A method for ultra-short-term forecasting of wind power considering volatility and periodicity

Abstract

The present invention discloses an ultra-short-term prediction method for wind power considering volatility and periodicity, which belongs to the technical field of renewable energy prediction, and includes the following steps: S1, establishing a wind power data mining method to obtain data volatility composition and periodic composition; S2, extracting fusion features from original wind power data through a deep learning method; S3, making predictions based on fusion features and volatility and periodic composition through an integrated neural network model; S4, determining the final ultra-short-term prediction method for wind power through the volatility composition and periodic composition obtained by S1, the fusion features of S2 and the integrated neural network of S3. The present invention adopts the above-mentioned ultra-short-term prediction method for wind power considering volatility and periodicity, introduces deep learning theory into wind power prediction, studies data mining methods and the application of deep learning methods in ultra-short-term wind power prediction problems, improves power grid stability and power supply reliability, and is of great significance to the development of diversified energy systems.

Description

Technical Field

[0001] The present invention relates to the technical field of renewable energy prediction, and in particular to an ultra-short-term prediction method for wind power taking into account volatility and periodicity. Background Art

[0002] With the rapid development of urbanization, energy demand is also growing. As a zero-emission renewable energy source, wind power can reduce air pollution, improve air quality, and promote the sustainable development of the economy and society. The use of wind power to generate electricity can reduce dependence on imported energy and improve the security of energy supply. In order to give full play to the potential of wind power, energy experts and scholars from various countries have been committed to wind power issues for many years, from the initial development of more efficient wind turbines to renewable energy systems that combine wind power and energy storage technologies. However, there are still areas to be studied in the research of wind power technology, and it is difficult to give accurate results for the existing ultra-short-term prediction of wind power.

[0003] In view of this, it is necessary to provide a method that can accurately predict the ultra-short-term changes in wind power, so as to facilitate timely adjustments to the power system. Summary of the invention

[0004] The purpose of this invention is to provide an ultra-short-term wind power prediction method taking into account volatility and periodicity, introduce deep learning theory into wind power prediction, study data mining methods and the application of deep learning methods in ultra-short-term wind power prediction problems, improve power grid stability and power supply reliability, which is of great significance to the development of diversified energy systems.

[0005] To achieve the above object, the present invention provides a method for ultra-short-term prediction of wind power considering volatility and periodicity, comprising the following steps:

[0006] S1. Establish a wind power data mining method to obtain data fluctuation composition and periodic composition;

[0007] S2, extracting fusion features from raw wind power data through deep learning methods;

[0008] S3,make predictions based on fusion features and fluctuation,periodic components through integrated neural network models;

[0009] S4, the final ultra-short-term wind power prediction method is obtained through the fluctuation composition and periodic composition obtained by S1, the fusion characteristics of S2 and the integrated neural network of S3.

[0010] Preferably, step S1 comprises the following steps:

[0011] S11. Determine the envelope mean: By calculating the envelope mean of the signal, the overall fluctuation of the signal is obtained. The specific calculation method is as follows:

[0012] ;

[0013] In the formula, e U (t) and e L (t) Represents the upper and lower envelopes of the signal, which are obtained by calculating the local maximum and local minimum of the signal; m ( t ) represents the envelope mean; t Indicates time;

[0014] S12. Obtaining local components of the signal: The local components of the signal are obtained by subtracting the envelope mean from the original signal. The specific calculation method is as follows:

[0015] ;

[0016] In the formula h ( t ) represents the local component of the signal; x(t) represents the original signal;

[0017] S13, determine whether the local component of the signal meets two conditions, the two conditions include: the number of extreme points in the entire signal range should be equal or the difference should not exceed 1; the mean value of the envelope at any point is 0; if the two conditions are not met, the local component is used as a new input signal to repeat steps S11 and S12; if the two conditions are met, the local component is considered valid and is subtracted from the original signal to obtain a new residual signal, and the calculation method is as follows:

[0018] ;

[0019] In the formula r 1 ( t ) represents the residual signal;

[0020] S14, obtain the approximate value of the original signal: take the remaining signal as the new input signal, repeat steps S11, S12, S13 until the remaining signal becomes a monotonic trend and cannot be decomposed any further, and obtain the approximate value of the original signal by adding all local components, which includes the fluctuation component and periodic component of the signal. The specific function is as follows:

[0021] ;

[0022] In the formula c i ( t ) represents the effective local component of the signal; ci ( t ) represents the effective local component of the signal and i From 1 to n The local components of all valid signals are summed up, i Indicates signal, i =1 refers to the first signal, n refers to the nth signal; r n ( t ) represents the final remaining signal.

[0023] Preferably, the step S2 specifically includes:

[0024] S21. Convolutional Neural Network

[0025] The convolutional neural network consists of an input layer, a convolutional layer, and an output layer. It mainly uses the convolution operation of the convolutional layer to extract the implicit pattern of the input data and obtain the fusion features. Its objective function is:

[0026] ;

[0027] Where: Indicates the number of feature information graphs; The corresponding i The weights of the jth incoming feature map and the jth output feature map; Corresponding to j The bias value of the output feature map; It corresponds to i An incoming feature map; The corresponding j Output characteristic diagram; It is the convolution operation of CNN; is the ReLU activation function;

[0028] S22, residual connection

[0029] The residual connection adds a cross-layer "shortcut" in some layers of the network, adding the input signal directly to the output signal instead of simply mapping it. Its function is as follows:

[0030] ;

[0031] In the formula, It corresponds to the final fusion feature; Corresponding convolution operation; Corresponds to the original input.

[0032] Preferably, the step S3 specifically includes:

[0033] S31, Long Short-Term Memory Neural Network

[0034] The long short-term memory neural network includes three gating mechanisms, namely the forget gate, input gate, and output gate. The functions are as follows:

[0035] ;

[0036] ;

[0037] ;

[0038] ;

[0039] ;

[0040] ;

[0041] In the formula, w xf The input gate gating unit corresponds to the external input x t The weight vector of xi is the forget gate gating unit corresponding to the external input x t The weight vector of w xo Is the output gate gating unit corresponding to the external input x t The weight vector of w xc Candidate cell states corresponding to external inputs x t The weight vector of It is the hidden state input of each gating unit corresponding to the previous moment h t-1 The weight vector of is the bias of the forget gate gating unit, is the bias of the input gate gating unit, is the bias of the output gate gating unit, is the bias of the candidate cell state; is the sigmoid function; is the tanh activation function, c t 、 、h t 、i t 、o t 、f t They correspond to the memory cells, candidate memory cells, hidden states, input gates, output gates, and forget gates of the long short-term memory neural network respectively;

[0042] S32. Data normalization: There is a big difference in the number of levels between wind power data and wind power characteristic data. Before the data is input into the model, the information with large differences between these values ​​must be processed into a unified dimension. The specific function is as follows:

[0043] ;

[0044] In the formula, x max is the maximum value of the sequence data; x min is the minimum value of the sequence data; x norm is the normalized sequence data;

[0045] S33, integrated network model: By stacking three layers of long short-term memory neural network, we can avoid the overfitting and gradient explosion problems that may occur in a single model. At the same time, we weight the output results of different long short-term memory neural networks to improve the model performance. The function is as follows:

[0046] ;

[0047] ;

[0048] ;

[0049] ;

[0050] In the formula, It is the operation function of the long short-term memory neural network; h t+1:t+n Represented by history k Time data h t-k+1:t The future predicted by long short-term memory network n The predicted value of the high-frequency part of wind power at each moment, and the same applies to the low-frequency part; It is the high-frequency part of the decomposed wind power data, corresponding to the fluctuation composition. 1, 2, and 3 represent the outputs of different layers of memory neural networks; w l is the weight vector corresponding to different lengths of short-term memory neural networks; b l is the bias of each LSTM neural network; w d is the fusion weight; ypre t+1 is the final output result.

