Wind turbine generator ultra-short-term output prediction method based on improved CEEMDAN-SVMD-LSTM
By improving CEEMDAN-SVMD-LSTM, the wind turbine data is cleaned and modally decomposed, and the LSTM network is built, which solves the problem of insufficient prediction accuracy of ultra-short-term output of wind power, and achieves more accurate prediction and stable operation of the power system.
Patent Information
- Application Number
- CN202510490058.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-18
- Publication Date
- 2025-08-12
Smart Images
Figure CN120473981A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of new energy power generation technology and power system automation technology, and particularly relates to an ultra-short-term output prediction method for wind turbines based on an improved CEEMDAN-SVMD-LSTM. Background Art
[0002] Against the backdrop of a global energy transition, wind power, as a clean, renewable energy source with abundant resources, widespread distribution, and environmental friendliness, has become a crucial component of energy strategy. However, the volatility and intermittency of wind power pose significant challenges to the stable operation and economic dispatch of power systems. While advances in meteorological, computing, and communications technologies have provided strong support for wind power forecasting, and policy support has driven the rapid development of the wind power industry, existing ultra-short-term wind power output forecasting methods still suffer from insufficient modal decomposition and inadequate capture of long-term dependencies, leading to inaccurate forecasts. Summary of the Invention
[0003] The purpose of the present invention is to provide a wind turbine ultra-short-term output prediction method based on an improved CEEMDAN-SVMD-LSTM, which can effectively utilize wind turbine SCADA data to predict the ultra-short-term output of intermittent and volatile renewable energy and improve the prediction accuracy.
[0004] The technical solution adopted by the present invention is to use an ultra-short-term wind turbine output prediction method based on an improved CEEMDAN-SVMD-LSTM method, with the following steps: S1. Obtain historical operation data of wind turbines; S2. Based on the wind speed-power data distribution, remove the abnormal operating condition distribution data of the wind turbine to obtain the wind speed-power data under normal operating conditions of the turbine; S3. Fill missing values in the wind speed-power data under normal operating conditions of the unit; S4, sampling the interpolated and filled data according to the sampling frequency, and performing modal decomposition on the sampled data using the improved CEEMDAN algorithm; S5, calculate the sample entropy of each modal component in S4; S6. Perform SVMD secondary decomposition on the modal components with larger entropy values; S7, synthesize the modal components decomposed by S4 and S6 into a new modal component set, and build a wind power prediction model based on LSTM network; S8. Input the real-time power data of the wind turbine generator set into the wind power prediction model to perform ultra-short-term rolling prediction of the wind turbine generator set power.
[0005] The present invention is also characterized in that The historical operation data of wind turbines include: wind turbine cut-in wind speed, rated wind speed, cut-out wind speed, and power data.
[0006] S2 is specifically: S21. Based on the distribution of wind speed-power data, perform the first data cleaning using the quartile method; S22. Perform a second cleaning using a longitudinal filtering method based on data discreteness.
[0007] S21 is specifically: First, remove the data points with wind speed less than the cut-in wind speed and greater than the cut-out wind speed, divide the wind speed axis into two wind speed segments: [cut-in wind speed, rated wind speed] and [rated wind speed, cut-out wind speed], and then divide the wind speed segment into several wind speed sub-intervals with equal intervals along the wind speed axis. The interval is generally 0.5m / s. After the division, sort the data of all wind speed sub-intervals from small to large according to the power value to obtain a sample arranged in ascending order. , in sequence , is the total number of samples, Represents a point in a sequence; Each wind speed sub-interval is then divided into four parts according to the quartile method. The values at the three split points are the lower quartiles. , median and upper quantile ; First calculate the median : (1) y is the number of samples in a single wind speed subinterval, k is a non-negative integer; Lower quantile and upper quantile Representation sample The value represented by the position of 25% of the data points before and after the middle separation. y =2 k ( k =0,1,2…), we have: (2) When y=4 k +1( k =0,1,2…), we have: (3) By calculating formula (2) and formula (3), the interquartile range ,Right now: (4) use Determine the inner limit range of a single wind speed sub-interval sample in the sequence, that is: (5) Where: The lower limit value of a single wind speed sub-interval sample determined by the quartile method; is the upper limit value of a single wind speed sub-interval sample; and is the weight.
