of each residual block is used to obtain the final output value of the Temporal Convolutional Network (TCN) using the following formula: (17) In the formula, This represents the final output value of TCN. Represents the ReLU activation function. This represents the output sequence of the Nth residual block. Represents the weight matrix. Indicates the bias coefficient; The final output value of the Temporal Convolutional Network (TCN) is input into the LSTM-Transformer model. The data processing procedure of the LSTM-Transformer model includes: Using the final output value as the input feature, it is mapped to a high-dimensional space to obtain the embedded time series. Based on the embedded time series, the initial representation matrix input into the encoder is obtained using the following formula: (19) (20) (twenty one) In the formula, The index value represents the time step, where, ; The index value represents the time step, where, P represents the position encoding matrix, and Z0 represents the initial matrix input to the encoder. Represents the embedded time series, d model This represents the dimension by which the input features are mapped to a higher-dimensional space; Based on the initial representation matrix shown, the process is performed layer by layer by a multi-layer encoder, with multi-head self-attention calculation performed at each layer to obtain the attention score for each head. The attention scores of each head are concatenated and linearly transformed, then passed through an LSTM model, and finally the matrix of the input feedforward network is obtained through residual connections and layer normalization. The calculation formula is as follows: (27) (28) (29) (30) In the formula, Indicates the first Output of size For multi-head output weight matrix, Representation layer normalization, This refers to computation using an LSTM model; Concat represents the concatenation operation. Let T represent the real number field, and T represent the total number of time steps. Indicates the first Attention score based on height This represents the multi-head attention output matrix. Indicates the first l A multi-head attention layer, Z l-1 H represents the output matrix after the encoder-decoder operation, and H represents the output matrix after the LSTM model operation. Based on matrix Then, perform feedforward neural network operations, and perform residual connections and layer normalization again to obtain the output matrix of one encoder layer; Based on the output matrix of each encoder layer, the low-frequency correction component is obtained using the following formula: (33) In the formula, This represents the output matrix of the Nth encoder. This represents the output weight matrix. This is a low-frequency correction component.

[0012] In one possible design, the high-frequency components are corrected using a high-frequency component correction method that matches errors between historical and future periods, resulting in high-frequency corrected components, including: The similarity score between a future NWP high-frequency component sequence and each historical NWP high-frequency component sequence is calculated using the following formula: (35) In the formula, P i,j Indicates the first in history i The first high-frequency component and the future... j The score value of each high-frequency component, A and B Indicates the first and second coefficients. Pcc ( X h,i , Y f,j ) indicates the historical number i The first high-frequency component and the future... j Pearson correlation coefficients for each high-frequency component NMAE ( X h,i , Yf,j ) indicates the historical number i The first high-frequency component and the future... j Normalized average absolute error of high-frequency air volume X h,i Indicates the first in history i High-frequency components Y f,j Indicates the future number j One high-frequency component; Based on calculations, the future nth [number] is determined using the following formula. w Ranking of scores for each time period k The high-frequency error sequence corresponding to the time series: (37) In the formula, For the future w Ranking of scores for each time period k The high-frequency error sequence corresponding to the time series. This indicates the search for the ranked k Time series, For the future period The similarity score between each time period and each time period in history; Calculate the similarity score and rank the top k The average value of the high-frequency error sequence corresponding to the time series is used as the error correction sequence for future high-frequency components.

[0013] Secondly, this application provides a day-ahead NWP wind speed frequency division correction device that considers spatiotemporal correlation and feature reconstruction, the device comprising: The wind farm screening module is configured to calculate the distance between the target wind farm and other wind farms, screen out wind farms whose distance from the target wind farm is less than a set distance, and calculate the score of each screened wind farm based on the historical actual wind speed of the screened wind farm and the historical actual wind speed of the target wind farm, and select the wind farms with the highest scores as reference wind farms. The actual wind speed correction module is configured to remove outliers from the actual wind speed-power curve using the quartile method under a multi-scale time window, arrange the wind speed data in descending order of quantity, fit an average wind speed-power curve, obtain the missing wind speed-power data, and correct the actual wind speed data of the target wind farm based on the missing wind speed-power data to obtain the corrected actual wind speed data; wherein, the actual wind speed-power curve is determined based on the historical actual wind speeds of the target wind farm and the reference wind farm. The sequence decomposition module is configured to decompose the wind speed sequence and wind power sequence of the target wind farm respectively, obtaining their respective wind speed sub-components and power sub-components; the wind speed sequence includes corrected historical actual wind speed data, historical NWP wind speed data, and future NWP wind speed data; each wind speed sub-component and power sub-component is reconstructed into a trend component, an oscillation component, and a random component; the trend component and the oscillation component are added to obtain the low-frequency component of the first decomposition; the random component is decomposed a second time to obtain the secondary decomposition sub-components; the trend component and the oscillation component in the secondary decomposition sub-components are added to the low-frequency component of the first decomposition to obtain the final low-frequency component; the remaining random component after the secondary decomposition is taken as the high-frequency component. The low-frequency component correction module is configured to construct a TCN-LSTM-Transformer model to correct the final low-frequency component, thereby obtaining the low-frequency corrected component. The high-frequency component correction module is configured to correct the high-frequency component by using a high-frequency component correction method that matches errors in historical and future periods, to obtain a high-frequency corrected component. The component superposition module is configured to add the low-frequency correction component and the high-frequency correction component to obtain the NWP correction result for the future time period.

[0014] Thirdly, embodiments of this application provide an electronic device, including: at least one processor and a memory; the memory stores computer execution instructions; the at least one processor executes the computer execution instructions stored in the memory, causing the at least one processor to execute the day-ahead NWP wind speed frequency division correction method considering spatiotemporal correlation and feature reconstruction as described in the first aspect and various possible designs of the first aspect.

[0015] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions. When a processor executes the computer-executable instructions, it implements the day-ahead NWP wind speed frequency division correction method considering spatiotemporal correlation and feature reconstruction as described in the first aspect and various possible designs of the first aspect.

