Ultra-short-term wind power forecasting method based on Gram matrix and convolutional neural network

Abstract

The present invention relates to a wind power ultra-short-term power prediction method based on a Gram matrix and a convolutional neural network. It is applicable to the field of wind power generation prediction. The technical solution adopted by the present invention is: a wind power ultra-short-term power prediction method based on a Gram matrix and a convolutional neural network, characterized in that: historical wind speed, historical wind direction and historical power are obtained, including wind speed, wind direction and power at n moments before the moment to be predicted; VMD decomposition is performed on the historical power data to obtain characteristic signals of m different center frequencies; the historical wind speed, historical wind direction and characteristic signals are normalized; data fusion is performed on the normalized historical wind speed, historical wind direction and characteristic signals, and a Gram matrix is ​​constructed based on the data obtained by the data fusion; the Gram matrix is ​​input into a trained wind power prediction model to obtain a power prediction result; the wind power prediction model is constructed based on a convolutional neural network.

Description

Technical Field

[0001] The invention relates to a wind power ultra-short-term power prediction method based on Gram matrix and convolutional neural network, which is applicable to the field of wind power generation prediction. Background Art

[0002] As a new energy source, wind energy is widely used and vigorously developed by various countries due to its unlimited reserves, safety, and cleanliness. The non-stationary nature of wind speed leads to the randomness and volatility of wind power generation, which brings challenges to the safe, stable and economic operation of large-scale wind power grid connection. The accuracy of wind power generation prediction needs to be improved.

[0003] Short-term and ultra-short-term forecasts can provide reliable power transient information for power dispatching and wind power grid connection safety. Therefore, wind power forecasting research mainly focuses on short-term and ultra-short-term wind power forecasts.

[0004] Wind power forecasting can be roughly divided into the following four types: (1) physical method; (2) statistical method; (3) deep learning method; and (4) hybrid method.

[0005] Due to its simple modeling and high accuracy, deep learning methods are widely used. At present, the main neural network structures commonly used in deep learning methods mainly include convolutional neural networks (CNN) and recurrent neural networks (RNN).

[0006] Convolutional neural networks are mainly used to process data signals with more than two dimensions (such as images), while recurrent neural networks are mainly used to process one-dimensional data signals, such as time series signals. Convolutional neural networks have achieved great success in the field of images, mainly due to the powerful feature extraction capabilities of their network structures. Convolutional neural networks can obtain multi-layer features of images, and can achieve the fusion of deep features and shallow features through residual structures, jump connections, etc., greatly improving the prediction ability of the network. Summary of the invention

[0007] The technical problem to be solved by the present invention is to provide a wind power ultra-short-term power prediction method based on Gram matrix and convolutional neural network.

[0008] The technical solution adopted by the present invention is: a method for ultra-short-term wind power prediction based on Gram matrix and convolutional neural network, characterized in that:

[0009] Obtain historical wind speed, historical wind direction and historical power, including the wind speed, wind direction and power at n moments before the moment to be predicted;

[0010] Perform VMD decomposition on the historical power data to obtain m characteristic signals with different center frequencies;

[0011] Normalize the historical wind speed, historical wind direction and characteristic signals;

[0012] Perform data fusion on the normalized historical wind speed, historical wind direction and characteristic signals, and construct a Gram matrix based on the data obtained by data fusion;

[0013] The Gram matrix is ​​input into a trained wind power prediction model to obtain a power prediction result; the wind power prediction model is constructed based on a convolutional neural network.

[0014] The data fusion of the normalized historical wind speed, historical wind direction and characteristic signal includes:

[0015] The historical wind speed and historical wind direction are normalized to obtain two 1*n vectors, and the characteristic signal is normalized to obtain m 1*n vectors;

[0016] The historical wind speed, historical wind direction and characteristic signal data are fused to obtain (m+2) 1*n vectors.

