Ultra-short-term wind power prediction method

Abstract

The invention belongs to the technical field of wind power prediction, and particularly discloses an ultra-short-term wind power prediction method, which comprises the following steps: S1, obtaining an original wind power sequence; s2, decomposing the wind power signal into a series of sub-signals by adopting a CEEMDAN signal decomposition algorithm; s3, denoising and reconstructing the IMF1 component by adopting an improved singular spectrum analysis (SSA) algorithm; s4, the permutation entropy of each component is calculated, the components are divided into high frequency, medium frequency and low frequency according to the permutation entropy, the high frequency is combined together, and the medium frequency and the low frequency are combined together; and S5, sending the data into the established TCN-GRU prediction network, and superposing the results of the two prediction networks to obtain a final prediction result. According to the ultra-short-term wind power prediction method, the prediction performance of the model is better than that of a traditional model and other models based on TCN, high-precision and high-stability prediction of the wind power plant power is achieved, and decision support is provided for operation management of the wind power plant.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power prediction, and in particular to an ultra-short-term wind power prediction method. Background Art

[0002] Wind power is a vital clean energy source today. However, due to the inherent randomness, volatility, and intermittency of wind energy, current wind power generation is highly uncertain. These limitations pose significant challenges to integrating large wind farms into modern power systems. Accurate wind power forecasting is essential for ensuring stable power system operation and promoting wind power consumption. To this end, we propose a novel wind power forecasting network.

[0003] Currently, there are two mainstream classification standards in the field of wind power forecasting. One is based on time, which is divided into ultra-short-term, short-term, medium-term, and long-term, as shown in Table 1; the other is based on the application method, which is divided into physical methods, statistical methods, and hybrid model methods.

[0004] Table 1 Prediction range in wind power forecast ;

[0005] Physical methods typically use numerical weather prediction (NWP) data to model wind turbines, applying predicted wind speeds to associated wind turbine power curves (typically provided by turbine manufacturers) to predict wind power generation. While these methods do not require historical data for training and offer good interpretability, they rely on physical data and are computationally complex. For example, Focken et al. created a physical prediction model that receives input from a weather forecast model and uses factors such as terrain as a boundary layer to predict power output for the next 48 hours.

[0006] Statistical models are generally based on establishing nonlinear and linear relationships between NWP data (such as wind speed, direction, and temperature) and power generation. To define this statistical relationship, historical data is used as training data. The model is then adjusted by comparing model predictions with online power measurements. The model can then be used to predict NWP forecasts and online measurements for the next few hours. Statistical methods can be further categorized into three categories: time series analysis, machine learning (ML), and deep learning. Within the context of time series analysis, AR, ARMA, and ARIMA models have emerged. With the development of artificial intelligence theory, machine learning and deep learning have been applied to wind power forecasting. For example, M. Nandana Jyothi et al. used an adaptive wavelet neural network to predict wind turbine power. They compared the proposed model with standard methods such as artificial neural networks and adaptive neuro-fuzzy inference systems, demonstrating that it outperformed standard methods. Zhewen Niu et al. achieved promising results using a gated recurrent unit (GRU) based on an attention mechanism to predict wind power.

[0007] Combining different forecasting methods is called a hybrid approach. The primary goal of this approach is to retain the strengths of each technique while improving overall accuracy. While combined forecasting may not always yield superior results, existing literature demonstrates that, in most cases, it can improve forecasting. Numerous approaches have been proposed based on combining different models. For example, in "Ultra-Short-Term Spatiotemporal Forecasting of Wind Power: A Deep Learning Approach," Hu T et al. used a CNN and a neural network together to forecast wind farm power. The proposed model considered not only temporal but also spatial influences, further improving forecast accuracy. For example, Wang J et al. proposed a PCA-SSA-VMD-BiLSTM network in the paper "Ultra-short-term offshore wind power forecasting based on principal component analysis-sparrow search algorithm-variational mode decomposition and bidirectional long short-term memory network" published in 2024. First, the principal component analysis algorithm (PCA) was used to screen the multivariate time series data to reduce the data dimension; secondly, the variable mode decomposition (VMD) optimized by the SSA algorithm was applied to adaptively decompose the wind power time series data into a set of different frequency components to eliminate the noise signal in the original data; on this basis, the hyperparameters of the BiLSTM model were optimized by integrating the SSA algorithm to obtain the final power prediction value, which was compared with models such as PCA-BP and PCA-LSTM. The results showed that the prediction result of the model proposed by the authors was the best.

