CN116401949A - Distributed photovoltaic resource output curve deduction method, system, equipment and medium - Google Patents
Distributed photovoltaic resource output curve deduction method, system, equipment and medium Download PDFInfo
- Publication number
- CN116401949A CN116401949A CN202310374894.XA CN202310374894A CN116401949A CN 116401949 A CN116401949 A CN 116401949A CN 202310374894 A CN202310374894 A CN 202310374894A CN 116401949 A CN116401949 A CN 116401949A
- Authority
- CN
- China
- Prior art keywords
- output
- time
- node
- space
- photovoltaic
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 53
- 238000010586 diagram Methods 0.000 claims abstract description 91
- 238000013507 mapping Methods 0.000 claims abstract description 25
- 238000005286 illumination Methods 0.000 claims abstract description 22
- 239000013598 vector Substances 0.000 claims description 57
- 230000006870 function Effects 0.000 claims description 30
- 238000004590 computer program Methods 0.000 claims description 21
- 230000015654 memory Effects 0.000 claims description 17
- 238000012549 training Methods 0.000 claims description 16
- 239000011159 matrix material Substances 0.000 claims description 12
- 230000007613 environmental effect Effects 0.000 claims description 11
- 230000008569 process Effects 0.000 claims description 11
- 238000012545 processing Methods 0.000 claims description 11
- 238000003860 storage Methods 0.000 claims description 10
- 230000005855 radiation Effects 0.000 claims description 8
- 230000004913 activation Effects 0.000 claims description 6
- 238000010276 construction Methods 0.000 claims description 6
- 230000007246 mechanism Effects 0.000 claims description 6
- 238000005457 optimization Methods 0.000 claims description 6
- 238000003825 pressing Methods 0.000 claims 1
- 238000000605 extraction Methods 0.000 abstract description 14
- 230000007774 longterm Effects 0.000 abstract description 6
- 238000013135 deep learning Methods 0.000 abstract description 4
- 238000013508 migration Methods 0.000 description 9
- 230000005012 migration Effects 0.000 description 9
- 238000013526 transfer learning Methods 0.000 description 7
- 235000019580 granularity Nutrition 0.000 description 3
- 230000008859 change Effects 0.000 description 2
- 238000010801 machine learning Methods 0.000 description 2
- 238000004519 manufacturing process Methods 0.000 description 2
- 238000005065 mining Methods 0.000 description 2
- 230000003287 optical effect Effects 0.000 description 2
- 238000000053 physical method Methods 0.000 description 2
- 238000010248 power generation Methods 0.000 description 2
- 238000007619 statistical method Methods 0.000 description 2
- 230000002123 temporal effect Effects 0.000 description 2
- 230000003044 adaptive effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000010219 correlation analysis Methods 0.000 description 1
- 238000007405 data analysis Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000002085 persistent effect Effects 0.000 description 1
- 230000006403 short-term memory Effects 0.000 description 1
- 230000036962 time dependent Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
- G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
- G06N3/0455—Auto-encoder networks; Encoder-decoder networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2113/00—Details relating to the application field
- G06F2113/04—Power grid distribution networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Software Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Mathematical Physics (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Medical Informatics (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Hardware Design (AREA)
- Geometry (AREA)
- Photovoltaic Devices (AREA)
Abstract
A distributed photovoltaic resource output curve deduction method, system, equipment and medium, wherein the method comprises the steps of constructing a weighted undirected time-space graph to characterize the relation between distributed photovoltaic resources; solving the time characteristics of each node in the weighted undirected time-space diagram; extracting the space-time characteristics of each node in the weighted undirected space-time diagram, capturing the output mode characteristics of each node in the space-time characteristics, splicing the time characteristics and the output mode characteristics of each node, inputting the spliced output mode characteristics into a multi-layer full-connection network, outputting output predicted values of distributed photovoltaic resources in corresponding areas by utilizing the full-connection network, and drawing a distributed photovoltaic resource output deduction curve. According to the method, multi-source data are comprehensively utilized, meteorological information and illumination information of a target region are combined to perform time feature extraction, a deep learning based on graph network driving is used for performing space feature extraction on a photovoltaic module of the target region, and a proper mapping relation is learned according to weather conditions of a long term and a short term so as to accurately deduce a future output curve of a distributed photovoltaic resource.
Description
Technical Field
The invention belongs to the technical field of multi-source data analysis and processing of power grids, and particularly relates to a distributed photovoltaic resource output curve deduction method, a system, equipment and a medium.
Background
Along with the adjustment of the electric energy generation proportion in each region, the power generation capacity of the traditional energy sources such as hydropower, thermal power and the like is continuously reduced, and how to accurately predict the output capacity of the photovoltaic system is important. The relationship between the output of the photovoltaic system and the output of other power source systems is reasonably coordinated, so that the safe and stable operation of the urban power grid can be ensured, stable and reliable power support is provided for the high-quality development of production life, economy and society, and further the goal of integrating the power grid and the load storage is realized. However, photovoltaic system power output is a complex nonlinear process, and how to effectively predict photovoltaic output is a very challenging task. Because the photovoltaic power generation output is influenced by multiple factors, under the condition that weather and illumination change or photovoltaic module parameters change, fluctuation with strong randomness can appear in the output of the photovoltaic module, and the electric energy output level of the whole network is influenced. Furthermore, the disturbances from weather and light received by photovoltaic modules in the same, similar areas tend to be similar, with a perceived local effect.
At present, solutions in photovoltaic power prediction at home and abroad are roughly divided into three categories: the physical method is used for deducing an output mode based on a model of solar radiation and a photovoltaic module; a statistical method class for predicting based on a statistical rule between input and output factors of the prediction model; and (3) a machine learning method, which is based on various models, and deduces the output of the photovoltaic module by optimizing an objective function. These power prediction schemes seem viable but there are still drawbacks in practical applications: one major drawback is that these schemes are overly focused on the model performance of short-term, small-scale deployments, but ignore the predictive performance of long-term output curves when actually deployed on a large scale.
In addition, the foregoing solution also has a series of drawbacks in model training. The physical method relies on detailed information of the photovoltaic module and the environment, has extremely high requirements on observation data, and cannot model the photovoltaic output under extreme weather conditions; the statistical method relies on a large amount of historical data information, and has poor robustness; the priori knowledge learned by the machine learning method is often obtained through long-term business data training in the existing mode, and reasonable output curve evaluation cannot be performed on short-term fine granularity indexes, especially when emergency conditions are met.
Disclosure of Invention
The invention aims to solve the problems in the prior art and provides a distributed photovoltaic resource output curve deduction method, a system, equipment and a medium, which can accurately and reasonably deduct the future output curve of the distributed photovoltaic resource under extreme weather conditions and emergency conditions and meet the prediction performance requirement of a long-term output curve in actual large-scale deployment.
In order to achieve the above purpose, the present invention has the following technical scheme:
in a first aspect, a distributed photovoltaic resource output curve deduction method is provided, including:
inputting geographic positions of distributed photovoltaic resources and historical photovoltaic output data under corresponding environmental conditions at different moments, and constructing a weighted undirected time-space diagram to characterize the relation among the distributed photovoltaic resources;
solving time characteristics of each node in the weighted undirected time-space diagram, wherein the time characteristics are distributed photovoltaic resource output intermediate prediction data under corresponding environmental conditions at different moments;
extracting space-time characteristics of each node in the weighted undirected space-time diagram, wherein the space-time characteristics are space characteristics comprising time characteristics, and the space characteristics are obtained from the weighted undirected space-time diagram; capturing the output mode characteristics of each node in the space-time characteristics, splicing the time characteristics and the output mode characteristics of each node, inputting the spliced output mode characteristics and the output mode characteristics into a multi-layer fully-connected network, outputting a target output predicted value of the distributed photovoltaic resource in a corresponding area by utilizing the fully-connected network, and drawing a distributed photovoltaic resource output deduction curve.
As a preferred solution, the step of constructing a weighted dead-space plot to characterize the relationship between the distributed photovoltaic resources includes:
defining a power grid undirected graph on a time node t as G t =<V,E t ,F t >Wherein V is a set of nodes, E t For the collection of edges,the characteristic values of all nodes at the time t are obtained;
calculating historical photovoltaic output data of the node i and the node j at the current momentAnd->Correlation coefficient between
Defining an adjacent matrix A to represent the interconnection relation between a photovoltaic resource node i and a node j, wherein the element A (i, j) in the adjacent matrix A takes the value as follows:
where τ is the average of the output correlation coefficients of all node pairs, euc t (i, j) is historical photovoltaic output data of node i and node j at time tAnd->Euclidean distance between them;
and establishing a connection relation on the power grid undirected graph by the formula to form a weighted undirected time-space graph representing the association degree of different photovoltaic nodes.