[0008] Pick ,if , it is determined to be abnormal data and removed; otherwise Retain and perform the first cleaning based on the quartile method on all wind speed-power to obtain the sequence , in sequence , is the total number of samples after the first cleaning, Represents a point in a sequence.
[0009] S22 is specifically: For the power value in each wind speed sub-interval, calculate its mean and variance, namely: (6) Where: For the The mean square error of the power of the subintervals; For the The amount of scatter in the interval; For the The first interval Power values; For the The power mean of the intervals; When the unit is operating normally, the wind speed and power scatter points are normally distributed around the power band, and the power mean value is As the center, To filter the upper and lower bounds, perform secondary data cleaning to obtain the wind speed-power data sequence under normal operating conditions of the unit. , in sequence , is the total number of wind speed-power data under normal working conditions, Represents a point in a sequence.
[0010] S3 specifically: Extract the wind speed-power data under normal operating conditions of the unit obtained in step S2 Power timing data in ,in is a time point in the time series, For the corresponding power values, the cubic spline interpolation method is used to interpolate and fill the missing values of the power time series data; set up is a cubic spline function, in each interval superior, Expressed as: (7) (8) According to the continuity condition of the spline function: , we get m-1 equations, By first-order derivative continuity , we get m-1 equations; By the continuity of the second-order derivative , we get m-1 equations; By natural boundary conditions , and solve to get the interpolation function S(X) , fill in the missing values of power time series data.
[0011] S4 is specifically: The decomposition steps of S41 and CEEMDAN are as follows: S411, first to complete the power timing data Add Gaussian white noise ,get V Second preprocessing sequence (9) Where, is the noise weight coefficient; S412, for each Perform EMD decomposition to obtain the first IMF component , and take its mean as the first IMF component of CEEMDAN decomposition , and get the first residual component : (10) (11) S413, sequence V The first EMD decomposition is obtained after CEEMDAN decomposition. d +1 IMF component: (12) (13) Where, is the first d IMF components; The first decomposition of CEEMDAN d +1 serving; For the d +1 residual component; S414, repeat steps S411-S413 until the decomposition is completed, the original power time series data The expression after CEEMDAN decomposition is: (14) Where: is the residual component after final decomposition; In S42 and CEEMDAN, the false components in the early decomposition will cause the real signal components to appear later. The CEEMDAN in S41 is improved to the ICEEMDAN algorithm. The complete wind turbine power time series data after interpolation and filling is sampled according to the sampling frequency, and then the ICEEMDAN algorithm is used to perform modal decomposition on the sampled power time series data. The decomposition steps of the improved ICEEMDAN algorithm are as follows: S421, through the local mean operator Extract the residual component and rewrite formula (9) as follows: (15) Where: is an operator that averages N local means; is an operator for finding the local mean through EMD; S422, the first IMF 1 serving contains: (16) S423, after subsequent component iterations, the q The residual components and IMF q Serving size: (17) (18) Decomposition by ICEEMDAN algorithm, the final decomposition result is: (19) The specific calculation process of S5 is: S51, for a length of U The modal components of , construct a vector of length L: (20) S52. Define vectors under the same pattern dimension and distance is the maximum absolute value of the difference between corresponding elements of two vectors: (twenty one) S53. Given similarity tolerance Tol , statistics satisfy The vector logarithm of , and calculate its mean: (twenty two) S54, increase the dimension to L +1, repeat steps S51~S53, and get ; S55. Calculate sample entropy: (twenty three).
[0012] S6 specifically: S61. For high entropy components Perform g decompositions, extracting one mode each time and residual components : (twenty four) Where: is the center frequency; is the penalty factor; S62, when the residual component The decomposition is terminated when the energy is lower than the threshold or the maximum number of iterations is reached. The final decomposition result is: (25).
[0013] S7 specifically: The modal components decomposed by S4 and S6 are synthesized into a new modal component set , LSTM is mainly composed of forget gate, input gate and output gate. The specific calculation method is as follows: The calculation formula of the forget gate is: (26) Where: is the weight matrix of the forget gate; is the hidden layer output; is the modal component of the current input; is the bias term; is the sigmoid function; The input gate calculation formula is: (27) (28) Where: and They are the activation function output values of sigmoid and tanh respectively; and is the weight matrix; and is the bias term.
