A short-term wind power forecasting method based on lightweight autoencoder network

Abstract

The present invention discloses a short-term wind power prediction method based on a lightweight autoencoder network, relates to the field of measurement technology, solves the technical problem of low wind power prediction accuracy, and its technical solution key points are to predict short-term wind power through a lightweight autoencoder network, which has the advantage of high feature learning efficiency and is more applicable to power prediction tasks of different wind turbines; it can effectively extract wind power information under complex weather conditions with high prediction accuracy; it has a small number of parameters, low required computing and storage costs, and a lightweight network structure, which can meet the "small, light, and fast" application requirements of the industrial Internet of Things for deep learning algorithms, and can be well applied to short-term wind power prediction of major wind farms.

Description

Technical Field

[0001] The present application relates to the field of measurement technology, and in particular to a short-term wind power prediction method based on a lightweight autoencoder network. Background Art

[0002] The natural randomness and volatility of wind power make it difficult and technically deficient to accurately predict its power generation. The lack of wind power prediction theory indirectly causes the phenomenon of wind abandonment. Therefore, accurate wind power prediction can not only improve the utilization efficiency of wind farms, but also serve the power dispatching department to ensure the safe, stable and economic operation of the power grid after a high proportion of wind power energy is connected. At present, domestic and foreign scholars have done a lot of research on wind power prediction. However, in the face of complex weather, wind power output fluctuates greatly and the accuracy of wind power prediction is low, which poses a severe challenge to the safe operation of the power grid. Summary of the invention

[0003] The present application provides a short-term wind power prediction method based on a lightweight autoencoder network, the technical purpose of which is to improve the accuracy of short-term wind power prediction under complex weather conditions and to better meet the needs of engineering applications.

[0004] The above technical objectives of this application are achieved through the following technical solutions:

[0005] A short-term wind power prediction method based on a lightweight autoencoder network, comprising:

[0006] S1: splicing the time series data in numerical weather forecast to form multi-dimensional wind power time series data;

[0007] S2: inputting the multi-dimensional wind power time series data into a first lightweight autoencoder network for training, extracting mutually coupled nonlinear features in the multi-dimensional wind power time series data, screening and strengthening the nonlinear features through an attention mechanism, and finally outputting wind power prediction data;

[0008] S3: performing loss calculation on the wind power data corresponding to the multi-dimensional wind power time series data and the wind power prediction data;

[0009] S4: Repeat S1 to S3 until the number of iterations reaches a preset maximum value or the loss is lower than a preset threshold, then the training of the first lightweight autoencoder network is completed, and a second lightweight autoencoder network is obtained;

[0010] S5: inputting numerical weather forecasts in the future into the second lightweight autoencoder network, and the second lightweight autoencoder network outputs short-term wind power prediction data;

[0011] The first lightweight autoencoder network includes an encoder and a decoder, the encoder includes three cascaded depth-separable convolutional layers and two fully connected layers, the decoder includes one fully connected layer and three depth-separable convolutional layers, and the convolutional structures of the encoder and the decoder are symmetrical;

[0012] The convolution operation of the depth-wise separable convolution layer is a one-dimensional depth-wise separable convolution, which includes depth-wise convolution and point-wise convolution. The parameter amount and calculation amount of the one-dimensional depth-wise separable convolution are L and L, respectively. k ×M+1×M×N k and L k ×M×L F +M×N k ×L F Among them, L F Indicates the input mapping length, L k represents the convolution kernel size, M represents the number of input mapping channels, and N k Indicates the number of convolution kernels.

[0013] The beneficial effects of the present application are as follows: the short-term wind power prediction method based on a lightweight autoencoder network described in the present application has the advantage of high feature learning efficiency, and is more applicable to power prediction tasks of different wind turbines; it can effectively extract wind power information under complex weather conditions, and has a high prediction accuracy; it has a small number of parameters, and requires low computing and storage costs, and a lightweight network structure, and can meet the "small, light, and fast" application requirements of the industrial Internet of Things for deep learning algorithms, and can be well applied to short-term wind power prediction of major wind farms. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 A flowchart of the method described in this application;

[0015] Figure 2 A schematic diagram of the structure of a lightweight autoencoding network according to a preferred embodiment of the present application;

[0016] Figure 3 This is a schematic diagram of the results of short-term wind power prediction data of the preferred embodiment of the present application. DETAILED DESCRIPTION

[0017] The technical solution of the present application will be described in detail below with reference to the accompanying drawings.