[0016] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the day-ahead NWP wind speed frequency division correction method considering spatiotemporal correlation and feature reconstruction as described in the first aspect and various possible designs of the first aspect.

[0017] The day-ahead NWP wind speed frequency division correction method, apparatus, device, and storage medium provided in this application, which consider spatiotemporal correlation and feature reconstruction, have at least the following beneficial effects: To overcome the shortcomings of existing technologies, this application provides a scientifically sound, physically meaningful, fast, efficient, simple, practical, and highly accurate method for correcting day-ahead NWP wind speed by frequency division. This method reflects the dynamic characteristics of the system and considers the spatiotemporal correlation between wind farms. It takes into account reference wind farm information associated with the target wind farm in a large-scale wind farm cluster, corrects abnormal data of actual wind speed, and establishes a mathematical model for selecting reasonable thresholds after time series decomposition and reconstruction. This model then obtains reasonable low-frequency and high-frequency components, and finally performs frequency division correction on the high-frequency and low-frequency components to obtain the final NWP wind speed correction result. This invention features scientific soundness, higher accuracy, clear physical meaning, speed, efficiency, and simplicity. The day-ahead NWP wind speed frequency division correction method considering spatiotemporal correlation and feature reconstruction shows an average reduction of 6.59% and 7.33% in RMSE and MAE, respectively, compared to the original method, demonstrating higher accuracy and more stable prediction performance. Attached Figure Description

[0018] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0019] Figure 1 A flowchart of a day-ahead NWP wind speed frequency division correction method considering spatiotemporal correlation and feature reconstruction is provided for embodiments of this application; Figure 2 A flowchart for correcting abnormal data of actual wind speed provided in the embodiments of this application; Figure 3 A flowchart for the correction of the low-frequency component of future NWP wind speed provided in this application embodiment; Figure 4 A flowchart for the correction of the low-frequency component of future NWP wind speed provided in this application embodiment; Figure 5 This is a schematic diagram illustrating the effect of NWP wind speed correction before and after the present application embodiment; Figure 6 A structural diagram of a day-ahead NWP wind speed frequency division correction device considering spatiotemporal correlation and feature reconstruction provided for embodiments of this application.

[0020] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0021] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0022] The collection, storage, use, processing, transmission, provision, and disclosure of financial data or user data involved in the technical solution of this application all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0023] It should be noted that in the embodiments of this application, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of this application. However, it does not mean that the applicant has used or necessarily used the solution.

[0024] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will be described below with reference to the accompanying drawings.

[0025] This application provides a method for frequency-division correction of day-ahead NWP wind speed considering spatiotemporal correlation and feature reconstruction. It includes steps such as selecting a reference wind farm, correcting the actual wind speed, reconstructing the time series, correcting the low-frequency component and the high-frequency component, simulation calculation, and error analysis. First, reference wind farms associated with the target wind farm are selected based on spatial distance constraints and similarity calculation formulas. Then, the historical actual wind speeds of these wind farms are corrected using an outlier correction method based on sliding multi-window and fitted average wind speed-power curves. Next, the historical actual wind speed, historical NWP wind speed, and future NWP wind speed of these wind farms are decomposed using the ICEEMDAN decomposition method. Low-frequency and high-frequency components are reconstructed using the permutation entropy corresponding to each sub-component. Finally, the high-frequency and low-frequency components are corrected using an artificial intelligence model and a similar high-frequency component error matching method to obtain the final NWP correction result. Specifically, as shown... Figure 1 As shown, the daytime NWP wind speed frequency division correction method considering spatiotemporal correlation and feature reconstruction includes the following steps S100-S600.

[0026] S100: Calculate the distance between the target wind farm and other wind farms, filter out wind farms whose distance from the target wind farm is less than the set distance, and calculate the score of each screened wind farm based on the historical actual wind speed of the screened wind farm and the historical actual wind speed of the target wind farm. Select the wind farms with the highest scores as reference wind farms.

[0027] In this embodiment, the purpose of step S100 is to select reference wind farms. For example, the distance between the target wind farm and other wind farms is calculated, and wind farms with a distance of less than 100 kilometers from the target wind farm are selected. Then, a scoring formula is used to calculate the score of each selected wind farm based on the historical actual wind speed of the first-selected wind farms and the historical actual wind speed of the target wind farm. The top three wind farms with the highest scores are selected as reference wind farms.

[0028] In some embodiments, a reference wind farm may be selected in the following manner: set up ,in, (1) in, This represents the straight-line distance between all wind farms and the target wind farm. This represents the straight-line distance between two wind farms. This is the Earth's radius, approximately 6371 kilometers. The geographical coordinates of the target wind farm are... , No. i The geographical coordinates of the reference wind farms are as follows Reference wind farms located 100 kilometers away from the target wind farm were screened, resulting in a number of wind farms meeting the criteria. ,in, The set of forecast wind speeds for wind farms within 100 kilometers is as follows: The actual wind speed set is .

[0029] To select a reasonable reference wind farm, the following scoring formula is used to score the actual wind speed of wind farms within a 100-kilometer radius of the target wind farm. The scoring criteria are as follows: (2) in, Represents normalized mutual information. express HuberLoss Loss value This represents the weight value.

[0030] Within 100 kilometers i The mutual information between a wind farm and a target wind farm can be defined as: (3) in, yes X and Y The joint probability distribution function, and and They are X and Y The marginal probability distribution function. In the case of continuous random variables, the above equation is transformed into a double integral calculation: (4) In equation (4), yes X and Y The joint probability density function, and and They are X and Y The marginal probability density function. Mutual information is X and Y Joint distribution relative to the assumption X and Y The inherent dependency between joint distributions in the independent case, although X and Y The relationship between them is not linear. For computational convenience, we use normalized mutual information to represent the similarity between the two time series.

[0031] (5) in, Representing variables X and Y Normalized mutual information between them, where X and Y The greater the similarity between them, The larger the value, the better. This represents the mutual information between two time series. and Representing variables respectively X and Y The entropy of these elements is calculated using the following formulas: (6) (7) In the formula, p ( x i )and p ( y i They are respectively X and Y The marginal probability distribution function.

