CN114511158B - Wind turbine power prediction method based on wake deflection effect and 2DJensen model - Google Patents

Wind turbine power prediction method based on wake deflection effect and 2DJensen model Download PDF

Info

Publication number
CN114511158B
CN114511158B CN202210412243.0A CN202210412243A CN114511158B CN 114511158 B CN114511158 B CN 114511158B CN 202210412243 A CN202210412243 A CN 202210412243A CN 114511158 B CN114511158 B CN 114511158B
Authority
CN
China
Prior art keywords
wind
wind turbine
wake
node
edge
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
CN202210412243.0A
Other languages
Chinese (zh)
Other versions
CN114511158A (en
Inventor
邱颖宁
柳靖
冯延晖
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.)
Nanjing University of Science and Technology
Original Assignee
Nanjing University of Science and 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 Nanjing University of Science and Technology filed Critical Nanjing University of Science and Technology
Priority to CN202210412243.0A priority Critical patent/CN114511158B/en
Publication of CN114511158A publication Critical patent/CN114511158A/en
Application granted granted Critical
Publication of CN114511158B publication Critical patent/CN114511158B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

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
    • G06N3/04—Architecture, e.g. interconnection topology
    • 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
    • H—ELECTRICITY
    • H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
    • H02J—CIRCUIT ARRANGEMENTS OR SYSTEMS FOR SUPPLYING OR DISTRIBUTING ELECTRIC POWER; SYSTEMS FOR STORING ELECTRIC ENERGY
    • H02J3/00—Circuit arrangements for ac mains or ac distribution networks
    • H02J3/003—Load forecast, e.g. methods or systems for forecasting future load demand
    • 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)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Health & Medical Sciences (AREA)
  • Economics (AREA)
  • General Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Strategic Management (AREA)
  • Human Resources & Organizations (AREA)
  • Software Systems (AREA)
  • Marketing (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Mathematical Physics (AREA)
  • Data Mining & Analysis (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • General Business, Economics & Management (AREA)
  • Tourism & Hospitality (AREA)
  • Artificial Intelligence (AREA)
  • Molecular Biology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Primary Health Care (AREA)
  • Water Supply & Treatment (AREA)
  • Public Health (AREA)
  • Development Economics (AREA)
  • Game Theory and Decision Science (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Operations Research (AREA)
  • Quality & Reliability (AREA)
  • Power Engineering (AREA)
  • Wind Motors (AREA)

Abstract

The invention discloses a wind turbine power prediction method based on wake deflection effect and a 2DJensen model, which collects and preprocesses SCADA data of a wind turbine in a wind field; constructing a relative position coordinate of each wind turbine based on the geographical position data of the wind turbines in the wind field; determining a node set and global attributes of a wind turbine in a wind field; determining an edge set and an adjacent matrix between wind machines in a wind field; constructing an input vector atlas; constructing an improved graph neural network model based on a wake deflection effect and a 2DJensen model; and performing parameter optimization according to the preprocessed SCADA data set and the input vector diagram, determining a wind turbine power prediction model in the wind field, and determining the predicted power of all wind turbine nodes in the wind field. The method can accurately predict the power output value of the wind turbine generator with a yaw error angle in the whole wind field under different wind conditions.

Description

Wind turbine power prediction method based on wake deflection effect and 2DJensen model
Technical Field
The invention belongs to the field of wind turbine generator power prediction, and particularly relates to a wind turbine power prediction method based on a wake deflection effect and a 2DJensen model.
Background
China is developing new energy industry vigorously, and wind energy becomes one of new energy with a great development prospect due to the advantages of wide distribution, no pollution and the like. With the continuous development of wind power generation technology, more and more large wind power plants appear. Due to factors such as cost and site, the large wind power plant has the conditions of mutual influence among wind turbines, complex and changeable internal wind field and the like. The power loss condition of the wind power plant can be analyzed through the predicted power, and the generating efficiency of the wind power plant can be effectively improved through the combination of the predicted power and a plant control system. Therefore, the method has important significance for accurately predicting the power of all wind turbines of the wind power plant through the wind condition information of the wind power plant. Generally, the wind power plant has abundant SCADA historical data, so that the improvement of the wind power plant power prediction accuracy through the historical data is also of great significance.
Disclosure of Invention
The invention aims to provide a wind turbine power prediction method based on a wake deflection effect and a 2DJensen model.
The technical solution for realizing the purpose of the invention is as follows: a wind turbine power prediction method based on wake deflection effect and a 2DJensen model comprises the following specific steps:
step 1, collecting and preprocessing SCADA data of a wind turbine in a wind field;
2, constructing a relative position coordinate of each wind turbine based on the geographical position data of the wind turbines in the wind field;
step 3, obtaining a node set and global attributes of the wind turbine in the wind field according to the preprocessed SCADA data set;
step 4, obtaining an edge set and an adjacency matrix between the wind machines in the wind field according to the relative position coordinates of the wind machines, the input wind direction at the current moment, the wake deflection effect and the 2DJensen model;
step 5, constructing an input vector graph set based on a node set, an edge set, global attributes and an adjacency matrix of a wind turbine in a wind field;
step 6, constructing an improved graph neural network model based on the wake deflection effect and the 2DJensen model, and determining the setting parameters of the graph neural network;
step 7, performing parameter optimization on the improved graph neural network according to the preprocessed SCADA data set and the input vector diagram by using an optimization algorithm and according to a loss function, and determining a graph neural network model after the parameter optimization is completed to serve as a wind turbine power prediction model in a wind field;
and 8, processing the input wind speed and wind direction by using a wind turbine power prediction model to obtain the predicted power of all wind turbine nodes in the wind field.
A wind turbine power prediction system based on a wake deflection effect and a 2DJensen model realizes wind turbine power prediction by utilizing the wind turbine power prediction method based on the wake deflection effect and the 2DJensen model.
A computer device comprises a storage, a processor and a computer program stored on the storage and capable of running on the processor, wherein when the processor executes the computer program, the wind turbine power prediction method based on the wake deflection effect and the 2DJensen model is utilized to realize wind turbine power prediction.
A computer-readable storage medium, on which a computer program is stored, which, when being executed by a processor, implements a wind turbine power prediction using the method for wind turbine power prediction based on wake deflection effect and 2DJensen model.
Compared with the prior art, the invention has the remarkable advantages that: the power of the wind turbine is predicted by combining the graph neural network, the wake deflection effect and the 2DJensen model, and the power output value of the wind turbine with a yaw error angle in the whole wind field under different wind conditions can be accurately predicted.
Drawings
FIG. 1 is a flow chart of a wind turbine power prediction method based on a wake deflection effect and a 2DJensen model.
FIG. 2 is a flow chart of the neural network construction of the present invention.
Fig. 3 is a schematic structural diagram of a network layer of an improved graph based on a wake deflection effect and a 2DJensen model.
FIG. 4 is a schematic diagram of a general architecture of a neural network layer according to the present invention.
FIG. 5 is a comparison graph of the predicted power value of the wind farm wind turbine generator, the predicted power value of the 2DJensen model and the actual power output value.
Detailed Description
In order to make the objects, technical solutions and advantages of the present application more apparent, the present application is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present application and are not intended to limit the present application.
As shown in FIG. 1, the wind turbine power prediction method based on the wake deflection effect and the 2DJensen model specifically comprises the following steps:
step 1, collecting and preprocessing data of a wind turbine in a wind field;
the method comprises the steps of collecting inflow wind speed, inflow wind direction, pitch angle, yaw error angle and output power of each wind turbine in an SCADA system of the wind turbine. The collected data are preprocessed, all data of the moment of the missing value in the data are removed, all data of the moment of the missing value in the data under the condition of power limitation are removed, all data of the moment of the collection error are removed if the wind speed is less than 0 or the active power is less than 0, and after the preprocessing, it is guaranteed that all wind turbines in the data set at the same moment need four characteristics of inflow wind speed, inflow wind direction, yaw error angle and output power.
The electricity limiting condition refers to the condition that the wind turbine does not achieve the advanced pitch control when the rated wind speed is reached yet, and aims to abandon part of captured wind energy to limit the power generation power, so that the data is removed only by judging whether the pitch angle is changed when the inflow wind speed reaches the rated wind speed.
2, constructing a relative position coordinate of each wind turbine based on the geographical position data of the wind turbines in the wind field;
and determining the position information of all wind turbines in the wind field according to the construction site selection of the wind field, and solving the relative position coordinates of each wind turbine in the wind field by taking a certain point in the wind field as a coordinate origin according to the relative azimuth angle and the relative distance between the wind turbines.
Step 3, obtaining a node set and global attributes of the wind turbine in the wind field according to the preprocessed data set;
the method comprises the steps of averaging the inflow wind directions of all wind turbines at a certain moment, judging that the wind flows through a first wind turbine in a wind field at the moment by using the obtained wind directions, and taking the wind speed and the wind direction of the wind turbine at the moment as global attributes. The node attributes in all wind turbines are set to be the wind speeds in the global attribute, and the node set in the wind turbine at the moment can be obtained.
Step 4, obtaining edge attributes and an adjacency matrix between the wind machines in the wind field according to the relative position coordinates of the wind machines, the input wind direction at the current moment, the wake deflection effect and the 2DJensen model;
based on the inflow wind direction in the global attribute at the moment and the yaw error angle between every two wind turbines, whether wake influence exists between every two wind turbines can be judged according to the wake deflection effect and the 2DJensen model, if so, a directed edge is established between the two wind turbines, the front wind turbine is a sending node, the back wind turbine is a receiving node, and the attribute of the edge is set as the transverse distance and the longitudinal distance of the back wind turbine in the wake deflected by the front wind turbine; if not, there is no edge between the two wind turbines. After traversing the relationship among all wind turbines in the wind field, the edge set and the adjacency matrix in the whole wind field can be obtained.
And when judging whether wake influence exists between every two wind turbines according to the wake deflection effect and the 2DJensen model, obtaining the relative positions and the azimuth angles of the two wind turbines according to the coordinates of the two wind turbines, obtaining the sequence of the wind passing through the wind turbines according to the inflow wind direction in the global attribute at the moment, and respectively defining the sequence as an upstream wind turbine and a downstream wind turbine. The specific method for judging whether the downstream wind turbine is influenced by the wake flow deflected by the upstream wind turbine and solving the influenced wake flow transverse distance and longitudinal distance under the condition of considering the wake flow deflection comprises the following steps:
defining an included angle theta between the connecting line of the upstream wind turbine generator and the downstream wind turbine generator and the central axis of the wake flow of the upstream wind turbine generatorLComprises the following steps:
Figure 606289DEST_PATH_IMAGE001
in the formula (I), the compound is shown in the specification,θ x in order to obtain an inflow wind direction angle,θ FWT the azimuth angle of the upstream unit relative to the downstream unit,β 1is the yaw error angle of the upstream wind turbine,ais a wind turbine axial induction factor;
the longitudinal distance between the section of the downstream wind turbine generator and the wake flow of the upstream wind turbine generator is defined asL l The calculation formula is：
Figure 54588DEST_PATH_IMAGE002
In the formula (I), the compound is shown in the specification,Ldthe distance between the cabin connecting lines of the wind turbine generator sets upstream and downstream is the distance between the cabin connecting lines of the wind turbine generator sets downstream and upstream;
and simultaneously defining the cross section of the downstream wind turbine generator and the wake flow transverse distance of the upstream wind turbine generator asL r The calculation formula is as follows:
Figure 99904DEST_PATH_IMAGE003