As a preferred solution, the step of solving the time characteristics of each node in the weighted dead-time space graph includes:
establishing an LSTM network model which is a long-short-time memory network model, and for node data x at each time step t Updating cell state C using forget gate t :
f t =σ(W f ·[h t-1 ,x t ]+b f )
Wherein f t Indicating a forgetting gate for controlling the state C of the cell at the previous moment t-1 For the current cell state C t Is a degree of influence of (a);representing a cell status update value,/-, for>For C t Is influenced by the input gate i t Performing control;
i t =σ(W i ·[h t-1 ,x t ]+b i )
The output gate o in the LSTM network model is pressed t Update and calculate output h of output unit t :
o t =σ(W o ·[h t-1 ,x t ]+b o )
h t =o t ×tanh(C t )
Definition e t 、c t 、h t 、w t 、r t Representing historical weather conditions, temperature, humidity, wind speed and wind direction data, sw t 、lw t 、kt t Representing historical illumination solar intensity, vertical illumination intensity, and solar radiation factor data, wherein t is a time step;
the LSTM network model is two layers, regression prediction is carried out on historical output data of the photovoltaic power station by using the LSTM network model at the time t, the goal is to output photovoltaic output data at the time t+1, and then the photovoltaic output data at the time t is fitted periodically by using photovoltaic output data, weather data and illumination data of the first n time slices;
by combining the output of the two layers of LSTM network models, based on historical photovoltaic output data, a mapping function T and a corresponding parameter set theta are fitted i Outputting the time characteristics of each nodeThe expression is as follows:
wherein T is 1 And T 2 And respectively corresponding to the two layers of structures of the selected LSTM network model.
As a preferred solution, the step of solving the time characteristics of each node in the weighted dead-time space graph further includes: learning N Source-LSTM models from the Source domain, the N Source-LSTM models being pre-trained from a sufficient amount of data in the corresponding Source domain; generating basic output from target domain Base-LSTM modelGeneration of Source output from N Source-LSTM models>
the attention network is designed based on a multi-layer perceptron, and the following formula is shown:
(e 1,X ,...,e N+1,X )=f α (X;θ α )
wherein X is input data, θ α As a learnable parameter, f (·) represents prediction by LSTM network model, (e) 1,X ,...,e N+1,X ) Is of dimension ofAn output of (2); the output of the multi-layer perceptron is converted into the weight of the model by a soft attention mechanism, and the attention network generates a plurality of nonzero weights omega according to the following mode i :
output of complete model based on attention networkN outputs +.defined as Source-LSTM model>And Base-LSTM model output +.>The expression is as follows:
Further, in the target domain, the parameters of the model are updated according to two functions:
One is a parameter of the Base-LSTM model, which is output by the target domainAnd target domain true value y real Calculated loss function->Updating; the other is the parameters in the attention network, output according to the model as a whole +.>And target domain true value y real Calculated loss functionCount->Updating;
two independent counter-propagating exists in the Base-LSTM network and the attention network, and the loss function is calculated by using average absolute error, and the expression is as follows:
where n is the number of photovoltaic nodes,for the set of all node output predicted values, y is the set of all node output true values, +.>Predicted value of output for ith node, y (i) For the true value of the i-th node force, Θ is the set of all the learnable parameters.
As a preferred solution, the step of extracting the spatio-temporal features of each node in the weighted dead space graph includes: building a space-time diagram self-encoder ST-GAE, weighting input weighted non-directional space-time diagram sequence Coding and reconstructing, and capturing a data mode in the weighted undirected space-time diagram; learning depth coding mapping relation->Weighted unoriented time space diagram sequence g t Mapped as a series of hidden layer spatial vectors Φ=<Z t-m ,…,Z t-1 ,Z t >At Z t In the ith column, the weighted undirected time-space diagram G t The ith node in (b)Space-time characteristics; learning depth decoding mapping relation by designing two decoder modules, respectively, including coding of feature level +.>Coding with edge level->
As a preferred solution, the step of capturing the output mode characteristics of each node in the space-time characteristics, splicing the time characteristics and the output mode characteristics of each node, inputting the spliced output mode characteristics and the output mode characteristics into a multi-layer fully-connected network, and outputting the target output predicted value of the distributed photovoltaic resource in the corresponding area by using the fully-connected network includes:
for a time window t e [ t ' -m, t ' up to a specific point in time t ] ']The corresponding hidden layer space vector is used for calculating the mean vector as the coding characteristic according to the following formula:the ith column Zi is a feature vector coded by the corresponding ith node; for all Z i 1.ltoreq.i.ltoreq.n, with sparse correlation vector calculated as follows>Is a dictionary of: /> Capturing a weighted undirected time-space diagram sequence +.>Component F of the i-th node of the input data i 1.ltoreq.i.ltoreq.n, corresponding significant spatiotemporal patterns, dictionary D and sparse correlation vector ai are applied by:
wherein x is t The vector is measured for the network load,fitting values for the corresponding vectors, +.>Representing a fitting operation; in model training, sparse correlation vector a is randomly initialized i And a dictionary D, and performing back propagation optimization based on errors of the output predicted value and the true value of the model; based on target hidden layer space variable Z t T '-m is not less than t and not more than t', n nodes are +.>Solving to obtain dictionary D and n sparse modes +.>Then map (F) i ,a i ) Obtaining the output mode characteristic of the node i at the current moment>
The time characteristics obtained by solvingAnd output mode characteristic p t Stitching complete feature F as a photovoltaic node t And inputting a multi-layer full-connection network, and outputting the output predicted value of the distributed photovoltaic resource of the corresponding region by using the ReLU activation function.
In a second aspect, a distributed photovoltaic resource output curve deduction system is provided, including:
the time-space diagram construction module is used for inputting the geographic position of each distributed photovoltaic resource and the historical photovoltaic output data under the corresponding environmental conditions at different moments, and constructing a weighted undirected time-space diagram to characterize the relation among the distributed photovoltaic resources;
the node time feature solving module is used for solving the time feature of each node in the weighted undirected time-space diagram, wherein the time feature is the distributed photovoltaic resource output intermediate prediction data under the corresponding environmental conditions at different moments;
the curve drawing module is used for extracting the space-time characteristics of each node in the weighted undirected space-time diagram, wherein the space-time characteristics are the space characteristics comprising the time characteristics, and the space characteristics are obtained from the weighted undirected space-time diagram; capturing the output mode characteristics of each node in the space-time characteristics, splicing the time characteristics and the output mode characteristics of each node, inputting the spliced output mode characteristics and the output mode characteristics into a multi-layer fully-connected network, outputting a target output predicted value of the distributed photovoltaic resource in a corresponding area by utilizing the fully-connected network, and drawing a distributed photovoltaic resource output deduction curve.
As a preferred solution, the space-time diagram construction module defines a grid undirected graph on the time node t as G t =<V,E t ,F t >Wherein V is a set of nodes, E t For the collection of edges,the characteristic values of all nodes at the time t are obtained;
calculating historical photovoltaic output data of the node i and the node j at the current momentAnd->Correlation coefficient between
Defining an adjacent matrix A to represent the interconnection relation between a photovoltaic resource node i and a node j, wherein the element A (i, j) in the adjacent matrix A takes the value as follows:
where τ is the average of the output correlation coefficients of all node pairs, euc t (i, j) is historical photovoltaic output data of node i and node j at time tAnd->Euclidean distance between them;
and establishing a connection relation on the power grid undirected graph by the formula to form a weighted undirected time-space graph representing the association degree of different photovoltaic nodes.
