Wind power ultra-short-term power set prediction method based on similar graph recognition

By screening key meteorological factors based on similar graph identification and gray correlation analysis, combining WRF mode and Informer model, the poor quality of observation data and low weather forecast accuracy in wind power ultra-short-term power prediction are solved, and higher precision wind power is achieved Power prediction improves the stability of the power system.

CN118916703AInactive Publication Date: 2024-11-08STATE GRID ECONOMIC TECH RES INST CO LTD
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202410928090.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-11
Publication Date
2024-11-08
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

In the ultra-short-term power forecast of wind power, the poor quality of observation data, the low weather forecast accuracy, and the intermittent and uncertainty of wind speed, the power generation power of wind power stations is difficult to accurately predict, affecting the reliability of power scheduling and transactions.

Method used

The similarity dates to be predicted are screened using a method based on similar graph recognition, combined with gray correlation analysis and WRF mode dynamic downscale processing, high-precision meteorological data are obtained through multi-mode weighted ensemble averaging method, and the wind power output prediction model is constructed using the Informer model. Identify the effect of meteorological variables on wind power.

Benefits of technology

It improves the accuracy and reliability of ultra-short-term power prediction of wind power, provides more accurate data support for power scheduling and transactions, and enhances the stability of the power system.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118916703A_ABST
    Figure CN118916703A_ABST
Patent Text Reader

Abstract

The invention discloses a similar graph identification-based wind power ultra-short-term power set prediction method, which comprises the following steps of: eliminating abnormal historical data in a historical data set by adopting a K-means method, and filling abnormal values of the eliminated historical data based on a k-neighbor value method; screening similar days of the to-be-predicted day based on a similar graph identification method, and taking the similar days as sample data of the wind power ultra-short-term prediction model; key meteorological factors are screened on the basis of a grey correlation analysis method and serve as meteorological factors in the wind power ultra-short-term prediction model; performing power downscaling processing based on a WRF mode, and obtaining future key meteorological data of the wind power station by adopting a multi-mode weighted set average method; and constructing a wind power ultra-short-term prediction model based on Informer to carry out ultra-short-term prediction on the wind power. The method can accurately recognize the influence of each meteorological variable on the wind power, facilitates the improvement of the prediction precision, and achieves the ultra-short-term wind power prediction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a wind power ultra-short-term power aggregate prediction method, and in particular to a wind power ultra-short-term power aggregate prediction method based on similarity graph recognition. Background Art

[0002] Promoting energy transformation and developing a modern energy structure have become the main means for countries around the world to reduce carbon emissions. With the continuous construction of my country's new power system and the rapid decline in the cost of renewable energy power generation, renewable energy, mainly wind and solar energy, will provide most of the future electricity demand. The World Energy Outlook 2021 predicts that the global installed capacity of renewable energy will reach 4,153GW in 2023, accounting for about 70%. According to statistics from the National Energy Administration, by the end of April 2024, the installed capacity of wind power will be about 460GW, a year-on-year increase of 20.6%.

[0003] With the increase in wind power penetration, its volatility and randomness bring huge challenges to power dispatching, peak load regulation and consumption. Accurate power forecasting can provide data support for power dispatching, peak load regulation and power trading, and is an important means to ensure the reliability and safety of the power system. In terms of wind power station power forecasting, current research still has problems such as poor quality of observation data and low accuracy of weather forecasts, which makes it difficult to accurately predict the power generation of wind power stations. In addition, the strong intermittent and uncertain characteristics of wind speed also cause the instability of ultra-short-term power output of wind power to a certain extent. The accurate prediction of ultra-short-term wind power is highly dependent on weather forecasts. Different data drivers will cause uncertainty in the results, and weather elements have certain similarities in time and space. Summary of the invention

[0004] In order to solve the shortcomings of the above-mentioned technology, the present invention provides a wind power ultra-short-term power aggregate prediction method based on similarity graph recognition.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: a wind power ultra-short-term power aggregate prediction method based on similarity graph recognition, comprising the following steps:

[0006] Step S1: Acquire historical data and organize the historical data into a historical data set; use the K-means method to remove abnormal historical data in the historical data set, and fill in the abnormal values ​​of the removed historical data based on the k-nearest neighbor method;

[0007] Step S2: Based on the similarity graph recognition method, similar days to the day to be predicted are selected as sample data for the ultra-short-term wind power prediction model;

[0008] Step S3: Select key meteorological factors based on the grey correlation analysis method as meteorological factors in the ultra-short-term wind power prediction model;

[0009] Step S4: performing dynamic downscaling processing on the forecast data of multiple meteorological factors based on the WRF model, and adopting the multi-model weighted ensemble average method to obtain the future weather forecast results of the wind power station, i.e., the future key meteorological data of the wind power station;

[0010] Step S5: Construct a wind power output prediction model based on Informer to perform ultra-short-term prediction of wind power. Use the day to be tested in step S2 as a training sample, input meteorological factors of similar days; and output a short-term prediction of wind power output.

[0011] Furthermore, in step S1, the K-means method is used to remove abnormal historical data in the historical data set, which specifically includes the following steps:

[0012] Step S11: clustering the historical data using the K-means method to obtain K-class data sets;

[0013] Step S12: Estimating the standard deviation of each obtained data set based on the Bessel formula; the Bessel formula is:

[0014]

[0015] Where s is the standard deviation; l i is the sample element to be estimated; is the sample mean;

[0016] Step S13: Determine whether the data in the data set is abnormal data by using the Raida criterion; the Raida criterion formula is:

[0017] |x v -x|≥3s;

[0018] Among them, x is the average value at the closest distance to the cluster center; x v is the distance between the detection data and the cluster center;

[0019] If the absolute value of the difference between the distance of the detection data from the cluster center and the average value at the closest distance to the cluster center is not less than 3 times the standard deviation, the detection data is abnormal data.

[0020] Furthermore, the outliers of the historical data removed in step S1 are filled based on the k-nearest neighbor method, that is, the outliers in step S13 are filled; the formula of the k-nearest neighbor method is:

[0021]

[0022] Among them, x i-k is the first k data of the abnormal data; x i+k is the k data after the abnormal data.

