Search

CN-122020007-A - Large-scale photovoltaic cluster short-term photovoltaic power prediction method

CN122020007ACN 122020007 ACN122020007 ACN 122020007ACN-122020007-A

Abstract

The invention discloses a large-scale photovoltaic cluster short-term photovoltaic power prediction method, which comprises the steps of calculating the importance of various numerical weather forecast features in a historical data set through a random forest algorithm, carrying out clustering division on photovoltaic power stations in a photovoltaic cluster by adopting a time dynamic graph neural network, constructing a topological graph network according to the geographical position coordinates of each photovoltaic power station and the numerical weather forecast features of the first position of importance ranking, obtaining an adjacent matrix, constructing a space-time attention mechanism network of the photovoltaic cluster, fusing the space-time attention mechanism network with the topological graph network, obtaining a photovoltaic power prediction model, training the photovoltaic power prediction model through the numerical weather forecast features of the first three positions of importance ranking and corresponding actual measured photovoltaic power, and obtaining the photovoltaic cluster predicted power through the numerical weather forecast features of future time periods. The method fully considers the spatial correlation between the adjacent photovoltaic power stations, has higher accuracy and stability, and has higher prediction accuracy compared with the prior art.

Inventors

  • ZHOU CHAOHONG
  • GU SHENHUI
  • ZHANG FAN
  • ZHAO ZEXI
  • LI XIAOFENG

Assignees

  • 中国石油天然气股份有限公司

Dates

Publication Date
20260512
Application Date
20241101

Claims (9)

  1. 1. The short-term photovoltaic power prediction method for the large-scale photovoltaic cluster is characterized by comprising the following steps of: Step 1, collecting a historical data set of a photovoltaic cluster, calculating the importance of various numerical weather forecast features in the historical data set through a random forest algorithm, and selecting the most important numerical weather forecast features as primary features of cluster division to classify the photovoltaic cluster; Step 2, clustering and dividing the photovoltaic power stations in the photovoltaic clusters into a plurality of sub-clusters by adopting a time dynamic graph neural network; Step 3, constructing a topological graph network according to the geographical position coordinates and the numerical weather forecast characteristics of the first position of the importance ranking of the photovoltaic power stations in each sub-cluster to obtain a corresponding adjacency matrix; step 4, constructing a space-time attention mechanism network of the photovoltaic cluster, and fusing the space-time attention mechanism network with the topological graph network in the step 3 to obtain a graph convolution network model fused with the space-time attention mechanism, namely a photovoltaic power prediction model; step 5, training a photovoltaic power prediction model through the numerical weather forecast features of the first three digits of the importance ranking and the corresponding actual measured photovoltaic power in the historical data set, and further obtaining the corresponding photovoltaic cluster prediction power of each sub-cluster through the numerical weather forecast features of the future time period; and 6, summing the predicted power of all the photovoltaic sub-clusters to obtain a final predicted result.
  2. 2. The method for predicting the short-term photovoltaic power of a large-scale photovoltaic cluster according to claim 1, wherein in the step 2, the photovoltaic power stations in the photovoltaic cluster are clustered and divided by adopting a time dynamic graph neural network, the input is the geographic position coordinates and the most important numerical weather forecast characteristics of each photovoltaic power station, and the output is the class number.
  3. 3. The method according to claim 2, wherein the time-dynamic graph neural network comprises two vectors Z and ψ of length d for each time period t, all elements being randomly initialized learnable parameters, and the initial adjacency matrix a of the time sequence is obtained by calculating the product of Z T and ψ: A=Z T ·Ψ (1) idx,idy=argtopk(A[:,:]),idx≠idy (2) A[-idx,-idy]=0 (3) wherein, Z= [ Z 1 ,z 2 ,…,z d ], Representing a random initialization of the embeddable node, wherein Z represents the geographical location coordinate information of the photovoltaic power plant, (x, y) represents a coordinate point, Z 1 represents the geographical location coordinate information of the 1 st photovoltaic power plant, Z 2 represents the geographical location coordinate information of the 2 nd photovoltaic power plant, Z d represents the geographical location coordinate information of the d-th photovoltaic power plant, ψ represents a numerical weather forecast feature, The 1 st numerical weather forecast feature in the time period t, The 2 nd numerical weather forecast feature in the time period t, The d-th numerical weather forecast feature in the time period t, argtopk (·) function can return the index of the first k maximum values of the adjacency matrix A, sparse adjacency matrix is obtained through formulas (2) - (3), only the elements with the first k maximum weights are reserved for the adjacency matrix of each time period, and other values are set to be zero; The dynamic graph neural network completely separates different time series data in the same dimension after the dynamic graph transformation, and the same vertex in different time periods generally aggregates information from different vertex sets, and the dynamic graph isomorphic network is defined as: Wherein, the Representing the output of the graph isomorphic network of node v at time t in the l-layer, Is a simple implementation of the dynamic map transformation, which is applicable in the case of t >2, the edge weights w ij are normalized to w ij , epsilon is a learning parameter, t represents the point in time, Representing a plurality of Splicing is performed, and furthermore, the expression of the adjacency matrix H l is as follows: wherein H l denotes an adjacency matrix, Epsilon l represents an error factor, I represents an identity matrix, MLP represents a multi-layer perceptron, H l-1 represents an output tensor of a first graph isomorphic network layer, H l-1 [t 1 :t T-1 represents data alignment during a second time period, and A passes through Normalizing, wherein D is a degree matrix of A; pooling by using convolution parameters, adopting a two-dimensional convolution neural network CNN layer, concentrating nodes into clusters according to a given parameter merging proportion, taking the nodes as channels for feature extraction in the CNN, distributing the kernel size of time convolution to the corresponding convolution kernel size, wherein X l represents an input node embedding tensor of a first layer, and calculating CNN of an output embedding tensor X l+1 to be expressed as: Where, is the effective 2D cross-correlation operator, N represents the number of nodes in or outside the pooling module, note that for layer I, N l is equal to the inner node, N l+1 is equal to the outer node, weight represents the weight ratio operation, bias represents the bias value; After the output tensor X l+1 is generated, the corresponding adjacency matrix is calculated, and the shape of the learnable weight W l is [ N l+1 ,N l , 1, kernel_size ], a vector V l ∈R l×k composed of the learnable parameters of the first layer is generated, so that a learnable allocation matrix can be calculated With rows corresponding to N l+1 nodes or clusters and columns corresponding to N l clusters, the given matrix M and adjacency matrix a (l) of input data at that layer, an output adjacency matrix a ( l+1 ) is generated using the following equation: Formulas (7) - (8) provide all steps of time map pooling TGP, in formula (7), X l+1 represents output cluster embedding after aggregate input embedding, in formula (8), a (l+1) represents connection relation and corresponding weight of new clusters, and furthermore, each element Representing the connection weight between i and j.