[0051] Preferably, step S4 comprises the following steps:

[0052] S41, input wind power data information sequence, its function is as follows:

[0053] Input={[ x t-k+1 , f t-k+1 ], …, [ x t , f t ], …, [ x t+n ]}

[0054] In the formula, f t-k+1 arrive f t It is the external characteristics of wind power data, such as wind speed data and wind direction information;

[0055] S42, data decomposition: decompose the original wind power data into high-frequency part and low-frequency part through data mining method, corresponding to the fluctuation component and periodic component of wind power data, and its function is as follows:

[0056] ;

[0057] In the formula, emd (•) is the empirical mode decomposition method, i.e., step S1 corresponds to the data mining method; ht - k +1: t+n It is the high-frequency part of wind power data; lt - k +1: t+n It is the low-frequency part of wind power data;

[0058] S43, obtain fusion features: for the external features of wind power data, use convolutional neural network to extract its fusion features, the function is as follows:

[0059] ;

[0060] In the formula, cnn (•) is the convolutional neural network operation expression; It is a fusion feature;

[0061] S44, model parameter update: The fusion feature is combined with the low-frequency part and the high-frequency part and input into the integrated long short-term memory neural network. After the prediction result is obtained, it is reconstructed and the iterative process is run. The specific function is as follows:

[0062] ;

[0063] ;

[0064] ;

[0065] In the formula, It is predicted n The high-frequency part prediction value of the future moment; l t+1:t+n It is predicted n The predicted value of the low-frequency part at a future moment; xpre t+1:t+n It is predicted n The wind power forecast value at a future moment; It is the operation function of the overall model of the integrated long short-term memory neural network; w h is the high-frequency weight corresponding to the long short-term memory neural network; w l is the low-frequency weight corresponding to the long short-term memory neural network; b h is the high frequency bias corresponding to the long short-term memory neural network; b l is the low frequency bias corresponding to the long short-term memory neural network; It is a reconstruction function that reconstructs the high-frequency part prediction value and the low-frequency part prediction value into the wind power prediction value.

[0066] Preferably, the model running iterative process in step S44 includes the following steps:

[0067] S441, Randomly initialize the integrated long short-term memory neural network model parameters: weight ω h , ω l and the bias vector b h , b l ;

[0068] S442, preliminarily set the size of the input timing to k , the input step size is 1, and the input training data sample [ h t-k+1 , f t-k+1 ], …, [ h t , f t ]and[ l t-k+1 , f t-k+1 ], …, [ l t , f t ];

[0069] S443, according to the initialization weight, bias vector and training data samples, after the model input at multiple moments, obtain each moment i The high-frequency part of the predicted value h i And the low frequency part prediction value l i ;

[0070] S444, weight fusion of different layers of long short-term memory neural network, and obtain the weights of each moment through reconstruction function i Wind power forecast value xpre t+1:t+n .

[0071] The beneficial effects of the present invention are: enhancing the power system's grasp of the volatility and intermittency of wind power. It can predict the changes in wind power output for the power system in advance, which helps the grid dispatcher to adjust the power generation plan in time to ensure the stable operation of the grid; it can help the power system operator to reasonably arrange the power transmission path, reduce the loss of transmission lines, and improve the efficiency of electric energy utilization; it can help power generation companies to better plan wind power operation and maintenance, improve the utilization rate of power generation equipment, reduce operating costs, and optimize power generation revenue; it helps to achieve the complementarity of wind power with other energy forms, realize the utilization of peak-valley electricity price differences, and smooth out fluctuations in power supply and demand; thereby greatly improving the predictability, stability and economy of wind power, and promoting the sustainable development of the wind power industry and the transformation to clean energy.

[0072] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 This is a schematic diagram of the working process of an embodiment of a method for ultra-short-term prediction of wind power considering volatility and periodicity of the present invention;

[0074] Figure 2 It is a diagram of constructing an ultra-short-term wind power prediction model through a deep learning method in an embodiment of an ultra-short-term wind power prediction method considering volatility and periodicity of the present invention. DETAILED DESCRIPTION

[0075] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0076] Unless otherwise defined, technical or scientific terms used in the present invention shall have the common meanings understood by one having ordinary skills in the field to which the present invention belongs.