[0032] Normalized mutual information can be used to determine the overall trend of two time series.

[0033] To determine the error between two time series, the following method is used: Huberloss The loss function is used to make the judgment. Huberloss The formula for calculating the loss function is as follows: (8) in, Indicates the time series between two time series HuberLoss Loss value y This represents the wind speed sequence of the target wind farm. f ( x ) represents the wind speed sequence of other wind farms. express HuberLoss Hyperparameters in the loss function. Based on experience, set... . HuberLoss It combines the advantages of loss function sets MAE (mean absolute error) and MSE (mean squared error), while reducing the sensitivity to outliers and enhancing the accuracy and robustness of judging the error value between two time series.

[0034] The score for each wind farm within a 100-kilometer radius is calculated using formula (9): (9) The set of rating values ​​is Select the top three wind farms by rating as reference wind farms. Let the predicted wind speed of the target wind farm be... The actual wind speed is The reference wind farm's forecast wind speed set is as follows: The set of actual wind speeds referenced from the wind farm is as follows The corresponding rating values ​​for the reference wind farms are as follows: The proportion of the reference wind farm relative to the target wind farm is as follows: (10) In equation (10), Indicates the first i The weights of each wind farm, Indicates the first i A score for a reference wind farm.

[0035] S200: Outliers in the actual wind speed-power curve are removed using the quartile method under a multi-scale time window. The wind speed data are arranged in descending order of quantity, and an average wind speed-power curve is fitted to obtain the missing wind speed-power data. The actual wind speed data of the target wind farm is corrected based on the missing wind speed-power data to obtain the corrected actual wind speed data. The actual wind speed-power curve is determined based on the historical actual wind speeds of the target wind farm and the reference wind farm.

[0036] Currently, actual wind speed data mainly comes from meteorological towers. However, in actual wind power projects, the quality of meteorological tower data is difficult to guarantee, especially in wind farms located at high altitudes and mountainous areas, which are frequently affected by changing weather patterns. Measuring equipment may malfunction or stop measuring, resulting in a large amount of missing and unreasonable wind speed data. Historical actual wind speeds are needed in the process of correcting data for weather forecasts. If the historical actual wind speeds are inaccurate, the corrected NWP (Non-Wave Speed-Power) may contain a large number of outliers and noise. Therefore, it is necessary to correct historical measured wind speeds. Based on this, this embodiment proposes a method for correcting anomaly data of historical measured wind speeds based on sliding multi-window and fitting an average wind speed-power curve. The actual wind speed data can be corrected through step S200. First, outliers in the actual wind speed-power curve are removed using the quartile method under a multi-scale time window. Then, the wind speed data are arranged in descending order of quantity. Finally, an average wind speed-power curve is fitted to obtain the missing wind speed-power data.

[0037] In some embodiments, such as Figure 2 The diagram shown is a flowchart of the abnormal data correction for actual wind speed provided in this application embodiment. Step S200 can be specifically implemented by the following steps S201-S206 to correct the actual wind speed data of the target wind farm.

[0038] S201: For the first time, the wind speed-power curve data is processed using a quartile window with a window size of 0.2 m / s to remove outliers. The 0.2 m / s window allows for more precise division of wind speed ranges, accommodating subtle changes and local features that may appear in the wind speed-power curve. In areas where wind speed changes slowly, outliers can be identified more accurately; in areas where wind speed changes rapidly, outliers can also be captured well, avoiding omissions due to an excessively large window.

[0039] S202: Perform quartile processing on all data in the wind speed-power curve to remove outliers that were not removed in the first step; after the small window quartile process, outlier processing is performed from a global perspective, which makes up for the shortcomings of small window processing that may cause outlier omissions.

[0040] S203: Perform a third outlier processing on the wind speed-power curve data using a 1 m / s window. After the above three steps of outlier processing, all outliers in the wind speed-power curve can be removed.

[0041] S204: Set the data window and data range. Arrange the wind speed data in descending order of quantity, and the power data should also change accordingly with the wind speed data. Assume the data window length is 10, remove redundant data from the wind speed-power data, and retain data that is a multiple of 10 in length.

[0042] S205: Fitting the average wind speed-power curve. Then, the wind speed-power data is sampled by dividing the data window, and the average values ​​of wind speed and power are calculated within each window. Finally, the average wind speed-power curve is fitted using the average wind speed and average power values.

[0043] S206: Supplement missing wind speed values. The missing wind speed data is obtained by using the historical actual power data and the fitted average wind speed-power curve after anomaly data processing, thus correcting the abnormal actual wind speed data.

[0044] S300: Decompose the wind speed sequence and wind power sequence of the target wind farm respectively to obtain their respective wind speed sub-components and power sub-components; the wind speed sequence includes corrected historical actual wind speed data, historical NWP wind speed data, and future NWP wind speed data; reconstruct each wind speed sub-component and power sub-component into trend components, oscillation components, and random components respectively; add the trend component and oscillation component to obtain the low-frequency component of the first decomposition; perform a second decomposition on the random component to obtain the second decomposition sub-components; add the trend component and oscillation component in the second decomposition sub-components to the low-frequency component of the first decomposition to obtain the final low-frequency component; use the remaining random component after the second decomposition as the high-frequency component.

[0045] In this embodiment, the improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) method is used to decompose the time series. Then, the complexity of each decomposed component is calculated using permutation entropy, and the components are reconstructed based on the permutation entropy values. The permutation entropy reconstructs the ICEEMDAN-decomposed components into random, oscillatory, and trend components. The key to reconstructing the time series components is determining the critical threshold. To ensure the reliability of the critical threshold, outliers in the permutation entropy set are first removed using the quartile method. Let the permutation entropy value without outliers be [value missing]. The number of permutation entropies is , The maximum value of the permutation entropy. This represents the minimum value of the permutation entropy. E 1 represents the permutation entropy of the first component; E 2 represents the permutation entropy of the second component; E i Indicates the first i The permutation entropy of each component E i+1 Indicates the first i +1 component permutation entropy, E j Indicates the first j The permutation entropy of each component E j+1 Indicates the first j +1 component permutation entropy; E n Indicates the first n The permutation entropy of each component.