[0023] Furthermore, in step S2, the similar days of the day to be predicted are screened based on the similar graph recognition method, which specifically includes the following steps:

[0024] Step S21: Set the i-th moment in a day to contain N meteorological elements and convert them into an N-dimensional meteorological vector v i ; Set the period of a moment to 15 minutes, and construct a meteorological vector group for one day, that is, convert it into 96 meteorological vector groups {v1, v2, ..., v 96};

[0025] Step S22: convert the inner products of the meteorological vector groups in step S21 into a 96×96 Gram matrix. The Gram matrix is:

[0026]

[0027] Among them, v i represents the i-th vector; v i ·v j Represents vector v i and v j The inner product of

[0028] Step S23: Perform maximum and minimum normalization on the Gram matrix of step S22 and map it to [0,1]. The formula for maximum and minimum normalization is:

[0029]

[0030] Among them, L i is the i-th element of the Gram matrix G; L' i For L i The maximum and minimum normalized results of L max is the maximum value among the elements of the Gram matrix; L min is the minimum value among the elements of the Gram matrix;

[0031] Step S24: Map the maximum and minimum normalization results to the color index interval of [1,255] to generate a picture with 96×96 pixels to generate a standardized sample picture;

[0032] Step S25: converting the meteorological data of each day into a 96×96 pixel picture, i.e., a comparison picture;

[0033] Step S26: Based on the structural similarity algorithm SSIM, the similarity comparison between the day to be predicted and each comparison graph of the training sample is evaluated, and the similar day corresponding to the similar graph is selected;

[0034] Step S27: Select M days that are most similar to the day to be tested from the historical data set as sample data using the structural similarity algorithm SSIM.

[0035] Furthermore, in step S24, the color index value corresponding to each element in the Gram matrix is ​​calculated, and the data is converted into an integer type for visualization in the picture; the formula for calculating the color index value corresponding to the element is:

[0036] k'=round(254L i )+1;

[0037] Where k' is the Gram matrix element L i The color index value of the

[0038] Further, the structural similarity algorithm in step S26 includes a brightness module algorithm, a contrast module algorithm and a structural module algorithm;

[0039] The calculation formula of the brightness module algorithm is:

[0040]

[0041] C1=(K1L) 2 ;

[0042]

[0043] Among them, x i is the pixel value of image x; N is the total number of pixels in image x; μ x is the average grayscale of image x; μ y is the average grayscale of image y; C1 is a constant used to avoid μ x 2 +μ y 2 Close to 0 will cause system instability; L is the gray level of the image, L = 255; K1 is a constant, K1 = 0.01; l(x, y) is the brightness contrast function of the two images;

[0044] The calculation formula of the contrast module algorithm is:

[0045]

[0046] C2=(K2L) 2 ;

[0047]

[0048] Among them, σ x is the variance of image x; K2 is a constant, K2=0.03; c(x,y) is the contrast function of the two images;

[0049] The calculation formula of the structural module algorithm is:

[0050]

[0051] Among them, σ xy is the covariance of the two images; s(x,y) is the structural contrast function of the two images;

[0052] The calculation formula of the structural similarity algorithm is:

[0053] SSIM(x,y)=[l(x,y)] α [c(x,y)] β [s(x,y)] γ ;

[0054] Set α = β = γ = 1; and C3 = C2 / 2; we get:

[0055]

[0056] If the calculated SSIM similarity is closer to 1, the two images are more similar, and if it is closer to -1, the two images are less similar.

[0057] Furthermore, in step S3, the key meteorological factors are screened based on the grey correlation analysis method, which specifically includes the following steps:

[0058] Step S31: converting meteorological data and power data into a meteorological factor matrix, where the meteorological factor matrix is:

[0059]

[0060] Among them, β is the total number of samples;

[0061] Step S32: Unify the dimensions of the meteorological factor matrix. The formula for unifying the dimensions is:

[0062]

[0063] Among them, β is the total number of samples;

[0064] The matrix after dimension unification is:

[0065]

[0066] Among them, β is the total number of samples;

[0067] Step S33: Calculate the difference between each meteorological factor sequence and the power sequence on each corresponding element, and calculate the correlation coefficient based on the difference between each corresponding element. The calculation formula of the correlation coefficient is:

[0068]

[0069] Among them, the value range of n is [1, N]; the value range of k is [1, β]; ρ is used to distinguish the correlation coefficient, and the value range of ρ is (0, 1);

[0070] Step S34: Based on the correlation coefficient, the correlation between each meteorological factor sequence and the power sequence is calculated. The correlation calculation formula is:

[0071]

[0072] Among them, the value range of k is [1,β];

[0073] Step S35: According to the correlation between each meteorological factor sequence and the power sequence, the key meteorological data that meets the similar day is selected.

[0074] Furthermore, in step S4, dynamic downscaling is performed on multiple sets of large-scale meteorological forecast data based on the WRF model, which specifically includes the following steps:

[0075] Step S41: perform WPS pre-processing in WRF mode; use geogrid decoder, ungrib decoder and metgrid decoder to process static ground data of the simulation area, the static ground data includes altitude, soil type and vegetation conditions, and interpolate the decompressed meteorological data to the selected simulation area to generate the initial field of the main processing module;

[0076] Step S42: using the main processing module real.exe and the executable program; real.exe reads the output file of WPS and combines it with the actual physical process scheme to provide an initial field for the integral operation; real.exe performs integral operation on the atmospheric physical process in the selected simulation area to extract the required key meteorological variables;

[0077] Step S43: using the spatial correlation coefficient R, the root mean square error RMSE, and the central standard deviation σ to verify the overall simulation capability of each forecast data set for key meteorological variables;

[0078] Step S44: using a multi-model weighted ensemble average method to obtain the future weather forecast results of the wind power station, calculate the interannual variation skill score IVS, and evaluate the skill of each model in recurring time variability;

[0079] Step S45: Calculate the weight W of each mode according to the calculated IVS of each mode, the ranking of the spatial correlation coefficient R, the root mean square error RMSE and the central standard deviation σ i ; The weight calculation formula is:

[0080]

[0081] Among them, W i is the weight of the pattern; R i is the comprehensive performance index of a single mode i, and a higher value indicates better mode skill; S i is the ranking of each mode; α is the number of modes.