[0008] In today's era of large models, multiple data sources, and artificial intelligence, people generally regard hybrid methods as the main innovation direction of predictive models. This design is proposed in this context. Summary of the Invention

[0009] The purpose of the present invention is to provide an ultra-short-term wind power prediction method, which solves the problems existing in the prior art, such as insufficient prediction accuracy; large computational loss due to separate prediction of each decomposed component; and the fact that the decomposed components may still be an unsteady sequence.

[0010] To achieve the above object, the present invention provides an ultra-short-term wind power prediction method, comprising the following steps: S1. Obtain the original wind power sequence; S2, using CEEMDAN signal decomposition algorithm to decompose the wind power signal into a series of sub-signals; S3, using the improved singular spectrum analysis algorithm SSA to denoise and reconstruct the IMF1 component; S4. Calculate the permutation entropy of each component separately, and classify the components into high, medium, and low frequencies according to the permutation entropy, merge the high frequencies together, and merge the medium and low frequencies together; S5. Send the data to the established TCN-GRU prediction network, and superimpose the results of the two prediction networks to obtain the final prediction result.

[0011] Preferably, S2 specifically includes the following steps: set up represents the original time series data, is obtained by the CEEMDAN method. IMF, Represents the first eigenmode functions, scalar coefficients Used to set the signal-to-noise ratio (SNR) at each stage, thereby determining the standard deviation of the added white noise; S21, the signal-to-noise ratio SNR is Different white noise sequences Add to original time series data Among them, t Indicates different time points. Indicates the added White noise, is the total number of times white noise is added; A new time series is constructed as follows: (1); in, ; S22. Obtain the first intrinsic mode function by the EMD iterative shift method and calculate the mean of this component as follows: (2); S23. Calculate the first residual: (3); S24, will ,in Further decompose until you get their first EMD mode, then calculate the second mode : (4); S25, calculate the residuals, ,…, K : (5); S26, continue to Decompose until the first EMD mode is obtained, and then calculate the modalities: (6); S27, go to step S25 and continue to calculate the next , until the residual components become monotone and EMD cannot be further decomposed, a total of K IMF, finally, the original data is decomposed and expressed as follows: (7).

[0012] Preferably, in S3, an improved singular spectrum analysis algorithm SSA is used to denoise and reconstruct the IMF1 component. The specific process is as follows: S31. Embedding: By selected window length , ,in Representing the total length of the time series, embedding the sampled time series in a vector dimension space, generating a series of lagged multidimensional data vectors, forming a trajectory matrix X, that is: (8); Where, , Represents the rows of the matrix; represents the columns of the matrix; S32, Singular Value Decomposition: Perform singular value decomposition and obtain the values ​​in descending order of size The eigenvalues ​​are , the corresponding eigenvector is ,make ,So The singular value decomposition of the trajectory matrix is ​​expressed as: (9); (10); S33, Grouping: in , Divide disjoint subsets , and sum the matrices of each group, let , then: (11); (12); Where, is the contribution rate of the group; S34, diagonal element averaging: Converted into its corresponding time series data, each set of data represents different characteristics of the original data, with matrix ,in, ,when season ,on the contrary, ,Right now: (13); In the formula, A time series satisfies All The Pearson correlation coefficient is used to search for the hyperparameters of the singular spectrum analysis algorithm.