[0046] The range-average value of the permutation entropy is defined as: (11) The range indicates the maximum discrete range of the permutation entropy, while the average range indicates the average discrete range of the permutation entropy. Next, the difference between adjacent permutation entropies is calculated using recursive operations to determine the critical thresholds for the random component, oscillatory component, and trend component. Let the critical thresholds for the random component and oscillatory component be... The critical thresholds for oscillating and trend components are: .when hour, The calculation formula is: (12) when hour, The calculation method is as follows and The average value between; when the calculated permutation entropy does not satisfy Then, the average entropy of the two permutations with the smallest number of permutations is taken. Therefore, The calculation formula is: (13) Passing the critical threshold and The permutation entropy is divided into three categories, and the sub-components of each category are reconstructed into random components, oscillatory components, and trend components. Finally, the trend components and oscillatory components are added together to obtain the low-frequency components, and the random components are named the high-frequency components.

[0047] S400: Construct a TCN-LSTM-Transformer model to correct the final low-frequency component, obtaining the low-frequency corrected component.

[0048] In some embodiments, such as Figure 3 The diagram shown illustrates the flowchart for correcting the low-frequency component of future NWP wind speed according to an embodiment of this application. This embodiment employs a TCN-LSTM-Transformer model to correct the low-frequency component. TCN is used to capture local patterns and multi-scale feature information from the original time series. LSTM is embedded in the Transformer to perform local continuity calibration and implicit position encoding on the output of the attention mechanism, enhancing the model's ability to model dynamic changes in the sequence. The model establishes a mapping between historical low-frequency NWP components and historical actual low-frequency wind speeds, and then obtains the corrected low-frequency NWP component for the future time period by inputting low-frequency NWP data for the future time period. The prediction model's time scale is 24 hours.

[0049] Temporal Convolutional Networks (TCNs) are deep learning models specifically designed for processing sequential data. They combine the parallel processing capabilities of Convolutional Neural Networks (CNNs) with the long-term dependency modeling capabilities of Recurrent Neural Networks (RNNs), making them a powerful tool for sequence modeling tasks. Compared to CNNs and RNNs, TCNs add causal convolutions and dilated convolutions. The core idea of ​​TCNs for processing time series data is to capture temporal dependencies through causal and dilated convolutions, while simultaneously using residual connections to address the training challenges of deep networks.

[0050] Causal convolution guarantees the temporal order, preventing the leakage of future information. Its calculation formula is: (14) In the formula, This represents the input time series. Indicates the convolution kernel weights, Indicates the convolution kernel weights, This represents the historical index of the convolution kernel.

[0051] When causal convolutions are stacked in multiple layers, the receptive field grows slowly, making it difficult to establish long sequence dependencies. TCN (Transient Convolutional Networks) introduces a dilation factor to construct dilated convolutions. This expands the receptive field through interval sampling, avoiding excessive stacking of causal convolution layers. The calculation method is as follows: (15) In the formula, This represents the expansion factor, used to control the sampling interval; x t-d·k Indicates the first Time series at several points in time.

[0052] when The dilated convolution degenerates into a regular causal convolution.

[0053] The residual blocks in TCN contain convolutional layers, activation functions, and skip connections, used to address the vanishing gradient problem in deep networks and accelerate the training process. Their calculation formula is: (16) Indicates the first n The output of each residual block This represents the activation function. Indicates the first n The input of each residual block, Let represent the residual function. Suppose that the TCN has . N If there are 1 residual block, then the final output value of TCN is: (17) In the formula, This represents the final output value of TCN. Represents the ReLU activation function. This represents the output sequence of the Nth residual block. Represents the weight matrix. This represents the bias coefficient.

[0054] The Transformer is a deep learning model architecture widely used in fields such as natural language processing and computer vision. It consists of an encoder and a decoder, with its core being a self-attention mechanism. This mechanism processes each element in the input sequence and computes its interactions with other elements in the sequence. This allows the model to consider all elements of the input sequence, not just the initial parts, when generating each output element. However, the Transformer model primarily focuses on global relationships, while LSTM is better able to handle local features and temporal dependencies. Therefore, we add an LSTM model after the multi-head attention mechanism module of the Transformer model. By combining the two, we can compensate for the Transformer's shortcomings in acquiring local information, enabling the model to understand the data more comprehensively.

[0055] Suppose the input time series X Include T There are *n* time steps, and the feature dimension of each time step is *n*. ,Right now The input features are mapped to a high-dimensional space with dimensions of . The embedded time series can be calculated using the following formula.

[0056] (18) In the formula, This represents the embedded time series. To embed the weight matrix, This is the bias vector.

[0057] Next, positional information is added to the time series using sine and cosine functions. The embedded time series is then combined with positional encoding to obtain the initial representation matrix input into the encoder.

[0058] (19) (20) (twenty one) In the formula, The index value represents the time step, where, ; The index value represents the time step, where, ; Represents the position encoding matrix, This represents the initial matrix input to the encoder.

[0059] The encoder is made by The same layers are stacked, each containing a multi-head self-attention network and a feed-forward network. The time-series transformations in the coding layer are shown below.

[0060] For the encoder l Time series of layer input Perform the following linear transformation: (twenty two) (twenty three) (twenty four) In the formula, These are the learnable weight matrices, and the dimensions of each attention head are: (25) in, This represents the number of attention heads. These represent the Query, Key, and Value matrices, respectively.

[0061] Next, the attention score is calculated using the following formula: (26) In the formula, The attention score matrix, This is a scaling factor to prevent the gradient from vanishing due to excessively large dot products. This is the activation function.

[0062] Next h The results from the header are concatenated and linearly transformed, then passed through an LSTM model, and finally the matrix of the input feedforward network is obtained through residual connections and layer normalization. The calculation formula is as follows: (27) (28) (29) (30) In the formula, Indicates the first Output of size For multi-head output weight matrix, Representation layer normalization means normalizing the feature vector at each time step. This refers to computation using an LSTM model. This represents the multi-head attention output matrix. Indicates the first l A multi-head attention layer, Z l-1 H represents the output matrix of the encoder-decoder, and H represents the output matrix after the LSTM model operation.