[0017] The step of constructing a Gram matrix based on the data obtained by data fusion includes:

[0018] After expanding the (m+2) 1*n vectors obtained by data fusion, (m+2)*n 1*1 vectors are obtained;

[0019] Construct the Gram matrix of (m+2)*n 1*1 vectors to obtain a ((m+2)*n)*((m+2)*n) matrix.

[0020] The training of the wind power prediction model includes:

[0021] The MSE loss function is used to calculate the loss between the power prediction result and the actual wind power, and the obtained network loss is back-propagated to update the weights of the wind power prediction model until the network converges to obtain the trained wind power prediction model.

[0022] The convolutional neural network adopts vgg19 or resnet.

[0023] A wind power ultra-short-term power prediction device based on Gram matrix and convolutional neural network, characterized by:

[0024] A historical data acquisition module is used to obtain historical wind speed, historical wind direction and historical power, including the wind speed, wind direction and power at n moments before the moment to be predicted;

[0025] A signal decomposition module is used to perform VMD decomposition on historical power data to obtain characteristic signals with m different center frequencies;

[0026] A normalization processing module is used to perform normalization processing on historical wind speed, historical wind direction and characteristic signals;

[0027] A Gram matrix construction module is used to perform data fusion on normalized historical wind speed, historical wind direction and characteristic signals, and construct a Gram matrix based on the data obtained by data fusion;

[0028] The power prediction module is used to input the Gram matrix into the trained wind power prediction model to obtain a power prediction result; the wind power prediction model is constructed based on a convolutional neural network.

[0029] A storage medium stores a computer program executable by a processor, wherein the computer program implements the steps of the wind power ultra-short-term power prediction method based on the Gram matrix and convolutional neural network when executed.

[0030] A wind power ultra-short-term power forecasting device comprises a memory and a processor, wherein the memory stores a computer program that can be executed by the processor, and is characterized in that when the computer program is executed, the steps of the wind power ultra-short-term power forecasting method based on the Gram matrix and convolutional neural network are implemented.

[0031] The beneficial effects of the present invention are as follows: the present invention obtains characteristic signals of multiple different center frequencies by VMD decomposition of historical power data, constructs a Gram matrix in combination with historical wind speed and historical wind direction, converts a one-dimensional wind power signal into a two-dimensional signal based on the Gram matrix, and uses a convolutional neural network combined with the two-dimensional signal to predict ultra-short-term wind power, thereby improving the accuracy of wind power prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 A flowchart of an embodiment. DETAILED DESCRIPTION

[0033] In the wind power prediction task, historical data containing multiple time points are obtained, including one-dimensional time series signals such as wind speed, wind direction, power and temperature. When using a two-dimensional convolutional neural network for wind power prediction, the data needs to be processed in a higher dimension. This example uses the method of constructing a Gram matrix to process the one-dimensional wind power data in a higher dimension.

[0034] This embodiment is a method for ultra-short-term wind power prediction based on Gram matrix and convolutional neural network, which specifically includes the following steps:

[0035] S1. Obtain historical wind speed, historical wind direction and historical power, including the wind speed, wind direction and power at n moments before the moment to be predicted.

[0036] Wind power generation is the conversion of wind energy into electrical energy. Wind speed and wind direction are closely related to wind power generation. The historical power of wind power generation has a time series correlation with the wind power at future moments. Therefore, this embodiment selects wind speed, wind direction and historical power as input features of the wind power prediction model.

[0037] Selection of time series length: In this example, the Pearson coefficient is used to measure the correlation between all variables in the original wind power data and wind power. The historical wind speed, wind direction and wind power at n moments with a Pearson coefficient greater than 0.8 (strong correlation) are selected. After calculation, n = 8.

[0038] S2. Perform VMD decomposition on the historical power data to obtain m (m is 8) characteristic signals with different center frequencies.

[0039] Variational mode decomposition (VMD) is a new adaptive signal processing method that has a good processing effect on non-stationary and nonlinear signals. VMD decomposition of historical power data can obtain multi-dimensional features of the data and enrich the extracted historical wind power information. After VMD decomposition of the historical power signal, 8 characteristic signals with different center frequencies are obtained.

