CN116341720A - Multi-fan wind speed and direction prediction method based on dynamic graph convolution and transformation - Google Patents
Multi-fan wind speed and direction prediction method based on dynamic graph convolution and transformation Download PDFInfo
- Publication number
- CN116341720A CN116341720A CN202310189178.4A CN202310189178A CN116341720A CN 116341720 A CN116341720 A CN 116341720A CN 202310189178 A CN202310189178 A CN 202310189178A CN 116341720 A CN116341720 A CN 116341720A
- Authority
- CN
- China
- Prior art keywords
- wind speed
- wind
- fan
- time
- prediction
- 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.)
- Withdrawn
Links
- 238000000034 method Methods 0.000 title claims abstract description 53
- 230000009466 transformation Effects 0.000 title claims abstract description 26
- 230000007246 mechanism Effects 0.000 claims abstract description 42
- 238000013528 artificial neural network Methods 0.000 claims abstract description 17
- 230000007774 longterm Effects 0.000 claims abstract description 12
- 238000005070 sampling Methods 0.000 claims abstract description 4
- 238000012549 training Methods 0.000 claims description 29
- 238000004364 calculation method Methods 0.000 claims description 19
- 238000006243 chemical reaction Methods 0.000 claims description 16
- 239000011159 matrix material Substances 0.000 claims description 15
- 230000008569 process Effects 0.000 claims description 15
- 238000006467 substitution reaction Methods 0.000 claims description 8
- 238000013527 convolutional neural network Methods 0.000 claims description 7
- 238000013507 mapping Methods 0.000 claims description 6
- 230000006870 function Effects 0.000 claims description 5
- 230000008859 change Effects 0.000 claims description 4
- 238000003062 neural network model Methods 0.000 claims description 4
- 238000012545 processing Methods 0.000 claims description 4
- 238000005096 rolling process Methods 0.000 claims description 4
- 230000004927 fusion Effects 0.000 claims description 3
- 238000011478 gradient descent method Methods 0.000 claims description 3
- 238000000844 transformation Methods 0.000 claims 1
- 238000013135 deep learning Methods 0.000 abstract description 4
- 238000004590 computer program Methods 0.000 description 16
- 230000000694 effects Effects 0.000 description 4
- 238000010586 diagram Methods 0.000 description 3
- 238000013501 data transformation Methods 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 238000009825 accumulation Methods 0.000 description 1
- 238000003491 array Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000011217 control strategy Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 230000002688 persistence Effects 0.000 description 1
- 238000012876 topography Methods 0.000 description 1
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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
-
- 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)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Business, Economics & Management (AREA)
- Mathematical Physics (AREA)
- Economics (AREA)
- Human Resources & Organizations (AREA)
- Strategic Management (AREA)
- Health & Medical Sciences (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Analysis (AREA)
- Pure & Applied Mathematics (AREA)
- General Business, Economics & Management (AREA)
- Tourism & Hospitality (AREA)
- Computational Mathematics (AREA)
- Marketing (AREA)
- Software Systems (AREA)
- Mathematical Optimization (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Quality & Reliability (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Molecular Biology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Primary Health Care (AREA)
- Water Supply & Treatment (AREA)
- Public Health (AREA)
- Operations Research (AREA)
- Entrepreneurship & Innovation (AREA)
- Game Theory and Decision Science (AREA)
- Development Economics (AREA)
- Algebra (AREA)
- Databases & Information Systems (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a multi-fan wind speed and direction prediction method based on dynamic graph convolution and transformation, which comprises the following steps: based on sensors and storage equipment at each fan of the wind power plant, sampling and storing wind speed and wind direction X at historical moment; modeling long-term space dependence and short-term space dependence at each fan of the wind power plant based on graph convolution and graph self-attention network, and constructing a dynamic graph convolution network to obtain space dependence information of each moment of the wind power plant; based on a self-attention mechanism of a transducer, the modeling of the time dependence relation of the wind speed and direction sequence is completed; embedding a graph convolution network into a transducer network, constructing a space-time sequence prediction neural network, modeling the space-time dependency relationship of a wind speed and wind direction sequence, and completing the wind speed and wind direction prediction of a plurality of fans; the wind power plant wind speed and direction prediction method based on the deep learning theory of the wind power plant storage equipment such as the historical wind speed and direction information, graph convolution and transformation improves the wind power plant wind speed and direction prediction accuracy.
Description
Technical Field
The invention belongs to the field of deep learning space-time sequence prediction, and relates to a multi-fan wind speed and direction prediction method based on dynamic graph convolution and transformation.