[0063] Next, the normalized result is processed using a feedforward neural network, followed by residual connections and layer normalization to obtain the output matrix of a single encoder layer. The calculation formula is as follows: (31) (32) In the formula, and This is the weight matrix. and This is the bias vector. This indicates the operation of the feedforward neural network model.

[0064] Let the number of encoder layers be . The encoder output is mapped to the target dimension, and the low-frequency correction component is obtained, which is the final predicted value. .

[0065] (33) In the formula, This represents the output matrix of the Nth encoder. This represents the output weight matrix.

[0066] S500: The high-frequency component is corrected by a high-frequency component correction method that matches errors in historical and future periods to obtain a high-frequency corrected component.

[0067] In some embodiments, such as Figure 4 The diagram shown is a flowchart of the low-frequency component correction process for future NWP wind speed provided in an embodiment of this application. Considering the poor predictive performance of high-frequency components, a high-frequency component correction method that matches errors between historical and future periods is proposed. The high-frequency component correction process includes the following steps S501-S505.

[0068] S501: Calculate the similarity score between a future NWP high-frequency component sequence and each historical NWP high-frequency component sequence. First, divide the historical and future NWP high-frequency components into time intervals of 96 points each. Let the historical period have... There are several time periods, and the future period has... There are several time periods. The component time intervals of a historical period can then be represented as... The time interval of the future period can be represented as .in, , The Pearson correlation coefficient measures the consistency of the trends between two time series, while the mean absolute error (MAE) measures the error between them. Combining these two metrics yields a formula used to measure the similarity between two time series. Wherein, .

[0069] (34) In the formula, Indicates the score value. A and B Represents the coefficient. This indicates the calculation of the Pearson correlation coefficient. This indicates the calculation of the normalized mean absolute error. According to the above formula, the... The future time period and the first The similarity scores for each historical period are as follows: (35) set up A = B =0.5, then the above expression can be simplified to: (36) Equation (36) uses a half-weighted sum of the difference between the Pearson correlation coefficient and the normalized mean absolute error, then superimposed it with a benchmark value. This ensures that when the data exhibits an effective linear correlation, i.e., a comprehensive index... The threshold can reasonably characterize the balance between basic correlation and error control in the model.

[0070] S502: Sort the similarity scores of historical time series and select the top 5 time series (with the following requirements). ), and renamed it. Let the future period be the first The similarity scores between each time period and each time period in history are as follows: .in, Then, the time series are sorted from highest to lowest similarity score, and the error time series corresponding to the top 5 time series are selected and named... The modeling process for this step is as follows: (37) In the formula, For the future w Ranking of scores for each time period k The high-frequency error sequence corresponding to the time series. This indicates the search for the ranked k The time series, where, .

[0071] S503: Calculate the average of the high-frequency error sequences corresponding to the top five time series with the highest similarity scores, and use them as the error correction sequences for future high-frequency components.

[0072] (38) S504: Through steps S501~S503, calculate the high-frequency component error correction sequence for all time periods in the future.

[0073] S505: Add the original time series of each time period in the future to the corresponding high-frequency error correction series to obtain the final high-frequency component correction result. Let the future... w The original sequence of the time period is The corresponding high-frequency component error correction sequence is Then the future number w Final revised results for the time period for: (39) S600: Add the low-frequency correction component and the high-frequency correction component to obtain the NWP correction result for the future time period.

[0074] To verify the feasibility and advancements of the method proposed in this application, a detailed explanation will be provided below using simulation examples. Using the historical actual wind speed, historical NWP wind speed, future NWP wind speed, and geographical coordinates of a large-scale wind farm cluster as simulation inputs, the corrected future NWP wind speed of the target wind farm is obtained through steps S100 to S600.

[0075] To verify the effectiveness of the proposed method, root mean square error and mean absolute error were used as evaluation metrics. Among them, Rm... MSE1 and R MAE1 R is used to evaluate the correction for NWP wind speed. MSE2 and R MAEE2 R is used to evaluate wind power prediction results. MSE1 and R MAE1 The calculation formula is as follows: (40) (41) In equations (40) and (41), m This indicates the duration of NWP wind speed. n This indicates the duration of wind power generation. It indicates the first i The actual wind speed at each point in time Indicates the first i Corrected wind speed at each time point.

[0076] R MSE2 and R MAEE2 The calculation method and R MSE1 and R MAE1 Similarly, it will not be elaborated here.

[0077] Step S100 obtains the reference wind farm corresponding to the target wind farm. Step S200 corrects the outliers of the historical actual wind speed of the target wind farm and the reference wind farm. Step S300 decomposes and reconstructs the historical actual wind speed, historical NWP wind speed and future NWP wind speed. Step S400 obtains the low-frequency component correction result of the future NWP wind speed of the target wind farm. Step S500 obtains the high-frequency component correction result of the future NWP wind speed of the target wind farm. The error evaluation standard formula (40)-formula (41) is used to calculate the error, and the root mean square error and mean absolute error are obtained. Taking the data of a large-scale wind farm group consisting of 100 wind farms in a certain area as an example, the sampling interval is 15 minutes. A certain wind farm is selected as the target wind farm, and the installed capacity of the wind farm is 130.5MW. The result of NWP wind speed correction is as follows. Figure 5As shown in the figure. Experimental verification shows that when the corrected NWP wind speed is compared with the original NWP and the actual wind speed, the corrected NWP wind speed RMSE is reduced by 0.4268 m / s and MAE is reduced by 0.2256 m / s compared with the original NWP.