[0013] Preferably, in S4, the permutation entropy of the signal components obtained in S3 is calculated, and the component signals are divided into three categories according to the corresponding thresholds, and the medium and low categories are merged to finally obtain two categories of high-frequency components and low-frequency components. The permutation_entropy function in the pyentrp library is used to calculate the value of the permutation entropy.

[0014] Preferably, in S5, the constructed TCN-GRU network is used for prediction, and the data is normalized by Min-Max Normalization before being input and scaled to [0, 1]; the data is changed into an input form suitable for the network, and the past 1 hour is used to predict the next 10 minutes.

[0015] Preferably, in S5, the basic structure of the temporal convolutional network TCN is: definition For the network The first residual block dilated convolution in time The state at time , the input of the residual block and Same, when and is the number of dilated convolutional layers in each residual block, the input is a single sample, It's time The output of the dilated convolution is is the output after adding the residual connection, which is calculated as follows: (14); (15); in, is the weight of the dilated convolution; and Is the bias of the dilated convolution. In each layer, the dilated convolution { }different; After completing each layer of dilated convolution operation, calculate the output of the skip connection residual block : (16); in, is the number of residual blocks, and then calculate the hidden layer state : (17); in, and They are the weights and biases of the hidden layer, and finally the prediction results are output through Softmax activation: (18); in, and are the hidden layer weights and biases respectively.

[0016] Preferably, in S5, the mathematical description of the GRU network is as shown in formula (19): (19); Where, 、 、 、 、 、 、 Represent the input vector, the state memory variable at the previous moment, the state memory variable at the current moment, the state of the update gate, the state of the reset gate, the state of the current candidate set, and the output vector at the current moment respectively; 、 、 、 They are respectively composed of update gate, reset gate, candidate set and output vector and The weight parameters of the connection matrix formed after multiplication; represents the identity matrix; represents the activation function; represents the hyperbolic tangent function; Using mean absolute error , mean absolute percentage error and root mean square error The performance of the prediction model is evaluated in the following form: (20); (twenty one); (twenty two); in, Time series length, and are the actual and predicted values.

[0017] Therefore, the present invention adopts the above-mentioned ultra-short-term wind power prediction method, and the beneficial effects are as follows: (1) The present invention improves the performance of the prediction model and reduces the impact of noise on the prediction results by decomposing non-stationary and highly fluctuating time series signals into a series of stable signals.

[0018] (2) The present invention reduces the computational cost of the model and improves the computational time by using permutation entropy to classify sub-signals.

[0019] (3) The present invention expands the receptive field of the sample through the causal convolution and dilated convolution of TCN, extracts the characteristics of wind power changes, and then extracts the temporal feature information of the sequence through GRU, and then predicts the wind power.

[0020] (4) The prediction performance of the proposed model is superior to that of traditional models and other TCN-based models, achieving high-precision and stable prediction of wind farm power, and providing decision support for wind farm operation and management.

[0021] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 This is a general framework diagram of an embodiment of an ultra-short-term wind power prediction method of the present invention; Figure 2 Schematic diagram of the temporal convolutional network (TCN) structure of an ultra-short-term wind power prediction method according to an embodiment of the present invention; Figure 3 This is a schematic diagram of a GRU network structure of an embodiment of an ultra-short-term wind power prediction method of the present invention; Figure 4 This is the original data for January of an embodiment of an ultra-short-term wind power prediction method of the present invention; Figure 5 This is a graph showing the results of CEEMDAN decomposition of January data in an embodiment of an ultra-short-term wind power forecasting method of the present invention; Figure 6 This is a flow chart of permutation entropy calculation in an embodiment of an ultra-short-term wind power prediction method of the present invention; Figure 7 This is a schematic diagram of prediction results of an embodiment of an ultra-short-term wind power prediction method of the present invention; Figure 8 It is an enlarged view of the frame portion in the forecast result schematic diagram of an embodiment of an ultra-short-term wind power forecasting method of the present invention. DETAILED DESCRIPTION

[0023] The technical solution of the present invention is further described below with reference to the accompanying drawings and embodiments.

