Wind power hybrid prediction method and system based on deep learning
Through the deep learning-based hybrid wind power prediction method, the problem of insufficient wind power power prediction accuracy is solved. Through outlier value processing, missing value filling and modal decomposition, combined with the CNN-LSTM model, higher precision wind power prediction is achieved.
Patent Information
- Application Number
- CN202311824233.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-12-27
- Publication Date
- 2025-08-01
AI Technical Summary
The accuracy of the power prediction of stroke power in the prior art is insufficient, especially when processing outliers and missing data.
The wind power hybrid prediction method based on deep learning is adopted, including outlier value removal, missing value filling, ensemble empirical modal decomposition and CNN-LSTM model construction. The abnormal data is filtered through the Gaussian function method and the quartile method, the missing data is processed using the interpolation method, and the modal decomposition is performed through the EEMD algorithm, and the power prediction is performed by combining the CNN and LSTM models.
It improves the accuracy of wind power power prediction, can effectively extract the spatial and temporal characteristics of wind power data, and enhances the accuracy of prediction.
Smart Images
Figure CN120414461A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of power systems, and particularly relates to a wind power hybrid prediction method and system. Background Art
[0002] With the increase of human social activities and the increasing development of science and technology, the problem of environmental protection has become increasingly prominent. At the same time, the quantity of non-renewable energy is gradually decreasing. Therefore, countries around the world have begun to pay more attention to renewable energy, especially high-clean renewable technology energy represented by wind energy. Wind energy resources have the characteristics of wide distribution and no pollution, and are one of the most promising renewable energies among all renewable energies. Under the background of the conventional energy crisis and the vigorous development of new energy, using wind power generation has become a research hotspot in the world today.
[0003] The wind power industry has developed rapidly, and the global wind power installed capacity has increased rapidly. Accelerating the development of the wind power industry has become the consensus and concerted action of the international community to promote energy transformation and development and address global climate change. Countries around the world have introduced industrial policies to encourage the development of wind power.
[0004] With the rapid development of artificial intelligence, artificial intelligence models for wind power prediction have also emerged. By learning strategies, a mapping relationship model between input and output is established, and then the output is predicted through the model. This method has been widely used in recent years and has achieved good prediction results. Summary of the Invention
[0005] The present invention proposes a wind power hybrid prediction method and system based on deep learning to solve the accuracy problem of wind power prediction.
[0006] The present invention adopts the following technical solutions:
[0007] A wind power hybrid prediction method based on deep learning, comprising the following steps:
[0008] Step 1: First, eliminate the outliers of wind power, and the outliers include the outliers during faults and curtailment of wind power;
[0009] Step 2: Fill in the missing wind power;
[0010] Step 3: Perform modal decomposition on the wind power data processed in Step 1 and Step 2 by using the ensemble empirical mode decomposition algorithm;
[0011] Step 4: Construct a CNN-LSTM power prediction model to predict the power in different modes, and perform power clustering to obtain the final predicted wind power.
[0012] Furthermore, the Gaussian function method and the quartile method are used to filter the abnormal data:
[0013] Q = [Q1 - 1.5σ, Q3 + 1.5σ]
[0014] Where Q1, Q, and Q3 are the first quartile (25%), the second quartile (50%), and the third quartile (75%) respectively.
[0015] Furthermore, interpolation method is used to process the missing data, and the process equation of the interpolation method is
[0016]
[0017] Where x is i-1 the wind power at the previous moment of x(t). i+1 x(t) is the wind power of the wind power generation at a later moment. If two or more consecutive missing values are within a range, the interpolation rule is to directly use the value at the previous moment as the interpolated missing value.
[0018] Furthermore, the ensemble empirical mode decomposition algorithm is used for mode decomposition, and the steps are as follows:
[0019] (1) Add the white noise time series to the original signal time series to construct a new time series.