[0078] This application also provides a day-ahead NWP wind speed frequency division correction device that considers spatiotemporal correlation and feature reconstruction, such as Figure 6 As shown, the day-ahead NWP wind speed frequency division correction device, which considers spatiotemporal correlation and feature reconstruction, includes: The wind farm screening module 601 is configured to calculate the distance between the target wind farm and other wind farms, screen out wind farms whose distance from the target wind farm is less than a set distance, and calculate the score of each screened wind farm based on the historical actual wind speed of the screened wind farm and the historical actual wind speed of the target wind farm, and select the wind farms with the highest scores as reference wind farms. The actual wind speed correction module 602 is configured to remove outliers from the actual wind speed-power curve using the quartile method under a multi-scale time window, arrange the wind speed data in descending order, fit an average wind speed-power curve to obtain the missing wind speed-power data, and correct the actual wind speed data of the target wind farm based on the missing wind speed-power data to obtain the corrected actual wind speed data; wherein, the actual wind speed-power curve is determined based on the historical actual wind speeds of the target wind farm and the reference wind farm. The sequence decomposition module 603 is configured to decompose the wind speed sequence and wind power sequence of the target wind farm respectively to obtain their respective wind speed sub-components and power sub-components; the wind speed sequence includes corrected historical actual wind speed data, historical NWP wind speed data, and future NWP wind speed data; each wind speed sub-component and power sub-component is reconstructed into a trend component, an oscillation component, and a random component respectively; the trend component and the oscillation component are added to obtain the low-frequency component of the first decomposition; the random component is decomposed a second time to obtain the secondary decomposition sub-components; the trend component and the oscillation component in the secondary decomposition sub-components are added to the low-frequency component of the first decomposition to obtain the final low-frequency component; the remaining random component after the secondary decomposition is taken as the high-frequency component. The low-frequency component correction module 604 is configured to construct a TCN-LSTM-Transformer model to correct the final low-frequency component and obtain the low-frequency corrected component. The high-frequency component correction module 605 is configured to correct the high-frequency component by using a high-frequency component correction method that matches errors in historical and future periods, to obtain a high-frequency corrected component. The component superposition module 606 is configured to add the low-frequency correction component and the high-frequency correction component to obtain the NWP correction result for the future time period.

[0079] This application provides an electronic device. The electronic device may include a processor and a memory, wherein the processor and the memory can communicate; exemplarily, the processor and the memory communicate via a communication bus.

[0080] The processor executes computer execution instructions stored in memory, causing the processor to perform the scheme in the above embodiments. The processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0081] The communication bus can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The system bus can be divided into address bus, data bus, control bus, etc. Transceivers are used to enable communication between database access devices and other computers (e.g., clients, read-write libraries, and read-only libraries). Memory may include Random Access Memory (RAM) and may also include Non-Volatile Memory.

[0082] The electronic device provided in this application embodiment can be the terminal device described in the above embodiments.

[0083] This application also provides a computer-readable storage medium storing computer instructions. When the computer instructions are executed on a computer, the computer performs the technical solution of the day-ahead NWP wind speed frequency division correction method considering spatiotemporal correlation and feature reconstruction described in the above embodiments.

[0084] This application also provides a computer program product, which includes a computer program stored in a computer-readable storage medium. At least one processor can read the computer program from the computer-readable storage medium. When the at least one processor executes the computer program, it can implement the technical solution of the day-ahead NWP wind speed frequency division correction method considering spatiotemporal correlation and feature reconstruction in the above embodiments.

[0085] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or modules, and may be electrical, mechanical, or other forms.

[0086] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to implement the solution of this embodiment according to actual needs.

[0087] Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The unit composed of the above modules can be implemented in hardware or in the form of hardware plus software functional units.

[0088] The integrated modules described above, implemented as software functional modules, can be stored in a computer-readable storage medium. These software functional modules, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods of the various embodiments of this application.

[0089] It should be understood that the aforementioned processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. A general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly manifested as execution by a hardware processor, or execution by a combination of hardware and software modules within the processor.

[0090] The memory may include high-speed RAM, and may also include non-volatile storage (NVM), such as at least one disk storage device, and may also be a USB flash drive, external hard drive, read-only memory, disk or optical disc, etc.

[0091] Buses can be Industry Standard Architecture (ISA) buses, Peripheral Component Interconnect (PCI) buses, or Extended Industry Standard Architecture (EISA) buses, etc. Buses can be categorized into address buses, data buses, control buses, etc.

[0092] The aforementioned storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The storage medium can be any available medium that can be accessed by a general-purpose or special-purpose computer.

[0093] An exemplary storage medium is coupled to a processor, enabling the processor to read information from and write information to the storage medium. Alternatively, the storage medium can be an integral part of the processor. The processor and storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and storage medium can exist as discrete components in an electronic control unit or main control device.

[0094] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0095] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A day-ahead NWP wind speed sub-frequency correction method considering spatiotemporal correlation and feature reconstruction, characterized in that, The method includes: Calculate the distance between the target wind farm and other wind farms, filter out wind farms whose distance from the target wind farm is less than a set distance, and calculate the score of each screened wind farm based on the historical actual wind speed of the screened wind farm and the historical actual wind speed of the target wind farm. Select the wind farms with the highest scores as reference wind farms. Outliers in the actual wind speed-power curve are removed using the quartile method under a multi-scale time window. The wind speed data are arranged in descending order of quantity, and the power data changes simultaneously with the position of the wind speed data. An average wind speed-power curve is fitted to obtain the missing wind speed and power data. Based on the missing wind speed and power data, the actual wind speed data of the target wind farm is corrected to obtain the corrected actual wind speed data. The actual wind speed-power curve is determined based on the historical actual wind speeds of the target wind farm and the reference wind farm. The wind speed sequence and wind power sequence of the target wind farm are decomposed to obtain their respective wind speed sub-components and power sub-components. The wind speed sequence includes corrected historical actual wind speed data, historical NWP wind speed data, and future NWP wind speed data. Each wind speed sub-component and power sub-component is reconstructed into a trend component, an oscillation component, and a random component, respectively. The trend component and the oscillation component are added to obtain the low-frequency component of the first decomposition. The random component is decomposed a second time to obtain the secondary decomposition sub-components. The trend component and the oscillation component in the secondary decomposition sub-components are added to the low-frequency component of the first decomposition to obtain the final low-frequency component. The remaining random component after the secondary decomposition is taken as the high-frequency component. A TCN-LSTM-Transformer model is constructed to correct the final low-frequency component, resulting in a low-frequency corrected component. The high-frequency components are corrected by a high-frequency component correction method that matches errors in historical and future periods, resulting in high-frequency corrected components. The low-frequency correction component and the high-frequency correction component are added together to obtain the NWP correction result for the future time period.