[0014] The cell state calculation formula is: (29) Where: is the current cell status; is the state of the previous cell.
[0015] The output gate calculation formula is: (30) (31) Where: Output of the output gate; is the weight matrix of the output gate; is the bias term; The prediction results of each modal component are superimposed as the final prediction result; S8 specifically includes: importing the real-time power data of the wind turbine into the power prediction model established in S7, predicting the next power point, and updating the predicted value to the power data set, using a sliding window with a step size of 1 to achieve ultra-short-term power rolling prediction of the wind turbine.
[0016] The beneficial effects of the present invention are: The improved CEEMDAN-SVMD-LSTM method for predicting ultra-short-term output of wind turbines of the present invention can effectively utilize wind turbine SCADA data. Through data preprocessing optimization, improved modal decomposition, secondary decomposition noise reduction and other measures, it can predict the ultra-short-term output of intermittent and fluctuating renewable energy, effectively improving prediction accuracy and model performance, and providing strong support for the stable operation and economic dispatch of the power system. It can also guide stations to adjust equipment inspection and maintenance plans in a timely manner, while assisting in maintaining power balance, optimizing grid dispatching operations, and improving the grid's anti-interference capability, thereby ensuring the reliable operation of the power system. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] Figure 1 This is a flow chart of the ultra-short-term output prediction method of a wind turbine generator system based on the improved CEEMDAN-SVMD-LSTM of the present invention; Figure 2 This is a flow chart of wind speed-power data processing in the ultra-short-term output prediction method for wind turbines of the present invention. DETAILED DESCRIPTION
[0018] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.
[0019] The improved CEEMDAN-SVMD-LSTM method for predicting ultra-short-term output of wind turbines of the present invention is as follows: Figure 1 As shown. First, the wind speed-power data is cleaned using the quartile method and the longitudinal filtering method based on data discreteness, and the missing values are filled by the cubic spline interpolation method to ensure the integrity of the power time series data. The power time series data is subjected to improved CEEMDAN-SVMD secondary mode decomposition, and an LSTM power prediction model is constructed. Each modal component after modal decomposition is input into the prediction model respectively, and the prediction results are superimposed to obtain the final prediction result. The process is as follows Figure 1 The specific implementation steps are as follows: S1. Obtain historical operation data of wind turbines.
[0020] The historical operation data of wind turbines include: wind turbine cut-in wind speed, rated wind speed, cut-out wind speed, and power data.
[0021] S2. Based on the distribution of wind speed-power data, the abnormal operating condition distribution data of the wind turbine is eliminated to obtain the wind speed-power data under normal operating conditions of the turbine, thereby improving the prediction accuracy of the model.
[0022] The specific process is as follows Figure 2 Shown, including: S21. First, remove the data points with wind speed less than the cut-in wind speed and greater than the cut-out wind speed, divide the wind speed axis into two wind speed segments: [cut-in wind speed, rated wind speed] and [rated wind speed, cut-out wind speed]. Then, divide the wind speed segment into several wind speed sub-intervals with equal intervals along the wind speed axis. The interval is generally 0.5 m / s. Sort all the wind speed sub-interval data after division by power value from small to large to obtain a sample arranged in ascending order. , in sequence , is the total number of samples, Represents a point in a sequence; Each wind speed sub-interval is then divided into four parts according to the quartile method. The values at the three split points are the lower quartiles. , median and upper quantile .
[0023] First calculate the median : (1) y is the number of samples in a single wind speed subinterval, k It is a non-negative integer that assists in calculation and represents the number of data. nWhether it is an odd or even number.
[0024] Then calculate the lower quantile and upper quantile : and Representation sample The values represented by the positions of 25% of the data points before and after the middle divide.
[0025] when y =2 k ( k =0,1,2…), there is (2) When y=4 k +1( k =0,1,2…), there is (3) By calculating formula (2) and formula (3), the interquartile range ,Right now: (4) In statistics, use Determine the inner limit range of a single wind speed sub-interval sample in the sequence, that is: (5) Where: The lower limit value of a single wind speed sub-interval sample determined by the quartile method; is the upper limit value of a single wind speed sub-interval sample; and is the weight.
[0026] Pick ,if , it is determined to be abnormal data and removed; otherwise Retain and perform the first cleaning based on the quartile method on all wind speed-power to obtain the sequence , in sequence , is the total number of samples after the first cleaning, Represents a point in a sequence.
[0027] S22. Perform a second cleaning using a longitudinal filtering method based on data discreteness.
[0028] For the power value in each wind speed sub-interval, calculate its mean and variance, namely: (6) Where: For the The mean square error of the power of the subintervals; For the The amount of scatter in the interval; For the The first interval Power values; For the The power mean of the interval.
[0029] When the unit is operating normally, the wind speed and power scatter points are normally distributed around the power band, that is, 95% of the data are within the confidence interval covered by the two mean square errors above and below the mean. As the center, To filter the upper and lower bounds, perform secondary data cleaning to obtain the wind speed-power data sequence under normal operating conditions of the unit. , in sequence , is the total number of wind speed-power data under normal working conditions, Represents a point in a sequence.
[0030] S3. Interpolate and fill missing values in the wind speed-power data under normal operating conditions of the unit.
[0031] Specifically, extract the wind speed-power data of the unit under normal operating conditions obtained in step S2: Power timing data in ,in is a time point in the time series, The cubic spline interpolation method is used to fill in the missing values of the power time series data to ensure the integrity of the power data time series.
[0032] set up is a cubic spline function, in each interval superior, It can be expressed as: (7) (8) According to the continuity condition of the spline function: , and we get m-1 equations.
[0033] By first-order derivative continuity , and we get m-1 equations.
[0034] By the continuity of the second-order derivative , and we get m-1 equations.
[0035] By natural boundary conditions , and solve to get the interpolation function S(X) , fill in the missing values of power time series data.
[0036] S4, the complete wind turbine power time series data after S3 interpolation and filling is sampled every 15 minutes, and the improved CEEMDAN algorithm is used to perform modal decomposition on the sampled power time series data. Decompose into modal components of different frequencies to reduce data complexity, speed up model prediction and improve prediction accuracy; S41, CEEMDAN (Complete Ensemble Empirical Mode Decomposition with Adaptive Noise) is a signal decomposition algorithm based on EEMD (Ensemble Empirical Mode Decomposition). It is primarily used to decompose complex signals into a series of intrinsic mode functions (IMFs) with different characteristic scales to better reflect the signal's characteristics at different frequency bands.
[0037] The CEEMDAN decomposition steps are as follows: 1) First, send power timing data Add Gaussian white noise ,get V Second preprocessing sequence (9) Where, is the noise weight coefficient.
[0038] 2) For each Perform EMD decomposition to obtain the first IMF component , and take its mean as the first IMF component of CEEMDAN decomposition , and get the first residual component .
[0039] (10) (11) 3) Perform the sequence V The first EMD decomposition is obtained after CEEMDAN decomposition. d +1 IMF component.
[0040] (12) (13) Where, is the first d IMF components; The first decomposition of CEEMDAN d +1 serving; For the d +1 residual component; 4) Repeat steps 1)-3) until the decomposition is complete, the original signal The expression after CEEMDAN decomposition is: (14) Where: is the residual component after the final decomposition.
[0041] The false components in the early decomposition of S42 and CEEMDAN will cause the real signal components to appear later. To this end, the ICEEMDAN algorithm is proposed. The complete wind turbine power time series data after S3 interpolation is sampled every 15 minutes, and then the ICEEMDAN algorithm is used to perform modal decomposition on the sampled power time series data. The specific decomposition steps are as follows: 1) Through the local mean operator Extract the residual component and rewrite formula (9) as follows: (15) Where: is an operator that averages N local means; is an operator that finds the local mean through EMD.
[0042] 2) The first IMF 1 serving contains: (16) 3) After subsequent component iterations, q The residual components and IMF q Serving size: (17) (18) Decomposition by ICEEMDAN algorithm, the final decomposition result is: (19) S5. Calculate the sample entropy of each modal component in S4.
[0043] Sample entropy is a key metric for measuring the complexity and irregularity of a time series, and is used to assess the probability of a signal generating new patterns. Lower sample entropy values indicate greater signal self-similarity and lower complexity; higher values indicate a more complex signal and a greater probability of generating new patterns.
[0044] 1) For a length ofU The modal components of , construct a vector of length L: (20) 2) Define vectors under the same pattern dimension and distance is the maximum absolute value of the difference between corresponding elements of two vectors: (twenty one) 3) Given similarity tolerance Tol , statistics satisfy The vector logarithm of , and calculate its mean: (twenty two) 4) Increase the dimension to L +1, repeat steps 1) to 3), and get .
[0045] 5) Calculate sample entropy: (twenty three).
[0046] S6. Perform SVMD secondary decomposition on the components with larger sample entropy to further reduce the sequence complexity and improve the prediction accuracy.
[0047] Successive Variational Mode Decomposition (SVMD) is an evolution of variational mode decomposition (VMD). VMD constructs a variational model to decompose a signal into a series of modal functions with different center frequencies. Each modal function is an amplitude-frequency modulated signal with a specific bandwidth. SVMD, on the other hand, decomposes the signal incrementally, extracting one modal function at a time. The remaining signal then serves as the input for the next decomposition. This adaptively determines the number of modes, avoiding the extraction of redundant modes, reducing computation time, and improving convergence speed.
[0048] 1) For high entropy components Perform g decompositions, extracting one mode each time and residual components : (twenty four) Where: is the center frequency; is the penalty factor.
[0049] 2) When the residual component The decomposition is terminated when the energy is lower than the threshold or the maximum number of iterations is reached. The final decomposition result is: (25) S7, synthesize the modal components decomposed by S4 and S6 into a new modal component set , build a wind power prediction model based on LSTM network.
[0050] Traditional RNNs encounter vanishing and exploding gradients when processing long sequences of data. This makes it difficult for the model to learn long-term dependencies in the time series, potentially leading to unstable training and even non-convergence. Long-Short Term Memory (LSTM) networks, by introducing a gating mechanism, can extract information about medium- and long-term temporal sequences, offering advantages over other neural networks in time series prediction.
[0051] LSTM is mainly composed of a forget gate, an input gate, and an output gate. The specific calculation method is as follows: 1) The calculation formula of the forget gate is: (26) Where: is the weight matrix of the forget gate; is the hidden layer output; is the modal component of the current input; is the bias term; is the sigmoid function.
[0052] 2) The input gate calculation formula is: (27) (28) Where: and They are the activation function output values of sigmoid and tanh respectively; and is the weight matrix; and is the bias term.
[0053] 3) The formula for calculating cell status is: (29) Where: is the current cell status; is the state of the previous cell.
[0054] 4) The output gate calculation formula is: (30) (31) Where: Output of the output gate; is the weight matrix of the output gate; is the bias term.
[0055] The prediction results of each modal component are superimposed as the final prediction result.
[0056] S8. Import the real-time power data of the wind turbine into the power prediction model established in S7, predict the next power point, and update the predicted value to the power data set. Use a sliding window with a step size of 1 to achieve ultra-short-term power rolling prediction of the wind turbine.
[0057] The improved CEEMDAN-SVMD-LSTM method for predicting ultra-short-term wind turbine output, based on data preprocessing optimization, improved modal decomposition, and secondary decomposition noise reduction, predicts the intermittent and volatile ultra-short-term output of renewable energy sources, effectively improving prediction accuracy and model performance. It can also guide stations to adjust equipment maintenance plans in a timely manner, while also assisting in maintaining power balance, optimizing grid dispatching operations, and enhancing grid anti-interference capabilities, ensuring reliable operation of the power system. This method helps renewable energy power generation companies reduce investment risks and promote renewable energy consumption, and is of great significance to the healthy and efficient development of the new energy industry.
Claims
1. The ultra-short-term output prediction method of wind turbines based on improved CEEMDAN-SVMD-LSTM is characterized by: Here are the steps: S1. Obtain historical operation data of wind turbines; S2. Based on the wind speed-power data distribution, remove the abnormal operating condition distribution data of the wind turbine to obtain the wind speed-power data under normal operating conditions of the turbine; S3. Fill missing values in the wind speed-power data under normal operating conditions of the unit; S4, sampling the interpolated and filled data according to the sampling frequency, and performing modal decomposition on the sampled data using the improved CEEMDAN algorithm; S5, calculate the sample entropy of each modal component in S4; S6. Perform SVMD secondary decomposition on the modal components with larger entropy values; S7, synthesize the modal components decomposed by S4 and S6 into a new modal component set, and build a wind power prediction model based on LSTM network; S8. Input the real-time power data of the wind turbine generator set into the wind power prediction model to perform ultra-short-term rolling prediction of the wind turbine generator set power.
2. The method for ultra-short-term wind turbine output prediction based on improved CEEMDAN-SVMD-LSTM according to claim 1 is characterized in that: The historical operation data of the wind turbine generator set includes: wind turbine generator set cut-in wind speed, rated wind speed, cut-out wind speed, and power data.
3. The method for ultra-short-term wind turbine output prediction based on improved CEEMDAN-SVMD-LSTM according to claim 1 is characterized in that: The S2 is specifically: S21. Based on the distribution of wind speed-power data, perform the first data cleaning using the quartile method; S22. Perform a second cleaning using a longitudinal filtering method based on data discreteness.
4. The method for ultra-short-term wind turbine output prediction based on improved CEEMDAN-SVMD-LSTM according to claim 3 is characterized in that: The S21 is specifically as follows: First, remove the data points with wind speed less than the cut-in wind speed and greater than the cut-out wind speed, divide the wind speed axis into two wind speed segments: [cut-in wind speed, rated wind speed] and [rated wind speed, cut-out wind speed], and then divide the wind speed segment into several wind speed sub-intervals with equal intervals along the wind speed axis. The interval is generally 0.5m / s. After the division, sort the data of all wind speed sub-intervals from small to large according to the power value to obtain a sample arranged in ascending order. , in sequence , is the total number of samples, Represents a point in a sequence; Each wind speed sub-interval is then divided into four parts according to the quartile method. The values at the three split points are the lower quartiles. , median and upper quantile ; First calculate the median : (1) y is the number of samples in a single wind speed subinterval, k is a non-negative integer; Lower quantile and upper quantile Representation sample The value represented by the position of 25% of the data points before and after the middle separation. y =2 k ( k =0,1,2…), we have: (2) When y=4 k +1( k =0,1,2…), we have: (3) By calculating formula (2) and formula (3), the interquartile range ,Right now: (4) use Determine the inner limit range of a single wind speed sub-interval sample in the sequence, that is: (5) Where: The lower limit value of a single wind speed sub-interval sample determined by the quartile method; is the upper limit value of a single wind speed sub-interval sample; and is the weight; Pick ,if , determined as abnormal data and removed; on the contrary Retain and perform the first cleaning based on the quartile method on all wind speed-power to obtain the sequence , in sequence , is the total number of samples after the first cleaning, Represents a point in a sequence.
5. The method for ultra-short-term wind turbine output prediction based on improved CEEMDAN-SVMD-LSTM according to claim 4 is characterized in that: The S22 is specifically: For the power value in each wind speed sub-interval, calculate its mean and variance, namely: (6) Where: For the The mean square error of the power of the subintervals; For the The amount of scatter in the interval; For the The first interval Power values; For the The power mean of the intervals; When the unit is operating normally, the wind speed and power scatter points are normally distributed around the power band, and the power mean value is As the center, To filter the upper and lower bounds, perform secondary data cleaning to obtain the wind speed-power data sequence under normal operating conditions of the unit. , in sequence , is the total number of wind speed-power data under normal working conditions, Represents a point in a sequence.
6. The method for ultra-short-term wind turbine output prediction based on improved CEEMDAN-SVMD-LSTM according to claim 5 is characterized in that: The S3 is specifically: Extract the wind speed-power data under normal operating conditions of the unit obtained in step S2 Power timing data in ,in is a time point in the time series, For the corresponding power values, the cubic spline interpolation method is used to interpolate and fill the missing values of the power time series data; set up is a cubic spline function, in each interval superior, Expressed as: (7) (8) According to the continuity condition of the spline function: , we get m-1 equations, By first-order derivative continuity , we get m-1 equations; By the continuity of the second-order derivative , we get m-1 equations; By natural boundary conditions , and solve to get the interpolation function S(X) , fill in the missing values of power time series data.
7. The method for ultra-short-term wind turbine output prediction based on improved CEEMDAN-SVMD-LSTM according to claim 1, characterized in that: The S4 is specifically: The decomposition steps of S41 and CEEMDAN are as follows: S411, first to complete the power timing data Add Gaussian white noise ,get V Second preprocessing sequence (9) Where, is the noise weight coefficient; S412, for each Perform EMD decomposition to obtain the first IMF component , and take its mean as the first IMF component of CEEMDAN decomposition , and get the first residual component : (10) (11) S413, sequence V The first EMD decomposition is obtained after CEEMDAN decomposition. d +1 IMF component: (12) (13) Where, is the first d IMF components; The first decomposition of CEEMDAN d +1 serving; For the d +1 residual component; S414, repeat steps S411-S413 until the decomposition is completed, the original power time series data The expression after CEEMDAN decomposition is: (14) Where: is the residual component after final decomposition; In S42 and CEEMDAN, the false components in the early decomposition will cause the real signal components to appear later. The CEEMDAN in S41 is improved to the ICEEMDAN algorithm. The complete wind turbine power time series data after interpolation and filling is sampled according to the sampling frequency, and then the ICEEMDAN algorithm is used to perform modal decomposition on the sampled power time series data. The decomposition steps of the improved ICEEMDAN algorithm are as follows: S421, through the local mean operator Extract the residual component and rewrite formula (9) as follows: (15) Where: is an operator that averages N local means; is an operator for finding the local mean through EMD; S422, the first IMF 1 serving contains: (16) S423, after subsequent component iterations, the q The residual components and IMF q Serving size: (17) (18) Decomposition by ICEEMDAN algorithm, the final decomposition result is: (19)。 8. The method for ultra-short-term output prediction of wind turbines based on improved CEEMDAN-SVMD-LSTM according to claim 1, characterized in that: The specific calculation process of S5 is: S51, for a length of U The modal components of , construct a vector of length L: (20) S52. Define vectors under the same pattern dimension and distance is the maximum absolute value of the difference between corresponding elements of two vectors: (21) S53. Given similarity tolerance Tol , statistics satisfy The vector logarithm of , and calculate its mean: (22) S54, increase the dimension to L +1, repeat steps S51~S53, and get ; S55. Calculate sample entropy: (23)。 9. The method for ultra-short-term wind turbine output prediction based on improved CEEMDAN-SVMD-LSTM according to claim 1, characterized in that: The S6 is specifically: S61. For high entropy components Perform g decompositions, extracting one mode each time and residual components : (24) Where: is the center frequency; is the penalty factor; S62, when the residual component The decomposition is terminated when the energy is lower than the threshold or the maximum number of iterations is reached. The final decomposition result is: (25)。 10. The method for ultra-short-term output prediction of wind turbines based on improved CEEMDAN-SVMD-LSTM according to claim 1, characterized in that: The S7 is specifically: The modal components decomposed by S4 and S6 are synthesized into a new modal component set , LSTM is mainly composed of forget gate, input gate and output gate. The specific calculation method is as follows: The calculation formula of the forget gate is: (26) Where: is the weight matrix of the forget gate; is the hidden layer output; is the modal component of the current input; is the bias term; is the sigmoid function; The input gate calculation formula is: (27) (28) Where: and They are the activation function output values of sigmoid and tanh respectively; and is the weight matrix; and is the bias term; The cell state calculation formula is: (29) Where: is the current cell status; is the state of the previous cell; The output gate calculation formula is: (30) (31) Where: Output of the output gate; is the weight matrix of the output gate; is the bias term; The prediction results of each modal component are superimposed as the final prediction result; The S8 is specifically as follows: importing the real-time power data of the wind turbine into the power prediction model established in S7, predicting the next power point, and updating the predicted value to the power data set, using a sliding window with a step size of 1 to achieve ultra-short-term power rolling prediction of the wind turbine.