CN117172413A - Power grid equipment operation state monitoring method based on multi-mode data joint characterization and dynamic weight learning - Google Patents
Power grid equipment operation state monitoring method based on multi-mode data joint characterization and dynamic weight learning Download PDFInfo
- Publication number
- CN117172413A CN117172413A CN202311154552.3A CN202311154552A CN117172413A CN 117172413 A CN117172413 A CN 117172413A CN 202311154552 A CN202311154552 A CN 202311154552A CN 117172413 A CN117172413 A CN 117172413A
- Authority
- CN
- China
- Prior art keywords
- power grid
- equipment
- grid equipment
- state
- data
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 67
- 238000012544 monitoring process Methods 0.000 title claims abstract description 47
- 238000012512 characterization method Methods 0.000 title claims abstract description 33
- 238000013528 artificial neural network Methods 0.000 claims abstract description 55
- 238000011156 evaluation Methods 0.000 claims abstract description 41
- 238000010586 diagram Methods 0.000 claims abstract description 18
- 230000036541 health Effects 0.000 claims abstract description 13
- 239000013598 vector Substances 0.000 claims description 117
- 230000003044 adaptive effect Effects 0.000 claims description 61
- 230000006870 function Effects 0.000 claims description 16
- 238000004364 calculation method Methods 0.000 claims description 15
- 238000007689 inspection Methods 0.000 claims description 13
- 238000012549 training Methods 0.000 claims description 13
- 238000006243 chemical reaction Methods 0.000 claims description 12
- 238000013461 design Methods 0.000 claims description 6
- 238000012423 maintenance Methods 0.000 claims description 5
- 230000005540 biological transmission Effects 0.000 claims description 4
- 230000008569 process Effects 0.000 claims description 4
- 238000004458 analytical method Methods 0.000 claims description 3
- 238000010276 construction Methods 0.000 claims description 3
- 230000004927 fusion Effects 0.000 claims description 3
- 238000012804 iterative process Methods 0.000 claims description 3
- 238000013507 mapping Methods 0.000 claims description 3
- 238000010801 machine learning Methods 0.000 abstract description 3
- 230000007547 defect Effects 0.000 abstract description 2
- 238000005259 measurement Methods 0.000 abstract description 2
- 238000005516 engineering process Methods 0.000 description 4
- 230000000694 effects Effects 0.000 description 3
- 238000013473 artificial intelligence Methods 0.000 description 2
- 238000003066 decision tree Methods 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 238000007637 random forest analysis Methods 0.000 description 2
- XEEYBQQBJWHFJM-UHFFFAOYSA-N Iron Chemical group [Fe] XEEYBQQBJWHFJM-UHFFFAOYSA-N 0.000 description 1
- 230000002159 abnormal effect Effects 0.000 description 1
- 238000012098 association analyses Methods 0.000 description 1
- 238000013145 classification model Methods 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 238000007405 data analysis Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000013210 evaluation model Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 230000003862 health status Effects 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000007500 overflow downdraw method Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 238000004804 winding Methods 0.000 description 1
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
Compared with the prior art, the invention solves the defect that the multi-source heterogeneous number joint characterization and unified application are difficult to establish the power grid equipment state measurement by utilizing the association relation among different parts of the equipment. The invention comprises the following steps: setting a distributed representation method of power grid equipment state data; the method comprises the steps of jointly representing multi-mode data of power grid equipment; constructing an internal topological relation diagram of the equipment; dynamic weight learning of the graph neural network; and (5) evaluating the state of the power grid equipment. The invention combines the machine learning classification method to carry out the health evaluation of the equipment state, can obviously improve the accuracy and the robustness of the evaluation method, and achieves the aim of monitoring the running state of the power grid equipment.
Description
Technical Field
The invention relates to the technical field of power grid equipment operation monitoring, in particular to a power grid equipment operation state monitoring method based on multi-mode data joint characterization and dynamic weight learning.
Background
With the development of artificial intelligence and big data technology, the application of the intelligent power grid equipment is more and more extensive and deep. The traditional power grid equipment management mainly adopts methods such as regular inspection, regular maintenance and post maintenance, and the like, and can ensure the safe operation of the power grid equipment to a certain extent, but lacks monitoring and prediction on the real-time state of the equipment, cannot discover the fault risk and abnormal condition of the equipment in advance, and has very limited effect and efficiency.
The power grid equipment state evaluation is to evaluate the running state and the health condition of the equipment by comprehensively analyzing multi-source data of the power grid equipment. The running state of the power grid equipment is affected by a plurality of factors, the association relation of different factors is contained in a large amount of multi-source heterogeneous power grid equipment state data, the power grid data is analyzed and processed by utilizing an artificial intelligence technology, intelligent prediction and optimization control of the power grid equipment can be realized, and the method has important significance for improving the running efficiency and economy of a power grid system.
And a large amount of multi-source heterogeneous data are generated in the daily monitoring process of the power grid equipment, such as real-time monitoring data, historical data, external data, GIS data and image data. The conventional power grid equipment state evaluation method often utilizes field experience knowledge to carry out statistical classification on data from different sources, and establishes a clear evaluation standard. Under the current big data background, the conventional method cannot mine the associated information among the data, and cannot well uniformly and effectively apply the data with different sources and structures. Deep learning technology is a typical representation of data intelligence and has good potential in data analysis modeling. The multi-source heterogeneous data representing the state of the power grid equipment can fully fuse effective characteristics in state data such as different images, characters, voices and the like to construct unified vector representation by using a deep learning-based distributed representation technology, the association relation among equipment parts is considered on the basis, the vector representation of the equipment state is optimized, the accurate and timely expression of the state of the power grid equipment is realized, the establishment of an intelligent and automatic power grid equipment state evaluation system is facilitated, and the intelligent management work of the power grid equipment is supported.