Background
Currently, the social demand for clean energy is increasing, and the clean energy industry is expanding. Wind energy is an important clean energy source, and the wind power industry scale is rapidly growing in recent years. However, as wind power resources are continuously developed, wind power projects are increasingly built in medium and low wind speeds and more complex terrain areas for a long time in the present and future, obvious local micro climates exist in wind fields, and the wind fields need to face more complex conditions in both the building period and the operating period, so that the traditional time sequence model has insufficient accuracy in predicting wind speed and wind direction and cannot provide more effective information for a fan control strategy. Improving the wind speed and direction prediction accuracy becomes a problem to be solved urgently.
Traditional time sequence prediction methods such as Autoregressive (AR), autoregressive moving average (ARIMA) and the like have high requirements on sequence stability and poor non-stationary sequence prediction effects. In recent years, the deep learning prediction model has been developed faster, and especially the application of the transducer in the field of time sequence prediction greatly improves the precision of long sequence prediction. However, the predictive effect of these methods still has room for improvement for spatio-temporal sequences with significant spatial dependencies.
For the problem of sequence prediction with spatial dependency, a spatio-temporal sequence prediction model has been developed. Compared with the traditional deep learning time sequence prediction model, the space-time sequence prediction model utilizes a convolution network and the like to represent space dependence, and the defect of the time sequence model on space dependence information is overcome. Particularly, in recent years, the development of a dynamic graph convolution network models the spatial dependency relationship changing along with time more accurately, and improves the prediction effect of complex space-time sequences.
Besides the time dependence, the wind speed and the wind direction in the wind farm have obvious spatial dependence. At the same time, this spatial relationship may change over time. The transformation and the dynamic graph convolution are combined, so that complex time relations in the sequence can be effectively captured, meanwhile, modeling space dependence can provide more information, and the wind speed and direction prediction effect of the wind power plant is effectively improved.
Disclosure of Invention
The invention aims to solve the problems in the prior art, and provides a multi-fan wind speed and direction prediction method based on dynamic graph convolution and transformation.
In order to achieve the purpose, the invention is realized by adopting the following technical scheme:
based on wind farm sensors and storage equipment, sampling and storing wind speed and wind direction X of each fan at historical moment;
modeling long-term space dependence and short-term space dependence at each fan of the wind power plant based on graph convolution (Graph convolution network, GCN) and graph self-attention network (Graph attention network, GAT), and constructing a dynamic graph convolution network to obtain space dependence relationship information of each moment of the wind power plant;
based on a self-attention mechanism of a transducer, the modeling of the time dependence relationship of the wind speed and direction sequence is completed.
The graph convolution network is embedded into a transducer, so that a space-time sequence prediction neural network is constructed, and space-time dependency modeling of the wind speed and direction sequence is realized.
Introducing a mask mechanism and approximate state substitution, and outputting multi-fan wind speed and direction prediction at multiple moments;
based on the stored historical moment data, training of the constructed neural network is completed, and the trained neural network is used for predicting wind speed and wind direction of each fan at future moment.
The improvement of the invention is that:
when the dynamic graph convolution network is constructed, the spatial dependence of wind speed and wind direction of the wind power plant is modeled as long-term global dependence and short-term local dependence, the long-term global dependence does not change with time, and the short-term local dependence is updated continuously with time. Based on this assumption, long-term global dependency adjacency matrix A is constructed separately global Short-term global dependency adjacency matrix A with time t t Obtaining a dynamic adjacency matrix at the moment t by using a multiplication model:
the symbol ". Alt represents the Hadamard product, i.e., the multiplication of the corresponding position elements. A is that global As a parameter of the network, obtained after model training is completed.Obtained through a graph self-attention mechanism, the calculation mode is as follows:
the input space-time sequence X epsilon R C×T×V Wherein C represents the characteristic number, T represents the input time step, and V represents the number of fans. Firstly, capturing local time characteristics by using a convolutional neural network (Convolutional neural network, CNN) fan by fan, namely, the convolution kernel size is k multiplied by 1, the padding of a model is set to be (k-1)/2, and finally, the local time characteristics are obtained as follows:
based on graph self-attention mechanism, obtaining short-term global dependency adjacency matrix A at t moment t :
X local,1 And (3) withCan be obtained by the formula (2), and performing a Softmax operation in columns to obtain the final +.>
Based on a self-attention mechanism of a transducer, the modeling of the time dependence relationship of the wind speed and direction sequence is completed. Because the self-attention mechanism is insensitive to the time sequence, firstly, the input sequence X is subjected to position coding, the input sequence X is coded by using a Position Embeddings method, the sequence coding modes at all fans are the same, and the coded result isWherein C is h Is the feature number. For a single fan sequence H :,:,v The self-attention mechanism calculation process is as follows:
wherein Q, K and V are all obtained through full connection, d k For Q, the feature number at each moment, i.e.
Q=W Q H :,:,v (5)
K=W K H :,:,v (6)
V=W V H :,:,v (7)
In order to increase the nonlinear modeling capability of the model, after self-attention mechanism operation, a full-connection layer fan-by-fan machine using a residual structure is added into the model to carry out nonlinear mapping. The specific process is that,
F :,:,v =FeedForward(Attention(Q,K,V)+H :,:,v ) (8)
the key step in the method is to embed a graph convolution network into a transducer, construct a space-time sequence prediction neural network, realize modeling of space-time dependency relationship of wind speed and wind direction sequences and complete multi-fan wind speed and wind direction prediction. At this time, for the self-attention mechanism calculation process described in (4), Q, K, V are all calculated by (5) - (7) fan-by-fan, and the spatial dependency relationship between fans is not utilized. Therefore, we replace the equations (5) - (7) with graph convolution operations to achieve spatial dependency modeling.
After the input sequence X is subjected to position coding, a coded space-time sequence is obtained The graph convolution operation process is that the space dependence is captured for the time sequence H time by time feature by feature, and fusion among the features is completed through 1X 1 convolution. For state H at time t :,t,: After the graph rolling operation, can obtain
G :,t,: =GraphConv(W,A t ,H :,t,: ) (9)
More specifically, each channel G q,t,: The calculation result is obtained by the following calculation process,
wherein W is a 1×1 convolution parameter, A t The dynamic adjacency matrix obtained by the formula (1). And (3) respectively replacing the formulas (5) - (7) by the formula (9), and thus completing the modeling of the spatial dependency relationship.
In order to realize one-time output of multi-step prediction results, a mask mechanism and approximate state substitution are introduced into the method, and meanwhile, multi-time multi-fan wind speed and wind direction prediction is output. The mask mechanism is mainly characterized in that when the self-attention mechanism operation is completed by using the formula (4) at each moment, input data only has the state of the moment and the moment before the moment, and the information after the moment is not relied on, so that the relation of 'front and back causality' is satisfied. When a plurality of moments are predicted by using a self-attention mechanism, the prediction result output each time is required to be used as input, and the prediction of the moments after iteration is completed. In order to increase the prediction speed, the final moment is copied L-1 times, and the model is input simultaneously with the connection of the original input sequence. And outputting the final L-step state of the model as a prediction result. Namely, the original input sequence is
X=[X :,1,: ,X :,2,: ,…,X :,T,: ]∈R C×T×V ,
After using the approximate substitution operation, the input sequence at this time is
X=[X :,1,: ,X :,2,: ,…,X :,T,: ,X :,T,: ,…,X :,T,: ]∈R C×(T+L-1)×V 。
After calculation by the K-layer self-attention mechanism, the final L-step state output by the model is used as a prediction result by using full-connection layer mapping.
Training of the graph volume transformers deep neural network is specifically as follows:
training dataset D, d= { X (i) ,Y (i) I=1, 2, …, M }, wherein, at t (i) -T+ 1 to T (i) Wind speed and direction information of each fan of wind power plant at moment +.> At t (i) +1 to t (i) Wind speed and direction information of each fan of the wind power plant at +L moment, wherein M is the total number of training samples;
because the wind speed and the wind direction are predicted simultaneously as a plurality of prediction tasks, the problem of inconsistent dimension and the like exists. The problem can be converted into a single task prediction problem by means of data transformation.Let the wind speed at a fan at the moment t be s t The wind direction is theta t The variable can be converted into (a) t ,b t )=(s t cosθ t ,s t sinθ t ). That is, the wind speed and direction information (s t ,θ t ) Is actually the converted variable (a t ,b t ) At the same time, the model prediction variable is (a t ,b t )。
When training the network, the random gradient descent method SGD is used for optimizing parameters of the network F, and an objective function to be optimized is shown as a formula (11):
after the network training is completed, the wind speed and wind direction conversion variable of a certain fan at the position can be predictedReversible conversion is carried out to obtain wind speed and direction information:
A multiple wind turbine wind speed and direction prediction system based on dynamic graph convolution and transformation, comprising:
the storage module is used for storing wind speed and wind direction information X acquired by each fan at the historical moment based on wind farm storage equipment;
the training module is used for processing the collected historical data through the method in claim 6, and training the graph convolution transformation neural network model based on the processed training data set;
the prediction module is used for completing the prediction of the future moment conversion variable by using a transformer graph convolution neural network based on the acquired historical moment wind speed and direction information conversion variable;
a reverse conversion module that converts the predicted variables into wind speed and direction information based on the formulas (12), (13) in claim 6.
A terminal device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, the processor implementing the steps of the method according to any one of claims 1-6 when the computer program is executed.
A computer readable storage medium storing a computer program which, when executed by a processor, implements the steps of the method according to any one of claims 1-6.
Compared with the prior art, the invention has the following beneficial effects:
the method is based on a dynamic graph rolling network, does not need to know information such as the topography, weather characteristics, fan distribution and the like of any wind power plant in advance, and constructs spatial dependency relations at different moments in real time in a data driving mode. Meanwhile, modeling of the time correlation relationship is achieved by means of a powerful transducer network. And embedding the dynamic graph convolution into a transformer to acquire space-time dependent information, so as to realize accurate prediction of wind speed and wind direction of multiple fans of the wind power plant.
Drawings
FIG. 1 is a flow chart of a method for predicting wind speed and direction of a multi-fan based on dynamic graph convolution and transformation of the present invention;
FIG. 2 is a neural network model diagram of the multi-fan wind speed and direction prediction method based on dynamic graph convolution and transformation of the invention;
FIG. 3 is a schematic diagram of a multi-temporal prediction implementation in a transducer using a mask mechanism;
FIG. 4 is an error curve of a model predicted wind speed at multiple moments in measured data of a wind farm;
FIG. 5 is a graph of error of model predicted wind direction at multiple moments in measured data of a wind farm;
FIG. 6 is a block diagram of a multi-fan wind speed and direction prediction system based on dynamic graph convolution and transformation according to the present invention.
Detailed Description
The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
Specific examples are given below.
Referring to fig. 1, the invention discloses a multi-fan wind speed and direction prediction method based on dynamic graph convolution and transformation, which comprises the following steps:
s101, based on wind farm sensors and storage equipment, sampling and storing wind speed and wind direction X of each fan at historical moment.
S102, constructing a dynamic graph convolution network based on graph convolution (Graph convolution network, GCN) and graph self-attention network (Graph attention network, GAT) to obtain spatial dependency relationship information of each moment of the wind power plant.
The spatial dependence of wind speed and wind direction of the wind farm is modeled as long-term global dependence and short-term local dependence, the long-term global dependence does not change with time, and the short-term local dependence is updated continuously with time. Based on this assumption, long-term global dependency adjacency matrix A is constructed separately global Short-term global dependency adjacency matrix with time tObtaining a dynamic adjacency matrix at the moment t by using a multiplication model:
the symbol ". Alt represents the Hadamard product, i.e., the multiplication of the corresponding position elements. A is that global As a parameter of the network, obtained after model training is completed.Obtained through a graph self-attention mechanism, the calculation mode is as follows:
the input space-time sequence X epsilon R C×T×V Wherein C represents the characteristic number, T represents the input time step, and V represents the number of fans. Firstly, capturing local time characteristics by using a convolutional neural network (Convolutional neural network, CNN) fan by fan, namely, the convolution kernel size is k multiplied by 1, the padding of a model is set to be (k-1)/2, and finally, the local time characteristics are obtained as follows:
based on graph self-attention mechanism, obtaining short-term global dependency adjacency matrix at t moment
X local,1 And (3) withCan be obtained by the formula (2), and performing a Softmax operation in columns to obtain the final +.>
S103, modeling the time dependence relationship of the wind speed and direction sequence based on a self-attention mechanism of the transducer.
Self-attentionThe mechanism is insensitive to the time sequence, firstly, the input sequence X is subjected to position coding, the Position Embeddings method is used for coding, the sequence coding modes at all fans are the same, and the coded result isWherein C is h Is the feature number. For a single fan sequence H :,:,v The self-attention mechanism calculation process is as follows:
wherein Q, K and V are all obtained through full connection, d k For Q, the feature number at each moment, i.e.
Q=W Q H :,:,v (5)
K=W K H :,:,v (6)
V=W V H :,:,v (7)
In order to increase the nonlinear modeling capability of the model, after self-attention mechanism operation, a full-connection layer fan-by-fan machine using a residual structure is added into the model to carry out nonlinear mapping. The specific process is that,
F :,:,v =FeedForward(Attention(Q,K,V)+H :,:,v ) (8)
s104, as shown in fig. 2, embedding a graph convolution network into a transducer to construct a space-time sequence prediction neural network so as to model the space-time dependency relationship of the wind speed and direction sequence.
And (3) for the self-attention mechanism calculation process described in the step (4), Q, K and V are all obtained through fan-by-fan calculation in the steps (5) - (7), and the spatial dependency relationship among the fans is not utilized. Therefore, we replace the equations (5) - (7) with graph convolution operations to achieve spatial dependency modeling.
After the input sequence X is subjected to position coding, a coded space-time sequence is obtained The graph convolution operation process is that the space dependence is captured for the time sequence H time by time feature by feature, and fusion among the features is completed through 1X 1 convolution. For state H at time t :,t,: After the graph rolling operation, can obtain
G :,t,: =GraphConv(W,A t ,H :,t,: ) (9)
More specifically, each channel G q,t,: The calculation result is obtained by the following calculation process,
wherein W is a 1×1 convolution parameter, A t The dynamic adjacency matrix obtained by the formula (1). And (3) respectively replacing the formulas (5) - (7) by the formula (9), and thus completing the modeling of the spatial dependency relationship.
S105, introducing a mask mechanism and replacing an approximate state, and outputting multi-fan wind speed and direction prediction at multiple moments.
The mask mechanism is mainly characterized in that when the self-attention mechanism operation is completed by using the formula (4) at each moment, input data only has the state of the moment and the moment before the moment, and the information after the moment is not relied on, so that the relation of 'front and back causality' is satisfied.
When a plurality of moments are predicted by using a self-attention mechanism, the prediction result output each time is required to be used as input, and the prediction of the moments after iteration is completed. Referring to fig. 5, the method copies the last moment L-1 times, connects with the original input sequence, and inputs the model at the same time. And taking the final L-step state output by the model as a prediction result. The method not only improves the prediction speed, but also avoids error accumulation.
The original input sequence is
X=[X :,1,: ,X :,2,: ,…,X :,T,: ]∈R C×T×V ,
After using the approximate substitution operation, the input sequence at this time is
X=[X :,1,: ,X :,2,: ,…,X :,T,: ,X :,T,: ,…,X :,T,: ]∈R C×(T+L-1)×V 。
After calculation by the K-layer self-attention mechanism, the final L-step state output by the model is used as a prediction result by using full-connection layer mapping.
S106, based on the stored historical moment data, training of the constructed neural network is completed, and the trained neural network is used for predicting wind speed and wind direction of each fan at future moment.
Training dataset D, d= { X (i) ,Y (i) I=1, 2, …, M }, wherein, at t (i) -T+ 1 to T (i) Wind speed and direction information of each fan of wind power plant at moment +.> At t (i) +1 to t (i) Wind speed and direction information of each fan of the wind power plant at +L moment, wherein M is the total number of training samples;
because the wind speed and the wind direction are predicted simultaneously as a plurality of prediction tasks, the problem of inconsistent dimension and the like exists. The problem can be converted into a single task prediction problem by means of data transformation. Let the wind speed at a fan at the moment t be s t The wind direction is theta t The variable can be converted into (a) t ,b t )=(s t cosθ t ,s t sinθ t ). That is, the wind speed and direction information (s t ,θ t ) Is actually the converted variable (a t ,b t ) At the same time, the model prediction variable is (a t ,b t )。
When training the network, the random gradient descent method SGD is used for optimizing parameters of the network F, and an objective function to be optimized is shown as a formula (11):
after the network training is completed, the wind speed and wind direction conversion variable of a certain fan at the position can be predictedReversible conversion is carried out to obtain wind speed and direction information:
Referring to fig. 4 and 5, fig. 4 and 5 are average absolute errors of the model under measured data for the predicted results of wind speed and wind direction at each fan at each time. The persistence method is to take the wind speed and direction at the last moment as wind speed and direction information of a future period of time, and the DGCTransformer is a designed graph convolution transducer model. The result shows that the graph convolution transducer model can effectively reduce the prediction error.
Referring to fig. 6, fig. 6 discloses a multi-fan wind speed and direction prediction system based on dynamic graph convolution and transformation, comprising:
the storage module is used for storing wind speed and wind direction information X acquired by each fan at the historical moment based on wind farm storage equipment;
the training module is used for processing the collected historical data through the method in claim 6, and training the graph convolution transformation neural network model based on the processed training data set;
the prediction module is used for completing the prediction of the future moment conversion variable by using a transformer graph convolution neural network based on the acquired historical moment wind speed and direction information conversion variable;
a reverse conversion module that converts the predicted variables into wind speed and direction information based on the formulas (12), (13) in claim 6.
The embodiment of the invention provides terminal equipment. The terminal device of this embodiment includes: a processor, a memory, and a computer program stored in the memory and executable on the processor. The steps of the various method embodiments described above are implemented when the processor executes the computer program. Alternatively, the processor may implement the functions of the modules/units in the above-described device embodiments when executing the computer program.
The computer program may be divided into one or more modules/units, which are stored in the memory and executed by the processor to accomplish the present invention.
The terminal equipment can be computing equipment such as a desktop computer, a notebook computer, a palm computer, a cloud server and the like. The terminal device may include, but is not limited to, a processor, a memory.
The processor may be a central processing unit (CentralProcessingUnit, CPU), but may also be other general purpose processors, digital signal processors (DigitalSignalProcessor, DSP), application specific integrated circuits (ApplicationSpecificIntegratedCircuit, ASIC), off-the-shelf programmable gate arrays (Field-ProgrammableGateArray, FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or the like.
The memory may be used to store the computer program and/or module, and the processor may implement various functions of the terminal device by running or executing the computer program and/or module stored in the memory and invoking data stored in the memory.
The modules/units integrated in the terminal device may be stored in a computer readable storage medium if implemented in the form of software functional units and sold or used as separate products. Based on such understanding, the present invention may implement all or part of the flow of the method of the above embodiment, or may be implemented by a computer program to instruct related hardware, where the computer program may be stored in a computer readable storage medium, and when the computer program is executed by a processor, the computer program may implement the steps of each of the method embodiments described above. Wherein the computer program comprises computer program code which may be in source code form, object code form, executable file or some intermediate form etc. The computer readable medium may include: any entity or device capable of carrying the computer program code, a recording medium, a U disk, a removable hard disk, a magnetic disk, an optical disk, a computer memory, a Read-only memory (ROM), a random access memory (RAM, randomAccessMemory), an electrical carrier signal, a telecommunication signal, a software distribution medium, and so forth. It should be noted that the computer readable medium contains content that can be appropriately scaled according to the requirements of jurisdictions in which such content is subject to legislation and patent practice, such as in certain jurisdictions in which such content is subject to legislation and patent practice, the computer readable medium does not include electrical carrier signals and telecommunication signals.
The foregoing is only a preferred embodiment of the present invention, but the scope of the present invention is not limited thereto, and any person skilled in the art, who is within the scope of the present invention, should make equivalent substitutions or modifications according to the technical scheme of the present invention and the inventive concept thereof, and should be covered by the scope of the present invention.
Claims (7)
1. The multi-fan wind speed and direction prediction method based on dynamic graph convolution and transformation is characterized by comprising the following steps of:
s101, sampling and storing wind speed and wind direction X of each fan at historical moment based on wind farm sensors and storage equipment;
s102, modeling long-term space dependence and short-term space dependence at each fan of a wind power plant based on graph convolution and graph self-attention network, and constructing a dynamic graph convolution network to obtain space dependence relation information of each moment of the wind power plant;
s103, modeling the time dependence relation of the wind speed and direction sequence based on a self-attention mechanism of a transducer;
s104, embedding a graph convolution network into a transducer to construct a space-time sequence prediction neural network so as to model the space-time dependency relationship of the wind speed and direction sequence;
s105, in order to realize one-time output of multi-step prediction results, a mask mechanism and approximate state substitution are introduced into the method, and multi-time multi-fan wind speed and wind direction prediction is output at the same time;
s106, based on the stored historical moment data, training of the constructed neural network is completed, and the trained neural network is used for predicting wind speed and wind direction of each fan at the future moment.
2. The method for predicting wind speed and direction of multiple wind turbines based on dynamic graph convolution and transformation according to claim 1, wherein the dynamic graph convolution network in step S102 is constructed as follows:
modeling the spatial dependence of wind speed and wind direction of a wind power plant into long-term global dependence and short-term local dependence, wherein the long-term global dependence does not change with time, the short-term local dependence is continuously updated with time, and respectively constructing a long-term global dependence adjacency matrix A according to the assumption global Short-term global dependency adjacency matrix A with time t t Obtaining a dynamic adjacency matrix at the moment t by using a multiplication model:
as indicated by the Hadamard product, the corresponding position elements were multiplied by one another, A global As parameters of the network, obtained after model training is completed,obtained through a graph self-attention mechanism, the calculation mode is as follows:
the input space-time sequence X epsilon R C×T×V Wherein C represents the characteristic number, T represents the input time step, and V represents the number of fans;
firstly, capturing local time characteristics by using a convolutional neural network fan by fan, namely, the convolution kernel size is k multiplied by 1, the padding of a model is set to be (k-1)/2, and finally, the local time characteristics are obtained as follows:
based on graph self-attention mechanism, obtaining short-term global dependency adjacency matrix at t moment
3. The method for predicting wind speed and wind direction of multiple fans based on dynamic graph convolution and transformation according to claim 1, wherein the transformation-based self-attention mechanism is used for modeling the time dependency relationship of the wind speed and wind direction sequence, specifically:
the self-attention mechanism is insensitive to time sequence order, firstly, the input sequence X is position coded, the Position Embeddings method is used for coding, the sequence coding modes at all fans are the same,obtain the result after coding asWherein C is h For a single fan sequence H as a feature number :,:,v The self-attention mechanism calculation process is as follows:
wherein Q, K and V are all obtained through full connection, d k For Q, the feature number at each moment, i.e.
Q=W Q H :,:,v (5)
K=W K H :,:,v (6)
V=W V H :,:,v (7)
in order to increase the nonlinear modeling capability of the model, after self-attention mechanism operation, a full-connection layer fan-by-fan machine using a residual structure is added into the model for nonlinear mapping, and the specific process is as follows
F :,:,v =FeedForward(Attention(Q,K,V)+H :,:,v ) (8)。
4. The multi-fan wind speed and direction prediction method based on dynamic graph convolution and transformation according to claim 1, wherein the graph convolution network is embedded into the transformation, a space-time sequence prediction neural network is constructed, the modeling of space-time dependency relationship of a wind speed and direction sequence is realized, and the multi-fan wind speed and direction prediction is completed, specifically:
for the self-attention mechanism calculation process described in (4), Q, K and V are all obtained by calculating fan by fan according to (5) - (7), and the spatial dependency relationship among fans is not utilized, so that the (5) - (7) is replaced by graph convolution operation, and the modeling of the spatial dependency relationship is realized;
after the input sequence X is subjected to position coding, a coded space-time sequence H epsilon is obtainedThe graph convolution operation process is that the space dependence is captured for the time sequence H time by time feature by feature, and the fusion among the features is completed through 1X 1 convolution, and the state H is the time t :,t,: After the graph rolling operation, can obtain
G :,t,: =GraphConv(W,A t ,H :,t,: ) (9)
More specifically, each channel G q,t,: The calculation result is obtained by the following calculation process,
wherein W is a 1×1 convolution parameter, A t And (3) replacing the formulas (5) - (7) with the formula (9) for the dynamic adjacency matrix obtained in the formula (1), and thus completing the modeling of the spatial dependency relationship.
5. The method for predicting wind speed and wind direction of multiple fans based on dynamic graph convolution and transformation according to claim 1, wherein in order to output multi-step prediction results once, a mask mechanism and approximate state substitution are introduced in the method, and meanwhile, multi-time multi-fan wind speed and wind direction prediction is output, specifically:
the mask mechanism is mainly characterized in that when the self-attention mechanism operation is completed by using the method (4) at each moment, input data only has the state of the moment and the moment before, and the input data does not depend on the information after the moment, so that the relation of 'front and back causality' is satisfied;
when a self-attention mechanism is utilized to predict a plurality of moments, the prediction result output each time is needed to be input again, the prediction of the moment after iteration is completed is performed, in order to improve the prediction speed, the last moment is copied for L-1 times, the last moment is connected with an original input sequence and is input into a model, and the last L-step state output by the model is the prediction result; namely, the original input sequence is
X=[X :,1,: ,X :,2,: ,…,X :,T,: ]∈R C×T×V ,
After using the approximate substitution operation, the input sequence at this time is
X=[X :,1,: ,X :,2,: ,…,X :,T,: ,X :,T,: ,…,X :,T,: ]∈R C×(T+L-1)×V ;
After calculation by the K-layer self-attention mechanism, the final L-step state output by the model is used as a prediction result by using full-connection layer mapping.
6. The method for predicting wind speed and direction of multiple wind turbines based on dynamic graph convolution and transformation according to claim 1, wherein the training of the graph convolution transformations deep neural network based on the collected data set is specifically as follows:
training dataset D, d= { X (i) ,Y (i) I=1, 2, …, M }, wherein,at t (i) -T+1 to T (i) Wind speed and direction information of each fan of wind power plant at moment +.>At t (i) +1 to t (i) Wind speed and direction information of each fan of wind power plant at +L moment, M is total number of training samples
Because the wind speed and the wind direction are predicted simultaneously as a plurality of prediction tasks, the problems such as dimension and the like are inconsistent, the problems are converted into single-task prediction problems by a data conversion mode, and the wind speed at a certain fan at the moment t is set as s t The wind direction is theta t The variable can be converted into (a) t ,b t )=(s t cos θ t ,s t sin θ t ) The method comprises the steps of carrying out a first treatment on the surface of the That is, the wind speed and direction information (s t ,θ t ) Is actually the converted variable (a t ,b t ) At the same time, the model prediction variable is (a t ,b t );
When training the network, the random gradient descent method SGD is used for optimizing parameters of the network F, and an objective function to be optimized is shown as a formula (11):
after the network training is completed, the wind speed and wind direction conversion variable of a certain fan at the position can be predictedReversible conversion is carried out to obtain wind speed and direction information:
7. The multi-fan wind speed and direction prediction system based on dynamic graph convolution and transformation is characterized by comprising:
the storage module is used for storing wind speed and wind direction information X acquired by each fan at the historical moment based on wind farm storage equipment;
the training module is used for processing the collected historical data through the method in claim 6, and training the graph convolution transformation neural network model based on the processed training data set;
the prediction module is used for completing the prediction of the future moment conversion variable by using a transformer graph convolution neural network based on the acquired historical moment wind speed and direction information conversion variable;
a reverse conversion module that converts the predicted variables into wind speed and direction information based on the formulas (12), (13) in claim 6.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202310189178.4A CN116341720A (en) | 2023-03-02 | 2023-03-02 | Multi-fan wind speed and direction prediction method based on dynamic graph convolution and transformation |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202310189178.4A CN116341720A (en) | 2023-03-02 | 2023-03-02 | Multi-fan wind speed and direction prediction method based on dynamic graph convolution and transformation |
Publications (1)
Publication Number | Publication Date |
---|---|
CN116341720A true CN116341720A (en) | 2023-06-27 |
Family
ID=86876744
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202310189178.4A Withdrawn CN116341720A (en) | 2023-03-02 | 2023-03-02 | Multi-fan wind speed and direction prediction method based on dynamic graph convolution and transformation |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN116341720A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116894384A (en) * | 2023-07-11 | 2023-10-17 | 湖北工业大学 | Multi-fan wind speed space-time prediction method and system |
-
2023
- 2023-03-02 CN CN202310189178.4A patent/CN116341720A/en not_active Withdrawn
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116894384A (en) * | 2023-07-11 | 2023-10-17 | 湖北工业大学 | Multi-fan wind speed space-time prediction method and system |
CN116894384B (en) * | 2023-07-11 | 2024-05-07 | 湖北工业大学 | Multi-fan wind speed space-time prediction method and system |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111126704B (en) | Multi-region precipitation prediction model construction method based on multi-graph convolution and memory network | |
CN114493050B (en) | Multi-dimensional fusion new energy power parallel prediction method and device | |
WO2022077767A1 (en) | Traffic flow prediction method and apparatus, computer device, and readable storage medium | |
Yahya et al. | Designing weather forecasting model using computational intelligence tools | |
CN113449919B (en) | Power consumption prediction method and system based on feature and trend perception | |
CN116341720A (en) | Multi-fan wind speed and direction prediction method based on dynamic graph convolution and transformation | |
CN113128666A (en) | Mo-S-LSTMs model-based time series multi-step prediction method | |
Yin et al. | A deep multivariate time series multistep forecasting network | |
CN117094451A (en) | Power consumption prediction method, device and terminal | |
CN116502774B (en) | Time sequence prediction method based on time sequence decomposition and Legend projection | |
CN107451684A (en) | Stock market's probability forecasting method based on core stochastic approximation | |
CN117787470A (en) | Time sequence prediction method and system based on EWT and integration method | |
CN117409578A (en) | Traffic flow prediction method based on combination of empirical mode decomposition and deep learning | |
CN117094431A (en) | DWTfar meteorological data time sequence prediction method and equipment for multi-scale entropy gating | |
CN117114168A (en) | Long-time-scale sea surface temperature intelligent forecasting method | |
Akiner et al. | Exploring the applicability of the experiment-based ANN and LSTM models for streamflow estimation | |
CN117493583B (en) | Method and system for generating flow operation sequence by combining event log and knowledge graph | |
CN114648152B (en) | Building energy consumption prediction method and system based on state constraint and time-frequency characteristics | |
CN118114057B (en) | Model training method, runoff prediction method, electronic device, and storage medium | |
CN116400430B (en) | Meteorological data prediction method and device, storage medium and electronic device | |
CN116933949B (en) | Water quality prediction method and system integrating hydrodynamic model and numerical model | |
CN115222024B (en) | Short-term photovoltaic power generation prediction method and system based on depth feature selection network | |
CN118036657A (en) | AIS data-based method for bidirectional RNN track completion method with encoding and decoding functions | |
Xie et al. | Grey-informed neural network for time-series forecasting | |
CN116629365A (en) | Knowledge graph completion method, system, equipment and medium based on knowledge distillation |
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 | ||
WW01 | Invention patent application withdrawn after publication |
Application publication date: 20230627 |
|
WW01 | Invention patent application withdrawn after publication |