[0082] Furthermore, the calculation formula of the spatial correlation coefficient R in step S43 is:

[0083]

[0084] The calculation formula of root mean square error RMSE is:

[0085]

[0086] The calculation formula for the centralized standard deviation σ is:

[0087]

[0088] Among them, A i is the observed value of the meteorological variable; S i is the simulated value of the meteorological variable;

[0089] The calculation formula of the IVS interannual variation skill score in step S44 is:

[0090]

[0091] Among them, std m is the standard deviation of the observed meteorological variable; std o is the standard deviation of the simulated meteorological variables.

[0092] Furthermore, in step S5, a wind power output prediction model based on the Informer model is constructed, including the following steps:

[0093] Step S51: taking similar days in the sample data as training sets and the days to be tested as validation sets; taking meteorological factor sequences and power sequences as inputs of the training sets; and constructing an ultra-short-term prediction model for power generation based on sample data based on historical data trained by the Informer model;

[0094] Step S52: Evaluate the accuracy of the ultra-short-term wind power forecasting model. The evaluation indicators include mean absolute percentage error (MAPE), root mean square error (RMSE), mean square error (MSE) and determination coefficient (R). 2 ;

[0095] The calculation formula for mean absolute percentage error MAPE is:

[0096]

[0097] The calculation formula of root mean square error RMSE is:

[0098]

[0099] The calculation formula for mean square error MSE is:

[0100]

[0101] Coefficient of determination R 2 The calculation formula is:

[0102]

[0103] Based on the daily forecast value of the output power of the ultra-short-term wind power forecast model, the daily accuracy of the daily forecast value is calculated. The calculation formula for the daily accuracy is:

[0104]

[0105] Among them, x i is the predicted value of power; i is the actual value of power; n is the number of data; P Mi is the actual power at the i-th moment; P Pi is the daily forecast value at the i-th moment; Cap is the rated capacity of the wind power station; n' is the number of samples in one day.

[0106] The present invention discloses a method for predicting wind power ultra-short-term power set based on similarity graph recognition. The method adopts the similarity graph recognition method to select similar days of the test day, and adopts the grey correlation analysis method to perform correlation analysis on multiple meteorological factors, so as to find out the key meteorological variables with the greatest correlation with the wind power sequence; and uses multiple sets of meteorological forecast data as driving sources, uses the WRF mode to perform dynamic downscaling to extract multiple sets of high-precision meteorological data; adopts the multi-mode weighted set method to obtain the future key meteorological data of the wind power station; and constructs a wind power ultra-short-term prediction model based on Informer to perform short-term prediction of wind power, which can accurately identify the influence of various meteorological variables on wind power, help to improve the prediction accuracy, and then realize ultra-short-term wind power prediction; effectively improves the reliability of wind power ultra-short-term power prediction, and provides technical support for power dispatching and peak regulation and power trading. BRIEF DESCRIPTION OF THE DRAWINGS

[0107] Figure 1 It is a flow chart of the present invention.

[0108] Figure 2 This is a flow chart of the WRF model dynamic downscaling in the present invention.

[0109] Figure 3This is a flow chart of building ultra-short-term wind power forecasting based on Informer in the present invention. DETAILED DESCRIPTION

[0110] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0111] like Figure 1 As shown in the flowchart of a method for predicting wind power ultra-short-term power aggregate based on similarity graph recognition, a method for predicting wind power ultra-short-term power aggregate based on similarity graph recognition comprises the following steps:

[0112] Step S1: Acquire historical data and organize the historical data into a historical data set; use the K-means method to remove abnormal historical data in the historical data set, and fill in the abnormal values ​​of the removed historical data based on the k-nearest neighbor method;

[0113] Step S2: Based on the similarity graph recognition method, similar days to the day to be predicted are selected as sample data for the ultra-short-term wind power prediction model;

[0114] Step S3: Select key meteorological factors based on the grey correlation analysis method as meteorological factors in the ultra-short-term wind power prediction model;

[0115] Step S4: performing dynamic downscaling processing on the forecast data of multiple meteorological factors based on the WRF model, and adopting the multi-model weighted ensemble average method to obtain the future weather forecast results of the wind power station, i.e., the future key meteorological data of the wind power station;

[0116] Step S5: Based on Informer, a wind power output prediction model, namely, an ultra-short-term wind power prediction model, is constructed to perform ultra-short-term wind power prediction, using the day to be tested in step S2 as a training sample, inputting meteorological factors of similar days; and outputting a short-term prediction of wind power output.

[0117] Specifically, in step S1, the K-means method is used to remove abnormal historical data in the historical data set, which specifically includes the following steps:

[0118] Step S11: clustering the historical data using the K-means method to obtain K-class data sets;

[0119] Step S12: Estimating the standard deviation of each obtained data set based on the Bessel formula; the Bessel formula is:

[0120]

[0121] Where s is the standard deviation; l i is the sample element to be estimated; is the sample mean;

[0122] Step S13: Determine whether the data in the data set is abnormal data by using the Raida criterion; the Raida criterion formula is:

[0123] |x v -x|≥3s;

[0124] Among them, x is the average value at the closest distance to the cluster center; x v is the distance between the detection data and the cluster center;

[0125] If the absolute value of the difference between the distance of the detection data from the cluster center and the average value at the closest distance to the cluster center is not less than 3 times the standard deviation, the detection data is abnormal data.

[0126] In step S1, the outliers of the historical data removed are filled based on the k-nearest neighbor method, that is, the outliers in step S13 are filled; the formula of the k-nearest neighbor method is:

[0127]

[0128] Among them, x i-k is the first k data of the abnormal data; x i+k is the k data after the abnormal data.