As a preferable scheme, the node time feature solving module establishes an LSTM network model, wherein the LSTM network model is a long-short-time memory network model, and for node data x at each time step t Updating cell state C using forget gate t :
f t =σ(W f ·[h t-1 ,x t ]+b f )
Wherein f t Indicating a forgetting gate for controlling the state C of the cell at the previous moment t-1 For the current cell state C t Is a degree of influence of (a);representing a cell status update value,/-, for>For C t Is influenced by the input gate i t Performing control;
i t =σ(W i ·[h t-1 ,x t ]+b i )
The output gate o in the LSTM network model is pressed t Update and calculate output h of output unit t :
o t =σ(W o ·[h t-1 ,x t ]+b o )
h t =o t ×tanh(C t )
Definition e t 、c t 、h t 、w t 、r t Representing historical weather conditions, temperature, humidity, wind speed and wind direction data, sw t 、lw t 、kt t Representing historical illumination solar intensity, vertical illumination intensity, and solar radiation factor data, wherein t is a time step;
the LSTM network model is two layers, regression prediction is carried out on historical output data of the photovoltaic power station by using the LSTM network model at the time t, the goal is to output photovoltaic output data at the time t+ 1, and then the photovoltaic output data at the time t is fitted periodically by using photovoltaic output data, weather data and illumination data of the first n time slices;
by combining the output of the two layers of LSTM network models, based on historical photovoltaic output data, a mapping function T and a corresponding parameter set theta are fitted i Outputting the time characteristics of each nodeThe expression is as follows:
wherein T is 1 And T 2 And respectively corresponding to the two layers of structures of the selected LSTM network model.
As a preferred embodiment, the The node time feature solving module also learns N Source-LSTM models from the Source domain, wherein the N Source-LSTM models are generated by pre-training a sufficient amount of data in the corresponding Source domain; generating basic output from target domain Base-LSTM modelGeneration of Source output from N Source-LSTM models>
the attention network is designed based on a multi-layer perceptron, and the following formula is shown:
(e 1,X ,...,e N+1,X )=f α (X;θ α )
wherein X is input data, θ α As a learnable parameter, f (·) represents prediction by LSTM network model, (e) 1,X ,...,e N+1,X ) Is of dimension ofAn output of (2); the output of the multi-layer perceptron is converted into the weight of the model by a soft attention mechanism, and the attention network generates a plurality of nonzero weights omega according to the following mode i :
based on the attention network, the complete model is inputOut ofN outputs +.defined as Source-LSTM model>And Base-LSTM model output +.>The expression is as follows:
in the target domain, the parameters of the model are updated according to two functions:
one is a parameter of the Base-LSTM model, which is output by the target domainAnd target domain true value y real Calculated loss function->Updating; the other is the parameters in the attention network, output according to the model as a whole +. >And target domain true value y real Calculated loss function->Updating;
two independent counter-propagating exists in the Base-LSTM network and the attention network, and the loss function is calculated by using average absolute error, and the expression is as follows:
where n is the number of photovoltaic nodes,for the set of all node output predicted values, y is the set of all node output true values, +.>Predicted value of output for ith node, y (i) For the true value of the i-th node force, Θ is the set of all the learnable parameters.
As a preferred solution, the step of extracting the spatio-temporal characteristics of each node in the weighted undirected space-time graph by the curve drawing module includes: building a space-time diagram self-encoder ST-GAE, weighting input weighted non-directional space-time diagram sequence Coding and reconstructing, and capturing a data mode in the weighted undirected space-time diagram; learning depth coding mapping relation->Weighted undirected time-space diagram sequence +.>Mapped as a series of hidden layer spatial vectors Φ=<Z t-m ,…,Z t-1 ,Z t >At Z t In the ith column, the weighted undirected time-space diagram G t Space-time characteristics of the ith node; learning depth decoding mapping relation by designing two decoder modules, respectively, including coding of feature level +.>Coding with edge level- >
As a preferred solution, the step of capturing the output mode characteristics of each node in the space-time characteristics by the curve drawing module, splicing the time characteristics and the output mode characteristics of each node, inputting the spliced output mode characteristics and the output mode characteristics into a multi-layer fully-connected network, and outputting the target output predicted value of the distributed photovoltaic resource in the corresponding area by using the fully-connected network includes:
for a time window t e [ t ' -m, t ' up to a specific point in time t ] ']The corresponding hidden layer space vector is used for calculating the mean vector as the coding characteristic according to the following formula:column i Z i The feature vector after being coded for the corresponding ith node; for all Z i 1.ltoreq.i.ltoreq.n, with sparse correlation vector calculated as follows>Is a dictionary of: /> Capturing a weighted undirected time-space diagram sequence g t Component F of the i-th node of the input data i 1.ltoreq.i.ltoreq.n, corresponding salient spatio-temporal patterns, dictionary D and sparse correlation vector a i The application is performed by the following way:
wherein x is t The vector is measured for the network load,fitting values for the corresponding vectors, +.>Representing a fitting operation; in model training, sparse correlation vector a is randomly initialized i And a dictionary D, and performing back propagation optimization based on errors of the output predicted value and the true value of the model; based on target hidden layer space variable Z t T '-m is not less than t and not more than t', n nodes are +.>Solving to obtain dictionary D and n sparse modes +.>Then map (F) i ,a i ) Obtaining the output mode characteristic of the node i at the current moment>
The time characteristics obtained by solvingAnd output mode characteristic p t Stitching complete feature F as a photovoltaic node t And inputting a multi-layer full-connection network, and outputting the output predicted value of the distributed photovoltaic resource of the corresponding region by using the ReLU activation function.
In a third aspect, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, where the processor implements the steps of the distributed photovoltaic resource output curve deduction method when executing the computer program.
In a fourth aspect, a computer readable storage medium is provided, where a computer program is stored, where the computer program when executed by a processor implements the steps of the distributed photovoltaic resource output curve deduction method.
Compared with the prior art, the first aspect of the invention has at least the following beneficial effects:
the relation among the distributed photovoltaic resources is characterized by using a weighted undirected time-space diagram, the relation among the photovoltaic resources is effectively constructed, reasoning is carried out from unstructured data, the time sequence and the non-time sequence characteristics of the relation are learned, and the dependence relation of different nodes on the power grid output in a local and overall mode is further characterized. Meanwhile, the time-space characteristics of each node in the weighted undirected time-space diagram are extracted, the output mode characteristics of each node in the time-space characteristics are captured, the time characteristics and the output mode characteristics of each node are spliced and used as the complete characteristics of the photovoltaic nodes to be input into a multi-layer full-connection network, the full-connection network is utilized to output the target output predicted value of the distributed photovoltaic resource of the corresponding region, and then the distributed photovoltaic resource output deduction curve is drawn. According to the method, the data of various sources can be comprehensively utilized, the time feature extraction is carried out by combining with the weather and illumination information of the target region, the space feature extraction is carried out on the photovoltaic module of the target region based on the deep learning algorithm driven by the graph network, extremely high observation data and a large amount of historical data are not required to be relied on, the output of the distributed photovoltaic resource under extreme weather conditions and sudden conditions can be accurately and reasonably predicted, and the prediction performance requirement of a long-term output curve when the distributed photovoltaic resource is actually deployed in a large scale is met.
It will be appreciated that the advantages of the second to fourth aspects may be found in the relevant description of the first aspect and are not repeated here.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings that are required for the embodiments or the description of the prior art will be briefly described below, it being obvious that the drawings in the following description are only some embodiments of the present application, and that other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.
FIG. 1 is a flow chart of a distributed photovoltaic resource output curve deduction method according to an embodiment of the invention;
FIG. 2 is a schematic diagram of a time feature extraction model based on an LSTM network according to an embodiment of the present invention;
FIG. 3 is a schematic diagram of an LSTM-based attention transfer learning model structure according to an embodiment of the present invention;
FIG. 4 is a schematic diagram of a structure of a ST-GAE (space-time diagram self-encoder) model according to an embodiment of the present invention;
FIG. 5 is a schematic diagram of a multi-layer fully connected network according to an embodiment of the present invention.
Detailed Description
In the following description, for purposes of explanation and not limitation, specific details are set forth, such as particular system configurations, techniques, etc. in order to provide a thorough understanding of the embodiments of the present application. It will be apparent, however, to one skilled in the art that the present application may be practiced in other embodiments that depart from these specific details. In other instances, detailed descriptions of well-known systems, devices, circuits, and methods are omitted so as not to obscure the description of the present application with unnecessary detail.
In addition, in the description of the present application and the appended claims, the terms "first," "second," "third," and the like are used merely to distinguish between descriptions and are not to be construed as indicating or implying relative importance.
Example 1
Referring to fig. 1, a distributed photovoltaic resource output curve deduction method according to an embodiment of the present invention includes the following steps:
s1: building a graph model of the output deduction system;
firstly, constructing a graph structure based on historical data of photovoltaic nodes, and constructing an undirected graph G under a time node t t =<V,E t ,F t >To characterize the situation of all nodes at time t, where V is the set of all photovoltaic nodes, E t Is a collection of edges, and F t Which is used to represent the eigenvalues of all nodes on the time node t. Because the output of the photovoltaic node is greatly influenced by the environment, the nodes with close geographic positions have larger correlation, and the invention regards the correlation between the two nodes with the correlation coefficient larger than the average level as the correlation, thereby constructing edges for the nodes and finally forming a weighted non-directional time space diagram.
S2: a temporal feature extraction scheme based on LSTM (long short time memory network), as shown in fig. 2;
in order to extract time-dependent features of input data, LSTM model is used to model time-sequence dependency relationship and make it persistent, for input data x t First, updating the cell state C according to the content of the forget gate t Forgetting door f t For controlling the state C of the cell at the previous moment t-1 To what extent the current cell state C is affected t Input gate i t Control unit state update valueFor C t To the output gate o in LSTM t Update and calculate output h of output unit t . And (3) designing a two-layer LSTM model to carry out regression prediction on the historical output data of the photovoltaic power station at the time t, and outputting the photovoltaic output data at the time t+ 1.
S3: attention transfer learning model based on LSTM (long short term memory network), as shown in fig. 3;
the LSTM model has the problem of insufficient available service data in an actual deployment scene, so that the time characteristics of model learning are inaccurate. For this purpose, the invention proposes an LSTM-based attention transfer learning model. Because the Source domain contains enough training data, the migration learning learns N Source-LSTM models from the Source domain, so that abundant time characteristics are modeled, and a prediction task can be better assisted; the Base-LSTM network of the target domain has low prediction accuracy due to limited data of the real training scene, so that the model of the source domain is migrated to the target domain through a learning and weight-distributing attention network, and the time feature extraction of the target domain is helped under the condition of limited data.
S4: a spatio-temporal feature extraction scheme based on ST-GAE (space-time diagram self-encoder), as shown in fig. 4;
based on the weighted undirected space-time diagram, firstly constructing a space-time diagram self-encoder ST-GAE to encode and reconstruct an input space-time diagram sequence, and learning a depth coding mapping relationThereby will->Mapped into a series of hidden layer spatial vectors Φ=<Z t-m ,…,Z t-1 ,Z t >Two decoder modules are then designed to learn depth decoding mappings, facilitating potential spatiotemporal features in the ST-GAE learning map through such an encoding-decoding architecture.
S5: future photovoltaic system output curve deduction based on space-time diagram dictionary learning;
based on the above ST-GAE, time window t E [ t '-m, t ]']The hidden layer space vector in (1) calculates the mean value Z thereof, wherein Z i Features encoded for the corresponding i-th node. Then calculate the correlation vector a with sparse i For capturing significant spatio-temporal patterns in the sequence of input spatio-temporal patterns. Solving to obtain a target hidden layer space variable Z t Then, n nodesOptimizing to obtain dictionary D and n sparse modes therein>Subsequently, by mapping (F i ,a i ) The characteristic sequence of the output mode for the node i at the current moment can be obtained>
Based on the attention migration model of the long-short-time memory network, the time-space diagram self-encoder and the dictionary learning model, fine granularity time-space characteristic information of the photovoltaic node is obtained, and finally the output mode characteristic p is spliced t And time characteristicsInput multi-layer full-connection network, as shown in figure 5, realizes the future time lightAnd (5) deduction of the output curve of the photovoltaic system.
According to the scheme provided by the invention, based on the actually deployed multi-source sensor data, the data of various sources can be comprehensively utilized, the time feature extraction is carried out by combining the meteorological information and the illumination information of a target region, the space feature extraction is carried out on the photovoltaic module of the target region based on a deep learning algorithm driven by a graph network, and the proper mapping relation is learned according to the weather conditions of a long term and a short term, so that the future output curve of the county distributed photovoltaic resource is deduced. Three major parts of the process according to the invention are described in detail below.
The photovoltaic system correlation analysis and modeling based on the space-time diagram model are carried out in the first part;
firstly, in order to analyze and model the correlation between distributed photovoltaic resources in a county scene, the invention constructs the relation between the county distributed photovoltaic resources based on a space-time diagram model, reasoning is carried out from unstructured data, and the time sequence and non-time sequence characteristics are learned, so that the dependence relationship of different nodes on the power grid output in a local and whole way is characterized. The scheme is characterized based on a series of time-space diagrams to effectively construct the relation between photovoltaic resources.
Specifically, the undirected graph of the power grid on the time node t is defined as G t =<V,E t ,F t >Wherein V is a set of nodes, E t Then it is the collection of edges,the characteristic values of all nodes at the time t are obtained. The output of the photovoltaic nodes is greatly influenced by the environment, photovoltaic units which are distributed nearer in geographic positions can share weather environments such as similar solar radiation, cloud layer measurement values, wind speeds and the like, the output correlation among photovoltaic resource nodes is described in one step, and when the high environmental correlation exists between the node i and the node j, the node i is connected with the node j. In order to characterize the correlation, historical photovoltaic output data of the node i and the node j at the current moment can be calculated based on cosine similarity and other methods>And->Correlation coefficient between
Furthermore, the defined adjacency matrix a represents the interconnection relationship between the photovoltaic resource node i and the node j. The element A (i, j) in the adjacent matrix A takes the following values:
where τ is the average of the output correlation coefficients of all node pairs, euc t (i, j) is historical photovoltaic output data of node i and node j at time tAnd->Euclidean distance between them. The present invention considers graph G at corresponding time when the historical output data correlation coefficient of two photovoltaic power stations is higher than the average value level t There is an association relationship. Based on the formula, the connection relation on the graph can be constructed, and finally the weighted undirected space-time graph representing the association of different photovoltaic nodes is formed.
The second part, time characteristic extraction and attention transfer learning based on LSTM (long short time memory) network;
2.1 extracting a model based on the time characteristics of the LSTM network;
based on the weighted undirected time-space diagram, the invention performs multidimensional mining on the time-space characteristics among the nodes so as to solve the time characteristics of the nodes in the diagram. Specifically, the LSTM (long and short time memory) network model is used to process the weighted dead space graph and further solve the time characteristics of the nodes. The model is presented for the number of nodes per time stepAccording to x t Updating cell state C using forget gate t :
f t =σ(W f ·[h t-1 ,x t ]+b f )
Wherein f t Indicating a forgetting gate for controlling the state C of the cell at the previous moment t-1 To what extent the current cell state C is affected t 。Representing a cell state update value for C t Is influenced by the input gate i t And controlling.
i t =σ(W i ·[h t-1 ,x t ]+b i )
Finally, for the output gate o in LSTM t Update and calculate output h of output unit t As a spatio-temporal feature of the node.
o t =σ(W o ·[h t-1 ,x t ]+b o )
h t =o t ×tanh(C t )
Considering that the state of the photovoltaic module is greatly related to the past information, the invention uses the LSTM network to extract the dependency relationship of time sequence data so as to realize the extraction of the time characteristics of the distributed photovoltaic scene. Definition e t 、c t 、h t 、w t 、r t Representing historical weather conditionsCondition, temperature, humidity, wind speed and wind direction data, sw t 、lw t 、kt t Historical illumination solar intensity, vertical illumination intensity, and solar radiation factor data are represented, where t is a time step. In order to realize the depiction of the output curve information of time-series data on different granularities, the invention constructs a two-layer LSTM model. At time t, regression prediction is performed on historical output data of a certain photovoltaic power station by using LSTM, and the aim is to output photovoltaic output data at time t+ 1. Subsequently, the photovoltaic output data at time t is fitted periodically using the photovoltaic output data, weather data, illumination data on the first n time slices, i.e., (t-1), (t-2), … (t-n). Furthermore, by combining the outputs of the two LSTM modules, the mapping function T and the corresponding parameter set theta are fitted based on the historical photovoltaic output data i Outputting the time characteristics of each node
Wherein T is 1 And T 2 Respectively correspond to the LSTM two-layer structure.
Particularly, considering that the proposed scheme has insufficient available service data in an actual deployment scene, so that LSTM model training is difficult to develop, the invention further provides an LSTM-based attention transfer learning model.
2.2 LSTM-based attention transfer learning model
Inspired by the application of attention and adaptive migration learning in deep learning, the invention proposes an LSTM model-based attention migration learning framework for transferring knowledge from a source domain trained LSTM model. The framework learns abundant space-time characteristics and dependency relations by utilizing the existing large-scale data set in a source domain, trains data with limited characteristics in the scene in a target domain, combines the two data through a attention network, updates parameters through a loss function, and completes a migration learning task.
For the Source domain part in the migration learning, N Source-LSTM models are learned from the Source domain. Considering that the LSTM model trained by the Source domain data can extract abundant time characteristics, the method has better performance on regression prediction tasks, and the migration learning framework of the Source-LSTM is designed to migrate the abundant time characteristic knowledge in the Source domain to the target domain.
For the target domain part, because the data in the target domain are all from real scenes, the Base-LSTM model trained under the condition of limited target domain data is often not high in prediction accuracy and needs to be combined with rich time features learned in the source domain. The N Source-LSTM models are generated by pre-training a sufficient amount of data in the corresponding Source domain. Base-LSTM model generation basic output N Source model Generation Source output +.>
The invention designs an attention network to output the modelAnd->And (5) performing weight processing. The attention network is designed based on a Multi-Layer Perceptron (MLP), which can be expressed as:
(e 1,X ,...,e N+1,X )=f α (X;θ α )
wherein X is input data, θ α As a learnable parameter, f (·) represents prediction by LSTM model, (e) 1,X ,...,e N+1,X ) Is of dimension ofIs provided. The output of the MLP is converted to weights of the model by the soft attention mechanism,the attention network may generate a plurality of non-zero weights ω i :
Output of complete model based on attention networkN outputs +.defined as Source-LSTM model>And Base-LSTM model output +.>Is a weighted sum of:
the present invention contemplates different methods to update parameters in a transfer learning and attention network. Source-LSTM models of Source domains are pre-trained from Source domain data, and their parameters remain unchanged throughout the training process of the migration model. In the target domain, the parameters of the model are updated according to two functions: one is the Base-LSTM parameter, output by the target domainAnd target domain true value y real Calculated loss function->Updating; one is the parameters in the attention network, output according to the model as a whole +. >And target domain true value y real Calculated loss functionUpdating.
There are two independent counter-propagating in the Base-LSTM network and the attention network. Both loss functions are calculated using the average absolute error:
where n is the number of photovoltaic nodes,for the set of all node output predicted values, y is the set of all node output true values, +.>Predicted value of output for ith node, y (i) For the true value of the i-th node force, Θ is the set of all the learnable parameters.
The third part, space-time feature extraction scheme and future photovoltaic system output curve deduction;
based on the weighted undirected time-space diagram of the photovoltaic node, the invention adopts ST-GAE (space-time diagram self-encoder) to extract the time-space characteristics among the nodes in the weighted undirected time-space diagram, and further adopts a fully connected network to realize the deduction of the output curve of the future photovoltaic system.
3.1 a spatio-temporal feature extraction scheme based on an ST-GAE (spatio-temporal graph self-encoder) network model;
first, a space-time diagram sequence of space-time diagram self-encoder ST-GAE pair input is established Encoding and reconstruction are performed to capture meaningful data patterns in the space-time diagram. In this process, learn depth coding mapping relation +. >Thereby will->Mapped as a series of hidden layer spatial vectors Φ=<Z t-m ,…,Z t-1 ,Z t >。Z t In the ith list view G t Spatio-temporal characteristics of the i-th node in (a). On the basis, two decoder modules are designed to respectively learn depth decoding mapping relation, wherein the depth decoding mapping relation comprises coding of characteristic level +.>Coding with edge level->These two encoding-decoding architectures help ST-GAE learn Z t A highly nonlinear spatiotemporal latent feature.
3.2, deducing a future photovoltaic system output curve based on space-time diagram dictionary learning;
next, the present invention performs deep mining on the time space diagram. For a time window t e [ t ' -m, t ' up to a specific point in time t ] ']Corresponding hidden layer space vector, calculating average value vector as coding characteristicIts ith column Z i And the feature vector is coded for the corresponding ith node. For all Z i (1.ltoreq.i.ltoreq.n) for which sparse correlation vectors are calculated>Dictionary of->Thereby capturing->Component F of the i-th node in the input data i (1.ltoreq.i.ltoreq.n) a corresponding significant spatiotemporal pattern. The aforementioned dictionary D and sparse correlation vector a i The method is applied by the following steps:
wherein x is t The vector is measured for the network load,fitting values for the corresponding vectors, +.>Representing a fitting operation; in model training, sparse correlation vector a is randomly initialized i And a dictionary D, and performing back propagation optimization based on errors of the output predicted value and the actual value of the model. Based on target hidden layer space variable Z t (t '-m.ltoreq.t.ltoreq.t') may be n nodes +.>Solving to obtain dictionary D and n sparse modes +.>Subsequently, by mapping (F i ,a i ) The output mode characteristic for the node i at the current moment can be obtained>
Finally, the time characteristics obtained by the solution are calculatedAnd output mode characteristic p t Stitching complete feature F as a photovoltaic node t And inputting a multi-layer full-connection network, and outputting the output predicted value of the distributed photovoltaic resource of the corresponding region by using the ReLU activation function.
Example 2
The invention also provides a distributed photovoltaic resource output curve deduction system, which comprises:
the time-space diagram construction module is used for inputting the geographic position of each distributed photovoltaic resource and the historical photovoltaic output data under the corresponding environmental conditions at different moments, and constructing a weighted undirected time-space diagram to characterize the relation among the distributed photovoltaic resources;
the node time feature solving module is used for solving the time feature of each node in the weighted undirected time-space diagram, wherein the time feature is the distributed photovoltaic resource output intermediate prediction data under the corresponding environmental conditions at different moments;
The curve drawing module is used for extracting the space-time characteristics of each node in the weighted undirected space-time diagram, wherein the space-time characteristics are the space characteristics comprising the time characteristics, and the space characteristics are obtained from the weighted undirected space-time diagram; capturing the output mode characteristics of each node in the space-time characteristics, splicing the time characteristics and the output mode characteristics of each node, inputting the spliced output mode characteristics and the output mode characteristics into a multi-layer fully-connected network, outputting a target output predicted value of the distributed photovoltaic resource in a corresponding area by utilizing the fully-connected network, and drawing a distributed photovoltaic resource output deduction curve.
In one possible embodiment, the time-space diagram construction module defines the grid undirected graph on the time node t as G t =<V,E t ,F t >Wherein V is a set of nodes, E t For the collection of edges,the characteristic values of all nodes at the time t are obtained;
calculating historical photovoltaic output data of the node i and the node j at the current momentAnd->Correlation coefficient between
Defining an adjacent matrix A to represent the interconnection relation between a photovoltaic resource node i and a node j, wherein the element A (i, j) in the adjacent matrix A takes the value as follows:
where τ is the average of the output correlation coefficients of all node pairs, euc t (i, j) is historical photovoltaic output data of node i and node j at time t And->Euclidean distance between them;
and establishing a connection relation on the power grid undirected graph by the formula to form a weighted undirected time-space graph representing the association degree of different photovoltaic nodes.
In one possible implementation, the node time feature solving module builds an LSTM network model for node data x at each time step t Updating cell state C using forget gate t :
f t =σ(W f ·[h t-1 ,x t ]+b f )
Wherein f t Indicating a forgetting gate for controlling the state C of the cell at the previous moment t-1 For the current cell state C t Is a degree of influence of (a);representing a cell status update value,/-, for>For C t Is controlled by the input gate it;
i t =σ(W i ·[h t-1 ,x t ]+b i )
The output gate o in the LSTM network model is pressed t Update and calculate output h of output unit t :
o t =σ(W o ·[h t-1 ,x t ]+b o )
h t =o t ×tanh(C t )
Definition e t 、c t 、h t 、w t 、r t Representing historical weather conditions, temperature, humidity, wind speed and wind direction data, sw t 、lw t 、kt t Representing historical illumination solar intensity, vertical illumination intensity, and solar radiation factor data, wherein t is a time step;
the LSTM network model is two layers, regression prediction is carried out on historical output data of the photovoltaic power station by using the LSTM network model at the time t, the goal is to output photovoltaic output data at the time t+ 1, and then the photovoltaic output data at the time t is fitted periodically by using photovoltaic output data, weather data and illumination data of the first n time slices;
By combining the outputs of the two-layer LSTM network model, the baseFitting a mapping function T and a corresponding parameter set theta to historical photovoltaic output data i Outputting the time characteristics of each nodeThe expression is as follows:
wherein T is 1 And T 2 And respectively corresponding to the two layers of structures of the selected LSTM network model.
In one possible implementation, the node temporal feature solution module further learns N Source-LSTM models from the Source domain, the N Source-LSTM models being generated by a sufficient amount of data pre-training in the respective Source domain; generating basic output from target domain Base-LSTM modelGeneration of Source output from N Source-LSTM models>
the attention network is designed based on a multi-layer perceptron, and the following formula is shown:
(e 1,X ,...,e N+1,X )=f α (X;θ α )
wherein X is input data, θ α As a learnable parameter, f (·) represents prediction by LSTM network model, (e) 1,X ,...,e N+1,X ) Is of dimension ofAn output of (2); the output of the multi-layer perceptron is converted into the weight of the model by a soft attention mechanism, and the attention network generates a plurality of nonzero weights omega according to the following mode i :
output of complete model based on attention networkN outputs +.defined as Source-LSTM model >And Base-LSTM model output +.>The expression is as follows:
in the target domain, the parameters of the model are updated according to two functions:
one is a parameter of the Base-LSTM model, which is output by the target domainAnd the target domain true value yr eal Calculated loss function->Updating; the other is the parameters in the attention network, output according to the model as a whole +.>And target domain true value y real Calculated loss function->Updating;
two independent counter-propagating exists in the Base-LSTM network and the attention network, and the loss function is calculated by using average absolute error, and the expression is as follows:
where n is the number of photovoltaic nodes,for the set of all node output predicted values, y is the set of all node output true values, +.>Predicted value of output for ith node, y (i) For the true value of the i-th node force, Θ is the set of all the learnable parameters.
In one possible implementation, the step of extracting the spatio-temporal features of each node in the weighted undirected time-space graph by the curve plotting module includes: building a space-time diagram self-encoder ST-GAE, weighting input weighted non-directional space-time diagram sequence Coding and reconstructing, and capturing a data mode in the weighted undirected space-time diagram; learning depth coding mapping relation- >Weighted undirected time-space diagram sequence +.>Mapped as a series of hidden layer spatial vectors Φ=<Z t-m ,…,Z t-1 ,Z t >At Z t In the ith column, the weighted undirected time-space diagram G t Space-time characteristics of the ith node; learning depth decoding mapping relation by designing two decoder modules, respectively, including coding of feature level +.>Coding with edge level->
In one possible implementation manner, the step of capturing the output mode characteristics of each node in the space-time characteristics by the curve drawing module, splicing the time characteristics and the output mode characteristics of each node, inputting the spliced output mode characteristics and the output mode characteristics into the multi-layer fully-connected network, and outputting the target output predicted value of the distributed photovoltaic resource of the corresponding area by using the fully-connected network comprises the following steps:
for a time window t e [ t ' -m, t ' up to a specific point in time t ] ']The corresponding hidden layer space vector is used for calculating the mean vector as the coding characteristic according to the following formula:column i Z i The feature vector after being coded for the corresponding ith node; for all Z i 1.ltoreq.i.ltoreq.n, with sparse correlation vector calculated as follows>Is a dictionary of: /> Capturing a weighted undirected time-space diagram sequence +.>Component F of the i-th node of the input data i 1.ltoreq.i.ltoreq.n, corresponding salient spatio-temporal patterns, dictionary D and sparse correlation vector a i The application is performed by the following way:
Wherein x is t The vector is measured for the network load,fitting values for the corresponding vectors, +.>Representing a fitting operation; in model training, sparse correlation vector a is randomly initialized i And a dictionary D, and performing back propagation optimization based on errors of the output predicted value and the true value of the model; based on target hidden layer space variable Z t T '-m is not less than t and not more than t', n nodes are +.>Solving to obtain dictionary D and n sparse modes +.>Then map (F) i ,a i ) Obtaining the output mode characteristic of the node i at the current moment>
The time characteristics obtained by solvingAnd output mode characteristic p t Stitching as a complete feature of photovoltaic nodesF t And inputting a multi-layer full-connection network, and outputting the output predicted value of the distributed photovoltaic resource of the corresponding region by using the ReLU activation function.
Example 3
The invention also provides an electronic device, which comprises a memory, a processor and a computer program stored in the memory and capable of running on the processor, wherein the processor realizes the steps of the distributed photovoltaic resource output curve deduction method when executing the computer program.
Example 4
Another embodiment of the present invention also proposes a computer readable storage medium storing a computer program, which when executed by a processor implements the steps of the distributed photovoltaic resource output curve deduction method.
The computer program comprises computer program code which may be in source code form, object code form, executable file or in some intermediate form, etc. The computer readable storage medium may include: any entity or device, medium, usb disk, removable hard disk, magnetic disk, optical disk, computer memory, read-only memory, random access memory, electrical carrier wave signals, telecommunications signals, software distribution media, and the like capable of carrying the computer program code. 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. For convenience of description, the foregoing disclosure shows only those parts relevant to the embodiments of the present invention, and specific technical details are not disclosed, but reference is made to the method parts of the embodiments of the present invention. The computer readable storage medium is non-transitory and can be stored in a storage device formed by various electronic devices, and can implement the execution procedure described in the method according to the embodiment of the present invention.
It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as a method, system, or computer program product. Accordingly, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.
The present invention is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the invention. It will be understood that each flowchart and/or block of the flowchart illustrations and/or block diagrams, and combinations of flowcharts and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
Finally, it should be noted that: the above embodiments are only for illustrating the technical aspects of the present invention and not for limiting the same, and although the present invention has been described in detail with reference to the above embodiments, it should be understood by those of ordinary skill in the art that: modifications and equivalents may be made to the specific embodiments of the invention without departing from the spirit and scope of the invention, which is intended to be covered by the claims.
Claims (15)
1. The distributed photovoltaic resource output curve deduction method is characterized by comprising the following steps of:
inputting geographic positions of distributed photovoltaic resources and historical photovoltaic output data under corresponding environmental conditions at different moments, and constructing a weighted undirected time-space diagram to characterize the relation among the distributed photovoltaic resources;
solving time characteristics of each node in the weighted undirected time-space diagram, wherein the time characteristics are distributed photovoltaic resource output intermediate prediction data under corresponding environmental conditions at different moments;
extracting space-time characteristics of each node in the weighted undirected space-time diagram, wherein the space-time characteristics are space characteristics comprising time characteristics, and the space characteristics are obtained from the weighted undirected space-time diagram; capturing the output mode characteristics of each node in the space-time characteristics, splicing the time characteristics and the output mode characteristics of each node, inputting the spliced output mode characteristics and the output mode characteristics into a multi-layer fully-connected network, outputting a target output predicted value of the distributed photovoltaic resource in a corresponding area by utilizing the fully-connected network, and drawing a distributed photovoltaic resource output deduction curve.
2. The method of claim 1, wherein the step of constructing a weighted dead-time graph to characterize the relationship between the distributed photovoltaic resources comprises:
Defining a power grid undirected graph on a time node t as G t =<V,E t ,F t >Wherein V is a set of nodes, E t For the collection of edges,at tEtching down characteristic values of all nodes;
calculating historical photovoltaic output data of the node i and the node j at the current momentAnd->Correlation coefficient between->
Defining an adjacent matrix A to represent the interconnection relation between a photovoltaic resource node i and a node j, wherein the element A (i, j) in the adjacent matrix A takes the value as follows:
where τ is the average of the output correlation coefficients of all node pairs, euc t (i, j) is historical photovoltaic output data of node i and node j at time tAnd->Euclidean distance between them;
and establishing a connection relation on the power grid undirected graph by the formula to form a weighted undirected time-space graph representing the association degree of different photovoltaic nodes.
3. The method of claim 1, wherein the step of solving the time characteristics of each node in the weighted dead-time graph comprises:
establishing an LSTM network model which is a long-short-time memory network model, and for node data x at each time step t Make the followingUpdating cell state C with forget gate t :
f t =σ(W f ·[h t-1 ,x t ]+b f )
Wherein f t Indicating a forgetting gate for controlling the state C of the cell at the previous moment t-1 For the current cell state C t Is a degree of influence of (a);representing a cell status update value,/-, for>For C t Is influenced by the input gate i t Performing control;
i t =σ(W i ·[h t-1 ,x t ]+b i )
The output gate o in the LSTM network model is pressed t Update and calculate output h of output unit t :
o t =σ(W o ·[h t-1 ,x t ]+b o )
h t =o t ×tanh(C t )
Definition e t 、c t 、h t 、w t 、r t Representing historical weather conditions, temperature, humidity, wind speed and wind direction data, sw t 、lw t 、kt t Representing historical illumination solar intensity, vertical illumination intensity, and solar radiation factor data, wherein t is a time step;
the LSTM network model is two layers, regression prediction is carried out on historical output data of the photovoltaic power station by using the LSTM network model at the time t, the goal is to output photovoltaic output data at the time t+1, and then the photovoltaic output data at the time t is fitted periodically by using photovoltaic output data, weather data and illumination data of the first n time slices;
by combining the output of the two layers of LSTM network models, based on historical photovoltaic output data, a mapping function T and a corresponding parameter set theta are fitted i Outputting the time characteristics of each nodeThe expression is as follows:
wherein T is 1 And T 2 And respectively corresponding to the two layers of structures of the selected LSTM network model.
4. The method of claim 3, wherein the step of solving the time characteristics of each node in the weighted dead-time graph further comprises: learning N Source-LSTM models from the Source domain, the N Source-LSTM models being pre-trained from a sufficient amount of data in the corresponding Source domain; generating basic output from target domain Base-LSTM model Generation of Source output from N Source-LSTM models>
the attention network is designed based on a multi-layer perceptron, and the following formula is shown:
(e 1,X ,…,e N+1,X )=f α (X;θ α )
wherein X is input data, θ α As a learnable parameter, f (·) represents prediction by LSTM network model, (e) 1,X ,...,e N+1,X ) Is of dimension ofAn output of (2); the output of the multi-layer perceptron is converted into the weight of the model by a soft attention mechanism, and the attention network generates a plurality of nonzero weights omega according to the following mode i :
output of complete model based on attention networkN outputs +.defined as Source-LSTM model>And Base-LSTM model output +.>The expression is as follows:
5. The method of claim 4, wherein in the target domain, parameters of the model are updated according to two functions:
one is a parameter of the Base-LSTM model, which is output by the target domainAnd target domain true value y real Calculated loss function->Updating; the other is the parameters in the attention network, output according to the model as a whole +.>And target domain true value y real Calculated loss function- >Updating;
two independent counter-propagating exists in the Base-LSTM network and the attention network, and the loss function is calculated by using average absolute error, and the expression is as follows:
6. The method for deriving a distributed photovoltaic resource output curve according to claim 4, wherein the step of extracting the spatio-temporal features of each node in the weighted undirected spatio-temporal graph comprises: building a space-time diagram self-encoder ST-GAE, weighting input weighted non-directional space-time diagram sequenceCoding and reconstructing, and capturing a data mode in the weighted undirected space-time diagram; learning depth coding mapping relation->Weighted undirected time-space diagram sequence +.>Mapped as a series of hidden layer spatial vectors Φ=<Z t-m ,…,Z t-1 ,Z t >At Z t In the ith column, the weighted undirected time-space diagram G t Space-time characteristics of the ith node; learning depth decoding mapping relation by designing two decoder modules, respectively, including coding of feature level +.>Coding with edge level- >
7. The method according to claim 6, wherein the step of capturing the output mode characteristics of each node in the space-time characteristics, splicing the time characteristics and the output mode characteristics of each node, inputting the spliced output mode characteristics and the output mode characteristics into a multi-layer fully-connected network, and outputting the target output predicted value of the distributed photovoltaic resource in the corresponding area by using the fully-connected network comprises the steps of:
for a time window t e [ t ' -m, t ' up to a specific point in time t ] ']The corresponding hidden layer space vector is used for calculating the mean vector as the coding characteristic according to the following formula:column i Z i The feature vector after being coded for the corresponding ith node; for all Z i 1.ltoreq.i.ltoreq.n, with sparse correlation vector calculated as follows>Is a dictionary of: /> Capturing a weighted undirected time-space diagram sequence +.>Component F of the i-th node of the input data i 1.ltoreq.i.ltoreq.n, corresponding salient spatio-temporal patterns, dictionary D and sparse correlation vector a i The application is performed by the following way:
wherein x is t The vector is measured for the network load,fitting values for the corresponding vectors, +.>Representing a fitting operation; in model training, sparse correlation vector a is randomly initialized i And a dictionary D, and performing back propagation optimization based on errors of the output predicted value and the true value of the model; based on target hidden layer space variable Z t T '-m is not less than t and not more than t', n nodes are +.>Solving to obtain dictionary D and n sparse modes +.>Then map (F) i ,a i ) Obtaining the output mode characteristic of the node i at the current moment>
The time characteristics obtained by solvingAnd output mode characteristic p t Stitching complete feature F as a photovoltaic node t And inputting a multi-layer full-connection network, and outputting the output predicted value of the distributed photovoltaic resource of the corresponding region by using the ReLU activation function.
8. A distributed photovoltaic resource output curve deduction system, comprising:
the time-space diagram construction module is used for inputting the geographic position of each distributed photovoltaic resource and the historical photovoltaic output data under the corresponding environmental conditions at different moments, and constructing a weighted undirected time-space diagram to characterize the relation among the distributed photovoltaic resources;
the node time feature solving module is used for solving the time feature of each node in the weighted undirected time-space diagram, wherein the time feature is the distributed photovoltaic resource output intermediate prediction data under the corresponding environmental conditions at different moments;
the curve drawing module is used for extracting the space-time characteristics of each node in the weighted undirected space-time diagram, wherein the space-time characteristics are the space characteristics comprising the time characteristics, and the space characteristics are obtained from the weighted undirected space-time diagram; capturing the output mode characteristics of each node in the space-time characteristics, splicing the time characteristics and the output mode characteristics of each node, inputting the spliced output mode characteristics and the output mode characteristics into a multi-layer fully-connected network, outputting a target output predicted value of the distributed photovoltaic resource in a corresponding area by utilizing the fully-connected network, and drawing a distributed photovoltaic resource output deduction curve.
9. The distributed photovoltaic resource output curve deduction system according to claim 8, wherein the time-space diagram construction module defines a grid undirected graph on a time node t as G t =<V,E t ,F t >Wherein V is a set of nodes, E t For the collection of edges,the characteristic values of all nodes at the time t are obtained;
calculating historical photovoltaic output data of the node i and the node j at the current momentAnd->Correlation coefficient between->
Defining an adjacent matrix A to represent the interconnection relation between a photovoltaic resource node i and a node j, wherein the element A (i, j) in the adjacent matrix A takes the value as follows:
where τ is the average of the output correlation coefficients of all node pairs, euc t (i, j) is historical photovoltaic output data of node i and node j at time tAnd->Euclidean distance between them;
and establishing a connection relation on the power grid undirected graph by the formula to form a weighted undirected time-space graph representing the association degree of different photovoltaic nodes.
10. The distributed photovoltaic resource output curve deduction system according to claim 8, wherein the node time feature solving module establishes an LSTM network model, the LSTM network model being a long and short memory network model, for the node data x at each time step t Updating cell state C using forget gate t :
f t =σ(W f ·[h t-1 ,x t ]+b f )
Wherein f t Indicating a forgetting gate for controlling the state C of the cell at the previous moment t-1 For the current cell state C t Is a degree of influence of (a);representing a cell status update value,/-, for>For C t Is influenced by the input gate i t Performing control;
i t =σ(W i ·[h t-1 ,x t ]+b i )
The output gate o in the LSTM network model is pressed t Update and calculate output h of output unit t :
o t =σ(W o ·[h t-1 ,x t ]+b o )
h t =o t ×tanh(C t )
Definition e t 、c t 、h t 、w t 、r t Representing historical weather conditions, temperature, humidity, wind speed and wind direction data, sw t 、lw t 、kt t Representing historical illumination solar intensity, vertical illumination intensity, and solar radiation factor data, wherein t is a time step;
the LSTM network model is two layers, regression prediction is carried out on historical output data of the photovoltaic power station by using the LSTM network model at the time t, the goal is to output photovoltaic output data at the time t+1, and then the photovoltaic output data at the time t is fitted periodically by using photovoltaic output data, weather data and illumination data of the first n time slices;
by combining the output of the two layers of LSTM network models, based on historical photovoltaic output data, a mapping function T and a corresponding parameter set theta are fitted i Outputting the time characteristics of each node The expression is as follows:
wherein T is 1 And T 2 And respectively corresponding to the two layers of structures of the selected LSTM network model.
11. The distributed photovoltaic resource output curve deduction system according to claim 10, wherein the node time feature solution module further learns N Source-LSTM models from Source domains, the N Source-LSTM models being generated by pre-training a sufficient amount of data in the respective Source domains; generating basic output from target domain Base-LSTM modelGeneration of Source output from N Source-LSTM models>
the attention network is designed based on a multi-layer perceptron, and the following formula is shown:
(e 1,X ,…,e N+1,X )=f α (X;θ α )
wherein X is input data, θ α As a learnable parameter, f (·) represents prediction by LSTM network model, (e) 1,X ,...,e N+1,X ) Is of dimension ofAn output of (2); the output of the multi-layer perceptron is converted into the weight of the model by a soft attention mechanism, and the attention network generates a plurality of nonzero weights omega according to the following mode i :
output of complete model based on attention networkN outputs +.defined as Source-LSTM model>And Base-LSTM model output +.>The expression is as follows:
In the target domain, the parameters of the model are updated according to two functions:
one is a parameter of the Base-LSTM model, which is output by the target domainAnd target domain true value y real Calculated loss function->Updating; another is a parameter in the attention network, according to a modelIntegral output->And target domain true value y real Calculated loss function->Updating;
two independent counter-propagating exists in the Base-LSTM network and the attention network, and the loss function is calculated by using average absolute error, and the expression is as follows:
12. The distributed photovoltaic resource output curve deduction system according to claim 11, wherein the step of extracting the spatio-temporal characteristics of each node in the weighted dead-time space graph by the curve drawing module includes: building a space-time diagram self-encoder ST-GAE, weighting input weighted non-directional space-time diagram sequenceCoding and reconstructing, and capturing a data mode in the weighted undirected space-time diagram; learning depth coding mapping relation- >Weighted undirected time-space diagram sequence +.>Mapped as a series of hidden layer spatial vectors Φ=<Z t-m ,…,Z t-1 ,Z t >At Z t In the ith column, the weighted undirected time-space diagram G t Space-time characteristics of the ith node; learning depth decoding mapping relation by designing two decoder modules, respectively, including coding of feature level +.>Coding with edge level->
13. The system according to claim 12, wherein the curve drawing module captures the output pattern characteristics of each node in the space-time characteristics, splices the time characteristics and the output pattern characteristics of each node, inputs the spliced output pattern characteristics into a multi-layer fully-connected network, and outputs the target output predicted value of the distributed photovoltaic resource in the corresponding area by using the fully-connected network, the method comprises:
for a time window t e [ t ' -m, t ' up to a specific point in time t ] ']The corresponding hidden layer space vector is used for calculating the mean vector as the coding characteristic according to the following formula:column i Z i The feature vector after being coded for the corresponding ith node; for all Z i 1.ltoreq.i.ltoreq.n, with sparse correlation vector calculated as follows>Is a dictionary of: /> Capturing a weighted undirected time-space diagram sequence +.>Component F of the i-th node of the input data i 1.ltoreq.i.ltoreq.n, corresponding salient spatio-temporal patterns, dictionary D and sparse correlation vector a i The application is performed by the following way:
wherein x is t The vector is measured for the network load,fitting values for the corresponding vectors, +.>Representing a fitting operation; in model training, sparse correlation vector a is randomly initialized i And a dictionary D, and performing back propagation optimization based on errors of the output predicted value and the true value of the model; based on target hidden layer space variable Z t T '-m is not less than t and not more than t', n nodes are +.>Solving to obtain dictionary D and n sparse modes +.>Then map (F) i ,a i ) Obtaining the output mode characteristic of the node i at the current moment>
The time characteristics obtained by solvingAnd output mode characteristic p t Stitching complete feature F as a photovoltaic node t And inputting a multi-layer full-connection network, and outputting the output predicted value of the distributed photovoltaic resource of the corresponding region by using the ReLU activation function.
14. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized by: the steps of the distributed photovoltaic resource output curve deduction method according to any one of claims 1 to 7 are realized when the processor executes the computer program.
15. A computer-readable storage medium storing a computer program, characterized in that: the computer program when executed by a processor performs the steps of the distributed photovoltaic resource output curve deduction method according to any one of claims 1 to 7.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202310374894.XA CN116401949A (en) | 2023-04-10 | 2023-04-10 | Distributed photovoltaic resource output curve deduction method, system, equipment and medium |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202310374894.XA CN116401949A (en) | 2023-04-10 | 2023-04-10 | Distributed photovoltaic resource output curve deduction method, system, equipment and medium |
Publications (1)
Publication Number | Publication Date |
---|---|
CN116401949A true CN116401949A (en) | 2023-07-07 |
Family
ID=87007066
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202310374894.XA Pending CN116401949A (en) | 2023-04-10 | 2023-04-10 | Distributed photovoltaic resource output curve deduction method, system, equipment and medium |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN116401949A (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117688109A (en) * | 2023-12-27 | 2024-03-12 | 山东大学 | Digital environment information system construction method based on nested frame representation |
CN117786372A (en) * | 2024-02-28 | 2024-03-29 | 北京岳能科技股份有限公司 | Distributed photovoltaic power generation data processing method and system based on machine learning |
-
2023
- 2023-04-10 CN CN202310374894.XA patent/CN116401949A/en active Pending
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117688109A (en) * | 2023-12-27 | 2024-03-12 | 山东大学 | Digital environment information system construction method based on nested frame representation |
CN117786372A (en) * | 2024-02-28 | 2024-03-29 | 北京岳能科技股份有限公司 | Distributed photovoltaic power generation data processing method and system based on machine learning |
CN117786372B (en) * | 2024-02-28 | 2024-05-17 | 北京岳能科技股份有限公司 | Distributed photovoltaic power generation data processing method and system based on machine learning |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109214575B (en) | Ultrashort-term wind power prediction method based on small-wavelength short-term memory network | |
CN116401949A (en) | Distributed photovoltaic resource output curve deduction method, system, equipment and medium | |
Kavousi-Fard | A hybrid accurate model for tidal current prediction | |
CN111461463B (en) | Short-term load prediction method, system and equipment based on TCN-BP | |
CN109583588B (en) | Short-term wind speed prediction method and system | |
Zhang et al. | Modeling IoT equipment with graph neural networks | |
CN116245033B (en) | Artificial intelligent driven power system analysis method and intelligent software platform | |
CN112100911B (en) | Solar radiation prediction method based on depth BILSTM | |
CN114021836B (en) | Multi-variable reservoir water inflow prediction system based on different angle fusion, training method and application | |
Li et al. | Evolutionary deep learning with extended Kalman filter for effective prediction modeling and efficient data assimilation | |
CN111141879B (en) | Deep learning air quality monitoring method, device and equipment | |
CN109886560A (en) | Distribution network transform measure and rate of qualified voltage index relevance method for digging and device | |
CN114265913A (en) | Space-time prediction algorithm based on federal learning on industrial Internet of things edge equipment | |
CN115131565A (en) | Histology image segmentation model based on semi-supervised learning | |
Wang et al. | A novel wind power prediction model improved with feature enhancement and autoregressive error compensation | |
CN114169251A (en) | Ultra-short-term wind power prediction method | |
CN113194493B (en) | Wireless network data missing attribute recovery method and device based on graph neural network | |
CN116720156A (en) | Weather element forecasting method based on graph neural network multi-mode weather data fusion | |
CN115376103A (en) | Pedestrian trajectory prediction method based on space-time diagram attention network | |
CN115080795A (en) | Multi-charging-station cooperative load prediction method and device | |
CN116862080B (en) | Carbon emission prediction method and system based on double-view contrast learning | |
Goui et al. | A Comparative Study of LSTM/GRU Models for Energy Long-Term Forecasting in IoT Networks | |
Arshad et al. | Wind power prediction using genetic programming based ensemble of artificial neural networks (GPeANN) | |
CN110020379B (en) | Link prediction method based on deep dynamic network embedded representation model | |
Yoon et al. | Prediction of machine learning base for efficient use of energy infrastructure in smart city |
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 |