However, since the monitoring data of the power grid equipment have different sources and formats, it is difficult to represent and analyze the different data by a unified method, and the application of the monitoring data of the power grid equipment is limited.
Disclosure of Invention
The invention aims to solve the defect that in the prior art, multi-source heterogeneous number joint characterization and unified application are difficult, and the association relation among different parts of equipment is utilized to establish power grid backup state measurement, and provides a power grid equipment running state monitoring method based on multi-mode data joint characterization and dynamic weight learning to solve the problems.
In order to achieve the above object, the technical scheme of the present invention is as follows:
a power grid equipment operation state monitoring method based on multi-mode data joint characterization and dynamic weight learning comprises the following steps:
setting a distributed representation method of power grid equipment state data: setting a method for carrying out distributed representation on data of different modes related to the power grid equipment state;
joint characterization of multi-modal data of power grid equipment: aligning the vector representations of the image and the voice to the vector representations of the words to realize the joint characterization of different modes;
building an internal topological relation diagram of equipment: constructing a topological relation diagram G (N, E) among internal components of the power grid equipment;
dynamic weight learning of graph neural network: training the graph neural network based on an equipment internal topological relation graph G (N, E) to analyze the influence of different equipment parts on the overall state evaluation of the power grid equipment;
evaluation of the state of the power grid equipment: monitoring the state of power grid equipment, acquiring power grid equipment state data, carrying out distributed representation and joint representation, mapping a joint representation vector of the power grid equipment into a node representing the power grid equipment in a graph neural network, and obtaining a state vector representation of corresponding equipment; after all the state monitoring data are mapped into the nodes, carrying out n times of iterative updating on the nodes of the graph neural network based on the weights corresponding to the current targets, and learning the information of the neighborhood nodes; and finally, representing the state of the whole equipment by using the state vector representation of each node in the graph neural network, and realizing intelligent state monitoring of the power grid equipment.
The distributed representation method for setting the power grid equipment state data comprises the following steps:
acquiring power grid equipment state data comprising text data, image data and voice data expressing equipment state,
the text data are sensor numerical information obtained by monitoring power grid equipment, file data of the equipment, state information of equipment inspection records, maintenance records of the equipment, account information of the equipment and data stored in a character string form, the image data are image information collected by an image sensor in monitoring the power grid equipment, daily monitored images of a camera, images taken by inspection of the equipment, images taken by inspection of an unmanned aerial vehicle and image information taken by an infrared camera, and the voice data are voice information collected by an audio sensor, daily working audio of the equipment and voice evaluated by inspection personnel;
sentence vectorization representation of text data by word2vec method, t=c for text sequence 1 ,c 2 ,…,c m Wherein c j The method is a j-th word in the sequence, a word2vec function is utilized to model a sequence T, a vector representation T of the sequence T is obtained, and the expression is as follows:
T＝word2vec(t)；
the image data is vectorized by image2vec, the voice data is vectorized by sound2vec, and the vector representation P of the image P and the vector representation V of the sound V are obtained, wherein the expression is as follows:
P＝image2vec(p)，V＝sound2vec(v)。
the joint characterization of the multi-mode data of the power grid equipment comprises the following steps:
acquiring a sample set D of three data sources passing through characters, images and sounds 1 、d 2 、D 3 Using distributed representation of power grid equipment state data, the converted vector set represents C T 、C P 、C v ；
Training an encoder of an adaptive dimension:
vector representation P of an image i ∈C P A new vector is obtained through the adaptive encoder, and training is as follows:
from { D 1 ,D 2 Get N in } 1 Samples of N 1 The samples are one-to-one corresponding to the image and the text, and are takenRepresenting text sample,/->Is the corresponding image sample, will be +.>Conversion to vector alpha using word2vec 1 Will->Conversion to vector alpha with image2vec 2 Design loss function calculation alpha 1 And alpha 2 Deviation of->Input to adaptive dimension decoder, < >>After passing through the sound adaptive dimension encoder, the sound adaptive dimension encoder is input into the adaptive dimension decoder, and the weight parameter of the adaptive dimension encoder is adjusted to realize the final alpha 1 And alpha 2 Keeping consistency;
vector V of audio i ∈C V Indicating that a new vector is obtained through the adaptive encoder; training is as follows:
from { D 1 ,D 3 N is taken 2 Samples of N 2 The samples are one-to-one corresponding to the audio and the text, and are takenRepresenting text sample,/->Is the corresponding audio sample, will be +.>Conversion to vector alpha using word2vec 1 Will->Conversion to vector alpha with sound2vec 3 Design loss function calculation alpha 1 And alpha 3 Deviation of->Input to adaptive dimension decoder, < >>The image is input to an adaptive dimension decoder after passing through an image adaptive dimension encoder, and weight parameters of the adaptive dimension encoder are adjusted to realize the final alpha 1 And alpha 3 Keeping consistency;
encoding and decoding distributed representation vectors of device text, image and sound data acquired in real time by using a trained adaptive dimension encoder;
text input to the adaptive dimension encoder, image input to the image adaptive dimension encoder, sound input to the sound adaptive dimension encoder, and output to obtain a unified representation alpha of the single-mode vector representations corresponding to the text, the image and the sound 1 、α 2 And alpha 3 Concatenating different unimodal unified vector representationsAlpha is used as a vector for fusion for subsequent analysis.
The construction of the device internal topological relation diagram comprises the following steps:
the association relation of different parts of the power grid equipment is represented by a graph structure, and a node N i Representing different equipment parts, edge E i Representing edges between different devices, e.g. by W i,j Representing a device node N i And N j The association relation between the two components is used for constructing a topological relation diagram G (N, E) between the internal components of the power grid equipment;
for W ij Uses the statistics of the node corresponding devicesThe calculation is performed according to the following calculation formula:
wherein Count (N) i ) And Count (N) j ) Respectively represent node N i And node N j The frequency of the corresponding equipment faults is taken as the association weight W between the nodes by taking the ratio of the frequency minimum value to the frequency maximum value i,j 。
The dynamic weight learning of the graph neural network comprises the following steps:
for any graph neural network node N, the initial state of the node N is expressed as alpha by using a single-mode unified vector of node monitoring data;
associated weight W ij Is used for updating the graph nodes in the dynamic weight learning process of the graph neural network, if the node N has K adjacent nodes, the corresponding adjacent nodes are { N } 1 ,N 2 ,…N K The monitoring information of the corresponding neighboring node is denoted as { alpha } 1 ,α 2 ,…α K The association weights of node N and neighboring nodes are { W }, respectively 1 ,W 2 ,…W K W, where W 1 ,W 2 ,…W K Corresponding to the associated weight W ij ；
........
The first round of iterative process obtains a new state vector representation alpha' of the node N, and the specific calculation is as follows:
Mini(A 1 ,A 2 ,...,A n )＝A i wherein i is such that A i For the sequence (A) 1 ,A 2 ,...,A n ) Subscript of minimum of (2); max (A) 1 ,A 2 ,...,A n )＝A i Wherein i is such that A i For the sequence (A) 1 ,A 2 ,...,A n ) Subscript of the maximum value of (v); count (N) represents the frequency of the equipment failure corresponding to node N;
and sequentially carrying out one round of message transmission on all nodes of the topological relation diagram G (N, E) to obtain a new set of vector representations, and sequentially repeating until the vector representations of all the nodes are converged.
The evaluation of the state of the power grid equipment comprises the following steps:
for the power grid equipment component S, text, image and sound state data of equipment are acquired, and vectors T are obtained by a distributed representation method respectively S 、P S 、V S Input adaptive dimension encoder, to image adaptive dimension encoder and sound adaptive dimension encoder to obtain a single-mode vector representation alpha T 、α P And alpha V Computing a joint characterization of S
Calculating the joint characterization of other parts of the power grid equipment by the method;
for all components of the power grid equipment, the joint representation of the components is used for representing the node states in the graph neural network G (N, E), the component S is represented by a node N in the graph, and the state vector of the node N is represented by the joint representation alpha of S;
the state vector of other parts of the power grid equipment in the graphic neural network G (N, E) is represented by the method;
when all components of the power grid equipment calculate to obtain a state vector at nodes in the graph neural network G (N, E), iteratively updating the nodes of the graph neural network by using the association weights among the nodes in the graph neural network G (N, E), learning the information of the neighborhood nodes, and calculating by the node N to obtain a new state vector to represent alpha';
calculating new state vector representations of all other nodes;
obtaining new state vector expression X of all nodes of the graph neural network, and establishing an association relationship between the X and the Y by combining an empirical state evaluation health grade Y endowed by an expert to the power grid equipment by using the fully-connected neural network to obtain a state evaluation function F of the power grid equipment;
Y＝F(X)；
and calculating new state vector representations X of all nodes by using a state evaluation function F of the power grid equipment at any moment to obtain a state evaluation health grade Y of the power grid equipment, so as to achieve the purpose of intelligent state monitoring of the power grid equipment.
Advantageous effects
Compared with the prior art, the method for monitoring the running state of the power grid equipment based on the multi-mode data joint characterization and dynamic weight learning adopts a multi-mode data fusion method to perform joint characterization on multi-source heterogeneous data, and on the basis of single-mode data vectorization, the data of different modes are fused into a unified form by utilizing encoder-decoder structures of different dimensions; meanwhile, the association relation of different parts in the equipment is considered, the topological relation in the equipment is learned by using a graph neural network, and different data sets are used for dynamically learning weights according to the state evaluation target, so that the high-efficiency extraction and representation of the multi-dimensional characteristics of the power grid equipment state are realized.
The invention combines the machine learning classification method to carry out the health evaluation of the equipment state, can obviously improve the accuracy and the robustness of the evaluation method, and achieves the aim of monitoring the running state of the power grid equipment.
Drawings
FIG. 1 is a process sequence diagram of the present invention;
FIG. 2 is a structural flow chart of the present invention;
FIG. 3 is a schematic sequential diagram of a distributed representation method according to the present invention;
FIG. 4 is a diagram of the topology of the internal components of the power grid device according to the present invention;
fig. 5 is a schematic structural diagram of a state evaluation method according to the present invention.
Detailed Description
For a further understanding and appreciation of the structural features and advantages achieved by the present invention, the following description is provided in connection with the accompanying drawings, which are presently preferred embodiments and are incorporated in the accompanying drawings, in which:
as shown in fig. 1 and fig. 2, the method for monitoring the running state of power grid equipment based on multi-mode data joint characterization and dynamic weight learning, disclosed by the invention, comprises the following steps:
the first step, a distributed representation method of power grid equipment state data is set: setting a method for carrying out distributed representation on data of different modes related to the power grid equipment state.
(1) Acquiring power grid equipment state data comprising text data, image data and voice data expressing equipment state,
the text data are sensor numerical information obtained by monitoring power grid equipment, file data of the equipment, state information of equipment inspection records, maintenance records of the equipment, account information of the equipment and data stored in a character string form, the image data are image information collected by an image sensor in monitoring of the power grid equipment, images monitored daily by a camera, images taken by inspection of the equipment, images taken by inspection of an unmanned aerial vehicle and image information taken by an infrared camera, and the voice data are voice information collected by an audio sensor, daily working audio of the equipment and voice evaluated by inspection personnel.