[0018] Figure 1 A flowchart of the method described in this application, such as Figure 1 As shown, the short-term wind power prediction method based on lightweight autoencoder network includes:

[0019] S1: The time series data in the numerical weather forecast (NWP) are spliced ​​to form multi-dimensional wind power time series data.

[0020] Specifically, the multidimensional wind power time series data includes time series data such as wind speed, humidity, temperature, wind direction and air pressure. The size of the multidimensional wind power time series data is N×L, and the resolution is 15 minutes; where L represents the time step, and N represents the number of data types in the multidimensional wind power time series data.

[0021] S2: Input the multi-dimensional wind power time series data into the first lightweight autoencoder network for training, extract the mutually coupled nonlinear features in the multi-dimensional wind power time series data, and screen and strengthen the non-linear features through the attention mechanism, and finally output the wind power prediction data.

[0022] Specifically, the lightweight autoencoder network includes an encoder and a decoder, and its structural diagram is as follows: Figure 2 As shown. The encoder includes three cascaded depthwise separable convolutional layers and two fully connected layers, and the fully connected layer is connected to the last depthwise separable convolutional layer. The decoder includes one cascaded fully connected layer and three depthwise separable convolutional layers, and the fully connected layer is connected to the first depthwise separable convolutional layer. The convolution structures of the encoder and decoder are symmetrical, and the sizes of the convolution kernels of the depthwise separable convolutional layers are, for example, 15×1, 7×1, and 3×1; the number of convolutional layers and the kernel size are not limited to this, and can be adjusted specifically according to the actual prediction task.

[0023] The convolution operation used by the lightweight autoencoder network is a one-dimensional depth-separable convolution, which decomposes the standard one-dimensional convolution into depth convolution and point convolution. On the basis of ensuring effective feature extraction, it greatly reduces the number of model parameters and calculations, and reduces the calculation and storage burden of model application. Assume L F Indicates the input mapping length, L k represents the convolution kernel size, M represents the number of input mapping channels, and N k represents the number of convolution kernels, then the number of parameters and the amount of calculation for standard convolution are L k ×M×N k and L k ×M×N k ×L F , while the number of parameters and the amount of computation for one-dimensional depth-separable convolution are L k ×M+1×M×N k and L k ×M×L F +M×N k ×L F .

[0024] In the process of extracting features through convolution operation, the channel attention mechanism is used to adaptively weight the extracted channel features, strengthen the features that are beneficial to prediction, weaken irrelevant and redundant features, and improve the effectiveness of feature learning.

[0025] S3: Perform loss calculation on the wind power data corresponding to the multi-dimensional wind power time series data and the wind power prediction data.

[0026] The wind power prediction data is an output one-dimensional vector with a size of 1×L, where L represents the time step.

[0027] The loss function used in loss calculation is the mean square loss MSE, which is expressed as:

[0028]

[0029] Among them, y t represents the wind power data corresponding to the multi-dimensional wind power time series data, represents the wind power prediction data.

[0030] S4: Repeat S1 to S3 until the number of iterations reaches a preset maximum value or the loss is lower than a preset threshold, then the training of the first lightweight autoencoder network is completed to obtain a second lightweight autoencoder network.

[0031] Generally speaking, when the mean square loss MSE is lower than a preset threshold, the training of the first lightweight autoencoder network ends, and at this time, the wind power data corresponding to the multi-dimensional wind power time series data and the wind power prediction data form an identity mapping.

[0032] S5: Inputting numerical weather forecasts for the future time into the second lightweight autoencoder network, and the second lightweight autoencoder network outputs short-term wind power prediction data.

[0033] As a specific implementation example, the measured data of a wind farm in Jiangsu from January 1 to December 31, 2020 is used as a data set. The acquired data includes time series such as wind speed, humidity, temperature, wind direction and wind power, with a time resolution of 15 minutes. The training set, validation set and test set suitable for the actual scenario of this method are found by cross-validation.

[0034] The short-term wind power prediction results are as follows: Figure 3 As shown. Figure 3 It can be seen that the prediction results are basically the same as the actual wind power result trends, and the predicted values ​​are highly consistent with the true values. Therefore, the model developed in this application has a relatively accurate short-term wind power prediction capability.

[0035] In summary, this application designs a short-term wind power prediction model of a lightweight autoencoder network based on a convolutional neural network. Compared with traditional deep learning methods, this application has better prediction performance in the face of complex weather, is more in line with the actual scenarios of wind farms, and meets the requirements of the industrial Internet of Things for deep learning algorithms.

[0036] It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as those generally understood by those skilled in the art to which this application belongs. It should also be understood that terms such as those defined in common dictionaries should be understood to have meanings consistent with the meanings in the context of the prior art, and will not be interpreted with idealized or overly formal meanings unless defined as herein.

[0037] The meaning of "and / or" described in this application means that the situations where each exists alone or both exist at the same time are included.

[0038] The term “connection” as used in this application may mean a direct connection between components or an indirect connection between components via other components.

[0039] Based on the above ideal embodiments of this application, the relevant staff can make various changes and modifications without departing from the technical concept of this application through the above description. The technical scope of this application is not limited to the content in the specification, and its technical scope must be determined according to the scope of the claims.

Claims

1. A short-term wind power prediction method based on a lightweight autoencoder network, characterized in that: include: S1: splicing the time series data in the numerical weather forecast to form multi-dimensional wind power time series data; S2: inputting the multi-dimensional wind power time series data into a first lightweight autoencoder network for training, extracting mutually coupled nonlinear features in the multi-dimensional wind power time series data, screening and strengthening the nonlinear features through an attention mechanism, and finally outputting wind power prediction data; S3: performing loss calculation on the wind power data corresponding to the multi-dimensional wind power time series data and the wind power prediction data; S4: Repeat S1 to S3 until the number of iterations reaches a preset maximum value or the loss is lower than a preset threshold, then the training of the first lightweight autoencoder network is completed, and a second lightweight autoencoder network is obtained; S5: inputting numerical weather forecasts in the future into the second lightweight autoencoder network, and the second lightweight autoencoder network outputs short-term wind power prediction data; The first lightweight autoencoder network includes an encoder and a decoder, the encoder includes three cascaded depth-separable convolutional layers and two fully connected layers, the decoder includes one fully connected layer and three depth-separable convolutional layers, and the convolutional structures of the encoder and the decoder are symmetrical; The convolution operation of the depth-wise separable convolution layer is a one-dimensional depth-wise separable convolution, which includes depth-wise convolution and point-wise convolution. The parameter amount and calculation amount of the one-dimensional depth-wise separable convolution are L and L, respectively. k ×M+1×M×N k and L k ×M×L F +M×N k ×L F Among them, L F Indicates the input mapping length, L k represents the convolution kernel size, M represents the number of input mapping channels, and N k Indicates the number of convolution kernels.

2. The prediction method according to claim 1, characterized in that: In step S1, the multidimensional wind power time series data includes time series data of wind speed, humidity, temperature, wind direction and air pressure, and the size of the multidimensional wind power time series data is N×L, and the resolution is 15 minutes; wherein L represents the time step, and N represents the number of data types in the multidimensional wind power time series data.

3. The prediction method according to claim 2, characterized in that: The wind power prediction data is a one-dimensional vector with a size of 1×L. In step S3, the loss function used in the loss calculation is the mean square loss MSE, which is expressed as: Among them, y t represents the wind power data corresponding to the multi-dimensional wind power time series data, represents the wind power prediction data, and L represents the time step.

4. The prediction method according to claim 3, characterized in that: In step S4, steps S1 to S3 are repeated until the number of iterations reaches a preset maximum value or the mean square loss MSE is lower than a preset threshold, and the training of the first lightweight autoencoder network is terminated.

Images