Short-term wind speed prediction method based on SSA-CEEMDAN-ITransform

Abstract

The invention discloses a short-term wind speed prediction method based on SSA-CEEMDAN-ITransform. The short-term wind speed prediction method is characterized by comprising the following steps: acquiring wind speed historical data; decomposing the wind speed historical data by using a singular spectrum analysis method to obtain a trend signal, a periodic signal and a noise signal; performing secondary mode decomposition on the periodic signal by using a completely integrated empirical mode decomposition algorithm and an adaptive noise algorithm, and performing de-noising processing on a noise signal by using a wavelet packet decomposition algorithm; and predicting the secondary decomposition sequence by using an ITransform model, and reconstructing a prediction result. According to the short-term wind speed prediction method based on the SSA-CEEMDAN-ITransform provided by the invention, the short-term wind speed can be effectively predicted.

Description

Technical Field

[0001] The present invention relates to the field of wind speed prediction, and in particular to a short-term wind speed prediction method based on SSA-CEEMDAN-ITransformer; Background Art

[0002] In recent years, due to extreme weather conditions, transmission lines have repeatedly experienced wind-induced tripping accidents, which has seriously affected the safe and reliable operation of the power grid. In order to cope with the impact of wind-induced flashover on transmission lines, it is necessary to propose and establish an accurate wind speed prediction method and model. The wind swing angle of the line is strongly correlated with the wind speed. Under known wind speed conditions, the wind swing amplitude change of the line can be obtained by combining the wind swing angle calculation method. Therefore, accurate wind speed prediction can provide early warning of the wind swing risk of the line. The accuracy of the wind speed prediction results will also directly affect the risk warning results of the line. Wind speed prediction has become an important research direction for risk warning. However, the existing short-term wind speed prediction methods do not effectively utilize the advantages of data decomposition methods, such as the literature [1]: Xiong, XF, Weng, SJ, Wang, J, An online early-warning method for wind swing discharge of the conductor toward the tangent tower and jumper toward the strain tower, IEEE TRANSACTIONS ON POWER DELIVERY, 2015, 30(1), 114-121. And no targeted processing is done on the wind speed characteristic components of different frequencies.

[0003] In addition, in terms of prediction model selection, the current advanced time series prediction technology has not been further evaluated and integrated to improve the overall prediction effect; such as reference 0: Wang, J, Xiong, XF, Li, Z, Wang, W, et al., Wind forecast-based probabilistic early warning method of wind swing discharge for OHTLs, IEEE TRANSACTIONS ON POWER DELIVERY, 2016, 31(5), 74-81. Summary of the invention

[0004] The purpose of this invention is to propose a short-term wind speed prediction method based on SSA-CEEMDAN-ITransformer in response to the frequent wind deviation tripping accidents of current transmission lines, which can effectively predict the short-term wind speed and provide a reference for related research on wind speed prediction;

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0006] A short-term wind speed prediction method based on SSA-CEEMDAN-ITransformer includes the following steps:

[0007] Step 1: Obtain historical wind speed data;

[0008] Step 2: Use the singular spectrum analysis method to decompose the historical wind speed data to obtain trend signals, periodic signals and noise signals;

[0009] Step 3: Use the fully integrated empirical mode decomposition algorithm and the adaptive noise algorithm to perform secondary mode decomposition on the periodic signal, and use the wavelet packet decomposition algorithm to denoise the noise signal;

[0010] Step 4: Use the ITransformer model to predict the secondary decomposition sequence and reconstruct the prediction results.

[0011] It also includes step 5: performing error analysis on the prediction results of the SSA-CEEMDAN-ITransformer model.

[0012] In step 1, the historical wind speed data is the maximum wind speed per hour.

[0013] In the step 2, the historical wind speed data is decomposed using a singular spectrum analysis method to obtain a trend signal, a periodic signal and a noise signal.