2. The method of claim 1, wherein the method is characterized by, Calculate the distance between the target wind farm and other wind farms, filter out wind farms whose distance to the target wind farm is less than a set distance, and calculate the score of each filtered wind farm based on the historical actual wind speed of the filtered wind farm and the historical actual wind speed of the target wind farm. Select the top few wind farms with the highest scores as reference wind farms, including: Obtain the set of straight-line distances between all wind farms and the target wind farm, denoted as: ,in, d 1. d 2 and d n This represents the straight-line distance between the target wind farm and the first, second, and nth wind farms, and the straight-line distance between the target wind farm and the i-th wind farm. d i The calculation formula is: (1) In the formula, It is the Earth's radius, and the geographical coordinates of the target wind farm are... , No. i The geographical coordinates of the reference wind farms are as follows , Based on a set distance, several wind farms that meet the conditions are selected from the set of straight-line distances; The actual wind speeds of the selected wind farms are scored using the following formula: (2) In the formula, Represents normalized mutual information. This represents the HuberLoss loss value. Indicates the weight value. s Indicates the score; The formula for calculating normalized mutual information is: (5) In the formula, Representing variables X and Y Normalized mutual information between them This represents the mutual information between two time series. and Representing variables respectively X and Y The entropy is calculated using the following formula: (6) (7) In the formula, p ( x i )and p ( y i They are respectively X and Y The marginal probability distribution function; The formula for calculating HuberLoss is as follows: (8) In the formula, This represents the HuberLoss loss value between two time series. y This represents the wind speed sequence of the target wind farm. f ( x () represents the wind speed sequence of other wind farms; This represents the hyperparameters in the HuberLoss loss function.

3. The day-ahead NWP wind speed frequency division correction method considering spatiotemporal correlation and feature reconstruction according to claim 1, characterized in that, Outliers in the actual wind speed-power curve were removed using the quartile method under a multi-scale time window. Wind speed data were arranged in descending order of quantity, and an average wind speed-power curve was fitted to obtain the missing wind speed-power data. Based on this missing data, the actual wind speed data of the target wind farm was corrected to obtain the corrected actual wind speed data, including: The data in the historical actual wind speed-power curve are subjected to the quartile method using a sliding window with a first predetermined wind speed range width to remove outliers within that window scale. The interquartile method was applied to all data in the historical actual wind speed-power curve to remove the remaining outliers; A sliding window with a width greater than the width of the first predetermined wind speed interval is used to apply the quartile method to the data in the historical actual wind speed-power curve to further remove outliers; The wind speed data after removing outliers is arranged in descending order, and the power data is adjusted synchronously with the corresponding wind speed data position; a data window of a predetermined length is set, and data is removed so that the remaining data length is an integer multiple of the predetermined length; The wind speed-power data is sampled by dividing the data window according to the predetermined length. The average wind speed and average power are calculated in each window. Based on the average wind speed and average power calculated from all windows, the average wind speed-power curve is fitted. Using the average wind speed power curve described in the steps, the corresponding wind speed value is derived from the historical actual power data to correct missing or abnormal historical actual wind speed data.

4. The day-ahead NWP wind speed frequency division correction method considering spatiotemporal correlation and feature reconstruction according to claim 1, characterized in that, The methods for decomposing the wind speed sequence and wind power sequence of the target wind farm to obtain their respective wind speed sub-components and power sub-components, or for performing a second decomposition on the random components, include: Empirical mode decomposition is performed on the input sequence to obtain sub-components; wherein, the input sequence includes the wind speed sequence, wind power sequence, or random components obtained from a single decomposition of the target wind farm; By reconstructing the individual components using permutation entropy, we obtain random components, oscillatory components, and trend components.

5. The day-ahead NWP wind speed frequency division correction method considering spatiotemporal correlation and feature reconstruction according to claim 4, characterized in that, By reconstructing the individual components using permutation entropy, we obtain random components, oscillatory components, and trend components, including: By removing outliers from the permutation entropy set using the quartile method, the permutation entropy value without outliers is obtained. ;in, E 1 represents the permutation entropy of the first component; E 2 represents the permutation entropy of the second component; E i Indicates the first i The permutation entropy of each component E i+1 Indicates the first i +1 component permutation entropy, E j Indicates the first j The permutation entropy of each component E j+1 Indicates the first j +1 component permutation entropy; E n Indicates the first n The permutation entropy of each component; The range-average value of the permutation entropy is calculated using the following formula. : (11) In the formula, The maximum value of the permutation entropy. This represents the minimum value of the permutation entropy. Let the critical threshold for the random component and the oscillating component be... The critical thresholds for oscillating and trend components are: ,when hour, The calculation formula is: (12) when hour, for and The average between; The calculation formula is: (13) Passing the critical threshold and The permutation entropy is divided into three categories, and the sub-components of each category are reconstructed into random components, oscillating components, and trend components.