[0077] like Figure 1-2 As shown, the present invention provides a method for ultra-short-term prediction of wind power considering volatility and periodicity, comprising the following steps:

[0078] S1. Establishing a data mining method for wind power data fluctuation composition and periodic composition, including the following steps:

[0079] S11. Determine the envelope mean: By calculating the envelope mean of the signal, the overall fluctuation of the signal can be obtained, so as to better understand the overall characteristics of the signal and help separate the local fluctuation and overall trend in the signal. The calculation method is as follows:

[0080] ;

[0081] In the formula, e U (t) and e L (t) Represents the upper and lower envelopes of the signal, which are obtained by calculating the local maximum and local minimum of the signal; m ( t ) represents the envelope mean; t Indicates time;

[0082] S12. Obtaining local components of the signal: The local components of the signal are obtained by subtracting the envelope mean from the original signal. The local components of the signal are the remainder after removing the overall trend and periodic components of the signal, which is helpful to reveal the detailed characteristics of the signal and make relevant predictions. The calculation method is as follows:

[0083] ;

[0084] In the formula, h ( t ) represents the local component of the signal; x(t) represents the original signal;

[0085] S13, determine whether the local component of the signal meets two conditions, the two conditions include: the number of extreme points in the entire signal range should be equal or the difference should not exceed 1; the mean value of the envelope at any point is 0; if the two conditions are not met, the local component is used as a new input signal to repeat steps S11 and S12; if the two conditions are met, the local component is considered valid and is subtracted from the original signal to obtain a new residual signal, and the calculation method is as follows:

[0086] ;

[0087] In the formula r 1 ( t ) represents the residual signal;

[0088] S14, obtain the approximate value of the original signal: take the remaining signal as the new input signal, repeat steps S11, S12, S13 until the remaining signal becomes a monotonic trend and cannot be decomposed any further, and obtain the approximate value of the original signal by adding all local components, which includes the fluctuation component and periodic component of the signal. The specific function is as follows:

[0089] ;

[0090] In the formula c i ( t ) represents the effective local component of the signal; c i ( t ) represents the effective local component of the signal and i From 1 to n The local components of all valid signals are summed up, i Indicates signal, i =1 refers to the first signal, n refers to the nth signal; r n ( t ) represents the final remaining signal.

[0091] S2. Extracting fusion features from the original wind power data by deep learning method, including the following steps:

[0092] S21. Convolutional neural network: The convolutional neural network consists of an input layer, a convolutional layer, and an output layer. It mainly uses the convolution operation of the convolutional layer to extract the implicit pattern of the input data and obtain the fusion features. Its objective function is:

[0093] ;

[0094] Where: Indicates the number of feature information graphs; The corresponding i The weights of the jth incoming feature map and the jth output feature map; Corresponding to j The bias value of the output feature map; It corresponds to i An incoming feature map; The corresponding j Output characteristic diagram; It is the convolution operation of CNN; is the ReLU activation function;

[0095] S22, residual connection: Residual connection adds a "shortcut" across layers in some layers of the network to directly add the input signal to the output signal instead of simply mapping. In this way, even if the mapping learned by the network is not the identity mapping, the residual part can be learned, making it easier to learn the identity mapping. Its function is as follows:

[0096] ;

[0097] In the formula, It corresponds to the final fusion feature; Corresponding convolution operation; Corresponds to the original input.

[0098] S3. Make predictions based on fusion features and different components through an integrated neural network model, including:

[0099] S31, Long Short-Term Memory Neural Network

[0100] The long short-term memory neural network includes three gating mechanisms, namely the forget gate, input gate, and output gate. The functions are as follows:

[0101] ;

[0102] ;

[0103] ;

[0104] ;

[0105] ;

[0106] ;

