CN114004152B - Multi-wind-field wind speed space-time prediction method based on graph convolution and recurrent neural network - Google Patents

Multi-wind-field wind speed space-time prediction method based on graph convolution and recurrent neural network Download PDF

Info

Publication number
CN114004152B
CN114004152B CN202111280781.0A CN202111280781A CN114004152B CN 114004152 B CN114004152 B CN 114004152B CN 202111280781 A CN202111280781 A CN 202111280781A CN 114004152 B CN114004152 B CN 114004152B
Authority
CN
China
Prior art keywords
wind
wind speed
field
graph
neural network
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202111280781.0A
Other languages
Chinese (zh)
Other versions
CN114004152A (en
Inventor
臧海祥
张烽春
刘冲冲
张越
刘璟璇
卫志农
孙国强
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Hohai University HHU
Original Assignee
Hohai University HHU
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Hohai University HHU filed Critical Hohai University HHU
Priority to CN202111280781.0A priority Critical patent/CN114004152B/en
Publication of CN114004152A publication Critical patent/CN114004152A/en
Application granted granted Critical
Publication of CN114004152B publication Critical patent/CN114004152B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
    • G06Q50/06—Energy or water supply
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2113/00—Details relating to the application field
    • G06F2113/06—Wind turbines or wind farms

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Business, Economics & Management (AREA)
  • Health & Medical Sciences (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • General Health & Medical Sciences (AREA)
  • Mathematical Physics (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • Data Mining & Analysis (AREA)
  • Computing Systems (AREA)
  • Computational Linguistics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biophysics (AREA)
  • Economics (AREA)
  • Strategic Management (AREA)
  • Human Resources & Organizations (AREA)
  • General Business, Economics & Management (AREA)
  • Marketing (AREA)
  • Tourism & Hospitality (AREA)
  • Water Supply & Treatment (AREA)
  • Primary Health Care (AREA)
  • Public Health (AREA)
  • Development Economics (AREA)
  • Game Theory and Decision Science (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Operations Research (AREA)
  • Quality & Reliability (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Computer Hardware Design (AREA)
  • Geometry (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses a multi-wind-field wind speed space-time prediction method based on a graph convolution and a cyclic neural network, and provides a graph convolution long-short-term memory neural network for processing data, aiming at the problem that a convolution neural network commonly used in the conventional wind speed space-time prediction method is difficult to effectively analyze wind speed data of a multi-wind field showing non-grid distribution in reality. Firstly, carrying out graph modeling on wind speed data of a plurality of wind fields based on a Pearson correlation coefficient, and constructing a wind speed graph signal sequence; then, the convolution of the graph is used for replacing the multiplication operation in the long-term and short-term memory neural network to construct the convolution of the graph and the long-term and short-term memory neural network; and finally, constructing a multi-wind-field wind speed space-time prediction model based on the graph convolution long-term short-term memory neural network and the transfer learning principle. The space-time prediction model provided by the invention has better point prediction and probability prediction performance, verifies that the fusion of the historical wind speed information of the adjacent wind field can be helpful for improving the accuracy of wind speed point prediction and probability prediction, and provides a new idea for short-term wind speed prediction of a multi-wind field.