(2) Sentence vectorization representation of text data by word2vec method, t=c for text sequence 1 ,c 2 ,…,c m Wherein c j The method is a j-th word in the sequence, a word2vec function is utilized to model a sequence T, a vector representation T of the sequence T is obtained, and the expression is as follows:
T＝word2vec(t)。
(3) The image data is vectorized by image2vec, the voice data is vectorized by sound2vec, and the vector representation P of the image P and the vector representation V of the sound V are obtained, wherein the expression is as follows:
P＝image2vec(p)，V＝sound2vec(v)。
secondly, joint representation of multi-mode data of power grid equipment: the vector representations of the image and the voice are aligned to the vector representations of the text to achieve joint characterization of the different modalities.
The multi-mode data joint characterization of the equipment is to perform unified vectorization representation on data of different sources and structures. The text t, the image p and the voice v are respectively vector-represented by respective vector tools. The text, the image and the voice form three types of data sources, and vector representations of the three types of data cannot be directly fused due to different vector representation spaces of different data. In order to extract and fuse the characteristics of three modes of data of words, images and voices, the method realizes the joint characterization of different modes by aligning the vector representation of the images and voices to the vector representation of the words. The joint representation of the multi-mode data of the power grid equipment is to convert the original text, image and audio data into vectors with the same format by using a unified model, and then analyze and evaluate the data of the power grid equipment based on the vectors, as shown in fig. 3.
(1) Acquiring a sample set D of three data sources passing through characters, images and sounds 1 、D 2 、D 3 Using distributed representation of power grid equipment state data, the converted vector set represents C T 、C P 、C V 。
(2) Training an encoder of an adaptive dimension:
a1 Vector representation P of image i ∈C P A new vector is obtained through the adaptive encoder, and training is as follows:
from { D 1 ,D 2 Get N in } 1 Samples of N 1 The samples are one-to-one corresponding to the image and the text, and are takenRepresenting text sample,/->Is the corresponding image sample, will be +.>Conversion to vector alpha using word2vec 1 Will->Conversion to vector alpha with image2vec 2 Design loss function calculation alpha 1 And alpha 2 Deviation of->Input to adaptive dimension decoder, < >>After passing through the sound adaptive dimension encoder, the sound adaptive dimension encoder is input into the adaptive dimension decoder, and the weight parameter of the adaptive dimension encoder is adjusted to realize the final alpha 1 And alpha 2 Keeping consistency;
a2 Vector V of audio i ∈C V Indicating that a new vector is obtained through the adaptive encoder; training is as follows:
from { D 1 ,D 3 N is taken 2 Samples of N 2 The samples are one-to-one corresponding to the audio and the text, and are takenRepresenting text sample,/->Is the corresponding audio sample, will be +.>Conversion to vector alpha using word2vec 1 Will->Conversion to vector alpha with sound2vec 3 Design loss function calculation alpha 1 And alpha 3 Deviation of->Input to adaptive dimension decoder, < >>The image is input to an adaptive dimension decoder after passing through an image adaptive dimension encoder, and weight parameters of the adaptive dimension encoder are adjusted to realize the final alpha 1 And alpha 3 And keep the same.
(3) Encoding and decoding distributed representation vectors of device text, image and sound data acquired in real time by using a trained adaptive dimension encoder;
text input to the adaptive dimension encoder, image input to the image adaptive dimension encoder, sound input to the sound adaptive dimension encoder, and output to obtain a unified representation alpha of the single-mode vector representations corresponding to the text, the image and the sound 1 、α 2 And alpha 3 Concatenating different unimodal unified vector representationsAlpha is used as a vector for fusion for subsequent analysis. The adaptive dimension encoder has two functions, namely, the vector generated by the single-mode distributed representation model is unified into a vector with consistent dimension, and the association of different mode data is learned by calculation of different mode vector representations.
Here, each time the image and audio are trained, the text is trained together. The text also has a corresponding adaptive encoder, see the accompanying drawings. When the images are trained, the weights of the text adaptive dimension encoder and the image adaptive dimension encoder are adjusted at the same time, and the audio frequency is the same method. The purpose is to keep the feature space consistent between the encoder output of images and audio and the encoder output of text.
Thirdly, building an internal topological relation diagram of the equipment: a topological relation diagram G (N, E) between the internal components of the grid device is constructed as shown in fig. 4.
(1) The association relation of different parts of the power grid equipment is represented by a graph structure, and a node N i Representing different equipment parts, edge E i Representing edges between different devices, e.g. by W i,j Representing a device node N i And N j And constructing a topological relation graph G (N, E) between the internal components of the power grid equipment according to the association relation between the internal components.
(2) For W ij The numerical value of the node is calculated by using the statistical value of the corresponding equipment of the node, and the calculation formula is as follows:
wherein Count (N) i ) And Count (N) j ) Respectively represent node N i And node N j The frequency of the corresponding equipment faults is taken as the association weight W between the nodes by taking the ratio of the frequency minimum value to the frequency maximum value i,j 。
Fourth, dynamic weight learning of the graph neural network: and training the graph neural network based on the equipment internal topological relation graph G (N, E) to analyze the influence of different equipment parts on the overall state evaluation of the power grid equipment.
The power grid equipment is a complex hardware equipment, such as main transformer equipment is a large system, and the power grid equipment specifically comprises different components such as high-low voltage bushings, oil tanks, gas relays, iron cores, windings, valves, switches, shells and the like, and each component has separate and specific functional division. From the perspective of complex systems, there is a correlation between different components, and when analyzing the health status of one component, the status of other components associated therewith can be comprehensively considered. The specific method is that a topological graph is established, the states of different devices are subjected to association analysis, association weights are obtained through calculation, and the states of adjacent parts of the target parts are utilized to update the target parts, so that the health evaluation of the power grid device is realized.
The dynamic weight learning of the graph neural network is to analyze the influence of different equipment parts on the overall state evaluation of the equipment by using an equipment internal topological relation graph G (N, E) constructed by domain expert knowledge. Taking the multisource heterogeneous monitoring information of each equipment component into consideration, and taking the fused monitoring information representation alpha as an initial representation of the graph neural network node. And meanwhile, the state information of different equipment parts is transmitted to other different parts by using a message transmission mechanism of the graphic neural network, and the node relation of the graphic neural network is trained by using the state evaluation marking information of the whole equipment to obtain a group of weight matrixes for evaluating the equipment.
(1) For any graph neural network node N, the initial state of the node N is expressed as alpha by using a single-mode unified vector of node monitoring data;
(2) Associated weight W ij Is used for updating the graph nodes in the dynamic weight learning process of the graph neural network, if the node N has K adjacent nodes, the corresponding adjacent nodes are { N } 1 ,N 2 ,...N K The monitoring information of the corresponding neighboring node is denoted as { alpha } 1 ,α 2 ,...α K The association weights of node N and neighboring nodes are { W }, respectively 1 ,W 2 ,...W K W, where W 1 ,W 2 ,...W K Corresponding to the associated weight W ij ；
........
The first round of iterative process obtains a new state vector representation alpha' of the node N, and the specific calculation is as follows:
Mini(A 1 ,A 2 ,...,A n )＝A i wherein i is such that A i For the sequence (A) 1 ,A 2 ,...,A n ) Subscript of minimum of (2); max (A) 1 ,A 2 ,...,A n )＝A i Wherein i is such that A i For the sequence (A) 1 ,A 2 ,...,A n ) Subscript of the maximum value of (v); count (N) represents the frequency of the equipment failure corresponding to node N;
and sequentially carrying out one round of message transmission on all nodes of the topological relation diagram G (N, E) to obtain a new set of vector representations, and sequentially repeating until the vector representations of all the nodes are converged.
Fifthly, evaluating the state of power grid equipment: monitoring the state of power grid equipment, acquiring power grid equipment state data, carrying out distributed representation and joint representation, mapping a joint representation vector of the power grid equipment into a node representing the power grid equipment in a graph neural network, and obtaining a state vector representation of corresponding equipment; after all the state monitoring data are mapped into the nodes, carrying out n times of iterative updating on the nodes of the graph neural network based on the weights corresponding to the current targets, and learning the information of the neighborhood nodes; and finally, representing the state of the whole equipment by using the state vector representation of each node in the graph neural network, and realizing intelligent state monitoring of the power grid equipment.
As shown in fig. 5, the device state evaluation is to use the result of the multi-mode data joint characterization to express the information of each node in the graph neural network, and construct a neural network full-connection layer for different device state evaluation targets and sample data, and train to generate a corresponding weight matrix. When the state of the equipment is updated, the corresponding vector is mapped into the corresponding node in the graph neural network, and the corresponding node is activated to obtain the state vector representation of the corresponding equipment. After all the state update data are mapped into the nodes, the nodes of the graph neural network are subjected to iterative update for a certain number of times based on the weights corresponding to the current targets, and the information of the neighborhood nodes is learned. And finally, representing the state of the whole equipment by using the state vector representation of each node in the graph neural network, and realizing intelligent state evaluation of the power grid equipment.
(1) For the power grid equipment component S, text, image and sound state data of equipment are acquired, and vectors T are obtained by a distributed representation method respectively S 、P S 、V S Input adaptive dimension encoder, to image adaptive dimension encoder and sound adaptive dimension encoder to obtain a single-mode vector representation alpha T 、α P And alpha V Computing a joint characterization of S
Calculating the joint characterization of other parts of the power grid equipment by the method;
(2) For all components of the power grid equipment, the joint representation of the components is used for representing the node states in the graph neural network G (N, E), the component S is represented by a node N in the graph, and the state vector of the node N is represented by the joint representation alpha of S;
the state vector of other parts of the power grid equipment in the graphic neural network G (N, E) is represented by the method;
(3) When all components of the power grid equipment calculate to obtain a state vector at nodes in the graph neural network G (N, E), iteratively updating the nodes of the graph neural network by using the association weights among the nodes in the graph neural network G (N, E), learning the information of the neighborhood nodes, and calculating by the node N to obtain a new state vector to represent alpha';
calculating new state vector representations of all other nodes;
(4) Obtaining new state vector expression X of all nodes of the graph neural network, and establishing an association relationship between the X and the Y by combining an empirical state evaluation health grade Y endowed by an expert to the power grid equipment by using the fully-connected neural network to obtain a state evaluation function F of the power grid equipment;
Y＝F(X)；
(5) And calculating new state vector representations X of all nodes by using a state evaluation function F of the power grid equipment at any moment to obtain a state evaluation health grade Y of the power grid equipment, so as to achieve the purpose of intelligent state monitoring of the power grid equipment.
The power grid equipment operation states are classified and evaluated as follows:
evaluation index:
the operation state evaluation model is used for evaluating the operation health state of the equipment, is a multi-classification model, and the evaluation results are classified by different health grades. The introduction evaluation index is as follows:
(1) Accuracy (Accuracy): the number of correctly rated and classified samples in a certain class is the proportion of the total number of all samples rated as the class by the model.
(2) Recall (Recall): the number of correctly evaluated samples in a certain class is the proportion of the total number of samples in all the classes.
(3) F1 value (F1-score): for the harmonic mean of the precision Accuracy and Recall, the harmonic mean is the inverse of the arithmetic mean of the inverse of the overall statistics.
Wherein TP (True Positive) is a true class, i.e., the instance is a positive class, and is predicted to be a positive class; FN (False Negative) is a pseudo-negative class, i.e. the instance is a positive class, but is predicted to be a negative class; (FP (False Positive) is a false positive class, i.e., the instance is a negative class, but is predicted to be a positive class; TN (True Negative) is a true negative class, i.e., the instance is a negative class, and is predicted to be a negative class.
The method comprises the following steps of:
the effect of running state health evaluation is carried out for comparison of different methods. The experiment adopts a decision tree method, a KNN method, a random forest method, an SVM method and the method for carrying out state evaluation. The experiment used 80% of the dataset as the training set, the remaining 20% of the dataset as the test set. The method is characterized in that different mode data are trained after multi-mode joint characterization and dynamic weight learning. The results of the comparison of the accuracy, recall, F1-score for the different methods are shown in Table 1:
table 1 comparison of the method according to the invention with other methods
| Method | Accuracy rate of | Recall rate of recall | F1-score |
| KNN | 0.72 | 0.75 | 0.74 |
| Decision tree | 0.80 | 0.82 | 0.80 |
| Random forest | 0.75 | 0.77 | 0.77 |
| SVM | 0.55 | 0.84 | 0.69 |
| The method of the invention | 0.82 | 0.85 | 0.83 |
Compared with the accuracy, recall rate and F1-score value results of different methods, the result data of the method is higher than that of other methods to different degrees. In a comprehensive view, the method provided by the invention is superior to the existing state evaluation method based on machine learning.
The foregoing has shown and described the basic principles, principal features and advantages of the invention. It will be understood by those skilled in the art that the present invention is not limited to the embodiments described above, and that the above embodiments and descriptions are merely illustrative of the principles of the present invention, and various changes and modifications may be made therein without departing from the spirit and scope of the invention, which is defined by the appended claims. The scope of the invention is defined by the appended claims and equivalents thereof.
Claims (6)
1. The utility model provides a power grid equipment operation state monitoring method based on multi-mode data joint characterization and dynamic weight learning, which is characterized by comprising the following steps:
11 A distributed representation method for setting power grid equipment state data: setting a method for carrying out distributed representation on data of different modes related to the power grid equipment state;
12 Joint characterization of multi-modal data of power grid equipment: aligning the vector representations of the image and the voice to the vector representations of the words to realize the joint characterization of different modes;
13 Construction of a topology relationship diagram inside the device: constructing a topological relation diagram G (N, E) among internal components of the power grid equipment;
14 Dynamic weight learning of graph neural network: training the graph neural network based on an equipment internal topological relation graph G (N, E) to analyze the influence of different equipment parts on the overall state evaluation of the power grid equipment;
15 Evaluation of the status of the network system: monitoring the state of power grid equipment, acquiring power grid equipment state data, carrying out distributed representation and joint representation, mapping a joint representation vector of the power grid equipment into a node representing the power grid equipment in a graph neural network, and obtaining a state vector representation of corresponding equipment; after all the state monitoring data are mapped into the nodes, carrying out n times of iterative updating on the nodes of the graph neural network based on the weights corresponding to the current targets, and learning the information of the neighborhood nodes; and finally, representing the state of the whole equipment by using the state vector representation of each node in the graph neural network, and realizing intelligent state monitoring of the power grid equipment.
2. The method for monitoring the running state of the power grid equipment based on the multi-mode data joint characterization and dynamic weight learning according to claim 1, wherein the method for setting the distributed representation of the power grid equipment state data comprises the following steps:
21 Acquiring power grid equipment status data including text data, image data and voice data representing equipment status,
the text data are sensor numerical information obtained by monitoring power grid equipment, file data of the equipment, state information of equipment inspection records, maintenance records of the equipment, account information of the equipment and data stored in a character string form, the image data are image information collected by an image sensor in monitoring the power grid equipment, daily monitored images of a camera, images taken by inspection of the equipment, images taken by inspection of an unmanned aerial vehicle and image information taken by an infrared camera, and the voice data are voice information collected by an audio sensor, daily working audio of the equipment and voice evaluated by inspection personnel;
22 Sentence vectorization representation of text data by word2vec method, t=c for text sequence 1 ,c 2 ,…,c m Wherein c j Is the j-th word in the sequence, using word2vec function pairsModeling the sequence T to obtain a vector representation T of the sequence T, wherein the expression is as follows:
T＝word2vec(t)；
23 Vectorizing image data with image2vec, vectorizing voice data with sound2vec to obtain vector representation P of image P and vector representation V of sound V, wherein the expression is as follows:
P＝image2vec(p)，V＝sound2vec(v)。
3. the method for monitoring the running state of the power grid equipment based on the multi-modal data joint characterization and dynamic weight learning according to claim 1, wherein the joint characterization of the multi-modal data of the power grid equipment comprises the following steps:
31 Acquiring a sample set D of three data sources of text, image and sound 1 、D 2 、D 2 Using distributed representation of power grid equipment state data, the converted vector set represents C T 、C P 、C V ；
32 Training the encoder of the adaptive dimension:
321 Vector representation P of image i ∈C P A new vector is obtained through the adaptive encoder, and training is as follows:
from { D 1 ,D 2 Get N in } 1 Samples of N 1 The samples are one-to-one corresponding to the image and the text, and are takenRepresenting text sample,/->Is the corresponding image sample, will be +.>Conversion to vector alpha using word2vec 1 Will->Conversion to vector alpha with image2vec 2 Design loss function calculation alpha 1 And alpha 2 Deviation of->Input to adaptive dimension decoder, < >>After passing through the sound adaptive dimension encoder, the sound adaptive dimension encoder is input into the adaptive dimension decoder, and the weight parameter of the adaptive dimension encoder is adjusted to realize the final alpha 1 And alpha 2 Keeping consistency;
322 Vector V of audio i ∈C V Indicating that a new vector is obtained through the adaptive encoder; training is as follows:
from { D 1 ,D 3 N is taken 2 Samples of N 2 The samples are one-to-one corresponding to the audio and the text, and are takenA sample of the text is represented and,is the corresponding audio sample, will be +.>Conversion to vector alpha using word2vec 1 Will->Conversion to vector alpha with sound2vec 3 Design loss function calculation alpha 1 And alpha 3 Deviation of->Input to adaptive dimension decoder, < >>The image is input to an adaptive dimension decoder after passing through an image adaptive dimension encoder, and weight parameters of the adaptive dimension encoder are adjusted to realize the final alpha 1 And alpha 3 Keeping consistency;
33 Using a trained adaptive dimension encoder to encode and decode distributed representation vectors of device text, image and sound data acquired in real time;
text input to the adaptive dimension encoder, image input to the image adaptive dimension encoder, sound input to the sound adaptive dimension encoder, and output to obtain a unified representation alpha of the single-mode vector representations corresponding to the text, the image and the sound 1 、α 2 And alpha 3 Concatenating different unimodal unified vector representationsAlpha is used as a vector for fusion for subsequent analysis.
4. The method for monitoring the running state of the power grid equipment based on multi-mode data joint characterization and dynamic weight learning according to claim 1, wherein the construction of the internal topological relation diagram of the equipment comprises the following steps:
41 The association relation of different parts of the power grid equipment is represented by a graph structure, and a node N i Representing different equipment parts, edge E i Representing edges between different devices, e.g. by W i,j Representing a device node N i And N j The association relation between the two components is used for constructing a topological relation diagram G (N, E) between the internal components of the power grid equipment;
42 For W) ij The numerical value of the node is calculated by using the statistical value of the corresponding equipment of the node, and the calculation formula is as follows:
wherein Count (N) i ) And Count (N) j ) Representing nodes respectivelyN i And node N j The frequency of the corresponding equipment faults is taken as the association weight W between the nodes by taking the ratio of the frequency minimum value to the frequency maximum value i,j 。
5. The method for monitoring the running state of the power grid equipment based on multi-modal data joint characterization and dynamic weight learning according to claim 1, wherein the dynamic weight learning of the graph neural network comprises the following steps:
51 For any graph neural network node N, the initial state of the node N is expressed as alpha by a single-mode unified vector of node monitoring data;
52 Associated weight W ij Is used for updating the graph nodes in the dynamic weight learning process of the graph neural network, if the node N has K adjacent nodes, the corresponding adjacent nodes are { N } 1 ,N 2 ,…N K The monitoring information of the corresponding neighboring node is denoted as { alpha } 1 ,α 2 ,…α K The association weights of node N and neighboring nodes are { W }, respectively 1 ,W 2 ,…W K W, where W 1 ,W 2 ,…W K Corresponding to the associated weight W ij ；
……
The first round of iterative process obtains a new state vector representation alpha' of the node N, and the specific calculation is as follows:
Mini(A 1 ,A 2 ,…,A n )＝A i wherein i is such that A i For the sequence (A) 1 ,A 2 ,…,A n ) Subscript of minimum of (2); max (A) 1 ,A 2 ,…,A n )＝A i Wherein i is such that A i For the sequence (A) 1 ,A 2 ,…,A n ) Subscript of the maximum value of (v); count (N) represents the frequency of the equipment failure corresponding to node N;
and sequentially carrying out one round of message transmission on all nodes of the topological relation diagram G (N, E) to obtain a new set of vector representations, and sequentially repeating until the vector representations of all the nodes are converged.
6. The method for monitoring the running state of the power grid equipment based on multi-mode data combined characterization and dynamic weight learning according to claim 1, wherein the evaluation of the running state of the power grid equipment comprises the following steps:
61 For the power grid equipment component S, text, image and sound state data of the equipment are acquired, and vectors T are respectively obtained by a distributed representation method S 、P S 、V S Input adaptive dimension encoder, to image adaptive dimension encoder and sound adaptive dimension encoder to obtain a single-mode vector representation alpha T 、α P And alpha V Computing a joint characterization of S
Calculating the joint characterization of other parts of the power grid equipment by the method;
62 For all components of the power grid device, the joint representation of the components is used for representing the node states in the graph neural network G (N, E), the component S is represented by a node N in the graph, and the state vector of the node N is represented by the joint representation alpha of S;
the state vector of other parts of the power grid equipment in the graphic neural network G (N, E) is represented by the method;
63 When all the components of the power grid equipment calculate to obtain a state vector at the nodes in the graph neural network G (N, E), the nodes of the graph neural network are iteratively updated by using the association weights among the nodes in the graph neural network G (N, E), the information of the neighborhood nodes is learned, and the node N calculates to obtain a new state vector to represent alpha';
calculating new state vector representations of all other nodes;
64 Obtaining new state vector representation X of all nodes of the graph neural network, and combining an empirical state evaluation health grade Y endowed by an expert to the power grid equipment, and establishing an association relationship between the X and the Y by using the fully-connected neural network to obtain a state evaluation function F of the power grid equipment;
Y＝F(X)；
65 For the power grid equipment at any moment, calculating new state vector representation X of all nodes by using a state evaluation function F of the power grid equipment to obtain state evaluation health grade Y of the power grid equipment, thereby achieving the purpose of intelligent state monitoring of the power grid equipment.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311154552.3A CN117172413B (en) | 2023-09-07 | 2023-09-07 | Power grid equipment operation state monitoring method based on multi-mode data joint characterization and dynamic weight learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311154552.3A CN117172413B (en) | 2023-09-07 | 2023-09-07 | Power grid equipment operation state monitoring method based on multi-mode data joint characterization and dynamic weight learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN117172413A true CN117172413A (en) | 2023-12-05 |
| CN117172413B CN117172413B (en) | 2024-05-14 |
Family
ID=88940854
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311154552.3A Active CN117172413B (en) | 2023-09-07 | 2023-09-07 | Power grid equipment operation state monitoring method based on multi-mode data joint characterization and dynamic weight learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117172413B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117421998A (en) * | 2023-12-18 | 2024-01-19 | 国网湖北省电力有限公司经济技术研究院 | Multi-mode data-based power transmission overhead line health state evaluation system |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113919540A (en) * | 2021-08-04 | 2022-01-11 | 鹏城实验室 | Method for monitoring running state of production process and related equipment |
| CN114818483A (en) * | 2022-04-14 | 2022-07-29 | 东南大学溧阳研究院 | Electromechanical disturbance positioning and propagation prediction method based on graph neural network |
| CN116071189A (en) * | 2022-12-29 | 2023-05-05 | 香港理工大学深圳研究院 | Main distribution network cooperative voltage and frequency control method based on graphic neural network intelligent agent |
| CN116155581A (en) * | 2023-02-07 | 2023-05-23 | 浙江大学 | Network intrusion detection method and device based on graph neural network |
| CN116316637A (en) * | 2023-03-20 | 2023-06-23 | 中国电力科学研究院有限公司 | Dynamic topology identification method, system, equipment and storage medium for power distribution network |
-
2023
- 2023-09-07 CN CN202311154552.3A patent/CN117172413B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113919540A (en) * | 2021-08-04 | 2022-01-11 | 鹏城实验室 | Method for monitoring running state of production process and related equipment |
| CN114818483A (en) * | 2022-04-14 | 2022-07-29 | 东南大学溧阳研究院 | Electromechanical disturbance positioning and propagation prediction method based on graph neural network |
| CN116071189A (en) * | 2022-12-29 | 2023-05-05 | 香港理工大学深圳研究院 | Main distribution network cooperative voltage and frequency control method based on graphic neural network intelligent agent |
| CN116155581A (en) * | 2023-02-07 | 2023-05-23 | 浙江大学 | Network intrusion detection method and device based on graph neural network |
| CN116316637A (en) * | 2023-03-20 | 2023-06-23 | 中国电力科学研究院有限公司 | Dynamic topology identification method, system, equipment and storage medium for power distribution network |
Non-Patent Citations (4)
| Title |
|---|
| HANSEN JB等: "Power Flow Balancing With Decentralized Graph Neural Networks", JOURNALS & MAGAZINES, vol. 38, no. 03, 1 August 2022 (2022-08-01), pages 2423 - 2433 * |
| ZHANG M等: "Graph Neural Network based Short-term Solar Irradiance Forcasting Model Considering Surrounding Meteorological Factors", IAS 58TH INDUSTRIAL AND COMMERCIAL POWER SYSTEMS TECHNICAL CONFERENCE, 17 May 2022 (2022-05-17), pages 1 - 9, XP034124104, DOI: 10.1109/ICPS54075.2022.9773879 * |
| 周晓燕 等: "基于图神经网络的智能电网监测系统的设计", 电脑知识与技术, vol. 18, no. 13, 5 May 2022 (2022-05-05), pages 77 - 78 * |
| 毛先胤 等: "基于电力大数据的多源异构数据融合技术研究与应用", 电力大数据, vol. 23, no. 08, 21 August 2020 (2020-08-21), pages 33 - 39 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117421998A (en) * | 2023-12-18 | 2024-01-19 | 国网湖北省电力有限公司经济技术研究院 | Multi-mode data-based power transmission overhead line health state evaluation system |
| CN117421998B (en) * | 2023-12-18 | 2024-03-12 | 国网湖北省电力有限公司经济技术研究院 | Multi-mode data-based power transmission overhead line health state evaluation system |
Also Published As
| Publication number | Publication date |
|---|---|
| CN117172413B (en) | 2024-05-14 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112632972B (en) | Method for rapidly extracting fault information in power grid equipment fault report | |
| CN110705873B (en) | Power distribution network running state portrait analysis method | |
| CN114580706A (en) | Power financial service wind control method and system based on GRU-LSTM neural network | |
| CN112100369A (en) | Semantic-combined network fault association rule generation method and network fault detection method | |
| CN113239186A (en) | Graph convolution network relation extraction method based on multi-dependency relation representation mechanism | |
| CN117172413B (en) | Power grid equipment operation state monitoring method based on multi-mode data joint characterization and dynamic weight learning | |
| CN116910633B (en) | Power grid fault prediction method based on multi-modal knowledge mixed reasoning | |
| CN117237559B (en) | Digital twin city-oriented three-dimensional model data intelligent analysis method and system | |
| CN115223049A (en) | Knowledge distillation and quantification technology for large model compression of electric power scene edge calculation | |
| CN113516379B (en) | Work order scoring method for intelligent quality inspection | |
| CN108447475A (en) | A kind of method for building up of the speech recognition modeling based on electric power dispatching system | |
| Bi et al. | Knowledge transfer for out-of-knowledge-base entities: Improving graph-neural-network-based embedding using convolutional layers | |
| CN116225760A (en) | Real-time root cause analysis method based on operation and maintenance knowledge graph | |
| CN116187508A (en) | Wind turbine generator system fault prediction method integrating knowledge graph and deep learning | |
| CN117993868B (en) | Dual-mode attention-based power grid engineering project audit early warning method and system | |
| CN115409122A (en) | Method, system, equipment and medium for analyzing concurrent faults of power transformation equipment | |
| CN117540035B (en) | RPA knowledge graph construction method based on entity type information fusion | |
| CN116258054A (en) | Intelligent diagnosis model construction method for constant potential instrument faults of negative protection system | |
| CN118194487A (en) | Automatic arrangement method, medium and system for circuit and electric equipment | |
| CN113705242A (en) | Intelligent semantic matching method and device for education consultation service | |
| CN117743595A (en) | Power monitoring event identification and intelligent analysis method based on graph neural network | |
| Guo et al. | Wind speed forecasting of genetic neural model based on rough set theory | |
| CN116304112A (en) | Intelligent monitoring method based on big data technology | |
| CN112529188B (en) | Knowledge distillation-based industrial process optimization decision model migration optimization method | |
| CN117133116B (en) | Traffic flow prediction method and system based on space-time correlation network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |