CN116070763A - New wind power plant wind power prediction method and system based on gradient evolution - Google Patents

New wind power plant wind power prediction method and system based on gradient evolution Download PDF

Info

Publication number
CN116070763A
CN116070763A CN202310101458.5A CN202310101458A CN116070763A CN 116070763 A CN116070763 A CN 116070763A CN 202310101458 A CN202310101458 A CN 202310101458A CN 116070763 A CN116070763 A CN 116070763A
Authority
CN
China
Prior art keywords
wind power
power plant
training
gradient
data
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.)
Pending
Application number
CN202310101458.5A
Other languages
Chinese (zh)
Inventor
孟安波
张海涛
容嘉瑜
冼梓康
陈黍
张展
李晨
殷豪
严柏平
罗坚强
范竞敏
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.)
Guangdong University of Technology
Original Assignee
Guangdong University of Technology
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 Guangdong University of Technology filed Critical Guangdong University of Technology
Priority to CN202310101458.5A priority Critical patent/CN116070763A/en
Publication of CN116070763A publication Critical patent/CN116070763A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • 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
    • 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
    • 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
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
    • Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
    • Y04S10/00—Systems supporting electrical power generation, transmission or distribution
    • Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications

Landscapes

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

Abstract

The invention provides a new wind power plant wind power prediction method and system based on gradient evolution, which relate to the technical field of wind power prediction, acquire wind power related data from a target wind power plant and an adjacent wind power plant thereof, process the wind power related data, then construct a space-time diagram convolution of the adjacent wind power plant and the target wind power plant to generate an antagonism network, effectively utilize the time and space relations between the adjacent wind power plant and the target wind power plant within a certain range, construct a similarity diagram, a correlation diagram and a distance diagram, fully excavate space-time characteristic generation data to realize data expansion, improve the quality of generated samples, realize the advantage complementation of different training modes based on a gradient evolution calculation frame, exchange fusion gradient information, solve the problems that gradient disappearance, gradient explosion and local optimization possibly exist in a single training mode of a gating circulation unit network constructed by a neural network, and improve the wind power prediction precision.