[0129] In step S2, the similar days to the day to be predicted are screened based on the similar graph recognition method, which specifically includes the following steps:

[0130] Step S21: Set the i-th moment in a day to contain N meteorological elements and convert them into an N-dimensional meteorological vector v i ; Set the period of a moment to 15 minutes, and construct a meteorological vector group for one day, that is, convert it into 96 meteorological vector groups {v1, v2, ..., v 96};

[0131] Step S22: convert the inner products of the meteorological vector groups in step S21 into a 96×96 Gram matrix. The Gram matrix is:

[0132]

[0133] Among them, v i represents the i-th vector; v i ·v j Represents vector v i and v j The inner product of

[0134] Step S23: Perform maximum and minimum normalization on the Gram matrix of step S22 and map it to [0,1]. The formula for maximum and minimum normalization is:

[0135]

[0136] Among them, L i is the i-th element of the Gram matrix G; L' i For L i The maximum and minimum normalized results of L max is the maximum value among the elements of the Gram matrix; L min is the minimum value among the elements of the Gram matrix;

[0137] Step S24: Map the maximum and minimum normalization results to the color index interval of [1,255] to generate a picture with 96×96 pixels to generate a standardized sample picture;

[0138] Step S25: converting the meteorological data of each day into a 96×96 pixel picture, i.e., a comparison picture;

[0139] Step S26: Based on the structural similarity algorithm SSIM, the similarity comparison between the day to be predicted and each comparison graph of the training sample is evaluated, and the similar day corresponding to the similar graph is selected;

[0140] Step S27: Select M days that are most similar to the day to be tested from the historical data set as sample data using the structural similarity algorithm SSIM.

[0141] In step S24, the color index value corresponding to each element in the Gram matrix is ​​calculated, and the data is converted into an integer type for visualization in the picture; the formula for calculating the color index value corresponding to the element is:

[0142] k'=round(254L i )+1;

[0143] Where k' is the Gram matrix element L i The color index value of the

[0144] In step S26, the structural similarity algorithm includes a brightness module algorithm, a contrast module algorithm and a structural module algorithm;

[0145] The calculation formula of the brightness module algorithm is:

[0146]

[0147] C1=(K1L) 2 ;

[0148]

[0149] Among them, x i is the pixel value of image x; N is the total number of pixels in image x; μ x is the average grayscale of image x; μ yis the average grayscale of image y; C1 is a constant used to avoid μ x 2 +μ y 2 Close to 0 will cause system instability; L is the gray level of the image, L = 255; K1 is a constant, K1 = 0.01; l(x, y) is the brightness contrast function of the two images;

[0150] The calculation formula of the contrast module algorithm is:

[0151]

[0152] C2=(K2L) 2 ;

[0153]

[0154] Among them, σ x is the variance of image x; K2 is a constant, K2=0.03; c(x,y) is the contrast function of the two images;

[0155] The calculation formula of the structural module algorithm is:

[0156]

[0157] Among them, σ xy is the covariance of the two images; s(x,y) is the structural contrast function of the two images;

[0158] The calculation formula of the structural similarity algorithm is:

[0159] SSIM(x,y)=[l(x,y)] α [c(x,y)] β [s(x,y)] γ ;

[0160] Set α = β = γ = 1; and C3 = C2 / 2; we get:

[0161]

[0162] If the calculated SSIM similarity is closer to 1, the two images are more similar, and if it is closer to -1, the two images are less similar.

[0163] In step S3, the key meteorological factors are screened based on the grey correlation analysis method, which specifically includes the following steps:

[0164] Step S31: converting meteorological data and power data into a meteorological factor matrix, where the meteorological factor matrix is:

[0165]

[0166] Among them, β is the total number of samples;

[0167] Step S32: Unify the dimensions of the meteorological factor matrix. The formula for unifying the dimensions is:

[0168]

[0169] Among them, β is the total number of samples;

[0170] The matrix after dimension unification is:

[0171]

[0172] Among them, β is the total number of samples;

[0173] Step S33: Calculate the difference between each meteorological factor sequence and the power sequence on each corresponding element, and calculate the correlation coefficient based on the difference between each corresponding element. The calculation formula of the correlation coefficient is:

[0174]

[0175] Among them, the value range of n is [1, N]; the value range of k is [1, β]; ρ is used to distinguish the correlation coefficient, and the value range of ρ is (0, 1);

[0176] Step S34: Based on the correlation coefficient, the correlation between each meteorological factor sequence and the power sequence is calculated. The correlation calculation formula is:

[0177]

[0178] Among them, the value range of k is [1,β];

[0179] Step S35: According to the correlation between each meteorological factor sequence and the power sequence, the key meteorological data that meets the similar day is selected.

[0180] In step S4, multiple sets of large-scale meteorological forecast data are dynamically downscaled based on the WRF model, which specifically includes the following steps:

[0181] Step S41: perform WPS pre-processing in WRF mode; use geogrid decoder, ungrib decoder and metgrid decoder to process static ground data of the simulation area, the static ground data includes altitude, soil type and vegetation conditions, and interpolate the decompressed meteorological data to the selected simulation area to generate the initial field of the main processing module;

[0182] Step S42: using the main processing module real.exe and the executable program; real.exe reads the output file of WPS and combines it with the actual physical process scheme to provide an initial field for the integral operation; real.exe performs integral operation on the atmospheric physical process in the selected simulation area to extract the required key meteorological variables;

[0183] Step S43: using the spatial correlation coefficient R, the root mean square error RMSE, and the central standard deviation σ to verify the overall simulation capability of each forecast data set for key meteorological variables;

[0184] Step S44: using a multi-model weighted ensemble average method to obtain the future weather forecast results of the wind power station, calculate the interannual variation skill score IVS, and evaluate the skill of each model in recurring time variability;

[0185] Step S45: Calculate the weight W of each mode according to the calculated IVS of each mode, the ranking of the spatial correlation coefficient R, the root mean square error RMSE and the central standard deviation σ i ; The weight calculation formula is:

[0186]

[0187] Among them, W i is the weight of the pattern; R i is the comprehensive performance index of a single mode i, and a higher value indicates better mode skill; S i is the ranking of each mode; α is the number of modes.

[0188] The calculation formula of the spatial correlation coefficient R in step S43 is:

[0189]

[0190] The calculation formula of root mean square error RMSE is:

[0191]

[0192] The calculation formula for the centralized standard deviation σ is:

[0193]

[0194] Among them, A i is the observed value of the meteorological variable; S i are simulated values ​​of meteorological variables.

[0195] The calculation formula of the IVS interannual variation skill score in step S44 is:

[0196]

[0197] Among them, stdm is the standard deviation of the observed meteorological variable; std o is the standard deviation of the simulated meteorological variables.

[0198] In step S5, a wind power output prediction model based on the Informer model is constructed, including the following steps:

[0199] Step S51: taking similar days in the sample data as training sets and the days to be tested as validation sets; taking meteorological factor sequences and power sequences as inputs of the training sets; and constructing an ultra-short-term prediction model for power generation based on sample data based on historical data trained by the Informer model;

[0200] Step S52: Evaluate the accuracy of the ultra-short-term wind power forecasting model. The evaluation indicators include mean absolute percentage error (MAPE), root mean square error (RMSE), mean square error (MSE) and determination coefficient (R). 2 ;

[0201] The calculation formula for mean absolute percentage error MAPE is:

[0202]

[0203] The calculation formula of root mean square error RMSE is:

[0204]

[0205] The calculation formula for mean square error MSE is:

[0206]

[0207] Coefficient of determination R 2 The calculation formula is:

[0208]

[0209] Based on the daily forecast value of the output power of the ultra-short-term wind power forecast model, the daily accuracy of the daily forecast value is calculated. The calculation formula for the daily accuracy is:

[0210]

[0211] Among them, x i is the predicted value of power; i is the actual value of power; n is the number of data; P Mi is the actual power at the i-th moment; P Pi is the daily forecast value at the i-th moment; Cap is the rated capacity of the wind power station; n' is the number of samples in one day.

[0212] Embodiment 1 provides a specific process for a wind power ultra-short-term power aggregate prediction method based on similarity graph recognition, including the following steps:

[0213] Step 1: Obtain historical data and organize the historical data into a historical data set; use the K-means method to remove abnormal historical data in the historical data set, and fill in the abnormal values ​​of the removed historical data based on the k-nearest neighbor method;

[0214] Step 11: Use K-means method to cluster the historical data to obtain K-class data sets;

[0215] Step 12: Estimate the standard deviation of each data set obtained based on the Bessel formula; the Bessel formula is:

[0216]

[0217] Where s is the standard deviation; l i is the sample element to be estimated; is the sample mean;

[0218] Step 13: Use the Raida criterion to determine whether the data in the data set is abnormal data; the Raida criterion formula is:

[0219] |x v -x|≥3s;

[0220] Among them, x is the average value at the closest distance to the cluster center; x v is the distance between the detection data and the cluster center;

[0221] If the absolute value of the difference between the distance of the detection data from the cluster center and the average value at the closest distance to the cluster center is not less than 3 times the standard deviation, the detection data is abnormal data.

[0222] In step 1, the outliers of the historical data removed are filled based on the k-nearest neighbor method, that is, the outliers in step S13 are filled; the formula of the k-nearest neighbor method is:

[0223]

[0224] Among them, x i-k is the first k data of the abnormal data; x i+k is the k data after the abnormal data.

[0225] Step 2: Based on the similarity graph recognition method, similar days to the day to be predicted are selected as sample data for the ultra-short-term wind power prediction model;

[0226] Step 21: Set the i-th moment of the day to contain N meteorological elements and convert them into an N-dimensional meteorological vector v i; Set the period of a moment to 15 minutes, and construct a meteorological vector group for one day, that is, convert it into 96 meteorological vector groups {v1, v2, ..., v 96};

[0227] Step 22: Convert the inner products of the meteorological vector groups in step S21 into a 96×96 Gram matrix. The Gram matrix is:

[0228]

[0229] Among them, v i represents the i-th vector; v i ·v j Represents vector v i and v j The inner product of

[0230] Step 23: Perform maximum and minimum normalization on the Gram matrix of step S22 and map it to [0,1]. The formula for maximum and minimum normalization is:

[0231]

[0232] Among them, L i is the i-th element of the Gram matrix G; L' i For L i The maximum and minimum normalized results of L max is the maximum value among the elements of the Gram matrix; L min is the minimum value among the elements of the Gram matrix;

[0233] Step 24: Map the maximum and minimum normalization results to the color index interval of [1,255] to generate a standardized sample image with 96×96 pixels;

[0234] Calculate the color index value corresponding to each element in the Gram matrix and convert the data into integer type for visualization in the picture; the formula for calculating the color index value corresponding to the element is:

[0235] k'=round(254L i )+1;

[0236] Where k' is the Gram matrix element L i The color index value of

[0237] Step 25: Convert the meteorological data of each day into a 96×96 pixel picture, i.e., a comparison picture;

[0238] Step 26: Based on the structural similarity algorithm SSIM, the predicted day is evaluated for similarity with each comparison graph of the training sample, and the similar day corresponding to the similar graph is selected; the structural similarity algorithm includes a brightness module algorithm, a contrast module algorithm and a structure module algorithm;

[0239] The calculation formula of the brightness module algorithm is:

[0240]

[0241] C1=(K1L) 2

[0242]

[0243] Among them, x i is the pixel value of image x; N is the total number of pixels in image x; μ x is the average grayscale of image x; μ y is the average grayscale of image y; C1 is a constant used to avoid μ x 2 +μ y 2 Close to 0 will cause system instability; L is the gray level of the image, L = 255; K1 is a constant, K1 = 0.01; l(x, y) is the brightness contrast function of the two images;

[0244] The calculation formula of the contrast module algorithm is:

[0245]

[0246] C2=(K2L) 2 ;

[0247]

[0248] Among them, σ x is the variance of image x; K2 is a constant, K2=0.03; c(x,y) is the contrast function of the two images;

[0249] The calculation formula of the structural module algorithm is:

[0250]

[0251] Among them, σ xy is the covariance of the two images; s(x,y) is the structural contrast function of the two images;

[0252] The calculation formula of the structural similarity algorithm is:

[0253] SSIM(x,y)=[l(x,y)] α [c(x,y)]β [s(x,y)] γ ;

[0254] Set α = β = γ = 1; and C3 = C2 / 2; we get:

[0255]

[0256] If the calculated SSIM similarity is closer to 1, the two images are more similar, and if it is closer to -1, the two images are less similar.

[0257] Step 27: Select M days that are most similar to the day to be tested from the historical data set using the structural similarity algorithm SSIM as sample data.

[0258] Step 3: Based on the grey correlation analysis method, key meteorological factors are selected as the determined meteorological factors in the ultra-short-term wind power prediction model;

[0259] Step 31: Convert the meteorological data and power data into a meteorological factor matrix. The meteorological factor matrix is:

[0260]

[0261] Among them, β is the total number of samples;

[0262] Step 32: Unify the dimensions of the meteorological factor matrix. The formula for dimension unification is:

[0263]

[0264] Among them, β is the total number of samples;

[0265] The matrix after dimension unification is:

[0266]

[0267] Among them, β is the total number of samples;

[0268] Step 33: Calculate the difference between each meteorological factor sequence and the power sequence on each corresponding element, and calculate the correlation coefficient based on the difference between each corresponding element. The calculation formula of the correlation coefficient is:

[0269]

[0270] Among them, the value range of n is [1, N]; the value range of k is [1, β]; ρ is used to distinguish the correlation coefficient, and the value range of ρ is (0, 1);

[0271] Step 34: Based on the correlation coefficient, calculate the correlation between each meteorological factor sequence and the power sequence. The correlation calculation formula is:

[0272]

[0273] Among them, the value range of k is [1,β];

[0274] Step 35: According to the correlation between each meteorological factor sequence and the power sequence, select key meteorological data that match similar days.

[0275] Step 4: Perform dynamic downscaling on multiple sets of large-scale meteorological forecast data based on the WRF model, and use the multi-model weighted ensemble average method to obtain the future weather forecast results of the wind power station, that is, the future key meteorological data of the wind power station; the system includes the following steps:

[0276] Step 41: Perform WPS pre-processing of the WRF mode; use the geogrid decoder, ungrib decoder and metgrid decoder to process the static ground data of the simulation area, the static ground data includes altitude, soil type and vegetation conditions, and interpolate the decompressed meteorological data to the selected simulation area to generate the initial field of the main processing module;

[0277] Step 42: Using the main processing module real.exe and the executable program; real.exe reads the output file of WPS and combines it with the actual physical process scheme to provide an initial field for the integral operation; real.exe performs integral operation on the atmospheric physical process in the selected simulation area to extract the required key meteorological variables;

[0278] Step 43: Use the spatial correlation coefficient R, root mean square error RMSE, and centralized standard deviation σ to verify the overall simulation capability of each forecast data set for key meteorological variables; the calculation formula of the spatial correlation coefficient R in step S43 is:

[0279]

[0280] The calculation formula of root mean square error RMSE is:

[0281]

[0282] The calculation formula for the centralized standard deviation σ is:

[0283]

[0284] Among them, A i is the observed value of the meteorological variable; S i is the simulated value of the meteorological variable;

[0285] Step 44: Use the multi-model weighted ensemble average method to obtain the future weather forecast results of the wind power station and calculate the interannual variation skill score IVS to evaluate the skill of each model in reproducing the time variability;

[0286] The calculation formula for the IVS interannual change skill score is:

[0287]

[0288] Among them, std m is the standard deviation of the observed meteorological variable; std o is the standard deviation of the simulated meteorological variables;

[0289] Step 45: Calculate the weight W of each mode based on the calculated IVS of each mode, using the ranking of the spatial correlation coefficient R, the root mean square error RMSE and the central standard deviation σ i ; The weight calculation formula is:

[0290]

[0291] Among them, W i is the weight of the pattern; R i is the comprehensive performance index of a single mode i, and a higher value indicates better mode skill; S i is the ranking of each mode; α is the number of modes.

[0292] Step 5: Based on Informer, a wind power output prediction model, i.e., an ultra-short-term wind power prediction model, is constructed to predict wind power in the ultra-short term. The day to be tested in step S2 is used as a training sample, and meteorological factors of similar days are input; the short-term prediction of wind power output is output;

[0293] In step 5, a wind power output prediction model based on the Informer model is constructed, which includes the following steps:

[0294] Step 51: Use similar days in the sample data as training sets and the days to be tested as validation sets; use meteorological factor sequences and power sequences as inputs of the training sets; and construct an ultra-short-term prediction model for power generation based on sample data based on historical data trained by the Informer model.

[0295] Step 52: Evaluate the accuracy of the ultra-short-term wind power forecasting model. The evaluation indicators include mean absolute percentage error (MAPE), root mean square error (RMSE), mean square error (MSE) and determination coefficient (R). 2 ;

[0296] The calculation formula for mean absolute percentage error MAPE is:

[0297]

[0298] The calculation formula of root mean square error RMSE is:

[0299]

[0300] The calculation formula for mean square error MSE is:

[0301]

[0302] Coefficient of determination R 2 The calculation formula is:

[0303]

[0304] Based on the daily forecast value of the output power of the ultra-short-term wind power forecast model, the daily accuracy of the daily forecast value is calculated. The calculation formula for the daily accuracy is:

[0305]

[0306] Among them, x i is the predicted value of power; i is the actual value of power; n is the number of data; P Mi is the actual power at the i-th moment; P Pi is the daily forecast value at the i-th moment; Cap is the rated capacity of the wind power station; n' is the number of samples in one day. It can accurately identify the impact of various meteorological variables on wind power, which helps to improve the prediction accuracy and thus achieve ultra-short-term wind power prediction; effectively improve the reliability of ultra-short-term wind power prediction, and provide technical support for power dispatching and peak regulation and power trading.

[0307] Example 2 provides an implementation method of WRF model dynamic downscaling for a wind power ultra-short-term power aggregate prediction method based on similarity graph recognition, such as Figure 2 The specific steps of WRF model dynamic downscaling in a wind power ultra-short-term power aggregate prediction method based on similarity graph recognition are shown in the flowchart of WRF model dynamic downscaling.

[0308] Step A1: Start collecting static terrain data;

[0309] Step A2: Input static terrain data to set the longitude and latitude of the wind power station and grid information (namelist.wps) for WPS pre-processing. WPS pre-processing is a pre-processing process for providing input for real data simulation. The model is defined by geogrid.exe, and the static terrain data is interpolated to the model network; meteorological data is extracted from the GRIB format file by ungrib.exe; finally, the meteorological data extracted by ungrib.exe is horizontally interpolated to the simulation domain grid defined by geogrid.exe by metgrid.exe;

[0310] Step A3: Carry out the simulation area experimental parameter plan (nam.elist.input) and use real.exe to execute the program;

[0311] Step A4: Execute the wrf.exe program to extract key meteorological data.

[0312] Embodiment 3 is a method for predicting wind power ultra-short-term power aggregate based on similarity graph recognition. It provides a prediction model based on Informer to construct wind power ultra-short-term prediction, such as Figure 3 As shown in the flowchart of constructing wind power ultra-short-term power prediction based on Informer in a wind power ultra-short-term power aggregate prediction method based on similarity graph recognition, the wind power ultra-short-term power prediction based on Informer specifically includes the following steps:

[0313] Step B1: Set the model input sequence and pre-process the data; collect historical meteorological data and wind power output data of the power station;

[0314] Step B2: Screen the abnormal data of historical meteorological data and wind power output data using the K-means method, fill in the abnormal data based on the k-nearest neighbor method, screen important meteorological elements based on the grey correlation, and construct a sample data set;

[0315] Step B3: Apply the Gram matrix and matrix color for mapping; perform structural similarity detection on the historical weather image and the weather image of the day to be tested; the structural similarity detection includes training samples and verification samples, taking N meteorological elements as the input of the training sample and the power generation as the output; based on the training sample, i.e. the training set, and the verification sample, i.e. the verification set, an ultra-short-term wind power prediction model is constructed based on Informer;

[0316] Step B4: Obtain wind power forecast results based on the ultra-short-term wind power forecast model, and evaluate the performance of the ultra-short-term wind power forecast model, including MAPE, RMSE, MSE, R 2 The daily accuracy of the daily forecast value is evaluated.

[0317] The above implementation modes are not limitations of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the scope of the technical solution of the present invention also belong to the protection scope of the present invention.

Claims

1. A wind power ultra-short-term power aggregate prediction method based on similarity graph recognition, characterized in that: The following steps are involved: Step S1: Acquire historical data and organize the historical data into a historical data set; use the K-means method to remove abnormal historical data in the historical data set, and fill in the abnormal values ​​of the removed historical data based on the k-nearest neighbor method; Step S2: Based on the similarity graph recognition method, similar days to the day to be predicted are selected as sample data for the ultra-short-term wind power prediction model; Step S3: Select key meteorological factors based on the grey correlation analysis method as meteorological factors in the ultra-short-term wind power prediction model; Step S4: performing dynamic downscaling processing on the forecast data of multiple meteorological factors based on the WRF model, and adopting the multi-model weighted ensemble average method to obtain the future weather forecast results of the wind power station, i.e., the future key meteorological data of the wind power station; Step S5: Based on Informe, a wind power output prediction model is constructed to perform ultra-short-term prediction of wind power. The test day in step S2 is used as a training sample, and the meteorological factors of similar days are input; the short-term prediction of wind power output is output.

2. The method for wind power ultra-short-term power aggregate prediction based on similarity graph recognition according to claim 1 is characterized in that: In step S1, the K-means method is used to remove abnormal historical data in the historical data set, which specifically includes the following steps: Step S11: clustering the historical data using the K-means method to obtain K-class data sets; Step S12: Estimating the standard deviation of each obtained data set based on the Bessel formula; the Bessel formula is: Where s is the standard deviation; l i is the sample element to be estimated; is the sample mean; Step S13: Determine whether the data in the data set is abnormal data by using the Raida criterion; the Raida criterion formula is: |x v -x|≥3s; Among them, x is the average value at the closest distance to the cluster center; x v is the distance between the detection data and the cluster center; If the absolute value of the difference between the distance of the detection data from the cluster center and the average value at the closest distance to the cluster center is not less than 3 times the standard deviation, the detection data is abnormal data.

3. The method for wind power ultra-short-term power aggregate prediction based on similarity graph recognition according to claim 1 is characterized in that: In step S1, the outliers of the historical data removed are filled based on the k-nearest neighbor method, that is, the outliers in step S13 are filled; the formula of the k-nearest neighbor method is: Among them, x i-k is the first k data of the abnormal data; x i+k is the k data after the abnormal data.

4. The method for wind power ultra-short-term power aggregate prediction based on similarity graph recognition according to claim 1 is characterized in that: In step S2, similar days to the day to be predicted are screened based on the similar graph recognition method, specifically comprising the following steps: Step S21: Set the i-th moment in a day to contain N meteorological elements and convert them into an N-dimensional meteorological vector v i ; Set the period of a moment to 15 minutes, and construct a meteorological vector group for one day, that is, convert it into 96 meteorological vector groups {v1, v2, ..., v 96 }; Step S22: convert the inner products of the meteorological vector groups in step S21 into a 96×96 Gram matrix. The Gram matrix is: Among them, v i represents the i-th vector; v i ·v j Represents vector v i and v j The inner product of Step S23: Perform maximum and minimum normalization on the Gram matrix of step S22 and map it to [0,1]. The formula for maximum and minimum normalization is: Among them, L i is the i-th element of the Gram matrix G; L' i For L i The maximum and minimum normalized results of L max is the maximum value among the elements of the Gram matrix; L min is the minimum value among the elements of the Gram matrix; Step S24: Map the maximum and minimum normalization results to the color index interval of [1,255] to generate a picture with 96×96 pixels to generate a standardized sample picture; Step S25: converting the meteorological data of each day into a 96×96 pixel picture, i.e., a comparison picture; Step S26: Based on the structural similarity algorithm SSIM, the similarity comparison between the day to be predicted and each comparison graph of the training sample is evaluated, and the similar day corresponding to the similar graph is selected; Step S27: Select M days that are most similar to the day to be tested from the historical data set as sample data using the structural similarity algorithm SSIM.

5. The method for wind power ultra-short-term power aggregate prediction based on similarity graph recognition according to claim 4 is characterized in that: In step S24, the color index value corresponding to each element in the Gram matrix is ​​calculated, and the data is converted into an integer type for visualization in the picture; the formula for calculating the color index value corresponding to the element is: k’=round(254L i )+1; Where k' is the Gram matrix element L i The color index value of the 6. The method for wind power ultra-short-term power aggregate prediction based on similarity graph recognition according to claim 4 is characterized in that: The structural similarity algorithm in step S26 includes a brightness module algorithm, a contrast module algorithm and a structural module algorithm; The calculation formula of the brightness module algorithm is: C1=(K1L) 2 ; Among them, x i is the pixel value of image x; N is the total number of pixels in image x; μ x is the average grayscale of image x; μ y is the average grayscale of image y; C1 is a constant used to avoid μ x 2 +μ y 2 Close to 0 will cause system instability; L is the gray level of the image, L = 255; K1 is a constant, K1 = 0.01; l(x, y) is the brightness contrast function of the two images; The calculation formula of the contrast module algorithm is: C2=(K2L) 2 ; Among them, σ x is the variance of image x; K2 is a constant, K2=0.03; c(x,y) is the contrast function of the two images; The calculation formula of the structural module algorithm is: Among them, σ xy is the covariance of the two images; s(x,y) is the structural contrast function of the two images; The calculation formula of the structural similarity algorithm is: SSIM(x,y)=[l(x,y)] α [c(x,y)] β [s(x,y)] γ ; Set α = β = γ = 1; and C3 = C2 / 2; we get: If the calculated SSIM similarity is closer to 1, the two images are more similar, and if it is closer to -1, the two images are less similar.

7. The method for wind power ultra-short-term power aggregate prediction based on similarity graph recognition according to claim 1 is characterized in that: The step S3 selects key meteorological factors based on the grey correlation analysis method, which specifically includes the following steps: Step S31: converting meteorological data and power data into a meteorological factor matrix, where the meteorological factor matrix is: Among them, β is the total number of samples; Step S32: Unify the dimensions of the meteorological factor matrix. The formula for unifying the dimensions is: Among them, β is the total number of samples; The matrix after dimension unification is: Among them, β is the total number of samples; Step S33: Calculate the difference between each meteorological factor sequence and the power sequence on each corresponding element, and calculate the correlation coefficient based on the difference between each corresponding element. The calculation formula of the correlation coefficient is: Among them, the value range of n is [1, N]; the value range of k is [1, β]; ρ is used to distinguish the correlation coefficient, and the value range of ρ is (0, 1); Step S34: Based on the correlation coefficient, the correlation between each meteorological factor sequence and the power sequence is calculated. The correlation calculation formula is: Among them, the value range of k is [1,β]; Step S35: According to the correlation between each meteorological factor sequence and the power sequence, the key meteorological data that meets the similar day is selected.

8. The method for wind power ultra-short-term power aggregate prediction based on similarity graph recognition according to claim 1 is characterized by: In step S4, dynamic downscaling is performed on multiple sets of large-scale meteorological forecast data based on the WRF model, which specifically includes the following steps: Step S41: perform WPS pre-processing in WRF mode; use geogrid decoder, ungrib decoder and metgrid decoder to process static ground data of the simulation area, the static ground data includes altitude, soil type and vegetation conditions, and interpolate the decompressed meteorological data to the selected simulation area to generate the initial field of the main processing module; Step S42: using the main processing module real.exe and the executable program; real.exe reads the output file of WPS and combines it with the actual physical process scheme to provide an initial field for the integral operation; real.exe performs integral operation on the atmospheric physical process in the selected simulation area to extract the required key meteorological variables; Step S43: using the spatial correlation coefficient R, the root mean square error RMSE, and the central standard deviation σ to verify the overall simulation capability of each forecast data set for key meteorological variables; Step S44: using a multi-model weighted ensemble average method to obtain the future weather forecast results of the wind power station, calculate the interannual variation skill score IVS, and evaluate the skill of each model in recurring time variability; Step S45: Calculate the weight W of each mode according to the calculated IVS of each mode, the ranking of the spatial correlation coefficient R, the root mean square error RMSE and the central standard deviation σ i ; The weight calculation formula is: Among them, W i is the weight of the pattern; R i is the comprehensive performance index of a single mode i, and a higher value indicates better mode skill; S i is the ranking of each mode; α is the number of modes.

9. The method for wind power ultra-short-term power aggregate prediction based on similarity graph recognition according to claim 8 is characterized in that: The calculation formula of the spatial correlation coefficient R in step S43 is: The calculation formula of root mean square error RMSE is: The calculation formula for the centralized standard deviation σ is: Among them, A i is the observed value of the meteorological variable; S i is the simulated value of the meteorological variable; The calculation formula of the IVS interannual variation skill score in step S44 is: Among them, std m is the standard deviation of the observed meteorological variable; std o is the standard deviation of the simulated meteorological variables.

10. The method for wind power ultra-short-term power aggregate prediction based on similarity graph recognition according to claim 1, characterized in that: In step S5, a wind power output prediction model based on the Informer model is constructed. The following steps are involved: Step S51: taking similar days in the sample data as training sets and the days to be tested as validation sets; taking meteorological factor sequences and power sequences as inputs of the training sets; and constructing an ultra-short-term prediction model for power generation based on sample data based on historical data trained by the Informer model; Step S52: Evaluate the accuracy of the ultra-short-term wind power forecasting model. The evaluation indicators include mean absolute percentage error (MAPE), root mean square error (RMSE), mean square error (MSE) and determination coefficient (R). 2 ; The calculation formula for mean absolute percentage error MAPE is: The calculation formula of root mean square error RMSE is: The calculation formula for mean square error MSE is: Coefficient of determination R 2 The calculation formula is: Based on the daily forecast value of the output power of the ultra-short-term wind power forecast model, the daily accuracy of the daily forecast value is calculated. The calculation formula for the daily accuracy is: Among them, x i is the predicted value of power; i is the actual value of power; n is the number of data; P Mi is the actual power at the i-th moment; P Pi is the daily forecast value at the i-th moment; Cap is the rated capacity of the wind power station; n' is the number of samples in one day.

Citation Information

Patent Citations

  • Short-term wind power prediction method based on similar days and FRS-SVM

    CN112132344A

  • Ultra-short-term wind power combined prediction method and system

    CN113505938A

  • Combined power station operation optimization method and system based on high-precision wind and light output prediction

    CN115714420A

  • Wind power prediction method and system and storable medium

    CN118199053A