[0014] The singular spectrum analysis method is used to decompose the historical wind speed data. The specific decomposition process is as follows:

[0015] 1) Embedding: Select an appropriate window length L and embed the original time series X = [x1, x2, ..., x N ] According to formula (1), the trajectory matrix is ​​obtained:

[0016]

[0017] Where Z is the trajectory matrix of the original time series arranged according to the window length, K = N-L+1, L∈[2,N], x k is the specific wind speed value at a certain moment, x L is the wind speed value at the window length position L, x N It is the last wind speed value of the original time series.

[0018] 2) Singular Value Decomposition (SVD): Through SVD, the trajectory matrix is ​​decomposed:

[0019] Z=Z1+Z2…+Z i …+Zd (2);

[0020] Where: Z d is the dth elementary matrix after trajectory matrix decomposition, Z i is the i-th elementary matrix after trajectory matrix decomposition, U i 、V i is the i-th feature triplet after singular value decomposition of the trajectory matrix Z.

[0021] 3) Grouping: Divide the above d sub-matrices into p disjoint groups I = {I1, I2, ..., I p}, Z ip is the intra-group matrix, and the composite matrix is:

[0022] Z=Z i1 +…+Z ip (3)

[0023] In the formula, Z i1 is the first within-group matrix.

[0024] 4) Reconstruction: Use the diagonal averaging formula to convert Z into z1,z2,...,z N The time series is reconstructed and the diagonal averaging formula is:

[0025]

[0026] In the formula, x k is the reconstructed wind speed sequence, q is the number of matrices within the group after grouping, p is some numerical representation in a matrix within a group, is a matrix value of the matrix Z within the group, L* is min(L,K), and K* is max(L,K).

[0027] In step 2, the parameters include the contribution value of each time series. The first k time series with cumulative contribution values ​​greater than 95% are classified as trend signals and periodic signals, among which time series 1 is a trend signal, time series 2-k are periodic signals, and the remaining series are classified as noise signals.

[0028] In step 3, the specific process of performing secondary modal decomposition on the periodic signal using the fully integrated empirical mode decomposition and adaptive noise algorithm is as follows:

[0029] 1) First, add N Gaussian white noises with a mean of 0 to the wind speed component x(t) to form N sequences to be decomposed x i (t)(i=1,2,…,N);

[0030] x i(t) = x(t) + εδ i (t) (5)

[0031] In the formula, x i (t) is the sequence to be decomposed, ε is the weight coefficient of white noise, δ i (t) is the Gaussian white noise during the i-th processing.

[0032] 2) For each sequence x to be decomposed i (t) Apply the EMD algorithm to decompose and obtain its intrinsic mode component (IMF) and residual component r1(t);

[0033]

[0034] Where IMF1(t) is the first intrinsic mode component; r1(t) is the first residual sequence, is the component of the first intrinsic mode component IMF.

[0035] 3) Then add noise to the jth (j=2,3,…,N) residual component obtained after decomposition, and continue EMD decomposition;

[0036]

[0037] r j (t) = r j-1 (t)-IMF j (t) (8);

[0038] In the formula, IMF j (t) is the jth intrinsic mode component; E1 is the first IMF component after EMD decomposition of the wind speed sequence, r j-1 (t) is the j-1th residual sequence, ε j-1 is the j-1th white noise weight coefficient, δ i (t) is the Gaussian white noise during the i-th processing, r j (t) is the jth residual sequence, r j-1 (t) is the j-1th residual sequence, IMF j (t) is the jth eigenmode component.

[0039] 4) Repeat step 3 until the termination condition is met: the number of extreme points of each residual component does not exceed 2; finally, the original signal sequence is decomposed into N intrinsic mode functions and a residual term R(t);

[0040]

[0041] In the formula, IMF n is the nth eigenmode component.

[0042] In step 3, the noise signal is denoised using the wavelet packet decomposition algorithm. The denoising principle is as follows:

[0043] Wavelet packet decomposition uses orthogonal mirror filters ({h 0k} k∈Z ,{h 1k} k∈Z ) and the two-scale equation to decompose the signal into multiple frequency bands, where the two-scale equation is:

[0044]

[0045] Where: W 2n (t) is the low-frequency part of the signal at the 2nth level (or scale), W 2n+1 (t) represents the high-frequency part of the signal at the 2n+1th level (or scale), h 0k and h 1k are the coefficients of the mirror filter.

[0046] When n = 0, W0(t) = φ(t) is the scaling function; and the function sequence corresponding to the scaling function {W n (t)} n∈Z That is the wavelet packet.

[0047] In step 4, the ITransformer model adopts the encoder architecture of Transformer, including embedding, projection and TrmBlock.

[0048] According to the historical series data X:n of wind speed, the future series of predicted variable Y:n can be expressed as:

[0049]

[0050] Where: is the high-dimensional vector representation of the historical series data of wind speed, K l is the input (or feature) of the lth layer. The model has l layers of TrmBlock blocks. From layer 0 to layer L-1, each layer processes the input and gradually improves the abstraction level of feature representation. is the predicted value of wind speed data at a future time point.

[0051] Where K = {k1, k2, ..., k M}∈R M×D contains M embedding tokens of dimension D; the embedding operation converts R T Dimension conversion to R D dimensional vector, the projection will be R D Dimension conversion to R S dimensional vector, both of which are accomplished by multi-layer perceptrons;

[0052] The ITransformer model is used to predict the secondary decomposition sequence and reconstruct the prediction results. Each group of time series is input into the ITransformer model, and the prediction results of each group are exported and accumulated to solve the final prediction result.

[0053] In step 5, the evaluation indicators of error analysis are mean absolute error, residual sum of squares, and mean relative error; the mean absolute error, residual sum of squares, and mean relative error are respectively expressed as:

[0054]

[0055] Where: E MAE is the mean absolute error, E MAPE is the residual sum of squares, E RMSE is the mean relative error, y t , are the final predicted value and true value at time t respectively; N' is the sequence length.

[0056] The present invention provides a short-term wind speed prediction method based on SSA-CEEMDAN-ITransformer, which has the following technical effects:

[0057] 1) The present invention decomposes the original wind speed data based on the singular spectrum analysis method, uses CEEMDAN to perform secondary decomposition on the periodic signal, and uses WPD to denoise the noise, thereby ensuring the reliability of the data.

[0058] 2) The present invention applies the ITransformer model to wind speed prediction. Its flexible attention mechanism and iterative refinement mechanism can effectively process long sequence data and complex time series patterns, and can achieve the best prediction effect of the data.

[0059] 3) Advantages of step 3: The CEEMDAN algorithm is an improved method based on ensemble empirical mode decomposition (EEMD); there is no need to set the number of decompositions K, which avoids the discontinuity of the intrinsic mode component (IMF); therefore, this paper chooses to use CEEMDAN to perform secondary decomposition of high-frequency periodic components; WPD is a signal processing technology developed based on multi-resolution analysis. Compared with traditional discrete wavelet transform, WPD processes noise signals and can improve the predictability of signals. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:

[0061] Figure 1 This is the wind speed prediction flow chart.

[0062] Figure 2 This is the original wind speed data graph.

[0063] Figure 3 It is the confidence function curve.

[0064] Figure 4 This is the simulation diagram of singular spectrum trend signal and periodic signal.

[0065] Figure 5 This is the noise signal decomposition diagram.

[0066] Figure 6 Make predictions for the test set.

[0067] Figure 7 This is the ITransformer model prediction architecture diagram. DETAILED DESCRIPTION

[0068] like Figure 1 As shown in the figure, the short-term wind speed prediction method based on SSA-CEEMDAN-ITransformer includes the following steps:

[0069] Step (1): Based on the data sent back by the micro-meteorological station monitoring device, obtain the historical wind speed data;

[0070] The wind speed data that needs to be obtained is the maximum wind speed per hour;

[0071] Step (2): Decompose the historical wind speed data using the singular spectrum analysis method to obtain trend signals, periodic signals and noise signals;

[0072] Singular spectrum analysis (SSA) is a data analysis method based on signal decomposition, which is used to extract trends, cycles and noise components in time series. It can effectively process nonlinear and non-stationary time series data. The basic idea of ​​SSA is to decompose the time series into a set of subsequences based on singular value decomposition, and then use these subsequences to reconstruct the original time series to analyze and extract features.

[0073] Trend signal: Trend signal is a slowly changing component in a time series, usually reflecting the overall trend or long-term changes of the data; it can be linear or nonlinear and is the smoothest part of the time series; therefore, it can be directly used for prediction;

[0074] Periodic signal: A periodic signal is a recurring pattern or periodic change in a time series;

[0075] Noise signal: Noise signals are irregular and random components in time series. These signals usually have no obvious patterns or regularities and contain unpredictable changes.

[0076] What needs to be obtained is the contribution value of each time series. The first k sequences with cumulative contribution values ​​greater than 95% are classified as trend signals and periodic signals, among which sequence 1 is the trend signal, the periodic signal is k-1 subsequences, and the rest of the sequences are classified as noise signals.

[0077] Step (3): Use the fully integrated empirical mode decomposition and adaptive noise algorithm (CEEMDAN) to perform secondary mode decomposition on the periodic signal, and use the wavelet packet decomposition algorithm (WPD) to denoise the noise signal;

[0078] CEEMDAN is a data analysis method for decomposing nonlinear and non-stationary signals. It is an improvement and extension of the empirical mode decomposition (EMD) method. CEEMDAN is mainly used to process time series data with nonlinear trends, periodicity and noise.

[0079] Wavelet packet decomposition (WPD) is a mathematical tool for signal processing that can provide information about the signal in both the time domain and the frequency domain. Unlike the traditional Fourier transform, it can process non-stationary signals and has advantages in analyzing the transient characteristics of signals.

[0080] Step (4): Use the ITransformer model to predict the secondary decomposition sequence and reconstruct the prediction results;

[0081] The ITransformer model is a neural network architecture based on the attention mechanism, which is mainly used to process sequence data. It is an improved model of Transformer, which aims to better process long sequence data and improve the parallel performance of the model.

[0082] Step (5): Use the mean absolute error (M = E MAE ), residual sum of squares (M = E MAPE ), mean relative error (M = E RMSE ) as the evaluation index to conduct error analysis on the prediction results of SSA-CEEMDAN-ITransformer model;

[0083] Formula (1) is the calculation formula for the evaluation indexes of the above three;

[0084]

[0085] In the formula, y t , are the final predicted value and true value at time t respectively; N' is the sequence length;

[0086] The present invention obtains historical wind speed data by transmitting data back from a micro-meteorological monitoring device, decomposes the wind speed using SSA, performs secondary decomposition of the periodic trend sequence using CEEMDAN, denoises the noise signal using WPD, and finally predicts the secondary modal decomposition sequence and the denoised sequence using ITransformer, and performs error analysis on the prediction results.

[0087] Example:

[0088] The simulation is carried out based on the historical wind speed of the actual transmission line micro-meteorological monitoring device. The data comes from a 330kV transmission line micro-meteorological station. There are 1561 wind speed data from April 1 to June 13 of a certain year. The resolution of wind speed is 1 hour. The original wind speed data is as follows Figure 2 As shown;

[0089] According to experience, the confidence level is set to 95%, which is Figure 3 It can be seen that when the number of subsequences is 6, it exceeds the empirical value of confidence (95%), and the subsequences greater than 6 are classified as noise signals; among them, IMF1 is a trend signal, and IMF2-IMF6 are periodic signals, as shown in the following figure. Figure 4 shown by Figure 4 It can be seen that the predictability of IMF1 is good, while the predictability of IMF2-IMF6 is poor. It is necessary to use CEEMDAN to perform secondary decomposition of the periodic trend series to improve the predictability;

[0090] Use WPD to denoise the noise signal, extract the available trend of the signal, and improve the accuracy of the data;

[0091] Depend on Figure 5 It can be seen that after the noise signal is decomposed by wavelet packets, more noise signals are removed and better predictability is achieved. Therefore, the two are selected to reconstruct the noise signal and use the ITransformer model for prediction. The prediction results of the test set are as follows: Figure 6 As shown;

[0092] Using E MAE 、E MAPE 、E RMSE As an evaluation index, the prediction results are evaluated. After calculating E MAE =0.235, E MAPE =6.59, E RMSE =0.107, and combined Figure 6 It can be seen from the prediction trend that this method can achieve accurate prediction of wind speed.

Claims

1. A short-term wind speed forecasting method based on SSA-CEEMDAN-ITransformer, characterized by: The following steps are involved: Step 1: Obtain historical wind speed data; Step 2: Use the singular spectrum analysis method to decompose the historical wind speed data to obtain trend signals, periodic signals and noise signals; Step 3: Use the fully integrated empirical mode decomposition algorithm and the adaptive noise algorithm to perform secondary mode decomposition on the periodic signal, and use the wavelet packet decomposition algorithm to denoise the noise signal; Step 4: Use the ITransformer model to predict the secondary decomposition sequence and reconstruct the prediction results.

2. The short-term wind speed forecasting method based on SSA-CEEMDAN-ITransformer according to claim 1 is characterized in that: It also includes step 5: performing error analysis on the prediction results of the SSA-CEEMDAN-ITransformer model.

3. The short-term wind speed forecasting method based on SSA-CEEMDAN-ITransformer according to claim 1 is characterized in that: In step 1, the historical wind speed data is the maximum wind speed per hour.

4. The short-term wind speed forecasting method based on SSA-CEEMDAN-ITransformer according to claim 1 is characterized in that: In the step 2, the historical wind speed data is decomposed using a singular spectrum analysis method to obtain a trend signal, a periodic signal and a noise signal; The singular spectrum analysis method is used to decompose the historical wind speed data. The specific decomposition process is as follows: 1) Embedding: Select an appropriate window length L and embed the original time series X = [x1, x2, ..., x N ] According to formula (1), the trajectory matrix is ​​obtained: Where Z is the trajectory matrix of the original time series arranged according to the window length, K = N-L+1, L∈[2,N], x k is the specific wind speed value at a certain moment, x L is the wind speed value at the window length position L, x N is the last wind speed value of the original time series; 2) Singular value decomposition (SVD): The trajectory matrix in formula (1) is decomposed by singular value decomposition (SVD): Z=Z1+Z2+…+Z i …+Z d (2); Where: Z i is the i-th elementary matrix after trajectory matrix decomposition, U i 、V i is the i-th feature triplet after singular value decomposition of trajectory matrix Z, Z d is the dth elementary matrix after trajectory matrix decomposition; 3) Grouping: Divide the above d sub-matrices into p disjoint groups I = {I1, I2, ..., I p }, Z ip is the intra-group matrix, and the composite matrix is: Z=Z i1 +…+From ip (3); In the formula, Z i1 is the first within-group matrix; 4) Reconstruction: Use the diagonal averaging formula to transform Z into z1, z2, ..., z N The time series is reconstructed and the diagonal averaging formula is: Where: x k is the reconstructed wind speed sequence, q is the number of matrices within the group after grouping, p is some numerical representation in a matrix within a group, is a matrix value of the matrix Z within the group, L * is min(L,K), K * is max(L,K).

5. The short-term wind speed prediction method based on SSA-CEEMDAN-ITransformer according to claim 4 is characterized in that: In step 2, the parameters include the contribution value of each time series. The first k time series with cumulative contribution values ​​greater than 95% are classified as trend signals and periodic signals, among which time series 1 is a trend signal, time series 2-k are periodic signals, and the remaining series are classified as noise signals.

6. The short-term wind speed forecasting method based on SSA-CEEMDAN-ITransformer according to claim 1, characterized in that: In step 3, the specific process of performing secondary modal decomposition on the periodic signal using the fully integrated empirical mode decomposition and adaptive noise algorithm is as follows: 1) Add N Gaussian white noises with a mean of 0 to the wind speed component x(t) to form N sequences to be decomposed x i (t)(i=1,2,…,N); x i (t)=x(t)+εδ i (t) (5); In the formula, x i (t) is the sequence to be decomposed, ε is the weight coefficient of white noise, δ i (t) is the Gaussian white noise during the i-th processing; 2) For each sequence x to be decomposed i (t) Apply the EMD algorithm to decompose and obtain its intrinsic mode component IMF and residual component r1(t); Where IMF1(t) is the first intrinsic mode component; r1(t) is the first residual sequence, is the component of the first intrinsic mode component IMF; 3) Add noise to the jth (j=2, 3, ..., N) residual component obtained after decomposition, and continue to perform EMD decomposition; r j (t)=r j-1 (t)-IMF j (t) (8); In the formula, IMF j (t) is the jth intrinsic mode component; E1 is the first IMF component after EMD decomposition of the wind speed sequence, r j-1 (t) is the j-1th residual sequence, ε j-1 is the j-1th white noise weight coefficient, δ i (t) is the Gaussian white noise during the i-th processing, r j (t) is the jth residual sequence, r j-1 (t) is the j-1th residual sequence, IMF j (t) is the jth eigenmode component; 4) Repeat step 3) until the termination condition is met: the number of extreme points of each residual component does not exceed 2; finally, the original signal sequence is decomposed into N intrinsic mode functions and a residual term R(t); In the formula, IMF n is the nth eigenmode component.

7. The short-term wind speed forecasting method based on SSA-CEEMDAN-ITransformer according to claim 1 is characterized in that: In step 3, the noise signal is denoised using the wavelet packet decomposition algorithm. The denoising principle is as follows: Wavelet packet decomposition uses orthogonal mirror filters ({h 0k } k∈Z ,{h 1k } k∈Z ) and the two-scale equation to decompose the signal into multiple frequency bands, where the two-scale equation is: Where: W 2n (t) is the low-frequency part of the signal at the 2n level, W 2n+1 (t) represents the high-frequency part of the signal at the 2n+1th level, h 0k and h 1k is the coefficient of the mirror filter; When n = 0, W0(t) = φ(t) is the scaling function; and the function sequence corresponding to the scaling function {W n (t)} n∈Z That is the wavelet packet.

8. The short-term wind speed forecasting method based on SSA-CEEMDAN-ITransformer according to claim 1 is characterized in that: In step 4, the ITransformer model adopts the encoder architecture of Transformer, including embedding, projection and TrmBlock; Based on the historical series data X:n of wind speed, the future series of the predicted variable Y:n is expressed as: Where: is the high-dimensional vector representation of the historical series data of wind speed, K l is the input of the lth layer. The model has l layers of TrmBlock blocks. From layer 0 to layer L-1, each layer processes the input and gradually improves the abstraction level of feature representation. is the predicted value of wind speed data at a future time point; Where K = {k1, k2, ..., k M }∈R M×D contains M embedding tokens of dimension D; the embedding operation converts R T Dimension conversion to R D dimensional vector, the projection will be R D Dimension conversion to R S dimensional vector, both of which are accomplished by multi-layer perceptrons; The ITransformer model is used to predict the secondary decomposition sequence and reconstruct the prediction results. Each group of time series is input into the ITransformer model, and the prediction results of each group are exported and accumulated to solve the final prediction result.

9. The short-term wind speed forecasting method based on SSA-CEEMDAN-ITransformer according to claim 1, characterized in that: In step 5, the evaluation indicators of error analysis are mean absolute error, residual sum of squares, and mean relative error; the mean absolute error, residual sum of squares, and mean relative error are respectively expressed as: Where: E MAE is the mean absolute error, E MAPE is the residual sum of squares, E RMSE is the average relative error, y t , are the final predicted value and true value at time t respectively; N' is the sequence length.

Images