Description

New wind power plant wind power prediction method and system based on gradient evolution
Technical Field
The invention relates to the technical field of wind power prediction, in particular to a method and a system for predicting wind power of a newly built wind power plant based on gradient evolution.
Background
Under the dual-carbon background, the development of new energy power generation becomes global trend, wind energy is one of clean energy, the wind energy has the characteristics of strong availability and large time sequence fluctuation, the scale of the wind energy is continuously enlarged, the economic and scheduling influence of the wind power large-scale grid connection on a power system is not negligible, and the accurate wind power prediction has important significance on the safety and stability of the whole power system.
The wind power plant with rich historical data can be predicted through mathematical modeling, machine learning and deep learning prediction models, and the problem of low prediction accuracy caused by lack of historical data of a newly built wind power plant is solved, the hidden relations of the adjacent wind power plant and the target wind power plant in time and space are considered, the time-space relation of the adjacent wind power plant to the target wind power plant is effectively quantized by utilizing the historical power data and the meteorological data owned by the adjacent wind power plant, and the data expansion is realized by utilizing space-time diagram convolution generation countermeasure network, so that the quality of the generated data is further improved. So far, the existing data generation method only considers the few sample historical data of the target wind power plant, and training samples are insufficient, so that the generated data is single, and the influence of the adjacent wind power plant is not fully considered, so that how to realize data expansion by utilizing the relationship between the adjacent wind power plant and the target wind power plant in time and space becomes a technical problem to be solved urgently.
In addition, the wind power time sequence and the wind speed time sequence have the characteristics of strong randomness and fluctuation, if the characteristic sequences are directly input into a prediction model, the difficulty of prediction can be increased by high-frequency components, and the prediction precision is reduced.
The method adopts evolutionary computation to optimize and generate the countermeasure network, so that the generation model can efficiently learn marginal distribution of original few data and generate new data with modal diversity and similar marginal distribution, so as to make up the limitation of original small-scale data, and has practical help for improving the precision of the low-data wind power prediction of the newly-built wind power plant; the method has the advantages that the Dense layer weight and the bias items in the BiGRU network are optimized by adopting the crisscross optimization algorithm, the model can be effectively avoided from sinking into local optimum and is helped to find out global optimum solutions, the method has an obvious effect on improving the wind power prediction precision of a small number of newly built wind power plants, but in the implementation process of the method, the parameter optimum solutions after the completion of training are searched from the outside of the model through the optimization algorithm, the problems of gradient disappearance, gradient explosion, local optimum and the like in the model training process are not substantially solved, the prediction precision of the model when the model is finally used for wind power prediction is adversely affected, and the wind power prediction precision cannot be ensured.
Disclosure of Invention
In order to solve the problem that the existing method for predicting the wind power of the newly-built small-sample wind power plant is low in prediction precision, the invention provides a method and a system for predicting the wind power of the newly-built wind power plant based on gradient evolution, and the precision of predicting the wind power of the newly-built small-sample wind power plant is improved.
In order to achieve the technical effects, the technical scheme of the invention is as follows:
a new wind power plant wind power prediction method based on gradient evolution comprises the following steps:
s1, acquiring wind power related data of a target wind power plant and a wind power plant adjacent to the target wind power plant, and processing the wind power related data;
s2, constructing a space-time diagram convolution generation countermeasure network of the adjacent wind power plant and the target wind power plant, and inputting the wind power related data processed in the S1 into the space-time diagram convolution generation countermeasure network to generate data to realize data expansion;
s3, wind power related data of an unprocessed target wind power plant and generated data based on a space-time diagram convolution generation countermeasure network are used as a source data set, the source data set is decomposed by utilizing multivariate mode decomposition, and an input feature matrix is formed based on the decomposed data;
s4, constructing a prediction model comprising a convolutional neural network and an evolution gating circulation unit network;
S5, taking the input feature matrix as the input of a prediction model, extracting features by using a convolutional neural network of the prediction model, and mining implicit space-time relations among the features by using an evolutionary gating circulating unit network of the prediction model;
s6, training a prediction model by utilizing data input into a feature matrix based on a gradient evolution calculation frame to obtain a trained prediction model for wind power prediction.
Preferably, the wind power related data in step S1 includes a wind power sequence, a wind speed sequence, a wind direction sequence and temperature data, and the processing process is as follows: abnormal value elimination and missing value filling are carried out on wind power sequence, wind speed sequence, wind direction sequence and temperature data, sine and cosine processing is adopted on the wind direction sequence, and wind direction sine sequence WD is obtained sin Cosine sequence WD of the sum and the wind direction cos 。
Preferably, when a space-time diagram convolution of an adjacent wind power plant and a target wind power plant is constructed to generate an countermeasure network, a multi-diagram convolution neural network is used as a generator of the space-time diagram convolution to generate the countermeasure network, wherein the generator comprises a similarity diagram GCN-SIM, a correlation diagram GCN-COR and a distance diagram GCN-DST of the target wind power plant and the adjacent wind power plant, the similarity diagram GCN-SIM, the correlation diagram GCN-COR and the distance diagram GCN-DST are overlapped to form a comprehensive diagram GCN-SYN, each diagram consists of wind power plant nodes v and edges epsilon, the nodes are expressed as G= (v, epsilon), each node represents an adjacent wind power plant, each edge represents a correlation relation between each adjacent wind power plant and the target wind power plant, and wind power related data of each wind power plant after processing is obtained;
Setting a similarity adjacent matrix corresponding to the similarity graph GCN-SIM as A sim The correlation adjacency matrix of the correlation graph (GCN-COR) is A cor The distance adjacency matrix of the distance graph (GCN-DST) is A dst Then the comprehensive adjacency matrix of the fusion graph (GCN-SYN) is A syn ＝A sim +A cor +A dst All matrices are normalized to (0, 1);
a layer of space-time convolution neural network is used as a discriminator for generating an countermeasure network by space-time diagram convolution, wind power related data processed by S1 is input into the space-time diagram convolution generation countermeasure network, and a distance adjacency matrix A is adopted dst And judging errors between the generated data and the original wind power related data as an adjacent matrix of the graph, guiding the generator to update iteratively, and generating high-quality data.
In the technical scheme, a space-time diagram convolution is adopted to generate a similar adjacent matrix A of an adjacent wind power plant and a target wind power plant within a certain range formed by an antagonism network sim Correlation adjacency matrix A cor Distance adjacency matrix A dst The three are respectively formed into a similarity graph (GCN-SIM), a correlation graph (GCN-COR) and a distance graph (GCN-DST), and then are fused to form a comprehensive graph (GCN-SYN), so that high-quality data are generated by comprehensively utilizing the space-time relations contained in different graphs, the expansion of the data is realized, the number and the quality of required samples are improved, and the wind power prediction accuracy is improved.
Preferably, the source data set is decomposed by using multivariate schema decomposition, and the process of forming the input feature matrix based on the decomposed data is as follows:
s31, set p (k) (t) represents a time sequence of wind power at t moment in a source data set, v m(k) (t)、v z(k) (t) respectively representing a warp wind speed time sequence and a band wind speed time sequence at the moment t in the source data set, and then satisfying the following conditions:
Figure BDA0004085586050000031
wherein x (t) comprises an intrinsic mode function u k Number n of subsequences of (t) imf ；
S32, decomposing the source data set by utilizing multi-element mode decomposition to obtain three groups of data respectively belonging to p (t) and v m (t)、v z Subsequence of (t)
Figure BDA0004085586050000032
The set of subsequences is used as one of the inputs to the prediction model;
s33, forming an input feature matrix X based on the decomposed data input The expression is:
Figure BDA0004085586050000041
wherein ,Pt-m 、
Figure BDA0004085586050000042
Tem t-m The power, radial wind speed, strip wind speed, wind direction sine, wind direction cosine and temperature of the wind farm at the moment t-m of the wind farm are respectively represented.
Preferably, the process of constructing the prediction model including the convolutional neural network and the evolutionary gating loop unit network in step S4 is as follows:
s41, constructing a convolutional neural network module to extract input features of an input feature matrix, wherein the convolutional neural network module comprises three layers of 2D convolutional layers and a full-connection layer, the convolutional kernels of the three layers of 2D convolutional layers are the same in size, the number of filters is 4, 8 and 16 respectively, the activating functions are ReLU functions, the filling modes are the same, the pooling is carried out in a maximum pooling mode, and the three layers of 2D convolutional layers are output by the full-connection layer;
S42, building a three-layer evolution gating circulating unit network, wherein the number of neurons of the three-layer evolution gating circulating unit network is 4, 8 and 16 respectively, the activation functions are all ReLU functions, the three-layer gating circulating unit network is connected with a full-connection layer, the full-connection layer activation function is the ReLU function, and the forward propagation formula of the three-layer evolution gating circulating unit network meets the following conditions:
Figure BDA0004085586050000043
wherein ,Wr 、W z 、W h 、U r 、U z 、U h As a weight parameter matrix, b r 、b z 、b h In order to bias the parameter matrix,
Figure BDA0004085586050000044
for matrix multiplication, σ is a Sigmod function, r t To reset the gate, z t For updating the door->
Figure BDA0004085586050000045
As candidate state of hidden layer at current moment, y t As the current implicit state, y t-1 Is the implicit state of the previous moment, x t Is the input state at the current time.
S43, the output end of the convolutional neural network module is connected with the input end of the three-layer evolution gate control circulation unit network. A predictive model is constructed.
Preferably, in step S5, when the convolutional neural network using the prediction model extracts features, the convolutional neural network module inputs the feature matrix X input The input multidimensional feature data in (1) is subjected to feature extraction, the dimension of an input layer of a convolutional neural network module is n multiplied by m multiplied by t, and the output is as follows:
Figure BDA0004085586050000051
wherein ,
Figure BDA0004085586050000052
representing the output of a convolutional neural network module, X i+n,j+m Representing input feature matrix X input The value of row n and column m, f cov (. Cndot.) represents the selection of an activation function, ω n,m Weights representing n rows and m columns of the convolution kernel, b n,m Is the convolution kernel deviation, k is the sliding window size.
Preferably, in step S6, the gradient evolution computing framework includes three training modules, specifically: the system comprises a gradient descent training module with a driving quantity, an acceleration gradient descent training module and a self-adaptive moment estimation training module, wherein in the gradient descent training module with the driving quantity, the parameter updating mode is as follows:
v dw ＝βv dw +(I-β)dW
v db ＝βv db +(1-β)db
W＝W-αv dw ，b＝b-αv db
wherein, alpha is learning rate, beta controls exponential weighted average, W is weight, and b is deviation;
in the acceleration gradient descent training module, the parameter updating mode is as follows:
S dW ＝βS dW +(1-β)dW 2
S db ＝β＝S db +(1-β)db 2
Figure BDA0004085586050000053
in the adaptive moment estimation training module, the parameter updating mode is as follows:
Figure BDA0004085586050000054
Figure BDA0004085586050000055
Figure BDA0004085586050000056
wherein ,β1 Weighting the average parameter, beta, for the control index in the momentum gradient descent training module 2 For the control exponential weighted average parameter in the acceleration gradient descent training module, ε is to prevent the denominator from being 0.
Preferably, when training the prediction model based on the gradient evolution computing framework, three training modules in the gradient evolution computing framework are adopted to respectively perform iterative training in the training process, and the method specifically comprises the following steps:
s61, after generating offspring once each iteration, carrying out quality evaluation on each offspring, setting a scoring mechanism according to prediction precision, wherein the prediction precision score is S p The expression is:
S P ＝ν RMSE
wherein ,νRMSE Lower represents more excellent offspring, according to v RMSE The optimal evolutionary offspring are selected out,
Figure BDA0004085586050000061
Figure BDA0004085586050000062
the method comprises the steps of respectively representing individuals with optimal performance in three training modules under the current iteration times, j representing algebra of the current evolution, S, R, A respectively representing training modes corresponding to the three training modules, namely a gradient descent training mode, an acceleration gradient descent training mode and a self-adaptive moment estimation training mode of driving quantity;
s62, after the individuals with the optimal performance in the three training modules are obtained respectively, respectively calculating the weight coefficients corresponding to the individuals according to the prediction precision scores in S61, wherein the expressions are as follows:
Figure BDA0004085586050000063
Figure BDA0004085586050000064
Figure BDA0004085586050000065
wherein ,k1 The weights representing the gradient descent training mode with momentum,
Figure BDA0004085586050000066
prediction accuracy score k representing a gradient descent training mode with momentum 2 Weights representing the way of training the gradient descent acceleration, +.>
Figure BDA0004085586050000067
Predictive accuracy score, k, representing the manner of training for acceleration gradient descent 3 Weights representing the adaptive moment estimation training pattern, +.>
Figure BDA0004085586050000068
A prediction accuracy score representing the adaptive moment estimation training mode;
s63, weighting and summing the obtained weight coefficients to obtain a comprehensive gradient, and updating parameters according to an updating formula of the parameters w and b, wherein the expression is as follows:
Figure BDA0004085586050000069
Figure BDA00040855860500000610
Figure BDA00040855860500000611
Figure BDA00040855860500000612
in the formula ,
Figure BDA00040855860500000613
and
Figure BDA00040855860500000614
Representing the integrated gradient of the parameters w and b, v, respectively dw 、v db Respectively represent the gradients of current parameters w and b in a gradient descent training mode with momentum, S dW 、S db Respectively represent the gradients of the current parameters w and b in the acceleration gradient descent training mode, +.>
Figure BDA00040855860500000615
Respectively representing gradients of current parameters w and b in a self-adaptive moment estimation training mode;
s64, calculating weight coefficients of the corresponding training methods according to the prediction precision scores, giving larger weights to the offspring in the training modes with good performance to the training modes corresponding to the three training modules, simultaneously keeping the offspring information with good performance to the other training modes to correct the offspring, fusing the gradient information of the three optimal offspring parameters under the current iteration times, and generating comprehensive gradient sum through evolution
Figure BDA0004085586050000071
and
Figure BDA0004085586050000072
And then, according to an updating formula of the parameters w and b, carrying out parameter updating by utilizing the comprehensive gradient to obtain updated parameters under the current iteration times, and repeating the operation in the next iteration as an initial solution of the next iteration to finish the training of the prediction model.
In the technical scheme, the parameters are updated by adopting the evolutionary gating circulating unit network to fuse three different model training methods, and the frame is calculated through evolutionaryThe frame advantages are complementary, gradient information of optimal offspring of the current iteration times is exchanged and fused, and comprehensive gradient is generated through evolution
Figure BDA0004085586050000073
and
Figure BDA0004085586050000074
The method has the advantages that the parameters are guided to be updated, different training methods are fused through the evolutionary computing framework in each training, the problems that gradient disappearance, gradient explosion and local optimization possibly exist in a single training mode of the GRU network built by the neural network can be solved, and the method has a certain practical significance for improving the network prediction precision of the gating circulating unit.
Preferably, during the training of the predictive model, the fitness function is defined as a root mean square error RMSE to reflect the error between the predicted value and the true value; when the trained prediction model is used for predicting the wind power of a newly built wind power plant, and taking wind power related data of the newly built wind power plant in a period of time before prediction as input of a prediction model, and outputting a wind power prediction result.
The application also provides a newly-built wind power plant wind power prediction system based on gradient evolution, wherein the system comprises:
the wind power data acquisition and preprocessing module is used for acquiring wind power related data of a target wind power plant and a wind power plant adjacent to the target wind power plant and processing the wind power related data;
the data expansion module is used for constructing a space-time diagram convolution generation countermeasure network of the adjacent wind power plant and the target wind power plant, inputting processed wind power related data into the space-time diagram convolution generation countermeasure network, and generating data to realize data expansion;
The data set decomposition module is used for decomposing the source data set by utilizing multi-element mode decomposition and forming an input feature matrix based on the decomposed data by taking the wind power related data of the target wind power plant before being processed and the generated data of the countermeasure network generated based on space-time diagram convolution as the source data set;
the prediction model construction module is used for constructing a prediction model comprising a convolutional neural network and an evolutionary gating circulating unit network;
the feature extraction and mining module is used for taking an input feature matrix as the input of a prediction model, extracting features by using a convolutional neural network of the prediction model, and mining implicit space-time relations among the features by using an evolutionary gating circulating unit network of the prediction model;
and the prediction model training module is used for training a prediction model by utilizing data in the input feature matrix based on the gradient evolution calculation frame to obtain a trained prediction model for wind power prediction.
Compared with the prior art, the technical scheme of the invention has the beneficial effects that:
the invention provides a new wind power prediction method and a system based on gradient evolution, which are used for acquiring wind power related data from a target wind power plant and an adjacent wind power plant thereof, processing the wind power related data, then constructing a space-time diagram convolution of the adjacent wind power plant and the target wind power plant to generate an antagonism network, effectively utilizing the time and space relations between the adjacent wind power plant and the target wind power plant within a certain range, constructing a similarity diagram, a correlation diagram and a distance diagram, fully excavating space-time characteristic generation data to realize data expansion, improving the quality of generated samples, realizing the advantage complementation of different training modes based on a gradient evolution calculation frame, exchanging and fusing gradient information, solving the problems of possible gradient disappearance, gradient explosion and local optimization in a single training mode of a gate control circulation unit network constructed by a neural network, and improving the wind power prediction precision.
Drawings
FIG. 1 shows a flow chart of a new wind power plant wind power prediction method based on gradient evolution, which is proposed in the embodiment 1 of the invention;
FIG. 2 is a schematic diagram showing a generator for generating an countermeasure network by convolution of a plurality of images with a neural network as a space-time diagram according to embodiment 2 of the present invention;
FIG. 3 shows a block diagram of a predictive model including a convolutional neural network and an evolutionarily gated loop unit network as proposed in example 2 of the present invention;
FIG. 4 is a diagram showing the evolutionary framework of the evolutionary gating loop unit network proposed in embodiment 2 of the present invention;
FIG. 5 shows a graph of the wind power prediction effect obtained by the wind power prediction method of the new wind farm based on gradient evolution, which is provided in embodiment 2 of the present invention;
fig. 6 shows a structure diagram of a new wind power plant wind power prediction system based on gradient evolution proposed in embodiment 3 of the present invention.
Detailed Description
The drawings are for illustrative purposes only and are not to be construed as limiting the present patent;
for better illustration of the present embodiment, some parts of the drawings may be omitted, enlarged or reduced, and do not represent actual dimensions;
it will be appreciated by those skilled in the art that some well known descriptions in the figures may be omitted.
The technical scheme of the invention is further described below with reference to the accompanying drawings and examples.
The positional relationship depicted in the drawings is for illustrative purposes only and is not to be construed as limiting the present patent;
example 1
The embodiment provides a new wind power plant wind power prediction method based on gradient evolution, a flow chart of the method is shown in fig. 1, and referring to fig. 1, the method comprises the following steps:
s1, acquiring wind power related data of a target wind power plant and a wind power plant adjacent to the target wind power plant, and processing the wind power related data; in this embodiment, the wind power related data includes a wind power sequence, a wind speed sequence, a wind direction sequence, and temperature data, and the processing process is as follows: abnormal value elimination and missing value filling are carried out on wind power sequence, wind speed sequence, wind direction sequence and temperature data, sine and cosine processing is adopted for the wind direction sequence, fluctuation can be reduced, and wind direction sine sequence WD is obtained sin Cosine sequence WD of the sum and the wind direction cos 。
S2, constructing a space-time diagram convolution generation countermeasure network of the adjacent wind power plant and the target wind power plant, and inputting the wind power related data processed in the S1 into the space-time diagram convolution generation countermeasure network to generate data to realize data expansion;
s3, wind power related data of an unprocessed target wind power plant and generated data based on a space-time diagram convolution generation countermeasure network are used as a source data set, the source data set is decomposed by utilizing multivariate mode decomposition, and an input feature matrix is formed based on the decomposed data;
S4, constructing a prediction model comprising a convolutional neural network and an evolution gating circulation unit network;
s5, taking the input feature matrix as the input of a prediction model, extracting features by using a convolutional neural network of the prediction model, and mining implicit space-time relations among the features by using an evolutionary gating circulating unit network of the prediction model;
s6, training a prediction model by utilizing data input into a feature matrix based on a gradient evolution calculation frame to obtain a trained prediction model for wind power prediction.
Example 2
When the space-time diagram convolution of the adjacent wind power plant and the target wind power plant is constructed to generate the countermeasure network in step S2, as shown in fig. 2, the generator for generating the countermeasure network by using the multi-diagram convolution neural network as the space-time diagram convolution includes a similarity diagram GCN-SIM, a correlation diagram GCN-COR and a distance diagram GCN-DST of the target wind power plant and the adjacent wind power plant, wherein the similarity diagram GCN-SIM, the correlation diagram GCN-COR and the distance diagram GCN-DST are superimposed to form a comprehensive diagram GCN-SYN, and a similarity adjacency matrix corresponding to the similarity diagram GCN-SIM is set as a sim The correlation adjacency matrix of the correlation graph (GCN-COR) is A cor The distance adjacency matrix of the distance graph (GCN-DST) is A dst Each graph consists of wind farm nodes v and edges epsilon, which are expressed as G= (v, epsilon), each node represents an adjacent wind farm, each edge represents the correlation relationship between each adjacent wind farm and the target wind farm, and the wind power correlation data of each processed wind farm are obtained.
The similarity adjacency matrix adopts a sparse DTW method to measure two wind power related data time sequences S of a wind power plant i and a wind power plant j i and Sj (mainly time series of wind power and wind speed)Similarly, the time series s can be defined as a power time series if the power calculation is good, a wind speed series if the wind speed is good, or both. Wherein the method comprises the steps of
Figure BDA0004085586050000101
Is provided with->
Figure BDA0004085586050000102
And
Figure BDA0004085586050000103
is two time sequences of wind farm i and wind farm j, P time steps in total, +.>
Figure BDA0004085586050000104
and
Figure BDA0004085586050000105
Is two points corresponding to the time step number t in the time sequence, and performs cumulative calculation on the distance between the time points, wherein the calculation expression is as follows:
Figure BDA0004085586050000106
two time sequences S i and Sj The DTW distance between them is expressed as follows:
Figure BDA0004085586050000107
the correlation adjacency matrix performs season-trend decomposition using RobustSTL, by which time series are decomposed into season and trend components, and for each time series, the output of RobustSTL is a trend decomposition sequence t= { T1, T2,... Calculating trend correlation c between time series i and j using pearson correlation coefficients Trd And seasonal relevance c Ses The method is characterized by comprising the following steps:
Figure BDA0004085586050000108
Figure BDA0004085586050000109
Wherein cov (T) i ,T j )、cov(ι i ,ι j ) Trend and seasonal covariance are represented, respectively, and σ represents standard deviation. Finally, the correlation adjacency matrix is formed by Hadamard product between the trend correlation matrix and the seasonal correlation matrix
Figure BDA00040855860500001010
Calculation of>
Figure BDA00040855860500001011
Is the correlation between wind farm i and wind farm j.
Distance adjacency matrix A dst The method is obtained by calculating the road network distance between two wind farms, and a matrix is constructed by thresholding Gaussian kernel functions, and the expression is as follows:
Figure BDA0004085586050000111
wherein ,
Figure BDA0004085586050000112
representing the link weight between wind farm i and wind farm j, dis (i, j) representing the length between wind farm i and wind farm j, v being the standard deviation of the distance, ε being the distance matrix A dst Is a threshold for the degree of sparseness of (a). The comprehensive adjacency matrix of the fusion map (GCN-SYN) is A syn ＝A sim +A cor +A dst All matrices are normalized to (0, 1);
wherein ,
Figure BDA0004085586050000113
Figure BDA0004085586050000114
a layer of space-time convolution neural network is used as a discriminator for generating an countermeasure network by space-time diagram convolution, wind power related data processed by S1 is input into the space-time diagram convolution generation countermeasure network, and a distance adjacency matrix A is adopted dst As the adjacency matrix of the graph, the error between the generated data and the original wind power related data is judged, the generator is guided to update in an iterative way, high-quality data is generated, the expansion of the data is realized, the number and the quality of required samples are improved, and the wind power prediction precision is improved.
Decomposing the source data set by utilizing multivariate mode decomposition, and forming an input feature matrix based on the decomposed data comprises the following steps:
s31, defining input data; let x (t) comprise the eigenmode function u k Number n of subsequences of (t) imf Expressed as:
Figure BDA0004085586050000121
wherein ,p(k) (t) represents a time sequence of wind power at t moment in a source data set, v m(k) (t)、v z(k) (t) represents a meridional wind speed time series and a banded wind speed time series at time t in the source data set, respectively;
s32, decomposing the source data set by utilizing multi-element mode decomposition to obtain three groups of data respectively belonging to p (t) and v m (t)、v z Subsequence of (t)
Figure BDA0004085586050000122
The set of subsequences is used as one of the inputs to the prediction model; the method comprises the following steps:
s321, calculating a single-side spectrum; first, hilbert-Huang transform (HHT) is applied to obtain ω k Analytical expression of each element in (3)
Figure BDA0004085586050000123
Corresponding center frequency omega k (t) can be defined by->
Figure BDA0004085586050000124
Determining L by modulating signal gradient function 2 The norms calculate the width of each mode.
S322, defining an optimization objective function; will single frequency component omega k For the whole vector
Figure BDA0004085586050000125
To find common frequency component omega of each channel in multiple oscillation modes k The method comprises the steps of carrying out a first treatment on the surface of the When extracting a plurality of oscillation modes, the sum of the extracted modes should be satisfied to be equal to the sum of the widths of the original signal, the extracted modes to be minimum; thus, constraint-related optimization problems for multi-VMDs are defined as:
Figure BDA0004085586050000126
Figure BDA0004085586050000127
wherein
Figure BDA0004085586050000128
Representing corresponding channels c and u k Analytical expression for each element in (t), for example>
Figure BDA00040855860500001210
Representing the time-dependent partial derivative.
S323, establishing a Lagrange expression of the variation problem, wherein the Lagrange expression is expressed as:
Figure BDA0004085586050000129
wherein alpha is penalty factor, and the meaning of other parameters is consistent with the previous;
s324, solving a corresponding unconstrained problem by using ADMM (Alternate Direction Method of Multipliers) in order to obtain decomposed signal components; all estimation modalities can be expressed as:
Figure BDA0004085586050000131
wherein ,
Figure BDA0004085586050000132
respectively represent x n (ω),λ(ω),u i,n (ω),
Figure BDA0004085586050000133
I is the current iteration number; the modal center frequency update formula is:
Figure BDA0004085586050000134
finally, p (t) and v are respectively obtained m (t)、v z Three subsequences of (t)
Figure BDA0004085586050000135
Based on the above procedure, finally, step S33 is performed, namely:
s33, forming an input feature matrix X based on the decomposed data input The expression is:
Figure BDA0004085586050000136
wherein ,Pt-m 、
Figure BDA0004085586050000137
Tem t-m The power, radial wind speed, strip wind speed, wind direction sine, wind direction cosine and temperature of the wind farm at the moment t-m of the wind farm are respectively represented.
Here is a generalized expression of the break down subsequence. If the wind power and the two wind speeds are decomposed for 5 times, the input features are the decomposed power for 5 times at the corresponding moment and the decomposed wind speed for 5 times, and the power and the wind speed at the corresponding moment are used.
In this embodiment, referring to the structure diagram shown in fig. 3, the process of constructing the prediction model including the convolutional neural network and the evolutionary gating loop unit network in step S4 is as follows:
s41, constructing a convolutional neural network module to extract input characteristics of an input characteristic matrix, wherein the convolutional neural network module comprises three layers of 2D convolutional layers and a full-connection layer, the convolutional kernels of the three layers of 2D convolutional layers are 2 in the same size, the number of filters is 4, 8 and 16 respectively, the number of neurons is 4, 8 and 16 respectively, the activation functions are ReLU functions, the filling modes are identical, the three layers of 2D convolutional layers are pooled in a maximum pooling mode, and the three layers of 2D convolutional layers are output by the full-connection layer;
s42, building a three-layer evolution gating circulating unit network, wherein the number of neurons of the three-layer evolution gating circulating unit network is 4, 8 and 16 respectively, the activation functions are all ReLU functions, the three-layer gating circulating unit network is connected with a full-connection layer, the full-connection layer activation function is the ReLU function, and the forward propagation formula of the three-layer evolution gating circulating unit network meets the following conditions:
Figure BDA0004085586050000141
wherein ,Wr 、W z 、W h 、U r 、U z 、U h As a weight parameter matrix, b r 、b z 、b h In order to bias the parameter matrix,
Figure BDA0004085586050000142
for matrix multiplication, σ is a Sigmod function, r t To reset the gate, z t For updating the door->
Figure BDA0004085586050000143
As candidate state of hidden layer at current moment, y t As the current implicit state, y t-1 Is the implicit state of the previous moment, x t Is the input state at the current time.
S43, the output end of the convolutional neural network module is connected with the input end of the three-layer evolution gate control circulation unit network. A predictive model is constructed.
In the step S5, when the convolutional neural network using the prediction model extracts the features, the convolutional neural network module inputs the feature matrix X input The input multidimensional feature data in (1) is subjected to feature extraction, the dimension of an input layer of a convolutional neural network module is n multiplied by m multiplied by t, and the output is as follows:
Figure BDA0004085586050000144
wherein ,
Figure BDA0004085586050000145
representing the output of a convolutional neural network module, X i+n,j+m Representing input feature matrix X input The value of row n and column m, f cov (. Cndot.) represents the selection of an activation function, ω n,m Weights representing n rows and m columns of the convolution kernel, b n,m Is the convolution kernel deviation, k is the sliding window size.
In step S6, the gradient evolution computing framework includes three training modules, specifically: the system comprises a gradient descent training module with a driving quantity, an acceleration gradient descent training module and a self-adaptive moment estimation training module, wherein in the gradient descent training module with the driving quantity, the parameter updating mode is as follows:
v dw ＝βv dw +(I-β)dW
v db ＝βv db +(1-β)db
W＝W-αv dw ，b＝b-αv db
wherein, alpha is learning rate, beta controls exponential weighted average, W is weight, and b is deviation;
In the acceleration gradient descent training module, the parameter updating mode is as follows:
S dW ＝βS dW +(1-β)dW 2
S db ＝β＝S db +(1-β)db 2
Figure BDA0004085586050000151
in the adaptive moment estimation training module, the parameter updating mode is as follows:
Figure BDA0004085586050000152
Figure BDA0004085586050000153
Figure BDA0004085586050000154
wherein ,β1 Weighting the average parameter, beta, for the control index in the momentum gradient descent training module 2 For the control exponential weighted average parameter in the acceleration gradient descent training module, ε is to prevent the denominator from being 0.
When the prediction model is trained based on the gradient evolution calculation frame, three training modules in the gradient evolution calculation frame are adopted to respectively carry out iterative training in the training process, and the method specifically comprises the following steps:
s61, after generating offspring once each iteration, carrying out quality evaluation on each offspring, setting a scoring mechanism according to prediction precision, wherein the prediction precision score is S p The expression is:
S P ＝ν RMSE
wherein ,νRMSE Lower represents more excellent offspring, according to v RMSE The optimal evolutionary offspring are selected out,
Figure BDA0004085586050000155
Figure BDA0004085586050000156
the method comprises the steps of respectively representing individuals with optimal performance in three training modules under the current iteration times, j representing algebra of the current evolution, S, R, A respectively representing training modes corresponding to the three training modules, namely a gradient descent training mode, an acceleration gradient descent training mode and a self-adaptive moment estimation training mode of driving quantity;
s62, after the individuals with the optimal performance in the three training modules are obtained respectively, respectively calculating the weight coefficients corresponding to the individuals according to the prediction precision scores in S61, wherein the expressions are as follows:
Figure BDA0004085586050000161
Figure BDA0004085586050000162
Figure BDA0004085586050000163
wherein ,k1 The weights representing the gradient descent training mode with momentum,
Figure BDA0004085586050000164
prediction accuracy score k representing a gradient descent training mode with momentum 2 Weights representing the way of training the gradient descent acceleration, +.>
Figure BDA0004085586050000165
Predictive accuracy score, k, representing the manner of training for acceleration gradient descent 3 Weights representing the adaptive moment estimation training pattern, +.>
Figure BDA0004085586050000166
A prediction accuracy score representing the adaptive moment estimation training mode;
s63, weighting and summing the obtained weight coefficients to obtain a comprehensive gradient, and updating parameters according to an updating formula of the parameters w and b, wherein the expression is as follows:
Figure BDA0004085586050000167
Figure BDA0004085586050000168
Figure BDA0004085586050000169
Figure BDA00040855860500001610
in the formula ,
Figure BDA00040855860500001611
and
Figure BDA00040855860500001612
Representing the integrated gradient of the parameters w and b, v, respectively dw 、v db Respectively represent the gradients of current parameters w and b in a gradient descent training mode with momentum, S dW 、S db Respectively represent the gradients of the current parameters w and b in the acceleration gradient descent training mode, +.>
Figure BDA00040855860500001617
Respectively representing gradients of current parameters w and b in a self-adaptive moment estimation training mode;
s64, calculating weight coefficients of the corresponding training methods according to the prediction precision scores, giving larger weights to the offspring in the training modes with good performance to the training modes corresponding to the three training modules, simultaneously keeping the offspring information with good performance to the other training modes to correct the offspring, fusing the gradient information of the three optimal offspring parameters under the current iteration times, and generating comprehensive gradient sum through evolution
Figure BDA00040855860500001613
and
Figure BDA00040855860500001614
And then, according to an updating formula of the parameters w and b, carrying out parameter updating by utilizing the comprehensive gradient to obtain updated parameters under the current iteration times, and repeating the operation in the next iteration as an initial solution of the next iteration to finish the training of the prediction model.
The process adopts an evolutionary gating loop unit network to fuse three different model training methods to update parameters, the advantage complementation of an evolutionary computing framework is adopted to exchange and fuse gradient information of optimal offspring of the current iteration times, and the evolutionary computing framework generates comprehensive gradients
Figure BDA00040855860500001615
and
Figure BDA00040855860500001616
The method has the advantages that the parameters are guided to be updated, different training methods are fused through the evolutionary computing framework in each training, the problems that gradient disappearance, gradient explosion and local optimization possibly exist in a single training mode of the GRU network built by the neural network can be solved, and the method has a certain practical significance for improving the network prediction precision of the gating circulating unit.
During the training of the predictive model, the fitness function is defined as the root mean square error RMSE to reflect the error between the predicted and the actual values.
Finally, when the trained prediction model is used for predicting the wind power of the newly built wind power plant, the wind power related data of the newly built wind power plant in a period of time before prediction is used as input of the prediction model, training iteration under different training modes is repeated until the preset iteration times are reached, iteration is stopped, a wind power prediction result is output, fig. 5 is a prediction effect diagram of the wind power obtained by the wind power prediction method of the newly built wind power plant based on gradient evolution in the embodiment, wherein a 'dotted line' legend represents a wind power true value, a 'solid line' legend represents the wind power prediction value, and it can be seen that an error between the prediction value and the true value is small, so that the wind power prediction method provided by the embodiment is high in prediction precision.
Example 3
The embodiment provides a newly built wind power plant wind power prediction system based on gradient evolution, referring to fig. 6, the system includes:
the wind power data acquisition and preprocessing module is used for acquiring wind power related data of a target wind power plant and a wind power plant adjacent to the target wind power plant and processing the wind power related data;
the data expansion module is used for constructing a space-time diagram convolution generation countermeasure network of the adjacent wind power plant and the target wind power plant, inputting processed wind power related data into the space-time diagram convolution generation countermeasure network, and generating data to realize data expansion;
the data set decomposition module is used for decomposing the source data set by utilizing multi-element mode decomposition and forming an input feature matrix based on the decomposed data by taking the wind power related data of the target wind power plant before being processed and the generated data of the countermeasure network generated based on space-time diagram convolution as the source data set;
the prediction model construction module is used for constructing a prediction model comprising a convolutional neural network and an evolutionary gating circulating unit network;
the feature extraction and mining module is used for taking an input feature matrix as the input of a prediction model, extracting features by using a convolutional neural network of the prediction model, and mining implicit space-time relations among the features by using an evolutionary gating circulating unit network of the prediction model;
The prediction model training module is used for training a prediction model by utilizing data in an input feature matrix based on a gradient evolution calculation frame to obtain a trained prediction model for wind power prediction
It is to be understood that the above examples of the present invention are provided by way of illustration only and are not intended to limit the scope of the invention. Other variations or modifications of the above teachings will be apparent to those of ordinary skill in the art. It is not necessary here nor is it exhaustive of all embodiments. Any modification, equivalent replacement, improvement, etc. which come within the spirit and principles of the invention are desired to be protected by the following claims.

Claims (10)

1. The new wind power plant wind power prediction method based on gradient evolution is characterized by comprising the following steps of:
s1, acquiring wind power related data of a target wind power plant and a wind power plant adjacent to the target wind power plant, and processing the wind power related data;
s2, constructing a space-time diagram convolution generation countermeasure network of the adjacent wind power plant and the target wind power plant, and inputting the wind power related data processed in the S1 into the space-time diagram convolution generation countermeasure network to generate data to realize data expansion;
S3, wind power related data of an unprocessed target wind power plant and generated data based on a space-time diagram convolution generation countermeasure network are used as a source data set, the source data set is decomposed by utilizing multivariate mode decomposition, and an input feature matrix is formed based on the decomposed data;
s4, constructing a prediction model comprising a convolutional neural network and an evolution gating circulation unit network;
s5, taking the input feature matrix as the input of a prediction model, extracting features by using a convolutional neural network of the prediction model, and mining implicit space-time relations among the features by using an evolutionary gating circulating unit network of the prediction model;
s6, training a prediction model by utilizing data input into a feature matrix based on a gradient evolution calculation frame to obtain a trained prediction model for wind power prediction.
2. The method for predicting wind power of a newly built wind power plant based on gradient evolution according to claim 1, wherein the wind power related data in step S1 includes a wind power sequence, a wind speed sequence, a wind direction sequence and temperature data, and the processing process is as follows: abnormal value elimination and missing value filling are carried out on wind power sequence, wind speed sequence, wind direction sequence and temperature data, sine and cosine processing is adopted on the wind direction sequence, and wind direction sine sequence WD is obtained sin Cosine sequence WD of the sum and the wind direction cos 。
3. The method for predicting wind power of newly built wind power plant based on gradient evolution according to claim 2, wherein when constructing a space-time diagram convolution of an adjacent wind power plant and a target wind power plant to generate an countermeasure network, a multi-diagram convolution neural network is used as a generator of the space-time diagram convolution to generate the countermeasure network, wherein the generator comprises a similarity diagram GCN-SIM, a correlation diagram GCN-COR and a distance diagram GCN-DST of the target wind power plant and the adjacent wind power plant, the similarity diagram GCN-SIM, the correlation diagram GCN-COR and the distance diagram GCN-DST are overlapped to form a comprehensive diagram GCN-SYN, each diagram consists of wind power plant nodes v and edges epsilon, and is expressed as G= (v, epsilon), each node represents one adjacent wind power plant, each edge represents the correlation relation between each adjacent wind power plant and the target wind power plant, and wind power related data of each wind power plant after processing is obtained;
setting a similarity adjacent matrix corresponding to the similarity graph GCN-SIM as A sim The correlation adjacency matrix of the correlation graph (GCN-COR) is A cor The distance adjacency matrix of the distance graph (GCN-DST) is A dst Then the comprehensive adjacency matrix of the fusion graph (GCN-SYN) is A syn ＝A sim +A cor +A dst All matrices are normalized to (0, 1);
a layer of space-time convolution neural network is used as a discriminator for generating an countermeasure network by space-time diagram convolution, wind power related data processed by S1 is input into the space-time diagram convolution generation countermeasure network, and a distance adjacency matrix A is adopted dst And judging errors between the generated data and the original wind power related data as an adjacent matrix of the graph, guiding the generator to update iteratively, and generating high-quality data.
4. The method for predicting wind power of a newly built wind farm based on gradient evolution according to claim 1, wherein the process of decomposing the source data set by using multivariate model decomposition and forming the input feature matrix based on the decomposed data is as follows:
s31, set p (k) (t) represents a time sequence of wind power at t moment in a source data set, v m(k) (t)、v z(k) (t) represents the warp direction wind at the time t in the source data setSpeed time sequence and banded wind speed time sequence, then satisfy:
Figure FDA0004085586040000021
wherein x (t) comprises an intrinsic mode function u k Number n of subsequences of (t) imf ；
S32, decomposing the source data set by utilizing multi-element mode decomposition to obtain three groups of data respectively belonging to p (t) and v m (t)、v z Subsequence of (t)
Figure FDA0004085586040000022
The set of subsequences is used as one of the inputs to the prediction model;
s33, forming an input feature matrix X based on the decomposed data input The expression is:
Figure FDA0004085586040000023
wherein ,Pt-m 、
Figure FDA0004085586040000024
Tem t-m The power, radial wind speed, strip wind speed, wind direction sine, wind direction cosine and temperature of the wind farm at the moment t-m of the wind farm are respectively represented.
5. The method for predicting wind power of a newly built wind farm based on gradient evolution according to claim 4, wherein the process of constructing the prediction model including the convolutional neural network and the evolution-gated loop unit network in step S4 is as follows:
S41, constructing a convolutional neural network module to extract input features of an input feature matrix, wherein the convolutional neural network module comprises three layers of 2D convolutional layers and a full-connection layer, the convolutional kernels of the three layers of 2D convolutional layers are the same in size, the number of filters is 4, 8 and 16 respectively, the activating functions are ReLU functions, the filling modes are the same, the pooling is carried out in a maximum pooling mode, and the three layers of 2D convolutional layers are output by the full-connection layer;
s42, building a three-layer evolution gating circulating unit network, wherein the number of neurons of the three-layer evolution gating circulating unit network is 4, 8 and 16 respectively, the activation functions are all ReLU functions, the three-layer gating circulating unit network is connected with a full-connection layer, the full-connection layer activation function is the ReLU function, and the forward propagation formula of the three-layer evolution gating circulating unit network meets the following conditions:
Figure FDA0004085586040000031
wherein ,Wr 、W z 、W h 、U r 、U z 、U h As a weight parameter matrix, b r 、b z 、b h In order to bias the parameter matrix,
Figure FDA0004085586040000032
for matrix multiplication, σ is a Sigmod function, r t To reset the gate, z t For updating the door->
Figure FDA0004085586040000033
As candidate state of hidden layer at current moment, y t As the current implicit state, y t-1 Is the implicit state of the previous moment, x t Is the input state at the current time.
S43, the output end of the convolutional neural network module is connected with the input end of the three-layer evolution gate control circulation unit network. A predictive model is constructed.
6. The method for predicting wind power of newly built wind power plant based on gradient evolution according to claim 5, wherein in step S5, when the convolutional neural network using the prediction model extracts features, the convolutional neural network module inputs a feature matrix X input The input multidimensional feature data in (1) is subjected to feature extraction, the dimension of an input layer of a convolutional neural network module is n multiplied by m multiplied by t, and the output is as follows:
Figure FDA0004085586040000034
wherein ,
Figure FDA0004085586040000035
representing the output of a convolutional neural network module, X i+n,j+m Representing input feature matrix X input The value of row n and column m, f cov (. Cndot.) represents the selection of an activation function, ω n,m Weights representing n rows and m columns of the convolution kernel, b n,m Is the convolution kernel deviation, k is the sliding window size.
7. The method for predicting wind power of a newly built wind farm with gradient evolution according to claim 1, wherein in step S6, the gradient evolution calculation frame comprises three training modules, specifically: the system comprises a gradient descent training module with a driving quantity, an acceleration gradient descent training module and a self-adaptive moment estimation training module, wherein in the gradient descent training module with the driving quantity, the parameter updating mode is as follows:
v dw ＝βv dw +(1-β)dW
v db ＝βv db +(1-β)db
W＝W-αv dw ，b＝b-αv db
wherein, alpha is learning rate, beta controls exponential weighted average, W is weight, and b is deviation;
in the acceleration gradient descent training module, the parameter updating mode is as follows:
S dW ＝βS dW +(1-β)dW 2
S db ＝β＝S db +(1-β)db 2
Figure FDA0004085586040000041
In the adaptive moment estimation training module, the parameter updating mode is as follows:
Figure FDA0004085586040000042
Figure FDA0004085586040000043
Figure FDA0004085586040000044
wherein ,β1 Weighting the average parameter, beta, for the control index in the momentum gradient descent training module 2 For the control exponential weighted average parameter in the acceleration gradient descent training module, ε is to prevent the denominator from being 0.
8. The method for predicting wind power of a newly built wind power plant with gradient evolution according to claim 7, wherein when the prediction model is trained based on the gradient evolution calculation frame, three training modules in the gradient evolution calculation frame are adopted to respectively perform iterative training in the training process, and the method is specifically as follows:
s61, after generating offspring once each iteration, carrying out quality evaluation on each offspring, setting a scoring mechanism according to prediction precision, wherein the prediction precision score is S p The expression is:
S P ＝ν RMSE
wherein ,νRMSE Lower represents more excellent offspring, according to v RMSE The optimal evolutionary offspring are selected out,
Figure FDA0004085586040000045
Figure FDA0004085586040000046
respectively represent the individuals with optimal performance in the three training modules under the current iteration times, j represents the algebra of the current evolution and S, R, A minutesThe training modes corresponding to the three training modules are respectively represented as a gradient descent training mode, an acceleration gradient descent training mode and a self-adaptive moment estimation training mode of the driving quantity;
S62, after the individuals with the optimal performance in the three training modules are obtained respectively, respectively calculating the weight coefficients corresponding to the individuals according to the prediction precision scores in S61, wherein the expressions are as follows:
Figure FDA0004085586040000051
Figure FDA0004085586040000052
Figure FDA0004085586040000053
wherein ,k1 The weights representing the gradient descent training mode with momentum,
Figure FDA0004085586040000054
prediction accuracy score k representing a gradient descent training mode with momentum 2 Weights representing the way of training the gradient descent acceleration, +.>
Figure FDA0004085586040000055
Predictive accuracy score, k, representing the manner of training for acceleration gradient descent 3 Weights representing the adaptive moment estimation training pattern, +.>
Figure FDA0004085586040000056
A prediction accuracy score representing the adaptive moment estimation training mode;
s63, weighting and summing the obtained weight coefficients to obtain a comprehensive gradient, and updating parameters according to an updating formula of the parameters w and b, wherein the expression is as follows:
Figure FDA0004085586040000057
Figure FDA0004085586040000058
Figure FDA0004085586040000059
Figure FDA00040855860400000510
in the formula ,
Figure FDA00040855860400000511
and
Figure FDA00040855860400000512
Representing the integrated gradient of the parameters w and b, v, respectively dw 、v db Respectively represent the gradients of current parameters w and b in a gradient descent training mode with momentum, S dW 、S db Respectively represent the gradients of the current parameters w and b in the acceleration gradient descent training mode, +.>
Figure FDA00040855860400000513
Respectively representing gradients of current parameters w and b in a self-adaptive moment estimation training mode;
s64, calculating weight coefficients of the corresponding training methods according to the prediction precision scores, giving larger weights to the offspring in the training modes with good performance to the training modes corresponding to the three training modules, simultaneously keeping the offspring information with good performance to the other training modes to correct the offspring, fusing the gradient information of the three optimal offspring parameters under the current iteration times, and generating comprehensive gradient sum through evolution
Figure FDA00040855860400000514
and
Figure FDA00040855860400000515
And then, according to an updating formula of the parameters w and b, carrying out parameter updating by utilizing the comprehensive gradient to obtain updated parameters under the current iteration times, and repeating the operation in the next iteration as an initial solution of the next iteration to finish the training of the prediction model.
9. The gradient evolution wind power prediction method for the newly built wind power plant according to claim 8, wherein in the process of training a prediction model, a fitness function is defined as a root mean square error RMSE so as to reflect an error between a predicted value and a true value, and when the trained prediction model is used for predicting the wind power of the newly built wind power plant, wind power related data of the newly built wind power plant in a period of time before prediction is used as input of the prediction model to output a wind power prediction result.
10. A newly-built wind power plant wind power prediction system based on gradient evolution, which is characterized by comprising:
the wind power data acquisition and preprocessing module is used for acquiring wind power related data of a target wind power plant and a wind power plant adjacent to the target wind power plant and processing the wind power related data;
the data expansion module is used for constructing a space-time diagram convolution generation countermeasure network of the adjacent wind power plant and the target wind power plant, inputting processed wind power related data into the space-time diagram convolution generation countermeasure network, and generating data to realize data expansion;
The data set decomposition module is used for decomposing the source data set by utilizing multi-element mode decomposition and forming an input feature matrix based on the decomposed data by taking the wind power related data of the target wind power plant before being processed and the generated data of the countermeasure network generated based on space-time diagram convolution as the source data set;
the prediction model construction module is used for constructing a prediction model comprising a convolutional neural network and an evolutionary gating circulating unit network;
the feature extraction and mining module is used for taking an input feature matrix as the input of a prediction model, extracting features by using a convolutional neural network of the prediction model, and mining implicit space-time relations among the features by using an evolutionary gating circulating unit network of the prediction model;
and the prediction model training module is used for training a prediction model by utilizing data in the input feature matrix based on the gradient evolution calculation frame to obtain a trained prediction model for wind power prediction.
CN202310101458.5A 2023-02-08 2023-02-08 New wind power plant wind power prediction method and system based on gradient evolution Pending CN116070763A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310101458.5A CN116070763A (en) 2023-02-08 2023-02-08 New wind power plant wind power prediction method and system based on gradient evolution

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310101458.5A CN116070763A (en) 2023-02-08 2023-02-08 New wind power plant wind power prediction method and system based on gradient evolution

Publications (1)

Publication Number Publication Date
CN116070763A true CN116070763A (en) 2023-05-05

Family

ID=86172918

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310101458.5A Pending CN116070763A (en) 2023-02-08 2023-02-08 New wind power plant wind power prediction method and system based on gradient evolution

Country Status (1)

Country Link
CN (1) CN116070763A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117933362A (en) * 2024-03-06 2024-04-26 广东工业大学 Short-term wind power prediction method and system for newly-built wind power plant

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117933362A (en) * 2024-03-06 2024-04-26 广东工业大学 Short-term wind power prediction method and system for newly-built wind power plant

Similar Documents

Publication Publication Date Title
Meng et al. Multi-fidelity Bayesian neural networks: Algorithms and applications
Zhang et al. Multi-head attention-based probabilistic CNN-BiLSTM for day-ahead wind speed forecasting
Li et al. Learning the aerodynamic design of supercritical airfoils through deep reinforcement learning
Jalali et al. An advanced short-term wind power forecasting framework based on the optimized deep neural network models
JP5888640B2 (en) Photovoltaic power generation prediction apparatus, solar power generation prediction method, and solar power generation prediction program
Ryberg et al. Metamodel-based multidisciplinary design optimization for automotive applications
CN111696345A (en) Intelligent coupled large-scale data flow width learning rapid prediction algorithm based on network community detection and GCN
CN111815033A (en) Offshore wind power prediction method based on RCNN and meteorological time sequence characteristics
CN108985515B (en) New energy output prediction method and system based on independent cyclic neural network
CN111292525A (en) Traffic flow prediction method based on neural network
Gad et al. A robust deep learning model for missing value imputation in big NCDC dataset
CN113190688A (en) Complex network link prediction method and system based on logical reasoning and graph convolution
CN112508244B (en) Multi-element load prediction method for user-level comprehensive energy system
CN115018193A (en) Time series wind energy data prediction method based on LSTM-GA model
CN112766603A (en) Traffic flow prediction method, system, computer device and storage medium
CN116070763A (en) New wind power plant wind power prediction method and system based on gradient evolution
CN112241802A (en) Interval prediction method for wind power
CN114139783A (en) Wind power short-term power prediction method and device based on nonlinear weighted combination
CN115713144A (en) Short-term wind speed multi-step prediction method based on combined CGRU model
CN115545159A (en) Average sea surface temperature forecasting method of deep neural network
CN115034432A (en) Wind speed prediction method for wind generating set of wind power plant
CN114742325A (en) Method and system for predicting land surface settlement during subway tunnel step method construction
CN114862015A (en) Typhoon wind speed intelligent prediction method based on FEN-ConvLSTM model
CN117636183A (en) Small sample remote sensing image classification method based on self-supervision pre-training
Smith et al. Multi-objective evolutionary recurrent neural network ensemble for prediction of computational fluid dynamic simulations

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