[0024] Unless otherwise defined, technical or scientific terms used in the present invention shall have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs.

[0025] The field of wind power prediction is mainly based on deep learning networks. Since deep learning has many improvement directions, this embodiment mainly introduces the technical improvement direction most closely related to the present invention - a decomposition-based hybrid model.

[0026] Wind farm power generation varies according to meteorological conditions and is therefore characterized by high volatility and randomness. Some researchers seek to decompose highly volatile data series into stationary series for forecasting. Currently, several decomposition-based methods have been used in wind power forecasting, including empirical mode decomposition (EMD), ensemble empirical mode decomposition (EEMD), complete ensemble empirical mode decomposition (CEEMD), and noise-adaptive complete ensemble empirical mode decomposition (CEEMDAN). Experiments have shown that hybrid decomposition models are effective in improving forecast accuracy and achieve better results than individual models. For example, Wang et al. used EMD as a data decomposition method and combined it with an Elman neural network (ENN) to predict wind speed. They used EMD to decompose the original data and used a neural network to build a prediction model for each subsequence. Finally, this model was applied to real-world data forecasting. Their results demonstrated that EMD-ENN can improve wind speed forecast accuracy. Hu et al. proposed a hybrid method for wind speed forecasting based on EEMD decomposition. In their method, the original time series is decomposed into several components using EEMD. Each component is predicted using an LSTM, and the LSTM hyperparameters are fine-tuned using Bayesian optimization (BO). Ren et al. decomposed wind farm data using empirical mode decomposition (EMD) and its improved algorithm and established a prediction model using an artificial neural network (ANN) and support vector regression (SVR). Their research demonstrated that CEEMDAN achieved better prediction performance than EMD, EEMD, and CEEMD.

[0027] While combining decomposition techniques with artificial intelligence can improve prediction accuracy, further improvement is still needed. Furthermore, the highest frequencies generated by the CEEMDAN technique still exhibit high volatility and contain noise, making the components difficult to predict. Most wind power prediction methods typically rely solely on a single data decomposition method without considering the highest frequency issue. For example, Wang et al. used only EMD for data decomposition, while Hu et al. used only EEMD. This inadequate handling of the complexity of the highest frequency sequence limits prediction accuracy. The highest frequency sequence issue can be addressed by applying the singular spectrum analysis (SSA) denoising technique to the highest frequency sequence generated by CEEMDAN. Since IMF1, the highest frequency sequence, is the most irregular subsequence and exhibits high randomness, SSA denoising can reduce the randomness in the IMF1 input data and remove noise, further improving prediction accuracy. SSA is a signal processing technique that can remove noise and irrelevant information from data. SSA has demonstrated excellent performance in various applications, such as fault diagnosis, electrocardiogram (ECG) signal denoising, and seismic data analysis. Some researchers have also attempted to use SSA to denoise wind data.

[0028] In view of the above reasons, this paper proposes a new wind power prediction network. It not only performs CEEMDAN decomposition on the original signal, but also uses an improved singular spectrum analysis (SSA) algorithm to denoise and reconstruct the IMF1. The reconstructed components are divided into two large components, and the two large components are predicted separately. Finally, the component prediction results are summed to obtain the final prediction result.

[0029] like Figure 1 As shown, a method for ultra-short-term wind power prediction includes the following steps: S1. Obtain the original wind power sequence, that is, obtain historical data of wind power.

[0030] The S2 and CEEMDAN signal decomposition algorithms decompose nonlinear and nonstationary data into a sequence of intrinsic mode functions (IMFs). They are based on the empirical mode decomposition (EMD) algorithm, which suffers from mode mixing. Therefore, EEMD is employed to address this issue. While effective, EEMD has some limitations, such as high computational cost. To overcome this limitation, the CEEMDAN algorithm was proposed. This algorithm enhances the EEMD algorithm by adding adaptive noise to the residual signal after EMD decomposition, rather than directly adding noise to the original signal.

[0031] The main difference between CEEMDAN and EEMD is their approach to adding the white noise component. EEMD decomposes each realization of the signal with noise into patterns independently, and the residuals obtained from each realization are independent of each other. In contrast, CEEMDAN adds noise to the residuals obtained at the previous iteration, rather than the original noise. CEEMDAN uses the noise patterns corresponding to the iterations obtained by EMD, rather than the noise itself. This approach results in adaptive noise that is averaged at each iteration and does not introduce additional input to the original signal. CEEMDAN is a significant improvement over EEMD, with lower computational cost and better decomposition results. Therefore, this embodiment adopts the CEEMDAN signal decomposition algorithm to decompose the wind power signal into a series of sub-signals, which specifically includes the following steps: set up represents the original time series data, is obtained by the CEEMDAN method. IMF, Represents the first eigenmode functions, scalar coefficients Used to set the signal-to-noise ratio (SNR) at each stage, thereby determining the standard deviation of the added white noise.

[0032] S21, the signal-to-noise ratio SNR is Different white noise sequences Add to original time series data Among them, Indicates different time points. Indicates the added White noise, is the total number of times white noise is added; A new time series is constructed as follows: (1); in, ; S22. Obtain the first intrinsic mode function by the EMD iterative shift method, and then calculate the mean of the component as follows: (2); S23. Calculate the first residual: (3); S24. To obtain the second mode ,Will ,in Further decompose until you get their first EMD mode, then calculate the second mode : (4); S25, calculate the residuals, ,…, K : (5); S26, continue to Decompose until the first EMD mode is obtained, and then calculate the modalities: (6); S27, go to step S25, repeat the process from step S25 to step S27, and continue to calculate the next , until the residual components become monotone and EMD cannot be further decomposed, a total of K IMF, finally, the original data is decomposed and expressed as follows: (7).

[0033] The decomposition results are as follows Figure 5 shown.

[0034] S3. After applying the CEEMDAN method to the original wind power time series, a series of IMF components are obtained. However, the decomposed IMF1 component still has strong fluctuations and contains noise. Therefore, the improved singular spectrum analysis algorithm SSA is used to denoise and reconstruct the IMF1 component. The specific process is as follows: 1) Decomposition S31. Embedding: By selected window length , ,in Representing the total length of the time series, the window length requires a time series of equal length. The sampled time series is embedded in a vector dimension space, thereby generating a series of lagged multidimensional data vectors to form a trajectory matrix X, that is: (8); Where, , Represents the rows of the matrix; represents the columns of the matrix; S32, Singular Value Decomposition: Perform singular value decomposition and obtain the values ​​in descending order of size The eigenvalues ​​are , the corresponding eigenvector is ,make ,So The singular value decomposition of the trajectory matrix is ​​expressed as: (9); (10); 2) Refactoring S33, Grouping: in , Divide disjoint subsets , and sum the matrices of each group, let , then: (11); (12); Where, is the contribution rate of the group; S34, diagonal element averaging: Converted into its corresponding time series data, each set of data represents different characteristics of the original data, with matrix ,in, ,when season ,on the contrary, ,Right now: (13); In the formula, A time series satisfies All The average value of .

[0035] To optimize the denoising effect, the Pearson correlation coefficient is used to search for the hyperparameters of the singular spectrum analysis algorithm. Here, if the Pearson correlation coefficient between the sum of the IMF1 components decomposed by the SSA algorithm and the original IMF1 components is greater than or equal to 0.95, the current result is used as the new IMF1.

[0036] S4. After executing S3, the processed power sequence is obtained. The permutation entropy of each component is calculated respectively, which represents the degree of fluctuation of each sequence. The components are divided into high, medium and low frequencies according to the permutation entropy. Then, the high frequencies are merged together, and the medium and low frequencies are merged together.

[0037] Specifically, the permutation entropy of the signal components obtained by S3 is calculated, and the component signals are divided into three categories according to the corresponding thresholds, and the medium and low categories are merged to finally obtain the high-frequency component and the low-frequency component. The detailed process of permutation entropy is as follows Figure 6 shown.

[0038] In this embodiment, the permutation_entropy function in the pyentrp library is used to calculate the value of the permutation entropy. Its parameters are the default function parameters. When the permutation entropy value is less than 0.33, it is set as a low-frequency component; when the permutation entropy value is greater than 0.33 and less than 0.67, it is a medium-frequency component; when the permutation entropy value is greater than 0.67, it is set as a high-frequency component. The medium and low frequency components are then added and merged to form the final low-frequency component, and the high-frequency component is still retained as the high-frequency component.

[0039] S5. Send the data to the established TCN-GRU prediction network, and superimpose the results of the two prediction networks to obtain the final prediction result.

[0040] Use the built TCN-GRU network for prediction. Before inputting the data, normalize the data using Min-MaxNormalization and scale it to [0, 1]. Furthermore, the data needs to be changed into an input form suitable for the network, using the past 1 hour to predict the next 10 minutes.

[0041] The basic structure and principles of TCN and GRU are introduced below.

[0042] Convolutional networks have been shown to perform well in extracting high-level features from structured data. The Temporal Convolutional Network (TCN) is a neural network model that utilizes causal convolution and dilated convolution to adapt to the temporal nature of time series data and provide insights into time series modeling. It forms residual blocks by stacking multiple layers of dilated convolutions and connecting them through residual connections. Multiple residual blocks are then used for feature extraction. Residual blocks are connected through Skips activated by the ReLU activation function, with the output being a Softmax.

[0043] 1) The basic structure of the time convolution network TCN is as follows Figure 2 As shown: definition For the network The first residual block dilated convolution in time The state at time , the input of the residual block and Same, when and is the number of dilated convolutional layers in each residual block, the input is a single sample, It's time The output of the dilated convolution is is the output after adding the residual connection, which is calculated as follows: (14); (15); in, is the weight of the dilated convolution; and Is the bias of the dilated convolution. In each layer, the dilated convolution { } is different; After completing each layer of dilated convolution operation, calculate the output of the skip connection residual block : (16); in, is the number of residual blocks, and then calculate the hidden layer state : (17); in, and They are the weights and biases of the hidden layer, and finally the prediction results are output through Softmax activation: (18); in, and are the hidden layer weights and biases respectively.

[0044] 2) The GRU network is an improved model of the long short-term memory network (LSTM). It optimizes the three gate functions of the long short-term memory network, merges the forget gate and input gate into an update gate, and mixes the neuron state with the hidden state. This can effectively alleviate the "gradient vanishing" problem in recurrent neural networks, reduce the number of parameters of the long short-term memory network unit, and shorten the training time of the model.

[0045] The basic structure of the GRU network is as follows Figure 3 As shown, the mathematical description is shown in formula (19): (19); Where, 、 、 、 、 、 、 Represent the input vector, the state memory variable at the previous moment, the state memory variable at the current moment, the state of the update gate, the state of the reset gate, the state of the current candidate set, and the output vector at the current moment respectively; 、 、 、 They are respectively composed of update gate, reset gate, candidate set and output vector and The weight parameters of the connection matrix formed after multiplication; represents the identity matrix; represents the activation function; represents the hyperbolic tangent function.

[0046] The core modules of the GRU network are the update gate and the reset gate. Input variables and the state record variables at the previous moment The partition matrix is ​​input to the update gate through a nonlinear transformation with Sigmoid, which determines the extent to which the state variables of the previous moment are brought into the current state. The reset gate controls the amount of information written to the candidate set last time. The last information is stored in Multiply The result of the current moment is recorded in Multiply In the result of , finally, their sum is the output at the current moment. The prediction results of the components are added together to get the final result.

[0047] Finally, using the mean absolute error , mean absolute percentage error and root mean square error The performance of the prediction model is evaluated in the following form: (20); (twenty one); (twenty two); in, Time series length, and are the actual and predicted values. Specific embodiment: This example uses a wind farm dataset located in a certain country. It uses data from January with a 10-minute interval. The first 80% of the data is used as the training dataset, and the remaining data is used as the test dataset. The model is developed using the training dataset, and the prediction performance is evaluated using the test dataset. Figure 4 From the raw data for January shown, we can see that the fluctuations are very large and it is not easy to predict.

[0049] The settings for this case are: the pyEMD package in Python is used to implement CEEMDAN, a custom SSA function is used to implement the SSA algorithm, Keras is used to implement GRU, Keras TCN is used to implement TCN, Adam is used as the optimizer, epochs is set to 60, batch size is set to 32, SSA embedding dimension is 15, and loss is mse.

[0050] This model is compared with TCN_LSTM, TCN_RNN, TCN_BPNN, GRU, LSTM, RNN and BPNN. The results are shown in Table 2 below: Table 2 Prediction results of different models ;

[0051] Judging from the above results, GRU is more suitable for the prediction of this experiment than LSTM, and the prediction accuracy is greatly improved after adding the decomposition algorithm, which shows that the decomposition algorithm is effective.

[0052] The prediction results of this embodiment are as follows Figure 7 、 Figure 8 As shown, Figure 8 for Figure 7 The enlarged picture of the middle frame shows that the network accuracy of the present invention is better than that of other comparison networks.

[0053] Therefore, the present invention adopts the above-mentioned ultra-short-term wind power prediction method. First, CEEMDAN is used to decompose the original data signal, and then the singular spectrum analysis algorithm with Pearson coefficient optimization is used to further denoise the sequence of noisy signals. On this basis, in order to reduce the accumulation of prediction errors and reduce computational consumption, the permutation entropy algorithm is introduced to classify the sub-signals. Then, the TCN network is used to extract the power change characteristics and capture the long-term dependence of the power sequence. At the same time, the GRU is used to extract the temporal feature information of the sequence, overcoming the shortcomings of LSTM that is difficult to adjust parameters and has a slow convergence speed.

[0054] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the same. 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 solutions of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for ultra-short-term wind power prediction, characterized in that: The following steps are involved: S1. Obtaining the original wind power sequence; S2, using CEEMDAN signal decomposition algorithm to decompose the wind power signal into a series of sub-signals; S3, using the improved singular spectrum analysis algorithm SSA to denoise and reconstruct the IMF1 component; S4, respectively calculating the permutation entropy of each component, and dividing the components into high, medium and low frequencies according to the permutation entropy, merging high frequencies together, and merging medium and low frequencies together; S5. Send the data to the built TCN-GRU prediction network, and superimpose the results of the two prediction networks to obtain the final prediction result.

2. The ultra-short-term wind power prediction method according to claim 1, characterized in that: S2 Specific The following steps are involved: set up represents the original time series data, is the first IMF, Represents the first eigenmode functions, scalar coefficients Used to set the signal-to-noise ratio (SNR) at each stage, thereby determining the standard deviation of the added white noise; S21, the signal-to-noise ratio SNR is Different white noise sequences Add to original time series data Among them, Indicates different time points. Indicates the added White noise, is the total number of times white noise is added; A new time series is constructed as follows: （1）； in, ; S22. Obtain the first intrinsic mode function by the EMD iterative shift method and calculate the mean of the component as follows: （2）； S23. Calculate the first residual: （3）； S24, will ,in Further decompose until you get their first EMD mode, then calculate the second mode : （4）； S25, calculate the The residuals, ,…, K : （5）； S26, continue Decompose until the first EMD mode is obtained, and then calculate the Modes: （6）； S27, go to step S25 and continue to calculate the next , until the residual components become monotone and the EMD cannot be further decomposed, a total of K IMF, finally, the original data is decomposed and expressed as follows: （7）。 3. The ultra-short-term wind power prediction method according to claim 2, characterized in that: In S3, an improved singular spectrum analysis algorithm SSA is used to denoise and reconstruct the IMF1 component. The specific process is as follows: S31, Embedding: By selected window length , ,in Representing the total length of the time series, embedding the sampled time series in a vector dimension space, generating a series of lagged multidimensional data vectors, forming a trajectory matrix X, that is: （8）； In the formula, , Represents the rows of a matrix; represents the columns of a matrix; S32, singular value decomposition: Perform singular value decomposition and obtain the values ​​in descending order of size. The characteristic values ​​are , the corresponding eigenvector is ,make ,So The singular value decomposition of the trajectory matrix is ​​expressed as: （9）； （10）； S33, Grouping: In , Divide Disjoint subsets , and sum the matrices of each group, let , then: （11）； （12）； In the formula, is the contribution rate of the group; S34, diagonal element averaging: Converted into its corresponding time series data, each set of data represents different characteristics of the original data, with a matrix ,in, ,when season ,on the contrary, ,Right now: （13）； In the formula, The time series meets All The Pearson correlation coefficient is used to search for hyperparameters of the singular spectrum analysis algorithm.

4. The ultra-short-term wind power prediction method according to claim 3 is characterized in that: In S4, the permutation entropy of the signal components obtained in S3 is calculated, and the component signals are divided into three categories according to the corresponding thresholds, and the medium and low categories are merged to finally obtain two categories of high-frequency components and low-frequency components. The permutation entropy value is calculated using the permutation_entropy function in the pyentrp library.

5. The ultra-short-term wind power prediction method according to claim 4, characterized in that: In S5, the built TCN-GRU network is used for prediction. Before the data is input, the data is normalized by Min-Max Normalization and scaled to [0,1]. The data is changed into an input form suitable for the network, and the past 1 hour is used to predict the next 10 minutes.

6. The ultra-short-term wind power prediction method according to claim 5, characterized in that: In S5, the basic structure of the temporal convolutional network TCN is: definition For the network The first residual block The dilated convolution in time The state at that time, the input of the residual block and Same, when and is the number of dilated convolutional layers in each residual block, the input is a single sample, It's time The output of the dilated convolution is is the output after adding the residual connection, which is calculated as follows: （14）； （15）； in, is the weight of the dilated convolution; and is the bias of the dilated convolution. In each layer, the dilated convolution { }different; After completing each layer of dilated convolution operation, calculate the output of the skip connection residual block : （16）； in, is the number of residual blocks, and then calculate the hidden layer state : （17）； in, and They are the weights and biases of the hidden layers, and finally the prediction results are output through Softmax activation: （18）； in, and are the hidden layer weights and biases respectively.

7. The ultra-short-term wind power prediction method according to claim 6, characterized in that: In S5, the mathematical description of the GRU network is shown in formula (19): （19）； In the formula, , , , , , , Respectively represent the input vector, the state memory variable at the previous moment, the state memory variable at the current moment, the state of the update gate, the state of the reset gate, the state of the current candidate set, and the output vector at the current moment; , , , They are respectively composed of the update gate, reset gate, candidate set and output vector pair and The weight parameters of the connection matrix formed after multiplication; represents the identity matrix; represents the activation function; represents the hyperbolic tangent function; Using mean absolute error , mean absolute percentage error and RMS error The performance of the prediction model is evaluated in the following form: （20）； （21）； （22）； in, The length of the time series, and are the actual and predicted values.

Images