[0020] Y n (t) = X(t) + u n (t)
[0021] Where n = 1, 2, …, N, and N is the number of tests;
[0022] (2) Based on the EMD algorithm, decompose the time series contaminated by noise into the sum of a series of intrinsic mode functions and residuals:
[0023]
[0024] Where m is the number of algorithm iterations and n is the number of integrations;
[0025] (3) Iterate steps (1) and (2) N times. In each iteration test, add different white noise time series u n (t) to the original signal time series;
[0026] (4) The intrinsic mode function Obtain the average value through N integration experiments:
[0027]
[0028] Furthermore, construct a CNN-LSTM power prediction model, including:
[0029] (1) Construct the network model of CNN as:
[0030]
[0031] f(F) = tanh(F)
[0032] Y = f(F)
[0033] where i, W i , B i and f(F) are the convolution kernel, weight, bias, and activation function, respectively;
[0034] (2) Construct the LSTM network model as follows:
[0035] f t = S(W f x t + U f h t-1 + b f )
[0036] i t = S(W i x t + U i h t-1 + b i )
[0037]
[0038]
[0039]
[0040] where S is the sigmoid activation function. W f represents the weight matrix of the input gate, b f represents the offset of the input gate. W i is the weight matrix of the input gate, b i is the offset entry of the input gate, W c is the weight matrix of the output gate, b c is the offset term of the output gate, W o is the weight matrix of the input gate, b o is the offset entry of the input gate.
[0041] A wind power hybrid prediction system based on deep learning, characterized by comprising:
[0042] Preprocessing module: First, eliminate the outliers of wind power, including the outliers during faults and curtailment of wind power, and fill in the missing wind power;
[0043] Modal decomposition module: Perform modal decomposition on the processed wind power data using the ensemble empirical mode decomposition algorithm;
[0044] Model construction module: Construct a CNN-LSTM power prediction model to predict the power under different modalities, and perform power clustering to obtain the final predicted wind power.
[0045] Further, the preprocessing module filters out abnormal data using the Gaussian function method and the quartile method:
[0046] Q = [Q1 - 1.5σ, Q3 + 1.5σ]
[0047] where Q1, Q, and Q3 are the first quartile 25%, the second quartile 50%, and the third quartile 75% respectively;
[0048] The interpolation method is used to process missing data, and the process equation of the interpolation method is
[0049] [[ID=IC=16]]
[0050] where x is i-1 (t) the wind power at the previous moment. x i+1 (t) is the wind power of the wind power at a later moment. If there are two or more consecutive missing values within a range, the interpolation rule is to directly use the value of the previous moment as the missing value for interpolation.
[0051] Further, the modal decomposition module uses the ensemble empirical mode decomposition algorithm for modal decomposition, and the steps are as follows:
[0052] (1) Add a white noise time series to the original signal time series to construct a new time series.
[0053] Y n (t) = X(t) + u n (t)
[0054] where n = 1, 2,..., N, and N is the number of tests;
[0055] (2) Based on the EMD algorithm, decompose the time series contaminated by noise into the sum of a series of intrinsic mode functions and residuals:
[0056]
[0057] where m is the number of algorithm iterations and n is the number of integrations;
[0058] (3) Iterate steps (1) and (2) N times. In each iteration test, add different white noise time series u n (t) to the original signal time series;
[0059] (4) Intrinsic mode function The average value is obtained through N integrated experiments:
[0060]
[0061] Furthermore, the model construction module constructs a CNN-LSTM power prediction model, including:
[0062] (1) The network model of CNN is constructed as:
[0063]
[0064] f(F) = tanh(F)
[0065] Y = f(F)
[0066] where i, W i , B i and f(F) are the convolution kernel, weight, bias, and activation function respectively;
[0067] (2) The network model of LSTM is constructed as:
[0068] f t = S(W f x t + U f h t-1 + b f )
[0069] i t = S(W i x t + U i h t-1 + b i )
[0070]
[0071]
[0072]
[0073] where S is the sigmoid activation function. W f represents the weight matrix of the input gate, b f represents the offset of the input gate. W i is the weight matrix of the input gate, b i is the offset entry of the input gate, W c is the weight matrix of the output gate, b c is the offset term of the output gate, W o is the weight matrix of the input gate, b o is the offset entry of the input gate.