6. The day-ahead NWP wind speed frequency division correction method considering spatiotemporal correlation and feature reconstruction according to claim 1, characterized in that, The TCN-LSTM-Transformer model includes a temporal convolutional network (TCN), an LSTM model, and a Transformer model. The LSTM model is placed after the multi-head attention mechanism module of the Transformer model to form the LSTM-Transformer model. The TCN-LSTM-Transformer model is used to correct the final low-frequency components, and the low-frequency corrected components are obtained by means of: The final low-frequency component is input into the temporal convolutional network (TCN), and the data processing procedure of the TCN includes: Temporal features are extracted layer by layer using causal convolution and dilated convolution. The calculation process is represented as follows: (15) In the formula, This represents the expansion factor, used to control the sampling interval. Indicates the first k Each convolutional kernel weight, Indicates the convolution kernel weights. This represents the historical index of the convolution kernel. x t-d·k Indicates the first Time series at each point in time y t Represents a convolutional layer t Output at any moment; The input and output of the convolutional layer are changed by N A stacked structure consisting of residual blocks, where the output of a single residual block is: (16) In the formula, Indicates the first n The output of each residual block This represents the activation function. Indicates the first n The input of each residual block, Represents the residual function; based on N The output of each residual block is used to obtain the final output value of the Temporal Convolutional Network (TCN) using the following formula: (17) In the formula, This represents the final output value of TCN. Represents the ReLU activation function. This represents the output sequence of the Nth residual block. Represents the weight matrix. Indicates the bias coefficient; The final output value of the Temporal Convolutional Network (TCN) is input into the LSTM-Transformer model. The data processing procedure of the LSTM-Transformer model includes: Using the final output value as the input feature, it is mapped to a high-dimensional space to obtain the embedded time series. Based on the embedded time series, the initial representation matrix input into the encoder is obtained using the following formula: (19) (20) (21) In the formula, The index value represents the time step, where, ; The index value represents the time step, where, P represents the position encoding matrix, and Z0 represents the initial matrix input to the encoder. Represents the embedded time series, d model This represents the dimension by which the input features are mapped to a higher-dimensional space; Based on the initial representation matrix shown, the process is performed layer by layer by a multi-layer encoder, with multi-head self-attention calculation performed at each layer to obtain the attention score for each head. The attention scores of each head are concatenated and linearly transformed, then passed through an LSTM model, and finally the matrix of the input feedforward network is obtained through residual connections and layer normalization. The calculation formula is as follows: (27) (28) (29) (30) In the formula, Indicates the first Output of size For multi-head output weight matrix, Representation layer normalization, This refers to computation using an LSTM model; Concat represents the concatenation operation. Let T represent the real number field, and T represent the total number of time steps. Indicates the first Attention score based on height This represents the multi-head attention output matrix. Indicates the first l A multi-head attention layer, Z l-1 H represents the output matrix of the encoder-decoder, and H represents the output matrix after the LSTM model operation. Based on matrix Then, perform feedforward neural network operations, and perform residual connections and layer normalization again to obtain the output matrix of one encoder layer; Based on the output matrix of each encoder layer, the low-frequency correction component is obtained using the following formula: (33) In the formula, This represents the output matrix of the Nth encoder. This represents the output weight matrix. This is a low-frequency correction component.

7. The day-ahead NWP wind speed frequency division correction method considering spatiotemporal correlation and feature reconstruction according to claim 6, characterized in that, The high-frequency components are corrected using a high-frequency component correction method that matches errors in historical and future periods, resulting in high-frequency corrected components, including: The similarity score between a future NWP high-frequency component sequence and each historical NWP high-frequency component sequence is calculated using the following formula: (35) In the formula, P i,j Indicates the first in history i The first high-frequency component and the future... j The score value of each high-frequency component; A and B Indicates the first and second coefficients. Pcc ( X h,i , Y f,j ) indicates the historical number i The first high-frequency component and the future... j Pearson correlation coefficients for each high-frequency component NMAE ( X h,i , Y f,j ) indicates the historical number i The first high-frequency component and the future... j Normalized average absolute error of high-frequency air volume X h,i Indicates the first in history i High-frequency components Y f,j Indicates the future number j One high-frequency component; Based on calculations, the future nth [number] is determined using the following formula. w Ranking of scores for each time period k The high-frequency error sequence corresponding to the time series: (37) In the formula, For the future w Ranking of scores for each time period k The high-frequency error sequence corresponding to the time series; This indicates the search for the ranked k Time series, For the future period The similarity score between each time period and each time period in history; Calculate the similarity score and rank the top k The average value of the high-frequency error sequence corresponding to the time series is used as the error correction sequence for future high-frequency components.

8. A day-ahead NWP wind speed frequency division correction device considering spatiotemporal correlation and feature reconstruction, characterized in that, The device includes: The wind farm screening module is configured to calculate the distance between the target wind farm and other wind farms, screen out wind farms whose distance from the target wind farm is less than a set distance, and calculate the score of each screened wind farm based on the historical actual wind speed of the screened wind farm and the historical actual wind speed of the target wind farm, and select the wind farms with the highest scores as reference wind farms. The actual wind speed correction module is configured to remove outliers from the actual wind speed-power curve using the quartile method under a multi-scale time window, arrange the wind speed data in descending order of quantity, and simultaneously change the power data according to the position of the wind speed data. An average wind speed-power curve is fitted to obtain the missing wind speed and power data. Based on the missing wind speed and power data, the actual wind speed data of the target wind farm is corrected to obtain the corrected actual wind speed data. The actual wind speed-power curve is determined based on the historical actual wind speeds of the target wind farm and a reference wind farm. The sequence decomposition module is configured to decompose the wind speed sequence and wind power sequence of the target wind farm respectively, obtaining their respective wind speed sub-components and power sub-components; the wind speed sequence includes corrected historical actual wind speed data, historical NWP wind speed data, and future NWP wind speed data; each wind speed sub-component and power sub-component is reconstructed into a trend component, an oscillation component, and a random component; the trend component and the oscillation component are added to obtain the low-frequency component of the first decomposition; the random component is decomposed a second time to obtain the secondary decomposition sub-components; the trend component and the oscillation component in the secondary decomposition sub-components are added to the low-frequency component of the first decomposition to obtain the final low-frequency component; the remaining random component after the secondary decomposition is taken as the high-frequency component. The low-frequency component correction module is configured to construct a TCN-LSTM-Transformer model to correct the final low-frequency component, thereby obtaining the low-frequency corrected component. The high-frequency component correction module is configured to correct the high-frequency component by using a high-frequency component correction method that matches errors in historical and future periods, to obtain a high-frequency corrected component. The component overlay module is configured to add the low-frequency correction component and the high-frequency correction component to obtain the NWP correction result for the future time period.

9. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes the computer execution instructions stored in the memory to implement the day-ahead NWP wind speed frequency division correction method considering spatiotemporal correlation and feature reconstruction as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the day-ahead NWP wind speed frequency division correction method considering spatiotemporal correlation and feature reconstruction as described in any one of claims 1-7.

Images