in the formula (I), the compound is shown in the specification,Ldthe distance between the cabin connecting lines of the wind turbine generator sets upstream and downstream is the distance between the cabin connecting lines of the wind turbine generator sets downstream and upstream;
the wake flow radius calculation formula of the cross section position where the downstream wind turbine generator is located is as follows:
Figure 760693DEST_PATH_IMAGE004
in the formula (I), the compound is shown in the specification,kas a function of the attenuation coefficient of the wake flow,r 0 is the radius of the wind wheel of the upstream wind turbine.
According to the calculated wake radiusR l Whether the downstream wind turbine is affected by the wake flow of the upstream wind turbine or not can be judged, and the specific method comprises the following steps:
1) if it is
Figure 355491DEST_PATH_IMAGE005
The downstream wind turbine is not influenced by the wake flow of the upstream wind turbine;
2) if it is
Figure 443533DEST_PATH_IMAGE006
The downstream wind turbine is influenced by the wake flow of the upstream wind turbine;
wherein the content of the first and second substances,r 0 ’ for wind-driven machinejThe projection length of the radius of the wind wheel is calculated by the following formula:
Figure 241725DEST_PATH_IMAGE007
in the formula (I), the compound is shown in the specification,β 2 is the yaw error angle of the downstream wind turbine.
Step 5, constructing an input vector graph set based on a node set, an edge set, global attributes and an adjacency matrix of a wind turbine in a wind field;
step 6, constructing an improved graph neural network model based on the wake deflection effect and the 2DJensen model, and determining the setting parameters of the graph neural network;
as shown in fig. 2, an improved graph neural network architecture based on wake deflection effect and a 2DJensen model is established, and a proper graph neural network layer number j is set according to the size of a wind field, and the j is suggested to be 2-4 layers. The first layer is an improved graph network layer based on a wake deflection effect and a 2DJensen model, when the graph neural network output graph is finally obtained, all wind turbine node attributes are extracted and input into the full connection layer, and the predicted power values of all wind turbines are obtained through output.
As shown in fig. 3, when an improved graph network layer based on the wake deflection effect and the 2DJensen model is established, the graph network layer includes three update functions, which are edge update functions respectively
Figure 706204DEST_PATH_IMAGE008
Updating function of node
Figure 640662DEST_PATH_IMAGE009
Global update function
Figure 915917DEST_PATH_IMAGE010
(ii) a Simultaneously comprises three aggregation functions which are adjacent edge aggregation functions
Figure 935825DEST_PATH_IMAGE011
Aggregation of all edges
Figure 938416DEST_PATH_IMAGE012
Aggregation of functions by all nodes
Figure 992960DEST_PATH_IMAGE013
(ii) a Finally, an edge weight distribution function based on a 2DJensen model wake flow loss factor is included
Figure 688383DEST_PATH_IMAGE014
. The input diagram is
Figure 461167DEST_PATH_IMAGE015
In which all input edges are included
Figure 251138DEST_PATH_IMAGE016
All input nodes
Figure 160188DEST_PATH_IMAGE017
And inputting the global attribute
Figure 26513DEST_PATH_IMAGE018
. In the improved graph network layer based on the wake deflection effect and the 2DJensen model, the updating sequence is as follows:
the updating of the edge is firstly carried out, and the updating of each directed edge needs to use the attribute of the adjacent sending node
Figure 21014DEST_PATH_IMAGE019
And attributes of the receiving node
Figure 630986DEST_PATH_IMAGE020
And the attribute of the edge
Figure 394543DEST_PATH_IMAGE021
Inputting all attributes to the edge update function simultaneously
Figure 182502DEST_PATH_IMAGE022
In, the edge update function is selected as a neural network
Figure 664299DEST_PATH_IMAGE023
Updated edge attributes
Figure 812383DEST_PATH_IMAGE024
The following:
Figure 430446DEST_PATH_IMAGE025
then, the function is distributed according to the edge weight
Figure 904153DEST_PATH_IMAGE026
And obtaining the due weight of each directed edge, wherein the weight enables each directed edge to play a role in different degrees in the subsequent node updating according to the rules of the physical wake model, and the weight value is as follows:
Figure 607667DEST_PATH_IMAGE027
finally, the final edge attribute after each directed edge is updated can be obtained
Figure 543131DEST_PATH_IMAGE028
The following were used:
Figure 546859DEST_PATH_IMAGE029
after all existing edges in the graph neural network layer are updated, the node attributes are updated. Each node
Figure 925887DEST_PATH_IMAGE030
All the adjacent edges are needed for updating
Figure 116697DEST_PATH_IMAGE031
Property of the node
Figure 872164DEST_PATH_IMAGE032
And global properties of the entire wind farm
Figure 215552DEST_PATH_IMAGE033
. First using neighboring edge aggregation function
Figure 765482DEST_PATH_IMAGE034
Aggregating the information of the adjacent S edges which have the connection relation with the node, wherein the aggregation function takes into account the wake superposition effect in the improved graph network layer based on the wake deflection effect and the 2DJensen model, and a summation mode is adopted to sum the attributes of all the adjacent edges to obtain the aggregated edge attribute, and the values of the aggregated edge attribute are as follows:
Figure 974746DEST_PATH_IMAGE035
then the aggregated side information
Figure 268324DEST_PATH_IMAGE036
Property of the local node
Figure 715486DEST_PATH_IMAGE037
And global properties of the entire wind farm
Figure 685585DEST_PATH_IMAGE038
Simultaneous input to node update function
Figure 850987DEST_PATH_IMAGE039
In (1), the node update function is selected as a neural network
Figure 948256DEST_PATH_IMAGE040
Updated node
Figure 515504DEST_PATH_IMAGE041
The following were used:
Figure 141657DEST_PATH_IMAGE042
all present in the neural network layer of the mapAfter the node is updated, the global attribute is finally updated. Global attribute updates require all side information, attributes of all nodes, and global attributes of the entire wind farm. First using all edge aggregation functions
Figure 794355DEST_PATH_IMAGE043
Aggregating the information of all edges, wherein the aggregation function usually adopts a summation mode, and sums the attributes of all edges to obtain the aggregated edge attribute, and the value of the aggregated edge attribute is as follows:
Figure 446048DEST_PATH_IMAGE044
then aggregate the functions using all nodes
Figure 602222DEST_PATH_IMAGE045
Aggregating the information of all nodes, wherein the aggregation function usually adopts a summation mode, and sums the attributes of all nodes to obtain the aggregated node attribute, and the values of the aggregated node attribute are as follows:
Figure 664856DEST_PATH_IMAGE046
finally, aggregated side information is obtained
Figure 804851DEST_PATH_IMAGE047
Aggregated node attributes
Figure 243922DEST_PATH_IMAGE048
And global properties of the entire wind farm
Figure 785762DEST_PATH_IMAGE049
Input to global update function at the same time
Figure 268565DEST_PATH_IMAGE050
In general, the node update function is selected as a neural network
Figure 895855DEST_PATH_IMAGE051
Updated global attributes
Figure 873039DEST_PATH_IMAGE052
The following were used:
Figure 3806DEST_PATH_IMAGE053
in a graph neural network layer, the global attribute is updated only once, and the output of the graph neural network layer is obtained after the update
Figure 673821DEST_PATH_IMAGE054
、
Figure 522829DEST_PATH_IMAGE055
And
Figure 788856DEST_PATH_IMAGE056
the updating and aggregation operations of the graph neural network layer do not change the connectivity of all nodes, i.e. the adjacency matrix does not change, and then outputs it
Figure 305288DEST_PATH_IMAGE054
、
Figure 880626DEST_PATH_IMAGE055
、
Figure 216929DEST_PATH_IMAGE056
And the adjacency matrix is used as an input attribute and is input into the next graph network layer.
In an improved graph network layer based on a wake deflection effect and a 2DJensen model, a 2DJensen model wake loss factor is used as an edge weight distribution function for distributing weights to edges
Figure 801494DEST_PATH_IMAGE057
. In the updating of the nodes, the wake deflection effect and the edge weight of the 2DJensen model are considered, and pairwise wind power is reflected more accuratelyPhysical connection between machines and wake impact level. Edge weight distribution function
Figure 906854DEST_PATH_IMAGE057
The principle is as follows:
first the 2DJensen model is as follows:
Figure 914079DEST_PATH_IMAGE058
in the formula (I), the compound is shown in the specification,
Figure 737679DEST_PATH_IMAGE059
for a longitudinal wake distance in the downstream wake of
Figure 860355DEST_PATH_IMAGE060
The wake transverse distance is
Figure 85800DEST_PATH_IMAGE061
The velocity of the wake flow at which the wind is coming,
Figure 2941DEST_PATH_IMAGE062
as a function of the attenuation coefficient of the wake flow,
Figure 313836DEST_PATH_IMAGE063
is the radius of the rotor of the upstream wind turbine,
Figure 990937DEST_PATH_IMAGE064
is an axial induction factor.
Defining wake deficit factor therein
Figure 805309DEST_PATH_IMAGE065
Comprises the following steps:
Figure 424509DEST_PATH_IMAGE066
the value of which reflects the longitudinal distance of the wake in the downstream wake of the flow
Figure 957121DEST_PATH_IMAGE060
The wake transverse distance is
Figure 687180DEST_PATH_IMAGE061
Wind speed deficiency degree of place, wake flow deficiency factor between wind turbines far apart
Figure 605326DEST_PATH_IMAGE065
Wake loss factor between small, closely spaced wind turbines
Figure 129849DEST_PATH_IMAGE065
Large, the definition can fully account for different wake influence degrees, so it is defined as the edge weight distribution function
Figure 149757DEST_PATH_IMAGE057
The wake flow loss factor among different wind turbines
Figure 417928DEST_PATH_IMAGE065
As an edge weight assignment function.
And the attribute of the directed edge between every two wind turbines is defined as:
Figure 206892DEST_PATH_IMAGE067
in the formula (I), the compound is shown in the specification,L l for a receiving node
Figure 902315DEST_PATH_IMAGE068
I.e. downstream wind turbine at the transmitting node
Figure 160253DEST_PATH_IMAGE069
Namely the longitudinal wake distance in the wake after deflection of the upstream wind turbine,L r for a receiving node
Figure 966535DEST_PATH_IMAGE070
I.e. downstream wind turbine at the transmitting node
Figure 875585DEST_PATH_IMAGE071
I.e. the lateral wake distance in the wake after deflection of the upstream wind turbine.
Edge-of-event weight distribution function
Figure 741910DEST_PATH_IMAGE057
The definition is as follows:
Figure 736410DEST_PATH_IMAGE072
as shown in fig. 4, when the subsequent j-th graph neural network layer connected to the first graph neural network layer is established, the graph network layer includes three update functions, which are edge update functions respectively
Figure 346383DEST_PATH_IMAGE073
Node update function