  4. 4. A method for predicting the short-term photovoltaic power of a large-scale photovoltaic cluster according to claim 3, wherein in step 4, a spatio-temporal attention mechanism network of the photovoltaic cluster is constructed, specifically as follows: Under the Seq2Seq framework, a spatio-temporal attention neural network is constructed, in an encoding-decoding structure, attention mechanisms are introduced from two dimensions of space and time, STAN provides a spatio-temporal context vector aligned directly with output variables, and a separate spatial attention parallel to the temporal attention is designed in the decoder layer to simultaneously pay attention to the most relevant time steps and the most important variables, wherein the input of the spatial and temporal attention is spatial embedding and temporal embedding, respectively, wherein the generation of the temporal and spatial embedding is mutually independent; Setting a multivariate time sequence X= [ X 1 ,X 2 ,…,X N ] T ∈R N×k ] containing N characteristics, wherein X is a two-dimensional matrix, k represents the length of an input sequence, X 1 represents the type of the 1 st numerical weather forecast characteristic, X 2 represents the type of the 2 nd numerical weather forecast characteristic, and X N represents the type of the N-th numerical weather forecast characteristic; For the ith numerical weather forecast feature The spatial embedding is calculated using a feed-forward neural network, The value of the kth time point of the ith numerical weather forecast feature is represented, and for input data X= [ X 1 ,X 2 ,…,X N ] T ], the embedded calculation results of all input variables are D= [ D 1 ,D 2 ,…,D N ] T ∈R N×k ,D 1 、D 2 、…、D N are respectively the calculation results after X 1 、X 2 、…、X N is input into a feedforward neural network; The feedforward neural network is adopted as an alignment model, the spatial attention weight is calculated, and for the output time step j, the ith spatial attention weight As shown in formulas (9), (10): Where h' j-1 ;d i ]∈R p+m ,h′ j-1 ∈R p represents the upper hidden layer state of the LSTM decoder, d i ∈R m represents the spatial embedding of the ith feature, b e represents the offset value of the spatial attention weight, Representing intermediate variables, we representing spatial attention weight values, further computing a spatial context vector g j from the spatial attention weights; for output time step j, computing the time attention weight corresponding to h t using the alignment model Is a function of h t and an implied layer state h' j-1 at the decoder, as shown in equation (11): Wherein h t represents the state of the current hidden layer, b a represents the bias value of the temporal attention weight, wa represents the temporal attention weight value, [ h' j-1 ;h t ]∈R p+m ], further computing the attention weight from the spatiotemporal context vector s j The spatial and temporal attentiveness depends on the hidden layer states of the LSTM decoder, LSTM G and LSTM S , which are defined as h ' G and c ' G , respectively, when calculating the spatial context vector g j , for the output time step j, the input to the spatial attentiveness is g j calculated from the decoder hidden layer state h ' G,j-1 , r G,j and the last time step Splicing to obtain Input to LSTM G to update hidden layer state h' G,j as shown in equation (12): r G,j represents the output result of the method, Representing r G,j and the last time step The value obtained by splicing is W G represents the weight value of the output result, and b G represents the offset value of the output result; For the temporal context vector s j of LSTM, the hidden layer and neuron states of LSTM are h ' S ∈R p and c ' S ∈R p , respectively, at time step j, h ' S,j-1 represents the temporal attention for computing the temporal context vector, stitching And r S,j and input to LSTM S update hidden layer state h' S,j as shown in equation (13): The last step, before predicting the photovoltaic power, is to connect the hidden state updates of the two LSTM to [ h ' G,j ;h′ S,j ],h′ G,j represents the hidden layer state of the decoder, h' S,j represents the temporal attention for calculating the temporal context vector.
  5. 5. The method of claim 4, wherein the graph structure in the graph convolution network model is obtained by a Laplacian matrix and eigenvalues, and the spectral convolution result on the graph is obtained by convolving the graph signal x ε R N×1 with the graph convolution kernel function Θ, Wherein, the Is a graph convolution operator, U is a fourier basis orthogonal matrix of laplace eigenvalue decomposition, Λ is a eigenvalue diagonal matrix, Λ=diag (λ 0 ,λ 1 ,…,λ N-1 ),λ 0 represents the 0 th eigenvalue, λ 1 represents the 1 st eigenvalue, λ N-1 represents the N-1 th eigenvalue, the fourier coefficients of the graph are obtained by fourier transformation x=u T x, L represents the laplace matrix, l=d-a, a is an adjacency matrix, and the diagonal matrix D e R N×N is a degree matrix, and the normalized form of L is shown in equation (15): Wherein I N is an identity matrix of dimension n×n, R N×N is a real matrix of dimension n×n; Wherein, the Is a chebyshev polynomial of order k, when k=0, When k=1, the number of the groups, Is an N x N real matrix, lambda max is the laplace maximum eigenvalue, theta k is the K-order chebyshev coefficient, K is the convolution kernel size, and the maximum radius of the convolution from the center node is determined.
  6. 6. The method of claim 5, wherein in order to determine the dynamic correlation of local space, the kth chebyshev inequality is used to predict the short-term photovoltaic power of the large-scale photovoltaic cluster And W are combined to obtain E is hadamard, W represents a weight value:
  7. 7. The method of claim 1, wherein the weather forecast values include irradiance, ground air temperature, air temperature 2 meters near the ground, high altitude cloud cover, hollow cloud cover, low altitude cloud cover, humidity 2 meters near the ground, wind speed, wind direction, and ground air pressure.
  8. 8. The method for predicting the short-term photovoltaic power of a large-scale photovoltaic cluster according to claim 1, wherein the most important numerical weather forecast feature is selected as the primary feature of cluster division to classify the photovoltaic cluster, and the most important numerical weather forecast feature is the numerical weather forecast feature with the greatest influence on the photovoltaic power.
  9. 9. The method for predicting the short-term photovoltaic power of a large-scale photovoltaic cluster according to claim 1, wherein training the photovoltaic power prediction model is preceded by a normalization process: Wherein x represents a feature vector before normalization, x' represents a feature vector after normalization, and x min and x max represent a maximum value and a minimum value of x respectively; In the prediction stage, the prediction result is restored to the original power interval according to an inverse normalization formula, wherein the inverse normalization formula is as follows: x=x'(x max -x min )+x min (21)。

Description

Large-scale photovoltaic cluster short-term photovoltaic power prediction method Technical Field The invention belongs to the technical field of photovoltaic power prediction, and relates to a short-term photovoltaic power prediction method of a large-scale photovoltaic cluster. Background In recent years, new energy power generation starts to replace thermal power generation, and in the future, it is becoming a main power generation. The photovoltaic power generation is used as a clean energy source, has the characteristics of reproducibility, greenness and the like, but also has the characteristics of high randomness, high fluctuation capacity, high adjustment difficulty and the like, and if the electric energy emitted by a large-scale photovoltaic cluster is directly integrated into a power grid, the fluctuation of a power system can be caused, a certain loss is caused to a user and a power plant, and the normal operation of the power system can be disturbed when serious, so that accidents such as large-scale power failure and the like are caused. Therefore, the photovoltaic cluster can predict the change and fluctuation of the photovoltaic power in advance by predicting the generated power, so that a dispatcher can conveniently determine a dispatching plan, and the running stability of a power system is further improved. In order to improve the photovoltaic cluster power prediction precision, the conventional method is to perform photovoltaic power prediction by means of a feature construction and artificial intelligent model, for example, a distributed photovoltaic cluster power prediction method, device, equipment and medium disclosed in China patent publication No. CN118487275A, a distributed photovoltaic cluster power prediction method based on depth similarity clustering disclosed in China patent publication No. CN118313496A, a photovoltaic cluster division-based photovoltaic cluster prediction precision method disclosed in China patent publication No. CN118157231A, a construction and prediction method and construction and prediction system disclosed in China patent publication No. CN117217357A, and a distributed photovoltaic cluster power prediction method and device disclosed in China patent publication No. CN114447916A, which are disclosed in 2024, 07, and a photovoltaic cluster division-based photovoltaic cluster prediction method disclosed in China patent publication No. CN118157231A, are disclosed in 2024, 12, and a construction and prediction system disclosed in China patent publication No. CN117217357A, and a distributed photovoltaic cluster power prediction method and device disclosed in 2022, but the photovoltaic power prediction methods disclosed in the above patents are not related to the space efficiency is lower than that of the related photovoltaic power station. Currently, as photovoltaic power stations are continuously expanded, the scheduling range of the photovoltaic power stations is continuously increased by schedulers, short-term photovoltaic power prediction methods aiming at large-scale photovoltaic clusters need to be researched, and the current photovoltaic cluster prediction methods generally comprise an integral method, an accumulation method and a cluster dividing method. The integral method and the accumulation method cannot utilize the space correlation information between the photovoltaic power stations, so that the prediction precision is low, and the Euclidean distance-based clustering method has larger dividing error for the photovoltaic cluster scene under the non-Euclidean distance, so that the prediction effect is influenced. In addition, the data of the photovoltaic power stations are data sets provided by a single photovoltaic power station, the data sets have certain divergence, the power output of adjacent photovoltaic power stations has certain spatial correlation, the existing photovoltaic power prediction method basically does not consider the spatial correlation between the adjacent photovoltaic power stations, and the characteristic construction method and the artificial intelligent model are complex, so that the prediction efficiency and the prediction accuracy are low. Disclosure of Invention The invention aims to provide a large-scale photovoltaic cluster short-term photovoltaic power prediction method, which solves the problems that the existing photovoltaic power prediction method basically does not consider the spatial correlation between adjacent photovoltaic power stations and the prediction result accuracy is lower. The technical scheme adopted by the invention is that the method for predicting the short-term photovoltaic power of the large-scale photovoltaic cluster comprises the following steps: Step 1, collecting a historical data set of a photovoltaic cluster, calculating the importance of various numerical weather forecast features in the historical data set through a random forest algorithm, and selecting the most important numerical wea