[0074] The present invention has the following beneficial effects:
[0075] In order to improve the prediction of short-term wind power, the present invention establishes a hybrid wind power prediction model. In this model, a quartile outlier data detection method is proposed, which can eliminate outliers between the input data and power to a certain extent. In addition, EEMD is used to decompose the wind power to reduce the non-stationary signal of the wind power. Considering the dependence of wind power data on time and space, CNN and LSTM models are used to mine the spatio-temporal variation characteristics of the wind power. Finally, experiments are carried out using the actual operation data of the wind turbines, and it is verified that the proposed model can fully extract features and improve the accuracy of wind power prediction. Description of the Drawings
[0076] Figure 1 A hybrid power prediction system provided by an embodiment of the present invention;
[0077] Figure 2 An outlier data detection method provided by an embodiment of the present invention;
[0078] Figure 3 An EEMD decomposition power diagram provided by an embodiment of the present invention. Detailed Embodiment
[0079] The following combines the drawings and embodiments to further describe in detail the specific embodiments of the present invention. The following embodiments are used to illustrate the present invention, but are not used to limit the scope of the present invention.
[0080] A wind power hybrid prediction method based on deep learning includes:
[0081] Step 1: Input historical data, such as wind speed, humidity, etc.
[0082] All prediction models are tested on MATLAB R2021b using an Intel(R) Core i5-4710HQ CPU@2.5GHz and 16GB of RAM. The data acquisition period is 15 minutes. The data information includes wind speed, wind direction, pressure, humidity, temperature, and wind force. The data acquisition time is from 0:00 on January 1, 2013 to 24:00 on December 31, 2013.
[0083] Step 2: Perform data processing according to historical power data.
[0084] Due to data transmission and sampling device failures, outlier data is detected to ensure a good data set. Combining the quartile theory, the data statistics are shown in Table 1.
[0085] Table 1
[0086] Type Quantity Original data statistics 35040 Outlier statistics 5471
[0087] Step 3: Perform EEMD decomposition on the original power.
[0088] Perform EEMD decomposition on the original wind power. The decomposition results of some data are as Figure 3 shown.
[0089] Step 4: Construct a CNN-LSTM to predict powers with different characteristics.
[0090] The model parameters are as follows: number of convolutional kernels: 4; convolutional kernel size: 2*2; number of hidden layers: 2; number of neurons in the hidden layers: 20, 10.
[0091] Step 5: Aggregate the prediction results. Finally, the power prediction results are given.
[0092] In this embodiment, the predicted time scales are 4 hours, 8 hours, and 12 hours respectively. The performance prediction is shown in Table 2.
[0093] Table 2
[0094]
[0095] In this model, a quartile abnormal data detection method is proposed, which can eliminate the outliers between the input data and the power to a certain extent.
[0096] The steps of the hybrid power prediction method proposed in this paper are as follows, and the structure is as Figure 1 shown:
[0097] Step 1: Input historical data such as wind speed, humidity, etc.
[0098] Step 2: Perform data processing according to the historical power data.
[0099] Step 3: Perform EEMD decomposition on the original power.
[0100] Step 4: Construct a CNN-LSTM to predict powers with different characteristics.
[0101] Step 5: Aggregate the prediction results. Finally, the power prediction results are given.
[0102] Step 6: Conduct model evaluation.
[0103] The present invention adopts the following technical solutions:
[0104] Step 1: Input historical data such as wind speed, humidity, etc.;
[0105] Step 2: Perform data processing according to the historical power data.
[0106] This patent uses the Gaussian function method and the quartile method to filter abnormal data.
[0107] Q=[Q1-1.5σ,Q3+1.5σ]
[0108] Where Q1, Q and Q1 are the first quartile (25%), the second quartile (50%) and the third quartile (75%) respectively, as shown in Figure 2 shown.