[0107] In the formula, w xf The input gate gating unit corresponds to the external input x t The weight vector of xi is the forget gate gating unit corresponding to the external input x t The weight vector of w xo Is the output gate gating unit corresponding to the external input x t The weight vector of w xc Candidate cell states corresponding to external inputs x t The weight vector of w h . is the hidden state input of each gating unit corresponding to the previous momenth t-1 The weight vector of is the bias of the forget gate gating unit, is the bias of the input gate gating unit, is the bias of the output gate gating unit, is the bias of the candidate cell state; is the sigmoid function; is the tanh activation function, c t 、 、h t 、i t 、o t 、f t They correspond to the memory cells, candidate memory cells, hidden states, input gates, output gates, and forget gates of the long short-term memory neural network respectively;

[0108] S32. Data normalization: There is a big difference in the number of levels between wind power data and wind power characteristic data. Before the data is input into the model, the information with large differences between these values ​​must be processed into a unified dimension. The specific function is as follows:

[0109] ;

[0110] In the formula, x max is the maximum value of the sequence data; x min is the minimum value of the sequence data; x norm is the normalized sequence data;

[0111] S33, integrated network model: By stacking three layers of long short-term memory neural network, we can avoid the overfitting and gradient explosion problems that may occur in a single model. At the same time, we weight the output results of different long short-term memory neural networks to improve the model performance. The function is as follows:

[0112] ;

[0113] ;

[0114] ;

[0115] ;

[0116] In the formula, It is the operation function of the long short-term memory neural network;h t+1:t+n Represented by history k Time data h t-k+1:t The future predicted by long short-term memory network n The predicted value of the high-frequency part of wind power at each moment, and the same applies to the low-frequency part; h. It is the high-frequency part of the decomposed wind power data, corresponding to the fluctuation composition. 1, 2, and 3 represent the outputs of different layers of memory neural networks; w l is the weight vector corresponding to different lengths of short-term memory neural networks; b l is the bias of each LSTM neural network; w d is the fusion weight; ypre t+1 is the final output result.

[0117] S4, determining the final ultra-short-term wind power prediction method through the fluctuation composition and periodic composition obtained in S1, the fusion characteristics of S2 and the integrated neural network of S3, including the following steps:

[0118] S41, input wind power data information sequence, its function is as follows:

[0119] Input={[ x t-k+1 , f t-k+1 ], …, [ x t , f t ], …, [ x t+n ]}

[0120] In the formula, f t-k+1 arrive f t It is the external characteristics of wind power data, such as wind speed data and wind direction information;

[0121] S42, data decomposition: decompose the original wind power data into high-frequency part and low-frequency part through data mining method, corresponding to the fluctuation component and periodic component of wind power data, and its function is as follows:

[0122] ;

[0123] In the formula, emd (•) is the empirical mode decomposition method, i.e., step S1 corresponds to the data mining method; ht - k +1: t+n It is the high-frequency part of wind power data; lt- k +1: t+n It is the low-frequency part of wind power data;

[0124] S43, obtain fusion features: for the external features of wind power data, use convolutional neural network to extract its fusion features, the function is as follows:

[0125] ;

[0126] In the formula, cnn (•) is the convolutional neural network operation expression; It is a fusion feature;

[0127] S44, model parameter update: The fusion feature is combined with the low-frequency part and the high-frequency part and input into the integrated long short-term memory neural network. After the prediction result is obtained, it is reconstructed and the iterative process is run. The specific function is as follows:

[0128] ;

[0129] ;

[0130] ;

[0131] In the formula, It is predicted n The high-frequency part prediction value of the future moment; l t+1:t+n It is predicted n The predicted value of the low-frequency part at a future moment; xpre t+1:t+n It is predicted n The wind power forecast value at a future moment; sl (∙) is the operation function of the overall model of the integrated long short-term memory neural network; w h is the high-frequency weight corresponding to the long short-term memory neural network; w l is the low-frequency weight corresponding to the long short-term memory neural network; b h is the high frequency bias corresponding to the long short-term memory neural network; b l is the low frequency bias corresponding to the long short-term memory neural network; It is a reconstruction function that reconstructs the high-frequency part prediction value and the low-frequency part prediction value into the wind power prediction value.