Figure 359208DEST_PATH_IMAGE074
Global update function
Figure 396434DEST_PATH_IMAGE075
(ii) a Simultaneously comprises three aggregation functions which are adjacent edge aggregation functions
Figure 878231DEST_PATH_IMAGE076
All edge aggregation function
Figure 26315DEST_PATH_IMAGE077
Aggregating functions of all nodes
Figure 909958DEST_PATH_IMAGE078
. The input diagram of the layer is
Figure 118085DEST_PATH_IMAGE079
Including all the j-1 layer output edges
Figure 837910DEST_PATH_IMAGE080
All j-1 level input nodes
Figure 524107DEST_PATH_IMAGE081
And a j-1 level output global property
Figure 262256DEST_PATH_IMAGE082
. In the graph neural network layer at layer j, the update sequence is as follows:
the updating of the edges is firstly carried out, and the updating of each directed edge needs to use the attribute of the adjacent sending node
Figure 641284DEST_PATH_IMAGE083
And attributes of the receiving node
Figure 832094DEST_PATH_IMAGE084
And the attribute of the edge
Figure 836828DEST_PATH_IMAGE085
Simultaneously inputting all attributes into the edge update function
Figure 429484DEST_PATH_IMAGE086
In, the edge update function is selected as a neural network
Figure 244993DEST_PATH_IMAGE087
Updated edge attributes
Figure 923099DEST_PATH_IMAGE088
The following:
Figure 951098DEST_PATH_IMAGE089
after all existing edges in the graph neural network layer are updated, the node attributes are updated. Each node
Figure 663839DEST_PATH_IMAGE081
All the adjacent edges are needed for updating
Figure 135403DEST_PATH_IMAGE085
Property of the node
Figure 566384DEST_PATH_IMAGE081
And global properties of the entire wind farm
Figure 398074DEST_PATH_IMAGE082
. First using neighboring edge aggregation function
Figure 699742DEST_PATH_IMAGE076
Aggregating the information of the adjacent S edges having connection relation with the node, wherein the aggregation function generally adopts an averaging method, and averages the attributes of all the adjacent edges to obtain the aggregated edge attribute, and the values of the aggregated edge attribute are as follows:
Figure 591475DEST_PATH_IMAGE090
then the aggregated side information is processed
Figure 509752DEST_PATH_IMAGE091
Property of the local node
Figure 394400DEST_PATH_IMAGE092
And global properties of the entire wind farm
Figure 816154DEST_PATH_IMAGE093
Simultaneous input to node update function
Figure 878788DEST_PATH_IMAGE094
In (1), the node update function is selected as a neural network
Figure 18783DEST_PATH_IMAGE095
Updated node
Figure 457854DEST_PATH_IMAGE096
The following were used:
Figure 734115DEST_PATH_IMAGE097
after all existing nodes in the graph neural network layer are updated, the global attributes are finally updated. Global attribute updates require all side information, attributes of all nodes, and global attributes of the entire wind farm. First using all edge aggregation functions
Figure 983962DEST_PATH_IMAGE098
The information of all edges is aggregated, the aggregation function usually adopts an averaging mode, and the attributes of all edges are averaged to obtain the aggregated edge attribute, and the value of the aggregated edge attribute is as follows:
Figure 611252DEST_PATH_IMAGE099
then aggregate the functions using all nodes
Figure 322856DEST_PATH_IMAGE100
The information of all nodes is aggregated, the aggregation function usually adopts an averaging mode, and the attributes of all nodes are summed to obtain the aggregated node attribute, and the value is as follows:
Figure 984782DEST_PATH_IMAGE101
finally, the aggregated side information
Figure 389218DEST_PATH_IMAGE102
Aggregated node attributes
Figure 753072DEST_PATH_IMAGE103
And global properties of the entire wind farm
Figure 268367DEST_PATH_IMAGE104
Input to global simultaneouslyUpdating functions
Figure 519220DEST_PATH_IMAGE105
In (1), the node update function is usually selected as a neural network
Figure 360137DEST_PATH_IMAGE106
Updated global attributes
Figure 696441DEST_PATH_IMAGE107
The following were used:
Figure 31738DEST_PATH_IMAGE108
in a graph neural network layer, the global attribute is updated only once, and the output of the graph neural network layer is obtained after the update
Figure 137097DEST_PATH_IMAGE109
、
Figure 148916DEST_PATH_IMAGE110
And
Figure 706936DEST_PATH_IMAGE111
the updating and aggregation operations of the graph neural network layer do not change the connectivity of all nodes, i.e. the adjacency matrix does not change, and then outputs it
Figure 829613DEST_PATH_IMAGE109
、
Figure 55058DEST_PATH_IMAGE110
、
Figure 221466DEST_PATH_IMAGE111
And the adjacency matrix is used as an input attribute and is input into the next graph network layer.
In input vector diagrams
Figure 532361DEST_PATH_IMAGE112
、
Figure 458729DEST_PATH_IMAGE113
、
Figure 273101DEST_PATH_IMAGE114
Obtaining an output vector diagram after the whole diagram neural network architecture
Figure 892301DEST_PATH_IMAGE115
In the power prediction of each wind turbine, only the wind turbine nodes are needed
Figure 424914DEST_PATH_IMAGE116
So that a full connection layer is established
Figure 374547DEST_PATH_IMAGE117
The node attribute in the final graph neural network output vector graph
Figure 574584DEST_PATH_IMAGE118
Input to the full connection layer
Figure 99106DEST_PATH_IMAGE119
And obtaining the predicted power of each wind turbine, wherein the predicted power is as follows:
Figure 853435DEST_PATH_IMAGE120
step 7, performing parameter optimization on the improved graph neural network according to the preprocessed data set and the input vector diagram by using an optimization algorithm and according to a loss function, and taking a graph neural network model with determined parameters after the optimization as a wind turbine power prediction model in a wind field;
in the process of training the neural network of the graph, firstly setting a loss function as a predicted power and a true valueMSEThe values are as follows:
Figure 387185DEST_PATH_IMAGE121
in the formula (I), the compound is shown in the specification,
Figure 690996DEST_PATH_IMAGE122
is composed ofkInputting a predicted value of the graph neural network power of the ith wind turbine in the wind field corresponding to the vector graph at a moment,
Figure 120840DEST_PATH_IMAGE123
is composed ofkInputting the actual power value of the ith wind turbine in the wind field corresponding to the vector diagram at a moment,sto be the total number of input data,mthe total number of wind turbines in the wind farm.
After the loss function is set, a proper optimization algorithm is selected, a gradient descent optimization algorithm such as Adam, SGD or RMSprop is suggested to be used, a proper learning rate is set, the graph neural network power prediction model can be optimized, when the accuracy of the model does not change any more, optimization is completed, and the graph neural network model with parameters determined after optimization is used as the wind turbine power prediction model in the wind field.
Step 8, processing the input wind speed and wind direction by using a wind turbine power prediction model to obtain the predicted power of all wind turbine nodes in a wind field;
and when the power of each wind turbine in the wind field corresponding to a certain wind speed and wind direction is predicted, the input wind speed and wind direction are taken as global attributes, the wind speed in the global attributes is set as the node attributes of all the wind turbines, a corresponding input vector diagram is generated according to the steps 3-5, and the optimized model is called to process the input vector diagram, so that the power output value of each wind turbine in the wind field can be obtained.
The invention further provides a wind turbine power prediction system of the graph neural network based on the wake deflection effect and the 2DJensen model improvement, and the wind turbine power prediction is realized by using the wind turbine power prediction method of the graph neural network based on the wake deflection effect and the 2DJensen model improvement.
A computer device comprises a storage, a processor and a computer program stored on the storage and capable of running on the processor, wherein when the processor executes the computer program, the wind turbine power prediction method based on wake deflection effect and 2DJensen model improved graph neural network is used for realizing wind turbine power prediction.
A computer-readable storage medium, on which a computer program is stored, which, when being executed by a processor, implements a wind turbine power prediction using the method for wind turbine power prediction based on wake deflection effect and a 2DJensen model improved graph neural network.
In order to verify the effectiveness of the scheme of the invention, the following experiment is carried out by using SCADA data of a real wind farm.
SCADA data of five wind turbines in a certain wind power plant are collected, the data comprise characteristics of inflow wind direction, inflow wind speed, yaw error angle, output power and the like, the unit numbers are A02, A03, A04, A05 and A06 respectively, and the geographical position information is shown in table 1. The five wind turbines arranged in the wind field are of the same type, the radius of an impeller of each wind turbine is 59m, the axial induction factor is 0.24, and the wake flow attenuation coefficient is 0.075.
TABLE 1 wind turbine position parameters
Figure 893624DEST_PATH_IMAGE125
Constructing an input vector atlas by utilizing collected SCADA data and position information of five wind turbines, using 70% of data in the input vector atlas for training a neural network model, setting all neural networks as three hidden layers, namely 128 neurons, 64 neurons and 32 neurons, setting an activation function as a ReLU function, selecting an Adam algorithm as an optimization algorithm, and selecting a learning rate of 3e-6And iterating 20000 steps. And obtaining a graph neural network model after the training is finished, using the residual 30% of data to measure the power prediction accuracy of the graph neural network model, and obtaining that the comprehensive error of the predicted power and the actual power is 15% for the A02 unit, which is better than 40% of the comprehensive error of the 2DJensen model, as shown in FIG. 5. The method is accurate and effective, and can accurately calculate the power output value of the wind turbine.
The technical features of the above embodiments can be arbitrarily combined, and for the sake of brevity, all possible combinations of the technical features in the above embodiments are not described, but should be considered as the scope of the present specification as long as there is no contradiction between the combinations of the technical features.
The above-mentioned embodiments only express several embodiments of the present application, and the description thereof is more specific and detailed, but not construed as limiting the scope of the present application. It should be noted that, for a person skilled in the art, several variations and modifications can be made without departing from the concept of the present application, which falls within the scope of protection of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims (6)

1. The wind turbine power prediction method based on the wake deflection effect and the 2DJensen model is characterized by comprising the following specific steps of:
step 1, collecting and preprocessing SCADA data of a wind turbine in a wind field;
2, constructing a relative position coordinate of each wind turbine based on the geographical position data of the wind turbines in the wind field;
step 3, obtaining a node set and global attributes of the wind turbine in the wind field according to the preprocessed SCADA data set;
step 4, obtaining an edge set and an adjacency matrix between the wind machines in the wind field according to the relative position coordinates of the wind machines, the input wind direction at the current moment, the wake deflection effect and the 2DJensen model;
step 5, constructing an input vector graph set based on a node set, an edge set, global attributes and an adjacency matrix of a wind turbine in a wind field;
step 6, constructing an improved graph neural network model based on the wake deflection effect and the 2DJensen model, and determining the setting parameters of the graph neural network;
step 7, performing parameter optimization on the improved graph neural network according to the preprocessed SCADA data set and the input vector diagram by using an optimization algorithm and according to a loss function, and determining a graph neural network model after the parameter optimization is completed to serve as a wind turbine power prediction model in a wind field;
step 8, processing the input wind speed and wind direction by using a wind turbine power prediction model to obtain the predicted power of all wind turbine nodes in a wind field;
step 4, obtaining an edge set and an adjacency matrix between the wind machines in the wind field according to the relative position coordinates of the wind machines, the input wind direction at the current moment, the wake deflection effect and the 2DJensen model, wherein the specific method comprises the following steps:
judging whether wake flow influence exists between every two wind turbines or not according to a wake flow deflection effect and a 2DJensen model based on an inflow wind direction and a yaw error angle between every two wind turbines in the global attribute at the moment, if so, establishing a directed edge between the two wind turbines, setting a foreground wind turbine as a sending node and a background wind turbine as a receiving node, and setting the attribute of the edge as the transverse distance and the longitudinal distance of the background wind turbine in the wake flow deflected by the foreground wind turbine; if not, no edge exists between the two wind turbines, and after traversing the relationship between all the wind turbines in the wind field, an edge set and an adjacent matrix in the whole wind field are obtained; the specific method for judging whether wake influence exists between every two wind turbines or not according to the wake deflection effect and the 2DJensen model comprises the following steps:
obtaining the relative positions and azimuth angles of the two wind turbines according to the coordinates of the two wind turbines, obtaining the sequence of the outlet wind passing through the wind turbines successively according to the inflow wind direction in the global attribute at the moment, respectively defining the sequence as an upstream wind turbine and a downstream wind turbine, and defining the included angle theta between the connection line of the upstream wind turbine and the downstream wind turbine and the central axis of the wake flow of the upstream wind turbineLComprises the following steps:
Figure DEST_PATH_IMAGE002
in the formula (I), the compound is shown in the specification,θ x in order to obtain an inflow wind direction angle,θ FWT the azimuth angle of the upstream unit relative to the downstream unit,β 1is the yaw error angle of the upstream wind turbine,ais a wind turbine axial induction factor;
the longitudinal distance between the section of the downstream wind turbine generator and the wake flow of the upstream wind turbine generator is defined asL l The calculation formula is as follows:
Figure DEST_PATH_IMAGE004
in the formula (I), the compound is shown in the specification,Ldthe distance between the cabin connecting lines of the wind turbine generator sets upstream and downstream is the distance between the cabin connecting lines of the wind turbine generator sets downstream and upstream;
and simultaneously defining the cross section of the downstream wind turbine generator set and the wake flow transverse distance of the upstream wind turbine generator set asL r The calculation formula is as follows:
Figure DEST_PATH_IMAGE006
in the formula (I), the compound is shown in the specification,Ldthe distance between the cabin connecting lines of the wind turbine generator sets upstream and downstream is the distance between the cabin connecting lines of the wind turbine generator sets downstream and upstream;
then, the wake radius calculation formula of the cross section position where the downstream wind turbine generator is located is as follows:
Figure DEST_PATH_IMAGE008
in the formula (I), the compound is shown in the specification,kas a function of the attenuation coefficient of the wake flow,r 0 the radius of the wind wheel of the upstream wind turbine generator set;
according to the calculated wake radiusR l Judging whether the downstream wind turbine is influenced by the wake flow of the upstream wind turbine or not, wherein the specific method comprises the following steps:
1) if it is
Figure DEST_PATH_IMAGE010
The downstream wind turbine is not influenced by the wake flow of the upstream wind turbine;
2) if it is
Figure DEST_PATH_IMAGE012
The downstream wind turbine is influenced by the wake flow of the upstream wind turbine;
wherein the content of the first and second substances,r 0 ’ for wind-driven machinejThe projection length of the radius of the wind wheel is calculated by the following formula:
Figure DEST_PATH_IMAGE014
in the formula (I), the compound is shown in the specification,β 2 the yaw error angle of the downstream wind turbine generator set;
step 6, constructing an improved graph neural network model based on the wake deflection effect and the 2DJensen model, determining the setting parameters of the graph neural network, and selecting 2-4 layers from the layer number of the graph neural network model, wherein the first layer is an improved graph network layer based on the wake deflection effect and the 2DJensen model, comprises three updating functions which are respectively edge updating functions
Figure DEST_PATH_IMAGE016
Updating function of node
Figure DEST_PATH_IMAGE018
Global update function
Figure DEST_PATH_IMAGE020
(ii) a Simultaneously comprises three aggregation functions which are adjacent edge aggregation functions
Figure DEST_PATH_IMAGE022
All edges aggregate function
Figure DEST_PATH_IMAGE024
Aggregation of functions by all nodes
Figure DEST_PATH_IMAGE026
(ii) a Edge weight distribution function based on 2DJensen model wake loss factor
Figure DEST_PATH_IMAGE028
；
Based on wake deflection effectIn the improved graph network layer of the 2DJensen model, the input graph is
Figure DEST_PATH_IMAGE030
In which all input edges are included
Figure DEST_PATH_IMAGE032
All input nodes
Figure DEST_PATH_IMAGE034
And inputting the global attribute
Figure DEST_PATH_IMAGE036
The update sequence is as follows:
the updating of the edges is firstly carried out, and the updating of each directed edge needs to use the attribute of the adjacent sending node
Figure DEST_PATH_IMAGE038
And attributes of the receiving node
Figure DEST_PATH_IMAGE040
And the attribute of the edge
Figure DEST_PATH_IMAGE042
Inputting all attributes to the edge update function simultaneously
Figure DEST_PATH_IMAGE044
In, the edge update function is selected as a neural network
Figure DEST_PATH_IMAGE046
Updated edge attributes
Figure DEST_PATH_IMAGE048
The following were used:
Figure DEST_PATH_IMAGE050
then, the function is distributed according to the edge weight
Figure DEST_PATH_IMAGE052
And obtaining the due weight of each directed edge, wherein the weight enables each directed edge to play a role in different degrees in the subsequent node updating according to the rules of the physical wake model, and the weight value is as follows:
Figure DEST_PATH_IMAGE054
finally, the final edge attribute after each directed edge is updated is obtained
Figure DEST_PATH_IMAGE056
The following were used:
Figure DEST_PATH_IMAGE058
after all existing edges in the graph neural network layer are updated, the node attributes are updated, and each node
Figure DEST_PATH_IMAGE060
All the adjacent edges are needed for updating
Figure DEST_PATH_IMAGE062
Property of the node
Figure DEST_PATH_IMAGE064
And global properties of the entire wind farm
Figure DEST_PATH_IMAGE066
First using the adjacent edge aggregation function
Figure DEST_PATH_IMAGE068
Aggregating the information of the adjacent S edges which have a connection relation with the node, wherein the adjacent edge aggregation function takes into account the wake superposition effect in the improved graph network layer based on the wake deflection effect and the 2DJensen model, and the attributes of all the adjacent edges are summed to obtain the aggregated edge attributes, and the values of the aggregated edge attributes are as follows:
Figure DEST_PATH_IMAGE070
then the aggregated side information
Figure DEST_PATH_IMAGE072
Property of the node
Figure DEST_PATH_IMAGE074
And global properties of the entire wind farm
Figure DEST_PATH_IMAGE076
Simultaneous input to node update function
Figure DEST_PATH_IMAGE078
In (1), the node update function is selected as a neural network
Figure DEST_PATH_IMAGE080
Updated node
Figure DEST_PATH_IMAGE082
The following were used:
Figure DEST_PATH_IMAGE084
after all existing nodes in the graph neural network layer are updated, the global attribute is required to be updated finally, the global attribute updating requires all the edge information, all the node attributes and the global attribute of the whole wind field, and all the edges are firstly utilized to aggregateFunction(s)
Figure DEST_PATH_IMAGE086
Aggregating the information of all edges, wherein the aggregation function of all edges adopts a summation mode, and the attributes of all edges are summed to obtain the aggregated edge attribute, and the value of the aggregated edge attribute is as follows:
Figure DEST_PATH_IMAGE088
then aggregate the functions using all nodes
Figure DEST_PATH_IMAGE090
Aggregating the information of all nodes, wherein the aggregation function of all nodes adopts a summation mode, and the attributes of all nodes are summed to obtain the aggregated node attribute, and the values are as follows:
Figure DEST_PATH_IMAGE092
finally, the aggregated side information
Figure DEST_PATH_IMAGE094
Aggregated node attributes
Figure DEST_PATH_IMAGE096
And global properties of the entire wind farm
Figure DEST_PATH_IMAGE098
Input to global update function at the same time
Figure DEST_PATH_IMAGE100
In general, the node update function is selected as a neural network
Figure DEST_PATH_IMAGE102
Updated global attributes
Figure DEST_PATH_IMAGE104
The following were used:
Figure DEST_PATH_IMAGE106
in a graph neural network layer, the global attribute is updated only once, and the output of the graph neural network layer is obtained after the update
Figure DEST_PATH_IMAGE108
、
Figure DEST_PATH_IMAGE110
And
Figure DEST_PATH_IMAGE112
the updating and aggregation operations of the graph neural network layer do not change the connectivity of all nodes, i.e. the adjacency matrix does not change, and then outputs it
Figure DEST_PATH_IMAGE108A
、
Figure DEST_PATH_IMAGE110A
、
Figure DEST_PATH_IMAGE112A
And the adjacency matrix is used as an input attribute and is input into the next graph network layer;
edge weight distribution function based on 2DJensen model wake loss factor
Figure DEST_PATH_IMAGE114
The method comprises the following steps:
the 2DJensen model was first established as follows:
Figure DEST_PATH_IMAGE116
in the formula (I), the compound is shown in the specification,
Figure DEST_PATH_IMAGE118
for the longitudinal wake distance in the downstream wake of
Figure DEST_PATH_IMAGE120
Transverse wake distance of
Figure DEST_PATH_IMAGE122
The velocity of the wake flow at which the wind is coming,
Figure DEST_PATH_IMAGE124
as a function of the attenuation coefficient of the wake flow,
Figure DEST_PATH_IMAGE126
is the radius of the rotor of the upstream wind turbine,
Figure DEST_PATH_IMAGE128
is an axial induction factor;
defining wake deficit factor therein
Figure DEST_PATH_IMAGE130
Comprises the following steps:
Figure DEST_PATH_IMAGE132
the value of which reflects the longitudinal wake distance in the downstream wake of
Figure DEST_PATH_IMAGE120A
Transverse wake distance of
Figure DEST_PATH_IMAGE122A
Wind speed deficiency degree of the wind turbine and wake flow deficiency factors between wind turbines far away from each other
Figure DEST_PATH_IMAGE130A
Wake loss factor between small, closely spaced wind turbines
Figure DEST_PATH_IMAGE130AA
Large, the definition fully accounts for different wake impact levels, so it is defined as an edge weight distribution function
Figure DEST_PATH_IMAGE114A
The wake flow loss factor among different wind turbines
Figure DEST_PATH_IMAGE130AAA
As an edge weight distribution function; and the attribute of the directed edge between every two wind turbines is defined as:
Figure DEST_PATH_IMAGE134
in the formula (I), the compound is shown in the specification,L l for a receiving node
Figure DEST_PATH_IMAGE136
I.e. downstream wind turbine at the transmitting node
Figure DEST_PATH_IMAGE138
Namely the longitudinal wake distance in the wake after deflection of the upstream wind turbine,L r for a receiving node
Figure DEST_PATH_IMAGE140
I.e. downstream wind turbine at the transmitting node
Figure DEST_PATH_IMAGE142
Namely the lateral distance of the wake flow in the wake flow after the deflection of the upstream wind turbine;
edge weight distribution function
Figure DEST_PATH_IMAGE144
The definition is as follows:
Figure DEST_PATH_IMAGE146
when the subsequent j-th graph neural network layer connected with the first graph neural network layer is established, the graph network layer comprises three updating functions which are respectively edge updating functions
Figure DEST_PATH_IMAGE148
Node update function
Figure DEST_PATH_IMAGE150
Global update function
Figure DEST_PATH_IMAGE152
(ii) a Simultaneously comprises three aggregation functions which are adjacent edge aggregation functions
Figure DEST_PATH_IMAGE154
All edge aggregation function
Figure DEST_PATH_IMAGE156
Aggregating functions of all nodes
Figure DEST_PATH_IMAGE158
；
The input diagram of the layer is
Figure DEST_PATH_IMAGE160
Including all the j-1 layer output edges
Figure DEST_PATH_IMAGE162
All the j-1 level input nodes
Figure DEST_PATH_IMAGE164
And a j-1 level output global property
Figure DEST_PATH_IMAGE166
In the graph neural network layer of the j-th layer, the update sequence is as follows:
the updating of the edges is firstly carried out, and the updating of each directed edge needs to use the attribute of the adjacent sending node
Figure DEST_PATH_IMAGE168
And attributes of the receiving node
Figure DEST_PATH_IMAGE170
And the attribute of the edge
Figure DEST_PATH_IMAGE172
Inputting all attributes to the edge update function simultaneously
Figure DEST_PATH_IMAGE174
In, the edge update function is selected as a neural network
Figure DEST_PATH_IMAGE176
Updated edge attributes
Figure DEST_PATH_IMAGE178
The following:
Figure DEST_PATH_IMAGE180
after all existing edges in the graph neural network layer are updated, the node attributes are updated, and each node
Figure DEST_PATH_IMAGE164A
All the adjacent edges are needed for updating
Figure DEST_PATH_IMAGE172A
Property of the node
Figure DEST_PATH_IMAGE164AA
And global properties of the entire wind farm
Figure DEST_PATH_IMAGE166A
First using the adjacent edge aggregation function
Figure DEST_PATH_IMAGE154A
Aggregating the information of the adjacent S edges which have a connection relation with the node, wherein the adjacent edge aggregation function adopts an averaging mode, and averages the attributes of all the adjacent edges to obtain the aggregated edge attribute, and the values of the aggregated edge attribute are as follows:
Figure DEST_PATH_IMAGE182
then the aggregated side information
Figure DEST_PATH_IMAGE184
Property of the node
Figure DEST_PATH_IMAGE186
And global properties of the entire wind farm
Figure DEST_PATH_IMAGE188
Simultaneous input to node update function
Figure DEST_PATH_IMAGE190
In (1), the node update function is selected as a neural network
Figure DEST_PATH_IMAGE192
Updated node
Figure DEST_PATH_IMAGE194
The following were used:
Figure DEST_PATH_IMAGE196
after all existing nodes in the graph neural network layer are updated, the global attribute is required to be updated finally, the global attribute updating requires all the edge information, all the node attributes and the global attribute of the whole wind field, and all the edge aggregation functions are firstly utilized
Figure DEST_PATH_IMAGE198
Aggregating the information of all edges, wherein the aggregation function of all edges adopts an averaging mode, and averaging the attributes of all edges to obtain the aggregated edge attribute, and the values are as follows:
Figure DEST_PATH_IMAGE200
then aggregate the functions using all nodes
Figure DEST_PATH_IMAGE202
Aggregating the information of all nodes, wherein the aggregation function of all nodes adopts an averaging mode, and the attributes of all nodes are summed to obtain the aggregated node attribute, and the value is as follows:
Figure DEST_PATH_IMAGE204
finally, the aggregated side information
Figure DEST_PATH_IMAGE206
Aggregated node attributes
Figure DEST_PATH_IMAGE208
And global properties of the entire wind farm
Figure DEST_PATH_IMAGE210
Input to global bit at the same timeNew function
Figure DEST_PATH_IMAGE212
In (1), the node update function is selected as a neural network
Figure DEST_PATH_IMAGE214
Updated global attributes
Figure DEST_PATH_IMAGE216
The following were used:
Figure DEST_PATH_IMAGE218
in a graph neural network layer, the global attribute is updated only once, and the output of the graph neural network layer is obtained after the update
Figure DEST_PATH_IMAGE220
、
Figure DEST_PATH_IMAGE222
And
Figure DEST_PATH_IMAGE224
the updating and aggregation operations of the graph neural network layer do not change the connectivity of all nodes, i.e. the adjacency matrix does not change, and then outputs it
Figure DEST_PATH_IMAGE220A
、
Figure DEST_PATH_IMAGE222A
、
Figure DEST_PATH_IMAGE224A
And the adjacency matrix is used as an input attribute and is input into the next graph network layer;
in input vector diagrams
Figure DEST_PATH_IMAGE226
、
Figure DEST_PATH_IMAGE228
、
Figure DEST_PATH_IMAGE230
Obtaining an output vector diagram after the whole diagram neural network architecture
Figure DEST_PATH_IMAGE232
In the power prediction of each wind turbine, only the wind turbine nodes are needed
Figure DEST_PATH_IMAGE234
So that a full connection layer is established
Figure DEST_PATH_IMAGE236
The node attribute in the final graph neural network output vector graph
Figure DEST_PATH_IMAGE238
Input to the full connection layer
Figure DEST_PATH_IMAGE240
And obtaining the predicted power of each wind turbine, wherein the values are as follows:
Figure DEST_PATH_IMAGE242
。
2. the method for predicting the power of the wind turbine based on the wake deflection effect and the 2DJensen model as claimed in claim 1, wherein in the step 1, SCADA data of the wind turbine in the wind field are collected and preprocessed, and the method comprises the following specific steps:
the method comprises the steps of collecting inflow wind speed, inflow wind direction, pitch angle, yaw error angle and output power of each wind turbine in an SCADA system of a wind turbine generator, preprocessing collected data, removing all data of missing values in the data at the moment, removing all data of the data at the moment under the condition of power limitation, removing all data of collection errors of which the wind speed is less than 0 or the active power is less than 0, and ensuring that all wind turbines in a data set at the same moment need four characteristics of inflow wind speed, inflow wind direction, yaw error angle and output power after preprocessing.
3. The method for wind turbine power prediction based on wake deflection effect and 2DJensen model as claimed in claim 1, wherein step 2, the relative position coordinates of each wind turbine are constructed based on the geographical position data of the wind turbine in the wind field, the method comprises:
and determining the position information of all wind turbines in the wind field according to the construction site selection of the wind field, and solving the relative position coordinates of each wind turbine in the wind field by taking a certain point in the wind field as a coordinate origin according to the relative azimuth angle and the relative distance between the wind turbines.
4. The method for predicting the power of the wind turbine based on the wake deflection effect and the 2DJensen model as claimed in claim 1, wherein in the step 3, the node set and the global attribute of the wind turbine in the wind field are obtained according to the preprocessed SCADA data set, and the method comprises the following specific steps:
the method comprises the steps of averaging the inflow wind directions of all wind turbines at a certain moment, judging that the wind at the moment flows through a first wind turbine in a wind field by using the obtained wind directions, setting the node attributes in all the wind turbines as the wind speeds in the global attributes by taking the wind speed and the wind direction of the wind turbine at the moment as the global attributes, and obtaining a node set in the wind field at the moment.
5. The method of claim 1, wherein in step 7, the improved graph neural network is optimized parametrically according to the preprocessed SCADA data set and the input vector diagram according to a loss function by using an optimization algorithm, wherein the loss function is the predicted power and the true valueMSEThe values are as follows:
Figure DEST_PATH_IMAGE244
in the formula (I), the compound is shown in the specification,
Figure DEST_PATH_IMAGE246
is composed ofkInputting a predicted value of the graph neural network power of the ith wind turbine in the wind field corresponding to the vector graph at a moment,
Figure DEST_PATH_IMAGE248
is composed ofkInputting the actual power value of the ith wind turbine in the wind field corresponding to the vector diagram at a moment,sto be the total number of input data,mthe total number of wind turbines in the wind field;
and after the loss function is set, selecting Adam, SGD or RMSprop algorithm, setting learning rate, optimizing the graph neural network power prediction model, finishing optimization when the precision of the model is not changed any more, and taking the graph neural network model with the parameters determined after the optimization as the wind turbine power prediction model in the wind farm.
6. The method for wind turbine power prediction based on wake deflection effect and 2DJensen model as claimed in claim 1, wherein step 8, the wind turbine power prediction model is used to process the input wind speed and wind direction to obtain the predicted power of all wind turbine nodes in the wind field, and the specific method is as follows:
when the power of each wind turbine in the wind field corresponding to a certain wind speed and wind direction is predicted, the input wind speed and wind direction are taken as global attributes, the wind speed in the global attributes is set as the node attributes of all the wind turbines, a corresponding input vector diagram is generated, an optimized graph neural network model is called to process the input vector diagram, and the power output value of each wind turbine in the wind field is obtained.
CN202210412243.0A 2022-04-19 2022-04-19 Wind turbine power prediction method based on wake deflection effect and 2DJensen model Active CN114511158B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210412243.0A CN114511158B (en) 2022-04-19 2022-04-19 Wind turbine power prediction method based on wake deflection effect and 2DJensen model

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210412243.0A CN114511158B (en) 2022-04-19 2022-04-19 Wind turbine power prediction method based on wake deflection effect and 2DJensen model

Publications (2)

Publication Number Publication Date
CN114511158A CN114511158A (en) 2022-05-17
CN114511158B true CN114511158B (en) 2022-06-21

Family

ID=81555364

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210412243.0A Active CN114511158B (en) 2022-04-19 2022-04-19 Wind turbine power prediction method based on wake deflection effect and 2DJensen model

Country Status (1)

Country Link
CN (1) CN114511158B (en)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116579479B (en) * 2023-05-15 2024-04-09 南京理工大学 Wind farm power ultra-short-term prediction method, system, computer and storage medium

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102663251A (en) * 2012-04-09 2012-09-12 华北电力大学 Physical prediction method for wind power station power based on computational fluid mechanics model
CN107506854A (en) * 2017-08-04 2017-12-22 国网浙江省电力公司经济技术研究院 A kind of 220kV Power grid structure planing methods for considering differentiation scene
CN109274121A (en) * 2018-11-15 2019-01-25 山东中车风电有限公司 A kind of wind power plant Optimization about control parameter method and system
CA3061717A1 (en) * 2018-11-16 2020-05-16 Royal Bank Of Canada System and method for a convolutional neural network for multi-label classification with partial annotations
CN111784041A (en) * 2020-06-28 2020-10-16 中国电力科学研究院有限公司 Wind power prediction method and system based on graph convolution neural network
CN112883652A (en) * 2021-03-12 2021-06-01 重庆大学 Design method of wind power plant active power off-line prediction controller considering dynamic wake flow

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112149363B (en) * 2020-09-03 2021-09-03 南京理工大学 Two-dimensional Jensen model and double-beam laser radar-based wake region fan power prediction method
CN114169614B (en) * 2021-12-09 2022-12-13 中国华能集团清洁能源技术研究院有限公司 Wind power plant optimal scheduling method and system based on wind turbine wake model optimization

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102663251A (en) * 2012-04-09 2012-09-12 华北电力大学 Physical prediction method for wind power station power based on computational fluid mechanics model
CN107506854A (en) * 2017-08-04 2017-12-22 国网浙江省电力公司经济技术研究院 A kind of 220kV Power grid structure planing methods for considering differentiation scene
CN109274121A (en) * 2018-11-15 2019-01-25 山东中车风电有限公司 A kind of wind power plant Optimization about control parameter method and system
CA3061717A1 (en) * 2018-11-16 2020-05-16 Royal Bank Of Canada System and method for a convolutional neural network for multi-label classification with partial annotations
CN111784041A (en) * 2020-06-28 2020-10-16 中国电力科学研究院有限公司 Wind power prediction method and system based on graph convolution neural network
CN112883652A (en) * 2021-03-12 2021-06-01 重庆大学 Design method of wind power plant active power off-line prediction controller considering dynamic wake flow

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
Physical Model Induced Neural Network for Wind Speed Prediction for Wind Turbines under Wake Effects;Jing Liu等;《《10th Renewable Power Generationconference（PRG 2021） 》;20211227;全文 *
考虑尾流效应的风电场短期功率空间预测模型;曾程等;《电力系统保护与控制》;20121216(第24期);全文 *

Also Published As

Publication number Publication date
CN114511158A (en) 2022-05-17

Similar Documents

Publication Publication Date Title
CN107785935B (en) The power grid wind electricity digestion capability appraisal procedure of meter and risk and Unit Combination
CN106487005A (en) A kind of Electric power network planning method considering T-D tariff
CN103996074A (en) CFD and improved PSO based microscopic wind-farm site selection method of complex terrain
CN105069236B (en) Consider the broad sense load joint probability modeling method of wind power plant node space correlation
CN111639870A (en) Power transmission network multi-target flexible planning method and system considering source load uncertainty
CN107316113A (en) A kind of Transmission Expansion Planning in Electric method and system
CN106849066A (en) A kind of regional wind power prediction method
CN113312735B (en) DMA partition method for urban water supply pipe network
CN110390467A (en) A kind of random ADAPTIVE ROBUST Optimization Scheduling of virtual plant distinguished based on key scenes
CN106684905A (en) Wind power plant dynamic equivalence method with wind power forecast uncertainty considered
CN108428017A (en) Wind power interval prediction method based on core extreme learning machine quantile estimate
CN114511158B (en) Wind turbine power prediction method based on wake deflection effect and 2DJensen model
CN113806947B (en) Offshore wind farm layout processing method, device and equipment
CN103366062A (en) Method for constructing core backbone grid structure based on BBO algorithm and power grid survivability
CN103440405A (en) Method and system for steady-state modeling of wind power plant based on measured data
CN111695666A (en) Wind power ultra-short term conditional probability prediction method based on deep learning
CN105529714B (en) Quick probability load flow calculation method based on normal distribution assemblage characteristic
CN115640963A (en) Offshore wind power access system robust planning method considering investment operation mode
CN112052544B (en) Wind farm collecting network design method, system, storage medium and computing equipment
Shu et al. Decentralised optimisation for large offshore wind farms using a sparsified wake directed graph
CN109066651A (en) The calculation method of wind-powered electricity generation-load scenarios limit transmitted power
CN104021315A (en) Method for calculating station service power consumption rate of power station on basis of BP neutral network
Wu et al. A synchronization methodology for 3D offshore wind farm layout optimization with multi-type wind turbines and obstacle-avoiding cable network
CN107908899A (en) A kind of wire selection system and selection method for Construction of Wind Power construction road
CN108736472B (en) Tidal current energy power generation field planning method considering reef influence

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