[0109] This patent uses interpolation to process missing data and perform effective data filling. The process equation of the interpolation method is
[0110]
[0111] where x i-1 (t) Wind power at the previous moment. x i+1 (t) is the wind power at the later time. If there are two or more consecutive missing values within a range, the interpolation rule is to directly use the value at the previous moment as the missing value to be interpolated.
[0112] Step 3: Perform EEMD decomposition on the original power.
[0113] Use the ensemble empirical mode decomposition algorithm to perform modal decomposition. The steps are as follows
[0114] (1) Add the white noise time series to the original signal time series to construct a new time series.
[0115] Y n (t) = X(t) + u n (t)
[0116] Where n = 1, 2, ..., N, N is the number of tests.
[0117] (2) Based on the EMD algorithm, the noise-contaminated time series is decomposed into a series of characteristic mode functions and the sum of residuals.
[0118]
[0119] Where m is the number of algorithm iterations and n is the number of integrations.
[0120] (3) Steps (1) and (2) are repeated N times. In each iteration test, different white noise time series u are added to the original signal time series. n (t).
[0121] (4) Modal function The average value is calculated through N integrated experiments.
[0122]
[0123] Step 4: Construct a CNN-LSTM to predict powers with different features.
[0124] (1) The network model of the CNN is constructed as
[0125]
[0126] f(F) = tanh(F)
[0127] Y = f(F)
[0128] where i, W i , B i and f(F) are the convolution kernel, weight, bias, and activation function, respectively.
[0129] (2) The network model of the LSTM is constructed as
[0130] f t = S(W f x t + U f h t-1 + b f )
[0131] i t = S(W i x t + U i h t-1 + b i )
[0132]
[0133]
[0134]
[0135] where S is the sigmoid activation function. W f represents the weight matrix of the input gate. b f represents the offset of the input gate. W i is the weight matrix of the input gate. b i is the offset entry of the input gate. W c is the weight matrix of the output gate. b c is the offset term of the output gate. W o is the weight matrix of the input gate. b o is the offset entry of the input gate.
[0136] Step 6: Conduct model evaluation.
[0137] To evaluate the performance of the model, standard evaluation rules were constructed to evaluate each prediction model reasonably and scientifically. Therefore, the mean absolute error (MAE), mean absolute percentage error (MAPE), and root mean square error (RMSE) were used to evaluate the prediction performance of the model, respectively, as follows:
[0138]
[0139]
[0140]
[0141] where y true and y pre are the actual power and the predicted power, respectively.
[0142] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the embodiments of the present invention, rather than to limit them; although the embodiments of the present invention have been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A wind power hybrid prediction method based on deep learning, characterized in that, Including the following steps: Step 1: First, eliminate the outliers of wind power, where the outliers include those during faults and curtailment of wind power; Step 2: Fill in the missing wind power; Step 3: Based on the wind power data processed in Step 1 and Step 2, perform modal decomposition using the ensemble empirical mode decomposition algorithm; Step 4: Construct a CNN-LSTM power prediction model to predict the power under different modes, and perform power clustering to obtain the final predicted wind power.
2. The prediction method according to claim 1, wherein Use the Gaussian function method and the quartile method to filter outlier data: Q = [Q1 - 1.5σ, Q3 + 1.5σ] where Q1, Q, and Q3 are the first quartile of 25%, the second quartile of 50%, and the third quartile of 75% respectively.
3. The prediction method according to claim 1, wherein Use the interpolation method to process the missing data, and the process equation of the interpolation method is where x i-1 (t) is the wind power at the previous moment. x i+1 (t) is the wind power of wind power generation at a later moment. If two or more missing values are consecutive within a range, the interpolation rule is to directly use the value at the previous moment as the missing value for interpolation.
4. The prediction method according to claim 1, wherein Perform modal decomposition using the ensemble empirical mode decomposition algorithm, and the steps are as follows: (1) Add a white noise time series to the original signal time series to construct a new time series. Y n y(t) = X(t) + u n (t) where n = 1, 2, …, N, and N is the number of tests; (2) Based on the EMD algorithm, decompose the time series contaminated by noise into the sum of a series of intrinsic mode functions and residuals: where m is the number of algorithm iterations and n is the number of integrations; (3) Steps (1) and (2) are iterated N times. In each iteration test, different white noise time series u n (t) are added to the original signal time series; (4) Modal function Find the average value through N integrated experiments:
5. The prediction method according to claim 1, wherein Construct a CNN-LSTM power prediction model, including: (1) The network model of CNN is constructed as: f(F) = tanh(F) where i, W i , B i and f(F) are the convolution kernel, weights, bias, and activation function respectively; Y = f(F) f t = S(W f x t + U f h t-1 + b f ) i t = S(W i x t + U i h t-1 + b i ) Among them, S is the S-shaped activation function. W f represents the weight matrix of the input gate, b f represents the offset of the input gate. W i is the weight matrix of the input gate, b i is the offset entry of the input gate, W c is the weight matrix of the output gate, b c is the offset term of the output gate, W o is the weight matrix of the input gate, b o is the offset entry of the input gate.
6. A wind power hybrid prediction system based on deep learning, characterized in that, (2) The network model of LSTM is constructed as: Including:
7. The wind power hybrid prediction system based on deep learning according to claim 6, wherein Preprocessing module: First, eliminate the outliers of wind power, where the outliers include those during faults and curtailment of wind power, and fill in the missing wind power; Modal decomposition module: Perform modal decomposition on the processed wind power data using the ensemble empirical mode decomposition algorithm; Where x i-1 (t) is the wind power at the previous moment. x i+1 (t) is the wind power of wind power in the later moment. If the missing values are two or more consecutive within a range, the interpolation rule is to directly use the value of the previous moment as the missing value for interpolation. Model construction module: Construct a CNN-LSTM power prediction model to predict the power under different modes, and perform power clustering to obtain the final predicted wind power. Y n Y(t) = X(t) + u n Y(t) The preprocessing module uses the Gaussian function method and the quartile method to filter outlier data: Q = [Q1 - 1.5σ, Q3 + 1.5σ] (3) Steps (1) and (2) are iterated N times. In each iteration test, different white noise time series u n (t) are added to the original signal time series; (4) Modal function Calculate the average value through N integrated experiments:
9. The wind power hybrid prediction system based on deep learning according to claim 6, wherein, where Q1, Q, and Q3 are the first quartile of 25%, the second quartile of 5%, and the third quartile of 75% respectively; Use the interpolation method to process the missing data, and the process equation of the interpolation method is where \(i\), \(W\) i , \(B\) i and \(f(F)\) are the convolution kernel, weights, bias, and activation function, respectively; 8. A wind power hybrid prediction system based on deep learning according to claim 6, characterized in that The modal decomposition module performs modal decomposition using the ensemble empirical mode decomposition algorithm, and the steps are as follows: (1) Add a white noise time series to the original signal time series to construct a new time series. where n = 1, 2, …, N, and N is the number of tests; (2) Based on the EMD algorithm, decompose the time series contaminated by noise into the sum of a series of intrinsic mode functions and residuals: where m is the number of algorithm iterations and n is the number of integrations; The model construction module constructs a CNN-LSTM power prediction model, including: (1) The network model of CNN is constructed as: f(F) = tanh(F) Y = f(F) (2) The network model of LSTM is constructed as: f t = S(W f x t + U f h t-1 + b f ) i t = S(W i x t + U i h t-1 + b i ) Among them, S is the S-shaped activation function. W f represents the weight matrix of the input gate, b f represents the offset of the input gate. W i is the weight matrix of the input gate, b i is the offset entry of the input gate, W c is the weight matrix of the output gate, b c is the offset term of the output gate, W o is the weight matrix of the input gate, b o is the offset entry of the input gate.