Description

Multi-wind-field wind speed space-time prediction method based on graph convolution and recurrent neural network
Technical Field
The invention relates to a wind speed prediction method, in particular to a multi-wind-field wind speed space-time prediction method based on a graph volume and a recurrent neural network.
Background
Wind energy is a clean and pollution-free renewable energy source, and has the characteristics of abundant resources, clean power generation process and the like, so that the wind energy is widely popularized and applied. However, due to the fluctuation and intermittency of wind speed, wind power has strong fluctuation and randomness. At present, wind speed spatio-temporal prediction methods generally include spatio-temporal prediction models based on correlation analysis and spatio-temporal prediction models based on convolutional neural networks. The time-space prediction model based on the correlation analysis judges the correlation degree between the historical wind speed of the adjacent wind field and the historical wind speed of the wind field to be predicted by using correlation analysis methods such as Pearson coefficients, mutual information and the like, extracts important features in the historical wind speed data of the adjacent wind field by setting a threshold value, and sends the important features and the historical wind speed data of the wind field to be predicted into the prediction model together for learning training. The method is simple to implement, but the space-time structure of the original multi-wind-field wind speed data is destroyed, so that the space-time characteristics in the original wind speed data cannot be extracted more fully. The time-space prediction model based on the convolutional neural network is used for analyzing and learning wind speed data of multiple wind fields by using the convolutional neural network, the wind speed data of the multiple wind fields are constructed into a two-dimensional map with channels through grid division, grid points on the two-dimensional map represent the relative positions of the wind fields, the longitudinal channel of each grid point is historical wind speed data of each wind field, the convolutional neural network simultaneously extracts the time-space characteristics of the wind speeds of the multiple wind fields from two layers of space and time through a convolutional kernel, and the extracted time-space characteristics can be directly sent to other models (such as a long-short term memory neural network) to be trained and learned, so that the time-space prediction of the wind speeds is realized. The time-space prediction model based on the convolutional neural network is applied to wind speed prediction, but the time-space prediction model has a problem that wind fields in actual engineering are not necessarily distributed in a grid shape, and the relative positions of the wind fields in a two-dimensional map are difficult to determine.
Disclosure of Invention
The purpose of the invention is as follows: the invention provides a multi-wind-field wind speed space-time prediction method based on a graph convolution and a cyclic neural network, aiming at the problem that the conventional convolutional neural network can only process data in a Euclidean space and is difficult to process data in a non-Euclidean space. The multi-wind-field prediction model based on the graph convolution recurrent neural network has strong feature extraction capability, can effectively process non-European data, and further improves the prediction accuracy of the wind speed of the multi-wind-field.
The technical scheme is as follows: the invention provides a multi-wind-field wind speed space-time prediction method based on a graph volume and a cyclic neural network, which comprises the following steps of:
step (1): obtaining position information and wind speed data of a plurality of wind fields in a certain area, carrying out graph modeling by using an undirected weighted graph based on the position information and the wind speed data, and carrying out spectrogram convolution calculation on the established undirected weighted graph to construct a wind speed graph signal sequence;
the anemogram signal sequence modeling method comprises the following steps:
for M wind farms in a region, its data is modeled using one undirected weighted graph G = (V, E, a), where V = { V, a) 1 ，v 2 ，…，v M And E is a set of edges, and A is an adjacency matrix reflecting the connection relation and the weight among all the vertexes. For each A ij e.A, the definition of which is shown below:
Figure BDA0003327862550000021
wherein d (v) i ，v j ) Representing a vertex v i And v j The distance between the wind field of the ith wind field and the wind field of the jth wind field,
Figure BDA0003327862550000022
representing the standard deviation of all pairs of vertex distances for removing the effect of dimension, PCC (v) i ，v j ) Representing a Pearson correlation coefficient calculated based on the ith wind field and the jth wind field historical wind speed data, wherein gamma is a threshold value of the Pearson coefficient when PCC (v i ，v j ) When gamma is greater than or equal to gamma, then v i And v j With edges e ij E is connected with E and has a weight of
Figure BDA0003327862550000023
When PCC (v) i ，v j ) When less than gamma, then v i And v j No edges are connected. The purpose of setting the gamma threshold is to connect wind field vertexes with strong correlation in order to obtain a sparse wind field image, so that a part of priori knowledge is provided for training and learning of a subsequent model, and the training time of the model is reduced.