[0132] The model running iteration process in step S44 includes the following steps:

[0133] S441, Randomly initialize the integrated long short-term memory neural network model parameters: weight ωh , ω l and the bias vector b h , b l ;

[0134] S442, preliminarily set the size of the input timing to k , the input step size is 1, and the input training data sample [ h t-k+1 , f t-k+1 ], …, [ h t , f t ]and[ l t-k+1 , f t-k+1 ], …, [ l t , f t ];

[0135] S443, according to the initialization weight, bias vector and training data samples, after the model input at multiple moments, obtain each moment i The high-frequency part of the predicted value h i And the low frequency part prediction value l i ;

[0136] S444, weight fusion of different layers of long short-term memory neural network, and obtain the weights of each moment through reconstruction function i Wind power forecast value xpre t+1:t+n .

[0137] Therefore, the present invention adopts the above-mentioned ultra-short-term wind power prediction method considering volatility and periodicity, introduces deep learning theory into wind power prediction, studies data mining methods and the application of deep learning methods in ultra-short-term wind power prediction problems, improves power grid stability and power supply reliability, and is of great significance to the development of diversified energy systems.

[0138] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for ultra-short-term wind power prediction considering volatility and periodicity, characterized in that: The following steps are involved: S1. Establish a wind power data mining method to obtain data fluctuation composition and periodic composition; S2, extracting fusion features from raw wind power data through deep learning methods; Specifically include: S21. Convolutional Neural Network The convolutional neural network consists of an input layer, a convolutional layer, and an output layer. The convolution operation of the convolutional layer is used to extract the implicit pattern of the input data and obtain the fusion features. The objective function is: ； Where: Indicates the number of feature information graphs; The corresponding The incoming feature map and The weight of each output feature map; Corresponding to The bias value of the output feature map; It corresponds to An incoming feature map; The corresponding Output characteristic diagram; It is the convolution operation of CNN; is the ReLU activation function; S22, residual connection The specific functions are as follows: ； In the formula, It corresponds to the final fusion feature; Corresponding convolution operation; S3. Make predictions based on fusion features and fluctuations and periodic components through an integrated neural network model; specifically: S31, Long Short-Term Memory Neural Network The long short-term memory neural network includes three gating mechanisms, namely the forget gate, input gate, and output gate. The functions are as follows: ； ； ； ； ； ； In the formula, is the forget gate gating unit corresponding to the external input The weight vector of The input gate gating unit corresponds to the external input The weight vector of Is the output gate gating unit corresponding to the external input The weight vector of is the candidate memory cell state corresponding to the external input The weight vector of Is the candidate memory cell state corresponding to the hidden state input at the previous moment The weight vector of is the input of the forget gate gating unit corresponding to the hidden state at the previous moment The weight vector of The input gate gating unit corresponds to the hidden state input at the previous moment The weight vector of The output gate gating unit corresponds to the hidden state input at the previous moment The weight vector of is the bias of the forget gate gating unit; is the bias of the input gate gating unit; is the bias of the candidate memory cell state; is the sigmoid function; is the tanh activation function; , , , , , They correspond to the memory cells, candidate memory cells, hidden states, input gates, output gates, and forget gates of the long short-term memory neural network respectively; S32. Data normalization There is a big difference in the order of magnitude between wind power data and wind power characteristic data. Before the data is input into the model, it is processed into a unified dimension. The function is as follows: ； In the formula, is the maximum value of the sequence data; is the minimum value of the sequence data; is the normalized sequence data; S33, Integrated Network Model Stack three layers of LSTM neural networks and perform weight fusion on the output results of different LSTM neural networks. The function is as follows: ； ； ； ； In the formula, It is the operation function of the long short-term memory neural network; Represented by history k Time data The future predicted by the long short-term memory network n The predicted value of the high-frequency part of wind power at each moment, and the same applies to the low-frequency part; It is the high-frequency part of the decomposed wind power data, corresponding to the fluctuation composition. 1, 2, and 3 represent the outputs of different layers of memory neural networks; is the weight vector corresponding to the first layer of long short-term memory neural network; is the weight vector corresponding to the second layer of long short-term memory neural network; is the weight vector corresponding to the third layer of long short-term memory neural network; is the bias of the first layer of LSTM neural network; is the bias of the second layer of LSTM neural network; is the bias of the third layer of LSTM neural network; is the fusion weight; The final output result; S4, obtaining the final wind power ultra-short-term prediction method through the fluctuation composition and periodic composition obtained in S1, the fusion characteristics of S2 and the integrated neural network of S3; comprising the following steps: S41, input wind power data information sequence, its function is as follows: ； In the formula, It is the external characteristic of wind power data; S42, data decomposition: decompose the original wind power data into high-frequency part and low-frequency part through data mining method, corresponding to the fluctuation component and periodic component of wind power data, and its function is as follows: ； In the formula, It is the empirical mode decomposition method, that is, step S1 corresponds to the data mining method; It is the high-frequency part of wind power data; It is the low-frequency part of wind power data; S43, obtain fusion features: for the external features of wind power data, use convolutional neural network to extract its fusion features, the function is as follows: ； In the formula, It is a fusion feature; S44, model parameter update: The fusion feature is combined with the low-frequency part and the high-frequency part and input into the integrated long short-term memory neural network. After the prediction result is obtained, it is reconstructed and the iterative process is run. The specific function is as follows: ； ； ； In the formula, It is predicted n The high-frequency part prediction value of the future moment; It is predicted n The predicted value of the low-frequency part at a future moment; It is predicted n The wind power forecast value at a future moment; It is the operation function of the overall model of the integrated long short-term memory neural network; is the high-frequency weight corresponding to the long short-term memory neural network; is the high frequency bias corresponding to the long short-term memory neural network; It is a reconstruction function that reconstructs the high-frequency part prediction value and the low-frequency part prediction value into the wind power prediction value.