[0040] S3. Normalize the historical wind speed, historical wind direction and characteristic signal. In order to eliminate the dimensional influence between the data and to facilitate the rapid convergence of the subsequent network training stage, the present invention performs min-max normalization on the historical wind speed, historical wind direction and decomposed wind power data, as shown in formula (1):

[0041]

[0042] After normalization, the characteristic signals of historical wind speed data, historical wind direction data, and historical wind power data are all within the range of [0, 1], which is conducive to the feature learning of the subsequent network.

[0043] S4. Perform data fusion on the normalized historical wind speed, historical wind direction and characteristic signal, and construct a Gram matrix based on the data obtained by data fusion.

[0044] After normalization, the historical wind speed data and historical wind direction data are obtained into two 1*8 vectors. After VMD decomposition and normalization, the historical power data are obtained into eight 1*8 vectors. After data fusion, ten 1*8 vectors are obtained.

[0045] After expanding the 10 1*8 vectors obtained by data fusion, 80 1*1 vectors are obtained. The Gram matrix of the 80 1*1 vectors is constructed to obtain an 80*80 matrix.

[0046] The matrix formed by calculating the inner product of any k vectors in the n-dimensional Euclidean space is called the Gram matrix of these k vectors. Through the Gram matrix, the original one-dimensional wind power data can be converted into two-dimensional data, which is convenient for the subsequent use of convolutional neural networks (CNN) for feature extraction and future information prediction.

[0047] S5. Input the Gram matrix into the trained wind power prediction model based on the convolutional neural network to obtain the power prediction result.

[0048] The two-dimensional Gram matrix obtained in step S4 is used as the input of the wind power prediction model. The data of the input model is similar to the two-dimensional image in the field of computer vision, so the structural network used can be the classic convolutional neural network in the field of image, such as vgg19-net, resnet.

[0049] In this embodiment, the MSE loss function is used to calculate the wind power prediction loss when training the electric power prediction model. The calculation is shown in formula (2):

[0050]

[0051] Among them, P true Represents the actual wind power, P pred Represents the wind power predicted by the network.

[0052] The calculated network loss is back-propagated, and the parameters of the wind power prediction model are updated until the prediction model converges; the final wind power prediction model is obtained for subsequent wind power prediction.

[0053] This embodiment combines historical wind speed, historical wind direction and historical power, constructs a Gram matrix for dimensionality increase and uses a convolutional neural network for prediction, which effectively improves the wind power prediction effect when the wind speed changes suddenly.

[0054] This embodiment also provides a wind power ultra-short-term power prediction device based on the Gram matrix and convolutional neural network, including a historical data acquisition module, a signal decomposition module, a normalization processing module, a Gram matrix construction module and a power prediction module.

[0055] In this example, the historical data acquisition module is used to obtain historical wind speed, historical wind direction and historical power, including the wind speed, wind direction and power at 8 moments before the moment to be predicted; the signal decomposition module is used to perform VMD decomposition on the historical power data to obtain characteristic signals with 8 different center frequencies; the normalization processing module is used to perform normalization processing on the historical wind speed, historical wind direction and characteristic signals; the Gram matrix construction module is used to perform data fusion on the normalized historical wind speed, historical wind direction and characteristic signals, and construct a Gram matrix based on the data obtained by data fusion; the power prediction module is used to input the Gram matrix into a trained wind power prediction model based on a convolutional neural network to obtain a power prediction result.

[0056] This embodiment also provides a storage medium on which a computer program that can be executed by a processor is stored. When the computer program is executed, the steps of the wind power ultra-short-term power prediction method based on the Gram matrix and convolutional neural network in this embodiment are implemented.

[0057] This embodiment also provides a wind power ultra-short-term power forecasting device, which has a memory and a processor, and the memory stores a computer program that can be executed by the processor. When the computer program is executed, the steps of the wind power ultra-short-term power forecasting method based on the Gram matrix and convolutional neural network in this embodiment are implemented.

Claims

1. A wind power ultra-short-term power prediction method based on Gram matrix and convolutional neural network, characterized by: Obtain historical wind speed, historical wind direction and historical power, including the wind speed, wind direction and power at n moments before the moment to be predicted; Perform VMD decomposition on the historical power data to obtain m characteristic signals with different center frequencies; Normalize the historical wind speed, historical wind direction and characteristic signals; Perform data fusion on the normalized historical wind speed, historical wind direction and characteristic signals, and construct a Gram matrix based on the data obtained by data fusion; Inputting the Gram matrix into a trained wind power prediction model to obtain a power prediction result; the wind power prediction model is constructed based on a convolutional neural network; The data fusion of the normalized historical wind speed, historical wind direction and characteristic signal includes: The historical wind speed and historical wind direction are normalized to obtain two 1*n vectors, and the characteristic signal is normalized to obtain m 1*n vectors; The historical wind speed, historical wind direction and characteristic signal data are integrated to obtain (m+2) 1*n vectors; The step of constructing a Gram matrix based on the data obtained by data fusion includes: After expanding the (m+2) 1*n vectors obtained by data fusion, (m+2)*n 1*1 vectors are obtained; Construct the Gram matrix of (m+2)*n 1*1 vectors to obtain a ((m+2)*n)*((m+2)*n) matrix.

2. The method for ultra-short-term wind power prediction based on Gram matrix and convolutional neural network according to claim 1 is characterized in that: The training of the wind power prediction model includes: The MSE loss function is used to calculate the loss between the power prediction result and the actual wind power, and the obtained network loss is back-propagated to update the weights of the wind power prediction model until the network converges to obtain the trained wind power prediction model.

3. The method for ultra-short-term wind power prediction based on Gram matrix and convolutional neural network according to claim 1 or 2, characterized in that: The convolutional neural network adopts vgg19 or resnet.

4. A wind power ultra-short-term power prediction device based on Gram matrix and convolutional neural network, characterized in that: A historical data acquisition module is used to obtain historical wind speed, historical wind direction and historical power, including the wind speed, wind direction and power at n moments before the moment to be predicted; A signal decomposition module is used to perform VMD decomposition on historical power data to obtain characteristic signals with m different center frequencies; A normalization processing module is used to perform normalization processing on historical wind speed, historical wind direction and characteristic signals; A Gram matrix construction module is used to perform data fusion on normalized historical wind speed, historical wind direction and characteristic signals, and construct a Gram matrix based on the data obtained by data fusion; A power prediction module, used for inputting the Gram matrix into a trained wind power prediction model to obtain a power prediction result; the wind power prediction model is constructed based on a convolutional neural network; The data fusion of the normalized historical wind speed, historical wind direction and characteristic signal includes: The historical wind speed and historical wind direction are normalized to obtain two 1*n vectors, and the characteristic signal is normalized to obtain m 1*n vectors; The historical wind speed, historical wind direction and characteristic signal data are integrated to obtain (m+2) 1*n vectors; The step of constructing a Gram matrix based on the data obtained by data fusion includes: After expanding the (m+2) 1*n vectors obtained by data fusion, (m+2)*n 1*1 vectors are obtained; Construct the Gram matrix of (m+2)*n 1*1 vectors to obtain a ((m+2)*n)*((m+2)*n) matrix.

5. A storage medium having stored thereon a computer program executable by a processor, characterized in that: When the computer program is executed, the steps of the wind power ultra-short-term power prediction method based on Gram matrix and convolutional neural network described in any one of claims 1 to 3 are implemented.

6. A wind power ultra-short-term power forecasting device, comprising a memory and a processor, wherein the memory stores a computer program executable by the processor, characterized in that: When the computer program is executed, the steps of the wind power ultra-short-term power prediction method based on Gram matrix and convolutional neural network described in any one of claims 1 to 3 are implemented.

Images