Will be located at vertex v i The wind speed of the wind field at time t is recorded as
Figure BDA0003327862550000024
The anemogram signal at time t can be recorded as
Figure BDA0003327862550000025
The space-time prediction to be realized by the invention is based on historical anemogram signal sequence x at a plurality of moments t-K+1 ，x t-K+2 ，…，x t To predict the anemogram signal x t+1 Wherein K is a natural number.
Further, in the step (1), a spectrogram convolution calculation is used for the established undirected weighted graph, wherein a calculation formula of the spectrogram convolution layer is as follows:
Figure BDA0003327862550000026
wherein Q represents the number of convolution kernels, Y ：，q Representing the output of the qth convolution kernel, X ：，p Representing the p-th channel map signal input, f is a convolution kernel g Representing a graph convolution operation, W q，p，： The parameter representing the qth convolution kernel corresponding to the pth channel map signal input, a (-) represents the activation function.
Step (2): the multiplication operation in the long-short term memory neural network is replaced by graph convolution calculation, a graph convolution long-short term memory neural network is constructed, and a multi-wind-field wind speed space-time point prediction model based on the graph convolution long-short term memory neural network is established on the basis; the multiplication operation in the long-short term memory neural unit is replaced by spectrogram convolution operation, and the method is improved to be a Graph convolution long-short term memory unit (GCLSTM cell), and the calculation formula is as follows:
I t ＝σ(W xi * g X t +W hi * g H t-1 +B i )
F t ＝σ(W xf * g X t +W hf * g H t-1 +B f )
O t ＝σ(W xo * g X t +W ho * g H t-1 +B o )
C t ＝F t ⊙C t-1 +I t ⊙tanh(W xc * g X t +W hc * g H t-1 +B c )
H t ＝O t ⊙tanh(C t )
wherein, I t 、F t 、O t Respectively representing the input gate, the forgetting gate and the output gate matrix at the time t, C t And H t Matrix representing the cell state and the hidden layer state at time t, X t Input matrix, W, representing long and short term memory cells at time t xi 、W xf 、W xo 、W xc Weight tensors W of input gate, forgetting gate, output gate and cell state in input layer neuron of long and short term memory network hi 、W hf 、W ho 、W hc Weight tensors of input gate, forget gate, output gate and cell state in neuron of hidden layer of long-short term memory network, respectively, B i 、B f 、B o And B c Bias matrices representing input gate, forget gate, output gate, and cell state, respectively g The graph convolution operation is shown.
And (3): training parameters of a multi-wind-field wind speed space-time point prediction model based on historical wind speed data of the multi-wind-field to obtain the multi-wind-field wind speed space-time point prediction model; the multi-wind-field wind speed space-time point prediction model based on the graph convolution long-term memory neural network comprises the following training error loss formula:
Figure BDA0003327862550000031
wherein M represents the number of samples in each batch of training set, M is the number of wind fields,
Figure BDA0003327862550000032
representing the actual value of the output of the s-th wind farm in the ith sample of each batch of training sets,
Figure BDA0003327862550000033
and (4) representing the predicted value of the output of the s-th wind field in the ith sample in each batch of training sets.
And (4): based on a transfer learning principle, transferring the point prediction model parameters trained in the step (3), constructing a multi-wind-field wind speed space-time probability prediction model, and further training to obtain the multi-wind-field wind speed space-time probability prediction model; the multi-wind-field wind speed space-time probability prediction model based on the graph convolution long-term memory neural network comprises a quantile loss formula of model training, wherein the quantile loss formula comprises the following components:
Figure BDA0003327862550000034
wherein q represents the number of quantiles,
Figure BDA0003327862550000035
denotes the quantile loss at the jth quantile, f s (W(τ j )，b(τ j )，X i ) Tau of the s wind field representing the output of the ith sample model in each batch of training sets j Wind speed prediction in quantile, W (τ) j ) And b (τ) j ) Is and quantile τ j The relevant model parameters, i.e. W and b parameters, X, of the fully-connected layer in the probabilistic predictive model i Representing the input of the ith sample in each batch of training sets. Here, based on the migration learning principle, the parameters of the GCLSTM layer in the point prediction model are migrated to the GCLSTM layer in the probabilistic prediction model, in order to reduce the time required for training the probabilistic prediction model.
And (5): and (3) constructing wind field information needing wind speed prediction into a wind speed diagram signal sequence through the step (1), sequentially inputting the wind speed diagram signal sequence to be predicted into the models obtained by training in the step (3) and the step (4), and obtaining the point prediction and probability prediction results of the wind speed of the multi-wind field.