2. The method for ultra-short-term wind power prediction considering volatility and periodicity according to claim 1 is characterized in that: The step S1 comprises the following steps: S11. Determine the envelope mean: By calculating the envelope mean of the signal, the overall fluctuation of the signal is obtained. The specific calculation method is as follows: ； In the formula, and Represents the upper and lower envelopes of the signal, which are obtained by calculating the local maximum and local minimum of the signal; represents the envelope mean; Indicates time; S12. Obtaining local components of the signal: The local components of the signal are obtained by subtracting the envelope mean from the original signal. The specific calculation method is as follows: ； In the formula, Represents the local components of the signal; represents the original signal; S13, determine whether the local component of the signal meets two conditions, the two conditions include: the number of extreme points in the entire signal range should be equal or the difference should not exceed 1; the mean value of the envelope at any point is 0; if the two conditions are not met, the local component is used as a new input signal to repeat steps S11 and S12; if the two conditions are met, the local component is considered valid and is subtracted from the original signal to obtain a new residual signal, and the calculation method is as follows: ； In the formula represents the remaining signal; S14, obtain the approximate value of the original signal: take the remaining signal as the new input signal, repeat steps S11, S12, S13 until the remaining signal becomes a monotonic trend and cannot be decomposed any further, and obtain the approximate value of the original signal by adding all local components, which includes the fluctuation component and periodic component of the signal. The specific function is as follows: ； In the formula Represents the effective local component of the signal and From 1 to The local components of all valid signals are summed; Indicates signal, Refers to the first signal, Indicates A signal; Represents the final remaining signal.

3. The method for ultra-short-term wind power prediction considering volatility and periodicity according to claim 1 is characterized in that: The model running iteration process in step S44 includes the following steps: S441, Randomly initialize the integrated long short-term memory neural network model parameters: weight , and the bias vector , ; S442, preliminarily set the size of the input timing to k , the input step size is 1, and the training data sample is passed in and ; S443, according to the initialization weight, bias vector and training data samples, after the model input at multiple moments, obtain each moment The high-frequency part of the predicted value And the low frequency part prediction value ; S444, weight fusion of different layers of long short-term memory neural network, and obtain the weights of each moment through reconstruction function Wind power forecast value .

Images