Photovoltaic power generation medium and long term online prediction method based on historical fluctuation information entropy extraction

Abstract

The invention discloses a photovoltaic power generation medium and long term online prediction method based on historical fluctuation information entropy extraction. The method comprises the following steps: performing frequency domain decomposition on a photovoltaic output historical sequence through an ensemble empirical mode method to obtain a mode component and a residual component; reconstructing the modal components into a vector group by using a phase-space reconstruction theory, and extracting historical fluctuation information entropy among the modal components; and comparing absolute values of historical fluctuation information entropy differences between adjacent modal components, adding the modal components of which the absolute values of the differences are smaller than a set threshold, and combining the modal components into one component. And splicing the combined modal component and residual component to obtain a reconstructed sequence. And finally, inputting the reconstructed sequence into the graph attention neural network, and predicting photovoltaic output. According to the method, long-term fluctuation trend characteristics in photovoltaic power generation are effectively captured through historical fluctuation information entropy, and the information amount is reduced through selective reconstruction of modal components. And the long-term prediction precision in photovoltaic power generation is improved.

Description

Technical Field

[0001] The present invention belongs to the field of photovoltaic power generation, and relates to feature extraction and prediction of historical data, and is specifically a medium- and long-term online prediction method based on historical fluctuation information entropy extraction. Background Art

[0002] As a clean and renewable energy technology, solar photovoltaic power generation has received widespread attention under the dual pressure of the current global energy crisis and environmental pollution. However, the output power of photovoltaic power generation is greatly affected by natural factors such as sunshine conditions, climate, and seasons, and has obvious volatility and uncertainty. The access of large-scale photovoltaic power generation may affect the stability of the power grid. The medium- and long-term forecast of photovoltaic power generation can help power station planning and resource assessment, ensure the balance of power system electricity, and improve the level of photovoltaic consumption at the macro level. However, the current medium- and long-term forecast methods for photovoltaic power generation do not fully utilize the characteristics of historical data in the frequency domain on the one hand, and do not fully consider the nonlinear dynamic characteristics of photovoltaic data on the other hand. Therefore, there is still a lot of room for improvement in the accuracy of prediction. Summary of the invention

[0003] In view of the shortcomings of the prior art, the present invention proposes a medium- and long-term online prediction method for photovoltaic power generation based on the extraction of historical fluctuation information entropy. The method decomposes the historical sequence of photovoltaic output in the frequency domain, extracts the historical information entropy of the frequency domain components, reconstructs the frequency domain components based on the historical information entropy, and uses a graph attention network to capture the correlation between the reconstructed sequences, so as to achieve medium- and long-term prediction of photovoltaic output.

[0004] A method for medium- and long-term online prediction of photovoltaic power generation based on historical fluctuation information entropy extraction specifically includes the following steps:

[0005] Step 1: Decompose the photovoltaic output history series in the frequency domain to obtain the modal component and residual component.

[0006] Step 2: Extract the historical fluctuation information entropy between modal components.

[0007] Step 3: Compare the absolute values ​​of the historical fluctuation information entropy differences between adjacent modal components, add the modal components whose absolute values ​​of the differences are less than the set threshold, and merge them into one component. The merged modal components and residual components are concatenated to obtain the reconstructed sequence.

[0008] Step 4: Input the reconstructed sequence into the graph attention neural network to predict the photovoltaic output.

[0009] The present invention has the following beneficial effects:

[0010] By extracting the frequency characteristics of photovoltaic historical data and the fluctuation information entropy it contains, the medium- and long-term fluctuation trend characteristics of photovoltaic power generation are effectively captured, and the frequency domain component data is selectively reconstructed to reduce the amount of information. The fluctuation characteristics are mapped to graph elements, making full use of the advantages of the graph attention mechanism in prediction. This method is superior to traditional time series prediction methods in improving the accuracy of medium- and long-term predictions of photovoltaic power generation. DETAILED DESCRIPTION

[0011] Step 1: Frequency domain decomposition

[0012] The ensemble empirical mode is used to perform frequency domain decomposition of the photovoltaic output history sequence. First, a Gaussian white noise n(t) is added to the historical photovoltaic power generation sequence x(t):

[0013] x′(t)=x(t)+n(t)

[0014] Where t represents time. Perform empirical mode decomposition on the noise sequence x′(t) to obtain the residual component r(t) and I frequency domain modal components c i (t):

[0015]

[0016] The above method is used to decompose the noise sequence x′(t) multiple times, and the modal components c obtained by multiple decompositions are calculated. i (t) and the mean of the residual component r(t), we get and

[0017] Step 2: Historical information entropy extraction

[0018] The modal components and residual components obtained by frequency domain decomposition of the photovoltaic power generation sequence contain local features of different time scales in the historical photovoltaic output sequence. However, indiscriminate prediction and superposition of too many components is not only cumbersome to calculate, but may also lead to error accumulation. This application uses a modal component reconstruction method based on historical information entropy, takes the information entropy between modal components as an indicator of complexity, classifies each modal component according to complexity, and selectively reconstructs it, reducing time and computational costs.

[0019] Using the phase space reconstruction theory, the embedding dimension d and the delay time τ are calculated, and the mean of each modal component obtained in step 1 is Reconstruct and obtain a vector group of size (N-(d-1)τ)*d, where the t-th row vector for:

[0020]

[0021] Where t=1,2,…,N-(d-1) τ, N is the total length of the photovoltaic power sequence.

[0022] Define the reconstruction vector The average value of AA t and relative amplitude RA t for:

[0023]

[0024]

[0025] Reconstruction vector If the elements in are sorted from largest to smallest, there are d! different permutations. Indicates the lth order of arrangement π l The probability of occurrence is l = 1, 2, ... d! . Let The initial value is 0, starting from t = 1 and ending at t = N-d + 1, traversing each reconstructed vector in the vector group in turn The order of arrangement, when the order is π l When, yes Make an update:

[0026]

[0027] Among them, α∈(0,1) is the weight used to adjust the mean value of the component amplitude and the deviation between the component amplitudes, and is taken as 0.5 in this embodiment. After the traversal is completed, the arrangement order π is calculated l The final probability p of appearing in the vector group * (π l ):

[0028]

[0029] According to the probability p * (π l ) Calculate the modal components of the photovoltaic power sequence The historical fluctuation information entropy H AAPE (i)

[0030]

[0031] The historical fluctuation information entropy reflects the complexity of each modal component. The frequency component with a larger historical fluctuation information entropy value represents the short-term component characteristics with stronger randomness in the photovoltaic power series, and the frequency component with a smaller historical fluctuation information entropy value represents the long-term change trend with stronger regularity in the photovoltaic power series.

[0032] Step 3: Sequence reconstruction

[0033] From the perspective of feature extraction, separately predicting and modeling components with similar complexity has limited effect on improving prediction accuracy and will greatly increase the workload of model training. In order to improve the efficiency of prediction model training, the historical fluctuation information entropy H of adjacent modal components is used AAPE (i) H AAPE The difference of (i+1) |H AAPE (i)-H AAPE (i+1)| is used as the judgment condition to reconstruct the frequency component. When the difference in historical fluctuation information entropy between adjacent modal components is greater than the set threshold s, it is considered that the complexity difference between the two is large, and the original modal component is retained. As two subsequences. When the difference in historical fluctuation information entropy between adjacent components is less than this value, it can be considered that the complexity between the components is close and the information contained in the components is close. The two components are merged and added as a subsequence. The residual component generally has a large difference with the adjacent component and is directly retained as a subsequence. The reconstructed sequence of the photovoltaic historical output sequence is obtained.

[0034] In this embodiment, the threshold s is set to 0.6.

[0035] Step 4: Photovoltaic output prediction

[0036] s4.1. Construct a graph attention neural network including graph data input layer, graph attention layer, fully connected layer and output layer. Take the reconstructed sequence obtained in step 3 as the node feature sequence, and each subsequence in the node feature sequence as a node in the graph data. Calculate any two subsequences S i , S j The mutual information I(S i ,S j ), and obtain the node adjacency matrix A as the edge information in the graph data:

[0037]

[0038] Where n is the number of subsequences in the node feature sequence; p(S i ,S j ) is a subsequence S i , S j The joint probability of i ), p(S j ) are subsequences S i , S j The marginal probability density function of .

[0039] s4.2. The attention mechanism of the graph attention layer is used to calculate the hidden representation of each node in the graph data, that is, the hidden features of the photovoltaic historical output sequence. Each layer outputs a new set of node features for a set of node features. In order for the model to have sufficient expressive power to transform the input features into higher-level features, at least one learnable linear transformation is required for each node. Therefore, a shared linear transformation parameterized by a weight matrix is ​​used in the graph attention layer. Calculate the characteristics of node j Features for node i The attention coefficient e ij :

[0040]

[0041] The number of heads for attention coefficient calculation can be taken as 3.

[0042] In order to make the coefficients easier to compare between different nodes, the attention coefficient e ij Normalize. Treat the attention mechanism of each layer of the graph attention network as a single-layer feedforward neural network, parameterized by a weight vector, and nonlinearized using a function F with a negative slope of 0.2:

[0043]

[0044] Where α ij is the attention mechanism coefficient of node i to node j, N i Represents the number of nodes in the graph data, that is, the number of subsequences in the reconstructed sequence of the PV historical output sequence in step 3.

[0045] s4.3. The output of the graph attention layer is integrated through the fully connected layer, and the output subsequences are superimposed to obtain the photovoltaic output at the time to be predicted.

Claims

1. A medium- and long-term online prediction method for photovoltaic power generation based on historical fluctuation information entropy extraction, characterized by: The specific steps include: Step 1: Decompose the photovoltaic output history series in the frequency domain to obtain the mean of the modal component and the mean of the residual component. , ; Step 2: Extract the historical fluctuation information entropy between modal components; Step 3: compare the absolute values ​​of the historical fluctuation information entropy differences between adjacent modal components, add the modal components whose absolute values ​​of the differences are less than the set threshold, and merge them into one component; concatenate the merged modal components and the residual components to obtain a reconstructed sequence; Step 4: Input the reconstructed sequence into the graph attention neural network to predict the photovoltaic output.

2. The method for mid- to long-term online prediction of photovoltaic power generation based on historical fluctuation information entropy extraction as claimed in claim 1, characterized in that: Using ensemble empirical modes to analyze historical photovoltaic power generation series Perform frequency domain decomposition to obtain the residual component and I frequency domain modal components : in, represents Gaussian white noise, Indicates time; Represents the noise sequence; the modal components obtained by multiple decompositions , residual component Take the average and get the mean of the modal component and the mean of the residual component , .

3. The method for medium- and long-term online prediction of photovoltaic power generation based on historical fluctuation information entropy extraction as claimed in claim 1, characterized in that: Using phase space reconstruction theory, calculate the embedding dimension d and delay time , for the modal component mean Reconstruct and get the size of ( )*d vector group, where the tth row vector for: in, , N is the historical photovoltaic power generation sequence Total length; Calculate the reconstruction vector The average and relative amplitude : Reconstruction vector The elements in are sorted in descending order. Indicates the first order of arrangement The probability of occurrence is l=1,2,…d!; let The initial value is 0, starting from t=1 to t= End, traverse each reconstructed vector in the vector group in turn The order of arrangement is as follows: When, yes Make an update: in, , used to adjust the weights of the component amplitude mean and the deviation between component amplitudes; When the traversal is completed, the order of arrangement is calculated The final probability of appearing in the vector group : According to probability Calculation of the modal components of the photovoltaic power series The historical fluctuation information entropy : 。 4. The method for mid- to long-term online prediction of photovoltaic power generation based on historical fluctuation information entropy extraction as claimed in claim 3, characterized in that: Set the weight α=0.

5.

5. The method for mid- to long-term online prediction of photovoltaic power generation based on historical fluctuation information entropy extraction as claimed in claim 1, characterized in that: Set when adjacent modal components , The historical fluctuation information entropy , The difference When it is less than 0.6, the two adjacent modal components are combined and added.

6. The method for mid- to long-term online prediction of photovoltaic power generation based on historical fluctuation information entropy extraction as claimed in claim 1, characterized in that: The specific steps of inputting the reconstructed sequence into the graph attention neural network to predict the photovoltaic output are: s4.

1. Construct a graph attention neural network including a graph data input layer, a graph attention layer, a fully connected layer, and an output layer. Take the reconstructed sequence obtained in step 3 as the node feature sequence, and each subsequence in the node feature sequence as a node in the graph data. Calculate the , Mutual information between , get the node adjacency matrix A as the edge information in the graph data; s4.

2. Use the attention mechanism of the graph attention layer to calculate the hidden representation of each node in the graph data, using a shared linear transformation parameterized by a weight matrix , compute nodes Features For Node Features The attention coefficient : Then the attention coefficient Normalize: In the formula Is a node For Node The attention mechanism coefficient, N i represents the number of nodes in the graph data, that is, the number of subsequences in the reconstructed sequence of the photovoltaic historical output sequence in step 3; s4.

3. The output of the graph attention layer is integrated through the fully connected layer, and the output subsequences are superimposed to obtain the photovoltaic output at the time to be predicted.

7. The method for mid- to long-term online prediction of photovoltaic power generation based on historical fluctuation information entropy extraction as claimed in claim 6, characterized in that: The neighbor matrix A is: Where n is the number of subsequences in the node feature sequence; For subsequence , The joint probability of , Subsequences , The marginal probability density function of .

8. A computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to execute the method according to any one of claims 1 to 7.

Images