Has the beneficial effects that: compared with the prior art, the technical scheme of the invention is based on the position information and the wind speed data of multiple wind fields in a certain area, the data is preprocessed by adopting graph modeling and spectrogram convolution calculation, different pooling combinations and corresponding training data pools are constructed by pooling, and an improved graph convolution wind speed prediction model is input, so that the space-time structure and the space-time characteristics of the wind speed data of the multiple wind fields can be deeply extracted, guidance is provided for the scheduling and demand response implementation of a wind power system, and the safe, stable and economic operation of the wind power generation system is ensured.
Drawings
FIG. 1 is a schematic flow chart of a multi-wind-field wind speed space-time prediction method based on a graph convolution and a recurrent neural network;
FIG. 2 is a schematic diagram of the modeling results of a wind speed data graph according to the present invention;
FIG. 3 is a schematic diagram of a convolutional long short term memory cell structure of the present invention;
FIG. 4 is a diagram of a GCLSTM-based multi-wind-field wind speed space-time prediction model structure;
FIG. 5 is a comparison graph of annual point prediction indexes of different models of each wind farm in the embodiment of the present invention, including corresponding graphs of the RMSE index, the MAE index, and the NSE index;
fig. 6 is a graph of the probability prediction effect of different models in each wind field during (a) spring, (b) summer, (c) autumn, and (d) winter periods, respectively, according to an embodiment of the present invention.
Detailed Description
The present invention is further illustrated by the following examples, which are intended to be purely exemplary and are not intended to limit the scope of the invention, as various equivalent modifications of the invention will occur to those skilled in the art upon reading the present disclosure and fall within the scope of the appended claims.
The invention provides a multi-wind-field wind speed space-time prediction method based on a graph convolution and a recurrent neural network, as shown in figure 1, the method comprises the following steps:
step (1): obtaining position information and wind speed data of a plurality of wind fields in a certain area, carrying out graph modeling on the data by using undirected weighted graphs, carrying out spectrogram convolution calculation on the established undirected weighted graphs, realizing preliminary analysis and processing of wind speed graph signals, and constructing a wind speed graph signal sequence.
Step (2): substituting the multiplication operation in the long-short term memory neural network for Graph convolution calculation, constructing a Graph convolution long-short term memory neural network, and building a multi-wind-field wind speed time-space-point prediction model based on the Graph convolution long-short term memory neural network (GCLSTM);
and (3): training parameters of a multi-wind-field wind speed space-time point prediction model based on historical wind speed data of the multi-wind-field to obtain the multi-wind-field wind speed space-time point prediction model;
and (4): migrating the point prediction model parameters trained in the step (3) based on a migration learning principle, constructing a multi-wind-field wind speed space-time probability prediction model, and further training to obtain the multi-wind-field wind speed probability prediction model;
and (5): and (3) constructing wind field information needing wind speed prediction into a wind speed diagram signal sequence through the step (1), sequentially inputting the wind speed diagram signal sequence to be predicted into the models obtained by training in the step (3) and the step (4), and obtaining the point prediction and probability prediction results of the wind speed of the multi-wind field.
The following describes in detail a specific implementation process of multi-wind field wind speed prediction using the method of the present invention with reference to specific embodiments. Taking 76 wind speed data of wind fields near Michigan lake in America as an example, the wind speed data from 1 month and 1 day in 2004 to 12 months and 31 days in 2006 are obtained according to the sampling frequency of 1 hour, and the geographical position information of the 76 wind fields is shown in Table 1. In order to test the generalization capability and the prediction precision of the model, the data of the last three days of each month is used as a test set, the rest data is used as a training set, wind speed data in 76 wind fields are predicted 6 hours in advance, and the used prediction error evaluation indexes are Nash efficiency coefficients ANSE, average absolute errors MAE and root mean square errors RMSE. The specific implementation steps are as follows:
in the step (1), the anemogram signal sequence modeling method and the graph convolution calculating method are as follows:
1.1 for M wind farms in a region, its data is modeled using an undirected weighted graph G = (V, E, a), where V = { V = 1 ，v 2 ，…，v M And E is a set of edges, and A is an adjacency matrix reflecting the connection relation and the weight among all the vertexes. For each A ij e.A, the definition of which is shown below:
Figure BDA0003327862550000051
wherein d (v) i ，v j ) Representing a vertex v i And v j The distance between the ith wind field and the jth wind field, i.e. the geographical distance between the ith wind field and the jth wind field,
Figure BDA0003327862550000052
representing the standard deviation of all pairs of vertex distances for removing the effect of dimension, PCC (v) i ，v j ) Representing a Pearson correlation coefficient calculated based on the ith wind field and the jth wind field historical wind speed data, wherein gamma is a threshold value of the Pearson coefficient when PCC (v i ，v j ) When gamma is greater than or equal to gamma, then v i And v j With edges e ij E is connected to E and has a weight of
Figure BDA0003327862550000061
When PCC (v) i ，v j ) When it is less than gamma, then v i And v j No edges are connected. The purpose of setting the gamma threshold is to connect wind field vertexes with strong correlation in order to obtain a sparse wind field image, so that a part of priori knowledge is provided for training and learning of a subsequent model, and the training time of the model is reduced.
Will bitAt vertex v i The wind speed of the wind field at time t is recorded as
Figure BDA0003327862550000062
The anemogram signal at time t can be recorded as
Figure BDA0003327862550000063
The space-time prediction to be realized by the invention is based on historical anemogram signal sequence x at a plurality of moments t-K+1 ，x t-K+2 ，…，x t To predict the anemogram signal x t+1 In the present invention, K is set to 6.
1.2, performing spectrogram convolution calculation on the established undirected weighted graph, wherein the calculation formula of the spectrogram convolution layer is as follows:
Figure BDA0003327862550000064
wherein Q represents the number of convolution kernels, Y ：，q Representing the output of the qth convolution kernel, X ：，p Representing the p-th channel map signal input, f is a convolution kernel g Representing a graph convolution operation, W q，p，： The parameters representing the q-th convolution kernel corresponding to the P-th channel map signal input, a (-) represents the activation function, and P represents the number of channels of the map signal input, i.e., having P vertex features.
This patent uses wind speed data from 1/2004 to 31/12/2005 for graph modeling, where the γ threshold is set to 0.85, and the finally obtained wind field graph contains 76 nodes, 1305 edges, as shown in fig. 2, which is drawn by a Spring Layout (Spring Layout). In addition, the data are also used for training and verifying the model, wherein the ratio of the training set to the verifying set is 9: 1, and wind speed data from 1 month and 1 day to 12 months and 31 days in 2006 are used as a testing set of the model.
The long and short term memory network is improved in the step (2), and the principle of the graph convolution long and short term memory neural network is as follows:
2.1 the long-short term memory neural network is a variant of the recurrent neural network, the long-short term memory unit is used to replace the hidden unit in the original recurrent neural network, the long-short term memory unit has three gate structures, which are the input gate, the forgetting gate and the output gate, and can effectively control the updating of the cell state and the hidden layer state, and the internal calculation formula is as follows:
i t ＝σ(ω xi x t +ω hi h t-1 +b i )
f t ＝σ(ω xf x t +ω hf h t-1 +b f )
o t ＝σ(ω xo x t +ω ho h t-1 +b o )
c t ＝f t ⊙c t-1 +i t ⊙tanh(ω xc x t +ω hc h t-1 +b c )
h t ＝o t ⊙tanh(c t )
wherein i t 、f t 、o t Respectively representing the vectors of an input gate, a forgetting gate and an output gate at the moment t, c t And h t Respectively representing the cell state and hidden layer state vectors at time t, x i Input vector, ω, representing long-short term memory unit at time t xi 、ω hi 、ω xf 、ω hf 、ω xo 、ω ho 、ω xc And ω hc As a weight matrix, b i 、b f 、b o And b c For the bias vector, σ (-) represents a sigmoid function, and tanh (-) represents a hyperbolic tangent function.
2.2 the patent replaces the multiplication operations in the long and short term memory neural unit with spectrogram convolution operations, and improves the multiplication operations into Graph convolution long and short term memory units (GCLSTM cells), and the basic calculation formula is as follows:
I t ＝σ(W xi * g X t +W hi * g H t-1 +B i )
F t ＝σ(W xf * g X t +W hf * g H t-1 +B f )
O t ＝σ(W xo * g X t +W ho * g H t-1 +B o )
C t ＝F t ⊙C t-1 +I t ⊙tanh(W xc * g X t +W hc *g H t-1 +B c )
H t ＝O t ⊙tanh(C t )
wherein, I t 、F t 、O t Respectively representing the input gate, the forgetting gate and the output gate matrix at the time t, C t And H t Matrix representing the state of the cell and the state of the hidden layer at time t, x t Input matrix, W, representing long-short term memory cells at time t xi 、W xf 、W xo 、W xc Weight tensors W of input gate, forgetting gate, output gate and cell state in input layer neuron of long and short term memory network hi 、W hf 、W ho 、W hc Weight tensors of input gate, forget gate, output gate and cell state in neuron of hidden layer of long-short term memory network, respectively, B i 、B f 、B o And B c Bias matrices representing input gate, forgetting gate, output gate and cell state, respectively g The graph convolution operation is shown. The basic structure of the convolutional long short term memory cell proposed in this patent is shown in FIG. 3.
Building a multi-wind-field wind speed space-time point prediction model based on a graph convolution long-term short-term memory neural network in the steps (2) and (3), building a probability prediction model based on a transfer learning principle in the step (4), wherein the basic structure of the model is shown in FIG. 4, and the error loss formula of the multi-wind-field wind speed space-time point prediction model training is as follows:
Figure BDA0003327862550000071
wherein M represents the number of samples in each batch of training set, M is the number of wind fields,
Figure BDA0003327862550000072
representing the actual value of the output of the s-th wind farm in the ith sample of each batch of training sets,
Figure BDA0003327862550000073
and (4) representing the predicted value of the output of the s-th wind field in the ith sample in each batch of training sets.
The quantile loss formula trained by the multi-wind-field wind speed space-time probability prediction model is as follows:
Figure BDA0003327862550000081
wherein q represents the number of quantiles, and the present embodiment is set to 21, i.e. 0.01,0.05,0.10.,. 0.95,0.99, mainly for reducing the model training time, X i Representing the input of the ith sample in each batch of training sets, f s (W(τ j )，b(τ j )，X i ) Tau of the s wind field representing the output of the ith sample model in each batch of training sets j Wind speed prediction in quantile, W (τ) j ) And b (τ) j ) Is and quantile τ j And relevant model parameters, namely parameters of a full connection layer in the probability prediction model. Based on the transfer learning principle, the parameters of the GCLSTM layer in the point prediction model are transferred to the GCLSTM layer in the probability prediction model, so that the time required by the training of the probability prediction model is reduced.
The method uses the Average Root Mean Square Error (ARMSE), the Nash efficiency coefficient (ANSE) and the Average absolute error (AMAE) of a plurality of wind fields to evaluate the effect of the time-space point prediction of the wind speed of the wind fields, and the expression is as follows:
Figure BDA0003327862550000082
Figure BDA0003327862550000083
Figure BDA0003327862550000084
wherein N is the number of samples in the test set.
This patent uses the Average Score (Average Quantile Score, AQS) of a plurality of wind fields to come the comprehensive evaluation model probability density prediction effect, and its expression is as follows:
Figure BDA0003327862550000085
wherein, the first and the second end of the pipe are connected with each other,
Figure BDA0003327862550000086
representing the s-th wind field τ corresponding to the i-th sample model output j And (5) predicting the wind speed under quantiles. The smaller the average fractional score is, the better the probability prediction effect is.
In order to verify the effectiveness of the proposed point prediction model, the patent selects ANN and LSTM as comparison models, and because ANN and LSTM cannot process the image signal samples, an ANN and LSTM model is established separately for each wind field. The method selects multi-wind-field wind speed data of 1 month, 4 months, 7 months and 10 months in 2006 to compare the prediction effects of the model in different seasons. The prediction indexes of different models in different seasons and all the year are shown in table 2. As can be seen from the table, the model provided by the patent has the optimal point prediction indexes in different seasons and all the year, which shows that the accuracy of wind speed prediction can be effectively improved by fully utilizing the historical wind speed information of the adjacent wind field. In addition, as can be seen from the dimensionless ANSE indexes, the model provided by the patent has the best prediction performance in winter, the ANSE index reaches 0.9335, the prediction performance in summer is the worst, the ANSE index is only 0.7753, other comparison models have similar characteristics, and the wind speed fluctuation and randomness in summer are stronger than those in other seasons possibly due to climate reasons, so that the predictability is relatively weak.
FIG. 5 illustrates the point predictors for the various models throughout the year for each wind farm. In the graph, the RMSE index curve and the MAE index curve of the model provided by the patent are obviously lower than those of the other two comparison models, and the NSE index curve is obviously higher than those of the two comparison models, so that the prediction accuracy of each wind field is improved by fusing the historical wind speed information of the adjacent wind fields.
In order to verify the effectiveness of the proposed probabilistic predictive model, the patent also chooses to use quantile loss optimized ANN and LSTM networks as comparison models. Table 3 shows the probability predictors for different models in different seasons and throughout the year. As can be seen from the table, the average probability indexes AQS of the model provided by the method in different seasons are optimal, and the model shows better probability prediction performance, so that the method is combined with the historical wind speed information of the adjacent wind field and is also beneficial to the probability prediction of the wind speed. It can also be seen from the table that the probabilistic prediction performance of the model proposed herein is optimal in the winter, while the probabilistic prediction performance in the summer is relatively poor.
Fig. 6 shows, by way of example, a wind field No. 1, a probabilistic prediction result of the model provided by the patent representing 5 days before a month in different seasons. As can be seen from the figure, the black dashed line representing the actual value mostly falls within the confidence interval in different seasons, which indicates that the model provided by the invention has better prediction reliability.
Geographical position information of 76 wind fields in area of table 1
Figure BDA0003327862550000101
TABLE 2 Point prediction indexes for different models in different seasons
Figure BDA0003327862550000111
TABLE 3 probability prediction index of different models in different Seasons (AQs)
Figure BDA0003327862550000112

Claims (6)

1. A multi-wind-field wind speed space-time prediction method based on a graph volume and a recurrent neural network is characterized by comprising the following steps:
step (1): obtaining position information and wind speed data of a plurality of wind fields in a certain area, carrying out graph modeling by using an undirected weighted graph based on the position information and the wind speed data, and carrying out spectrogram convolution calculation on the established undirected weighted graph to construct a wind speed graph signal sequence; the method for carrying out graph modeling by using the undirected weighted graph based on the position information and the wind speed data comprises the following steps:
for M wind farms in a certain area, a non-directional weighted graph G = (V, E, a) to model its data, where V = { V = { V = 1 ，v 2 ，…，v M Is a set of vertices, each vertex representing a wind field of the area, E is a set of edges, A is an adjacency matrix, for each A ij e.A, the definition of which is shown below:
Figure FDA0003886777410000011
wherein d (v) i ，v j ) Representing a vertex v i And v j The distance between the wind field of the ith wind field and the wind field of the jth wind field,
Figure FDA0003886777410000012
representing the standard deviation of all pairs of vertex distances for removing the effect of dimension, PCC (v) i ，v j ) Representing a Pearson correlation coefficient calculated based on the ith wind field and the jth wind field historical wind speed data, wherein gamma is a threshold value of the Pearson coefficient;
step (2): the multiplication operation in the long and short term memory neural network is replaced by graph convolution calculation, a graph convolution long and short term memory neural network is constructed, and a multi-wind-field wind speed space-time point prediction model based on the graph convolution long and short term memory neural network is established on the basis; the multiplication operation in the long-short term memory neural network is replaced by graph convolution calculation to construct the graph convolution long-short term memory neural network, and the calculation formula is as follows:
I t ＝σ(W xi * g X t +W hi * g H t-1 +B i )
F t ＝σ(W xf * g X t +W hf * g H t-1 +B f )
O t ＝σ(W xo * g X t +W ho * g H t-1 +B o )
Figure FDA0003886777410000013
Figure FDA0003886777410000014
wherein, I t 、F t 、O t Respectively representing the input gate, the forgetting gate and the output gate matrix at the time t, C t And H t Matrix representing the state of the cell and the state of the hidden layer at time t, X t Input matrix, W, representing long-short term memory cells at time t xi 、W xf 、W xo 、W xc Weight tensors W of input gate, forgetting gate, output gate and cell state in input layer neuron of long and short term memory network hi 、W hf 、W ho 、W hc Weight tensors of input gate, forget gate, output gate and cell state in neuron of hidden layer of long-short term memory network, respectively, B i 、B f 、B o And B c Bias matrices representing input gate, forget gate, output gate, and cell state, respectively g Representing a graph convolution operation;
and (3): training parameters of a multi-wind-field wind speed space-time point prediction model based on historical wind speed data of the multi-wind-field to obtain the multi-wind-field wind speed space-time point prediction model;
and (4): migrating the parameters of the multi-wind-field wind speed space-time point prediction model trained in the step (3) based on a migration learning principle, constructing a multi-wind-field wind speed space-time probability prediction model, and further training to obtain the multi-wind-field wind speed space-time probability prediction model;
and (5): and (3) constructing wind field information needing wind speed prediction into a wind speed diagram signal sequence through the step (1), sequentially inputting the wind speed diagram signal sequence to be predicted into the models obtained by training in the step (3) and the step (4), and obtaining the point prediction and probability prediction results of the wind speed of the multi-wind field.
2. The multi-wind-field wind speed space-time prediction method based on the graph volume and the recurrent neural network as claimed in claim 1, wherein: constructing the anemogram signal sequence described in step (1) to be located at the vertex v i The wind speed of the wind field at the moment t is recorded as
Figure FDA0003886777410000021
The anemogram signal at time t is recorded as
Figure FDA0003886777410000022
The anemogram signal sequence at several historical moments is x t-K+1 ，x t-K+2 ，…，x t Wherein K is a natural number, and predicting an anemogram signal x at the next moment by using the anemogram signal sequence t+1 。
3. The multi-wind-field wind speed space-time prediction method based on graph convolution and recurrent neural network as claimed in claim 1, wherein the step (1) uses a spectrogram convolution calculation on the established undirected weighted graph, wherein the calculation formula of the spectrogram convolution layer is as follows:
Figure FDA0003886777410000023
wherein Q represents the number of convolution kernels, Y ：，q To representOutput of the qth convolution kernel, X ：，p Representing the p-th channel map signal input, f is a convolution kernel g Representing a graph convolution operation, W q，p，： The parameter representing the q-th convolution kernel corresponding to the P-th channel map signal input, a (-) represents the activation function, and P represents the number of channels of the map signal input.
4. The method for multi-wind-field wind speed space-time prediction based on graph convolution and recurrent neural network as claimed in claim 1, wherein the error loss formula of the model training of the multi-wind-field wind speed space-time point prediction model in step (3) is:
Figure FDA0003886777410000024
wherein M represents the number of samples in each batch of training set, M is the number of wind fields,
Figure FDA0003886777410000031
representing the actual value of the output of the s-th wind farm in the ith sample of each batch of training sets,
Figure FDA0003886777410000032
and (4) representing the predicted value of the output of the s-th wind field in the ith sample in each batch of training sets.
5. The multi-wind-field wind speed spatiotemporal prediction method based on graph convolution and recurrent neural network as claimed in claim 1, wherein the multi-wind-field wind speed spatiotemporal probability prediction model in step (4) is trained by quantile loss formula as follows:
Figure FDA0003886777410000033
wherein q represents the number of quantiles, M represents the number of samples in each batch of training set, M is the number of wind fields,
Figure FDA0003886777410000034
represents the quantile loss of the jth quantile,
Figure FDA0003886777410000035
representing the actual value of the s-th wind field output in the ith sample of the training set, f s (W(τ j )，b(τ j )，X i ) Representing the quantile tau of the s wind field output by the ith sample model in each batch of training sets j Lower wind speed prediction, W (τ) j ) And b (τ) j ) Is the fractional number tau in a multi-wind field wind speed space-time probability prediction model j W and b parameters, X, of the underlying fully-connected layer i Representing the input of the ith sample in each batch of training sets.
6. The multi-wind-field wind speed space-time prediction method based on the graph convolution and the recurrent neural network as claimed in claim 1, wherein the step (4) of migrating the point prediction model parameters trained in the step (3) based on the migration learning principle is to migrate the parameters of the graph convolution long and short term memory neural network layer in the point prediction model to the graph convolution long and short term memory neural network layer in the probabilistic prediction model based on the migration learning algorithm.
CN202111280781.0A 2021-10-29 2021-10-29 Multi-wind-field wind speed space-time prediction method based on graph convolution and recurrent neural network Active CN114004152B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111280781.0A CN114004152B (en) 2021-10-29 2021-10-29 Multi-wind-field wind speed space-time prediction method based on graph convolution and recurrent neural network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111280781.0A CN114004152B (en) 2021-10-29 2021-10-29 Multi-wind-field wind speed space-time prediction method based on graph convolution and recurrent neural network

Publications (2)

Publication Number Publication Date
CN114004152A CN114004152A (en) 2022-02-01
CN114004152B true CN114004152B (en) 2022-11-22

Family

ID=79926091

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111280781.0A Active CN114004152B (en) 2021-10-29 2021-10-29 Multi-wind-field wind speed space-time prediction method based on graph convolution and recurrent neural network

Country Status (1)

Country Link
CN (1) CN114004152B (en)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114580509B (en) * 2022-02-21 2022-10-28 国家海洋环境预报中心 Sea wave macroscopic characteristic quantity prediction system based on convolution length memory network
CN114298444B (en) * 2022-03-09 2022-05-27 广东电网有限责任公司佛山供电局 Wind speed probability prediction method, device, equipment and storage medium

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112529282A (en) * 2020-12-03 2021-03-19 清华大学 Wind power plant cluster short-term power prediction method based on space-time graph convolutional neural network
CN113516271A (en) * 2021-04-05 2021-10-19 东北电力大学 Wind power cluster power day-ahead prediction method based on space-time neural network

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110929962A (en) * 2019-12-13 2020-03-27 中国科学院深圳先进技术研究院 Traffic flow prediction method and device based on deep learning
CN114220271B (en) * 2021-12-21 2023-06-30 南京理工大学 Traffic flow prediction method, equipment and storage medium based on dynamic space-time diagram convolution circulation network

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112529282A (en) * 2020-12-03 2021-03-19 清华大学 Wind power plant cluster short-term power prediction method based on space-time graph convolutional neural network
CN113516271A (en) * 2021-04-05 2021-10-19 东北电力大学 Wind power cluster power day-ahead prediction method based on space-time neural network

Also Published As

Publication number Publication date
CN114004152A (en) 2022-02-01

Similar Documents

Publication Publication Date Title
CN111784041B (en) Wind power prediction method and system based on graph convolution neural network
CN111860982A (en) Wind power plant short-term wind power prediction method based on VMD-FCM-GRU
CN114004152B (en) Multi-wind-field wind speed space-time prediction method based on graph convolution and recurrent neural network
CN110648014B (en) Regional wind power prediction method and system based on space-time quantile regression
CN112862090B (en) Air temperature forecasting method based on deep space-time neural network
CN111815033A (en) Offshore wind power prediction method based on RCNN and meteorological time sequence characteristics
CN113627093B (en) Underwater mechanism trans-scale flow field characteristic prediction method based on improved Unet network
CN115374995A (en) Distributed photovoltaic and small wind power station power prediction method
CN114330935B (en) New energy power prediction method and system based on multiple combination strategies integrated learning
CN113449919B (en) Power consumption prediction method and system based on feature and trend perception
CN112966853A (en) Urban road network short-term traffic flow prediction method based on space-time residual error mixed model
CN114462718A (en) CNN-GRU wind power prediction method based on time sliding window
CN112686376A (en) Node representation method based on timing diagram neural network and incremental learning method
CN115983494A (en) Short-term wind power prediction method and system for newly-built small-sample wind power plant
CN116345555A (en) CNN-ISCA-LSTM model-based short-term photovoltaic power generation power prediction method
CN114065646B (en) Energy consumption prediction method based on hybrid optimization algorithm, cloud computing platform and system
CN114862023A (en) Distributed photovoltaic power prediction method and system based on four-dimensional point-by-point meteorological forecast
CN111783688B (en) Remote sensing image scene classification method based on convolutional neural network
CN116826734A (en) Photovoltaic power generation power prediction method and device based on multi-input model
CN116579408A (en) Model pruning method and system based on redundancy of model structure
CN112734073A (en) Photovoltaic power generation short-term prediction method based on long and short-term memory network
CN116721327A (en) Neural network architecture searching method based on generalization boundary
CN115834161A (en) Power grid false data injection attack detection method of artificial intelligence four-layer architecture
CN114386666A (en) Wind power plant short-term wind speed prediction method based on space-time correlation
CN115222024B (en) Short-term photovoltaic power generation prediction method and system based on depth feature selection network

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant