CN117148034A - Power distribution network fault line selection method based on empirical wavelet transformation and graph annotation force network - Google Patents
Power distribution network fault line selection method based on empirical wavelet transformation and graph annotation force network Download PDFInfo
- Publication number
- CN117148034A CN117148034A CN202310994871.9A CN202310994871A CN117148034A CN 117148034 A CN117148034 A CN 117148034A CN 202310994871 A CN202310994871 A CN 202310994871A CN 117148034 A CN117148034 A CN 117148034A
- Authority
- CN
- China
- Prior art keywords
- distribution network
- power distribution
- signal
- line selection
- empirical wavelet
- 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
- 230000009466 transformation Effects 0.000 title claims abstract description 25
- 238000010187 selection method Methods 0.000 title claims abstract description 21
- 238000000034 method Methods 0.000 claims description 38
- 238000000354 decomposition reaction Methods 0.000 claims description 11
- 230000007246 mechanism Effects 0.000 claims description 11
- 230000008569 process Effects 0.000 claims description 11
- 238000001228 spectrum Methods 0.000 claims description 11
- 230000008602 contraction Effects 0.000 claims description 6
- 238000011176 pooling Methods 0.000 claims description 5
- 238000001914 filtration Methods 0.000 claims description 4
- 238000003062 neural network model Methods 0.000 claims description 3
- 238000013528 artificial neural network Methods 0.000 abstract description 12
- 230000001052 transient effect Effects 0.000 abstract description 8
- 238000004422 calculation algorithm Methods 0.000 abstract description 6
- 238000004458 analytical method Methods 0.000 abstract description 2
- 238000012360 testing method Methods 0.000 abstract description 2
- 239000010410 layer Substances 0.000 description 18
- 230000006870 function Effects 0.000 description 12
- 238000004364 calculation method Methods 0.000 description 8
- 238000010586 diagram Methods 0.000 description 5
- 238000012545 processing Methods 0.000 description 4
- 239000013598 vector Substances 0.000 description 4
- 230000004913 activation Effects 0.000 description 3
- 238000010276 construction Methods 0.000 description 3
- 230000007704 transition Effects 0.000 description 3
- 230000000694 effects Effects 0.000 description 2
- 230000004927 fusion Effects 0.000 description 2
- 238000002347 injection Methods 0.000 description 2
- 239000007924 injection Substances 0.000 description 2
- 238000013507 mapping Methods 0.000 description 2
- 238000010606 normalization Methods 0.000 description 2
- 238000005192 partition Methods 0.000 description 2
- 238000003672 processing method Methods 0.000 description 2
- 238000012795 verification Methods 0.000 description 2
- 238000012935 Averaging Methods 0.000 description 1
- 101100499229 Mus musculus Dhrsx gene Proteins 0.000 description 1
- 108010076504 Protein Sorting Signals Proteins 0.000 description 1
- 230000002776 aggregation Effects 0.000 description 1
- 238000004220 aggregation Methods 0.000 description 1
- 230000032683 aging Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 238000009413 insulation Methods 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 238000010248 power generation Methods 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
- 239000002356 single layer Substances 0.000 description 1
- 239000000243 solution Substances 0.000 description 1
- 238000000547 structure data Methods 0.000 description 1
- 238000012549 training Methods 0.000 description 1
- 230000001131 transforming effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
- G01R31/08—Locating faults in cables, transmission lines, or networks
- G01R31/081—Locating faults in cables, transmission lines, or networks according to type of conductors
- G01R31/086—Locating faults in cables, transmission lines, or networks according to type of conductors in power transmission or distribution networks, i.e. with interconnected conductors
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R19/00—Arrangements for measuring currents or voltages or for indicating presence or sign thereof
- G01R19/175—Indicating the instants of passage of current or voltage through a given value, e.g. passage through zero
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
- G01R31/08—Locating faults in cables, transmission lines, or networks
- G01R31/088—Aspects of digital computing
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
- Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
- Y04S10/00—Systems supporting electrical power generation, transmission or distribution
- Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications
- Y04S10/52—Outage or fault management, e.g. fault detection or location
Landscapes
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Theoretical Computer Science (AREA)
- Locating Faults (AREA)
Abstract
A power distribution network fault line selection method based on empirical wavelet transformation and graph annotation force network belongs to the technical field of power systems and automation thereof. According to the invention, the characteristic of self-adaptive frequency band division of the empirical wavelet transformation is utilized, the characteristic analysis is carried out on the zero sequence current signal of the power distribution network, and the frequency band distributed in a concentrated way of extracting transient information is utilized, so that the correct judgment on the fault feeder line is facilitated; the input features are converted into higher-level features based on the graph attention neural network and classified, so that fault line selection is realized. Experimental results prove that the empirical wavelet transformation algorithm can well divide the frequency band containing the zero sequence current data when the distributed power distribution network fails, extract the characteristic components by combining the envelope entropy, perform learning test on the characteristic components through the graph attention neural network algorithm, and return the feeder types, wherein the feeder types are two classifications of failed feeders and sound feeders, so that the fault line selection of the distributed power distribution network is realized.
Description
Technical Field
The invention belongs to the technical field of power systems and automation thereof, and particularly relates to a power distribution network fault line selection method based on empirical wavelet transformation and a graph annotation force network.
Background
The distribution line has the characteristics of complex erection environment, wide distribution range, numerous nodes and the like. Under the influence of conditions such as line insulation aging, bad weather, the distribution network fault frequently occurs, influences the normal work of load, if untimely excision probably leads to the conflagration for power equipment damages, influences the distribution network reliability and stability, threatens personal safety. The fault feeder line of the power distribution network is rapidly and accurately determined and timely cut off, and the method has important significance for maintaining safe and stable operation of the power distribution network. And as the distribution network is continuously connected with the distributed energy, the distribution system is easy to generate operation fluctuation due to the instability and intermittence of the power generation of the distributed energy. The distributed power supply access changes the current distribution when faults occur, and meanwhile, the zero sequence current amplitude of most power distribution network faults is smaller, the fault characteristic quantity is weak, and the line selection difficulty is higher under the strong noise background.
The traditional power distribution network fault line selection method mainly comprises the following steps: 1) The signal injection method needs to inject additional signals into the power distribution network to realize line selection, and the method has good robustness, but is not suitable for high-resistance fault line selection, and in addition, because of the need of an additional signal generator, the primary equipment investment is relatively high. 2) The steady-state characteristic component method mainly comprises a group ratio amplitude-phase method, a zero sequence reactive power method, a negative sequence current method and the like. Although steady-state characteristics are stable, the characteristic information quantity is often insufficient, and fault characteristics are weak. 3) The transient characteristic component method is characterized in that when a distribution line breaks down, the electric signal transient component contains rich fault information, the fault characteristic is obvious, and the transient characteristic component method comprises a transient energy method, a DC attenuation component method, a first half-wave method and the like. Transient fault characteristics are more abundant than steady state characteristic information quantity, but are easily interfered by electromagnetic noise, harmonic signals and the like in the power distribution network.
There is a need in the art for a new solution to this problem.
Disclosure of Invention
The technical problems to be solved by the invention are as follows: the power distribution network fault line selection method based on the empirical wavelet transformation and the graph annotation meaning network is used for solving the problem that the traditional power distribution network fault line selection method is not suitable for high-resistance fault line selection and has high primary equipment investment; the fault characteristic information quantity is insufficient; the fault characteristics are complex, and the technical problems of signal interference, difficulty in accurate extraction and the like exist.
The power distribution network fault line selection method based on the empirical wavelet transformation and the graph annotation force network comprises the following steps of:
step one, power distribution network zero sequence current signal decomposition and characteristic component selection based on empirical wavelet transformation
Performing empirical wavelet decomposition on zero sequence current signals of all measuring points acquired by a power distribution network, arranging all MRA components obtained by decomposition from low to high according to frequency, acquiring the envelope entropy of all MRA components by a signal characteristic component selection method based on the envelope entropy, and selecting the MRA component corresponding to the minimum envelope entropy as the characteristic component of the zero sequence current signals;
step two, adopting a graph attention neural network model to learn the relationship among nodes by weighting attention of neighborhood nodes so as to realize fault line selection of the power distribution network
And finally, outputting a classification result through a Softmax classifier, wherein the output result is 0 for representing a sound feeder, the output result is 1 for representing a fault feeder, and the fault feeder of the distribution network containing distributed energy sources is selected.
The specific method for performing empirical wavelet decomposition on the zero sequence current signals of each measuring point in the first step comprises the following steps: firstly, selecting a zero sequence current signal of a power distribution network as input data, carrying out Fourier transform on the data, then carrying out self-adaptive cutting on the frequency spectrum of the signal, constructing a series of filters in different frequency bands, filtering the signal and reconstructing the signal to correspondingly obtain a series of MRA components with tight support characteristics.
The signal characteristic component selection method of the envelope entropy reflects the fluctuation degree of the envelope of different signal components, the envelope entropy is obtained by carrying out Hilbert demodulation on the component signals, and the envelope entropy value is calculated in a mode of combining the envelope signals with the information entropy and is used for representing the characteristic component with the largest fault characteristic information, so that the complexity of the signal is measured.
The attention layer processes the relation among the nodes, calculates the attention weight of each node and the neighborhood nodes thereof through an attention mechanism, thereby obtaining the characteristics among the nodes and capturing the relevance among the data; and the image pooling layer performs edge contraction on the image structure, selects reserved or discarded nodes through fractional computation, selects a node subset with the highest characteristic information content, and generates a more representative new image.
Through the design scheme, the invention has the following beneficial effects:
the invention can accurately extract fault characteristics so as to rapidly realize fault line selection and cutting. The empirical wavelet transformation is a non-stationary signal processing method, by cutting the fourier spectrum of a signal into continuous intervals, filtering the continuous intervals by using a tightly supported wavelet filter bank, and finally obtaining a set of amplitude modulation and frequency modulation components by signal reconstruction. According to the method, the signal spectrum is decomposed into different frequency bands, so that localized signal characteristics are realized, the frequency bands can be selected in a self-adaptive mode, and the defect of modal aliasing caused by discontinuous time-frequency scale of the signal is avoided. The method provided by the invention is mainly characterized in that the characteristic of self-adaptive frequency band division of empirical wavelet transformation is utilized to perform characteristic analysis on the zero sequence current signal of the power distribution network, and the frequency band distributed in a concentrated way of extracting transient information is utilized to facilitate the subsequent correct judgment on the fault feeder line. The envelope entropy is a nonlinear time sequence signal processing method, and after the envelope spectrum is obtained by performing Hilbert demodulation on the signal, the complexity and the irregularity of the signal are quantitatively analyzed by utilizing the information entropy, so that the randomness of the signal is measured.
The graph attention neural network is a neural network which introduces a self-attention mechanism to process graph structure data, and the node information is selectively weighted for neighborhood nodes to realize signal data classification, so that the expression capacity of a model is improved, and the classification is accurate and efficient. The method provided by the invention is based on the graph attention neural network to convert the input characteristics into higher-level characteristics and classify the higher-level characteristics, thereby realizing fault line selection.
The power distribution network model containing the distributed power supply is built in PSCAD simulation software, so that simulation and verification are carried out on the method. Experimental results prove that the empirical wavelet transformation algorithm can well divide the frequency band containing the zero sequence current data when the distributed power distribution network fails, extract the characteristic components by combining the envelope entropy, perform learning test on the characteristic components through the graph attention neural network algorithm, and return the feeder types, wherein the feeder types are two classifications of failed feeders and sound feeders, so that the fault line selection of the distributed power distribution network is realized.
Drawings
The invention is further described with reference to the drawings and detailed description which follow:
fig. 1 is a flow chart of a power distribution network fault line selection method based on empirical wavelet transformation and a graph annotation force network.
Fig. 2 is a schematic diagram of the division of fourier axes in an embodiment of the power distribution network fault line selection method based on empirical wavelet transform and graph injection force network according to the present invention.
Fig. 3 is a schematic diagram of attention coefficient calculation in an embodiment of a power distribution network fault line selection method based on empirical wavelet transform and a graph attention network according to the present invention.
Fig. 4 is a flowchart illustrating the operation of the multi-head attention network in an embodiment of the power distribution network fault line selection method based on the empirical wavelet transform and the attention network of the present invention.
Detailed Description
The power distribution network fault line selection method based on the empirical wavelet transformation and the graph annotation force network is shown in fig. 1, and comprises the following steps:
step one, power distribution network zero sequence current signal decomposition and characteristic component selection based on empirical wavelet transformation
The first step comprises a zero sequence current data input process of an S1 power distribution network database; s2, performing a data signal decomposition process by using the empirical wavelet.
And S2, in the process of decomposing the data signals by using the empirical wavelet, processing the input signals by applying Fourier transformation, obtaining a group of boundary points after carrying out self-adaptive frequency band division on the processed signals by using a scale space method, establishing a filter bank by using a boundary sequence and the empirical wavelet, and finally obtaining different MRA component signals in a time domain and a frequency domain through filtering and reconstruction operation. Taking original data I (t) of a power distribution system as input, and obtaining a frequency-division band component I after empirical wavelet transformation processing j (t) as output, the empirical wavelet transform divides the signal band from low to high, while suppressing modal aliasing with good anti-noise interference capability, comprising the specific steps of:
s21 frequency-band divides the input signal.
And (3) applying Fourier transform to the acquired power distribution network zero sequence current signals, and setting a Fourier spectrum support interval within a range of [0, pi ].
Assuming that the zero sequence current signal is composed of N single component components, the support interval of the Fourier spectrum is divided into N continuous intervals during the processing, and the boundary of each interval is expressed as omega n ，ω n-1 Represents the n-1 th segment interval boundary; the band partition boundaries total n+1, where ω 0 ＝0，ω π Pi. A schematic of the segmentation of the fourier axis is shown in fig. 2.
The frequency band partition is Λ n The following formula is shown:
Λ n ＝[ω n-1 ,ω n ],n＝1,2,…,N
the total support interval formula for the fourier axis is as follows:
s22, constructing an empirical wavelet function and an empirical scale function.
Empirical wavelet function ψ n The formula (ω) is as follows:
empirical scale functionThe formula is as follows:
wherein beta (x) is [0,1 ]]Arbitrary function satisfying K-order derivative in interval, β (x) =x 4 (35-84x+
70x 2 -20x 3 ) The method comprises the steps of carrying out a first treatment on the surface of the ω represents the absolute value of the angular frequency; τ n A filter half-bandwidth representing an nth segment interval; τ n+1 A filter half-bandwidth representing the n+1th segment interval; omega n+1 Indicating the n+1st segment interval boundary.
S23, detail coefficient and approximation coefficient calculation.
Detail coefficientFrom the warpWavelet-checking function psi n The inner product of (t) and the zero sequence current signal I (t) is generated as follows:
approximation coefficientFrom empirical scale functions->The inner product of the zero sequence current signal I (t) is generated, and the formula is as follows:
wherein, psi is n (omega) andrespectively is psi n (t) and->F [. Cndot.]And F -1 [·]Respectively representing fourier transform and its inverse, τ represents the half-bandwidth of the filter, and I (ω) represents a frequency domain signal obtained by fourier transforming the time domain signal I (t).
S24, reconstructing a signal.
The zero sequence current signal is reconstructed according to the following formula:
the zero sequence current signal I (t) is decomposed by EWT to obtain the single component I of amplitude modulation and frequency modulation j (t) (j=1, 2,3, …), the frequency is expressed as follows:
in the step S21, ω is determined n The method comprises the following steps:
s21.1, determining the maximum number.
For Fourier spectrum of the zero sequence current signal I (t), assuming that M maximum points exist in the spectrum in total by utilizing an algorithm, and carrying out descending order arrangement on the M maximum points, if M is more than or equal to N, N is the number of mode components, indicating that enough maximum values are found and can be used for spectrum division, and reserving the first N-1 maximum values at the moment; if M < N, the preset value of the number of mode components in the signal is larger, the selected M maximum values are reserved, and the parameter N is reset, so that n=m.
S21.2 mode number N value estimation.
Searching M maximum value points of the Fourier spectrum of the zero sequence current signal, and forming a set of M maximum value pointsFor M in a collection i The values are arranged in descending order to give M 1 >M 2 >M 3 >…>M M Setting a threshold M M +α(M 1 -M M ) At this time, the number of maximum points above the threshold is the value of N, where α is the relative amplitude ratio.
S21.3 constructs boundaries and transitions.
For the zero sequence current signal I (t), the corresponding angular frequency omega of the first N maximum values in the frequency domain is set n (n=1, …, N), then the frequency band of the empirical wavelet transform divides the boundaryTo facilitate the construction of the filter in subsequent operations, the transition width is defined as T n ＝2τ n These transitions are bounded by a band boundary omega n Centered, where τ n Half-bandwidth, τ, of filter for nth segment interval n ＝γω n (0＜γ＜1)，/>Gamma represents a scaling factor;
and secondly, adopting a graph attention neural network model to learn the relationship among nodes by weighting attention of the neighborhood nodes, and realizing fault line selection of the power distribution network. The method comprises the steps of envelope entropy calculation, graph database construction and input to a graph neural network, wherein the graph neural network comprises a graph attention layer, a pooling layer and a full connection layer, and finally, a classification result is output through a Softmax classifier.
The envelope entropy can be used for calculating the flatness of the signal envelope so as to reflect the complexity of the signal, wherein the envelope entropy is smaller when the instantaneous envelope value of the signal is distributed intensively, and the envelope entropy is larger when the instantaneous envelope value is distributed uniformly; if the noise in the MRA component is more and the characteristic information is less after the EWT is decomposed, the sparsity of the signal is weaker, the corresponding envelope entropy is larger, and otherwise, the envelope entropy value is smaller. Envelope entropy value E p Calculated with the following formula:
wherein H represents Hilbert transform of the signal I (j); a (j) is represented as a signal I (j) (j=1, 2, …, m) and p is obtained by normalizing a (j) by an envelope signal sequence obtained by hilbert demodulation j 。
According to MRA components obtained by decomposing signals through empirical wavelet transformation, calculating and obtaining envelope entropy under different MRA components, and taking the MRA component corresponding to the minimum envelope entropy value as a characteristic component of an original signal, namely:
Y＝f(min(E pj ))
wherein min (E pj ) And f is the corresponding relation between the envelope entropy and the MRA component.
Graph database construction process: q measuring points of the power distribution network are used as nodes of a graph database, line connection among the measuring points is used as an edge, and signals after MRA component normalization corresponding to minimum envelope entropy are used as node characteristics h u A graph data sample of the zero sequence current is thus created and the graph network is named G.
Fault line selection is carried out according to the extracted characteristics: when a distribution network line fails, fault zero-sequence current flows along the distribution line to the network. Because of the numerous nodes of the power distribution network, the topological structure is complex, and the zero sequence current at each measuring point and the fault point form differential nonlinear mapping. The topological structure of the distribution network and the extracted characteristic component information of the signals are fused by the graph attention neural network, the measuring points are used as nodes of the graph attention neural network, the lines between the measuring points are used as edges of the graph attention neural network, the characteristic components corresponding to the minimum envelope entropy of the signals are used as node characteristics by EWT decomposition, and the structural relevance deep learning component characteristics of each measuring point are utilized, so that a fault line is selected.
The technical scheme of the invention also comprises the following steps of:
s6.1, constructing a data diagram.
Q current measuring points exist in the power distribution network, topology connection is carried out according to the positions of the measuring points in the power distribution network, and q node diagram data of fault zero-sequence current information can be formed. Max-Min normalization is carried out on MRA characteristic components corresponding to the minimum envelope entropy, and node characteristics h of fault zero sequence current data are obtained u 。
S6.2, calculating a single attention mechanism of the attention layer.
The schematic diagram of the attention coefficient calculation is shown in fig. 3.
For vertex u, its neighbor nodes (v ε N) are computed one by one u ) And similarity coefficient between itself:
e uv ＝a([Wh u ||Wh v ]),v∈N u
h u and h v The characteristics of the central node and the adjacent nodes thereof respectively. W is used to map features, enhance the expressive power of features, [/i ]]Means to splice high-dimensional features, combine the mapped features, and map into a real number by a (). v denotes a single neighbor node of vertex u; n (N) u A set of all neighbor nodes representing vertex u; after the mapping is realized by the single-layer feedforward neural network, the attention coefficient is solved by the following formula:
and taking the calculated attention coefficient result as an aggregation weight, and fusing the characteristics of the neighbor nodes:
wherein h is u ' is a new feature of vertex u after fusion with a neighboring node, and δ is an activation function.
S6.3, calculating the multi-head attention mechanism of the attention layer.
In order to stabilize the self-attention learning process, the attention mechanism is extended to a multi-head attention mechanism, and K independent attention mechanism calculation is utilized, wherein each attention mechanism function is only responsible for one subspace in the final output sequence.
Wherein pi represents the splicing operation,for being by kth attention mechanism->Calculated normalized attention coefficient, W k For the weight matrix of the corresponding input linear transformation, the final returned output h u ' K ' will consist of KF ' features for each node.
The predictive layer of the multi-headed attention network uses an averaging operation with a delay applied to the final nonlinear function, the operation being shown in fig. 4, expressed as:
wherein σ represents a Softmax function;
and calculating the importance degree of each neighbor node to the central node by using an attention mechanism, so that the whole information of the graph network is obtained from the local information. When each node updates the feature output, the attention coefficient calculation is needed to be carried out on the neighbor nodes, so that corresponding weights are distributed to each neighbor node, the emphasis is placed on the nodes which are considered to be more useful for the model, and the nodes with lower importance are ignored. Each node is calculated independently from its neighbor nodes and is assigned a weight of arbitrary value.
The input of the graph attention layer is a node feature vector set, and the expression is:
h l ＝{h l,1 ,h l,2 ,…,h l,q }
wherein h is l Inputting a first layer of a graph annotation force network; h is a l,q Is the input feature vector for node q in the first layer.
The output of the graph attention layer is a new set of node feature vectors, expressed as:
h′ l ＝{h′ l,1 ,h′ l,2 ,…,h′ l,q }
wherein h' l The output of the first layer obtained after the activation function is the input of the first layer (1+1); h's' l,q Is the output eigenvector of node q in the first layer.
Pooling process of pooling layer edge shrinkage: to the contraction edgex＝{z u ,z v Introduction of a new node z e And a new connecting edge, so that z e And all nodes z u Or z v Adjacent. Z is omitted from the figure u And z v And all sides thereof. The contraction process may be denoted as G/x, and each contraction operation may decrease the number of nodes of the graph network by 1. To avoid different edges from impinging on the same node, an edge set contraction G/X 'is defined, where the set X' = { X 1 ,x 2 ,…,x n }∈X。
The edge shrink pool computes the raw score of each edge using a simple linear combination of the cascade node features to determine whether to preserve the nodes connected to the edge. For the edge from node u to node v, the calculation formula of the original score r is:
r(x uv )＝W(h′ u ||h′ v )+b
to compare the scores of the edges, the scores of the adjacent edges of the nodes are normalized, and the formula is as follows:
s uv ＝0.5+softmax r*v (r uv )
the adjacent nodes with highest scores and not merged are sequentially subjected to shrinkage fusion, and node characteristics are updated by adding and summing the fused new nodes, namely,
the process of the full connection layer: the full connection layer formula is as follows.
Wherein k is w Represent the weight, d w Representing a w-dimensional input vector, b representing the bias, g being the output, i.e. the eigenvalue sought.
The feature g calculates the probability distribution of the category to which the line belongs through a Softmax activation function, and the probability formula of the Softmax is as follows:
g 1 an output value representing the 1 st node; g 2 An output value representing the 2 nd node;
to minimize the loss value of the model, a two-class cross entropy loss function is selected, expressed as follows:
wherein L is x Represents the cross entropy of sample x; y is x A label representing sample x, negative class 0, positive class 1; p is p x Representing the probability that sample x is predicted to be a positive class.
And outputting a classification result: output results 0 and 1 represent a sound feeder and a faulty feeder, respectively.
The fault line selection effect of the invention compared with the fault line selection effect of the traditional method is shown in the following table:
proved by verification, the method realizes fault line selection of the power distribution network containing the distributed power supply. By the method of empirical wavelet transformation, fault zero-sequence current data information of the power distribution network can be effectively decomposed, MRA components can be adaptively cut in different frequency bands, and transient steady state characteristics of zero-sequence current signals can be effectively represented. By the method of envelope entropy calculation, the MRA component with the most fault characteristic information is effectively selected, the corresponding relation between the fault information of the power distribution network and the self-adaptive frequency band is enhanced, the burden of the signal component corresponding to the frequency band with the less fault characteristic information content on subsequent learning is avoided, and the data processing efficiency is improved. Through a graph attention network mechanism, complex nonlinear association of network topology and zero sequence current data of the distribution network is deeply combined, and the fault line selection of the distribution network is converted into two classification problems of a fault feeder and a sound feeder, so that the fault line selection of the distribution network under the distributed energy access is realized. The method has strong feasibility, can be combined with an actual distribution network for training, has strong algorithm reliability, can adapt to various power distribution network fault types, and has significance in application in actual engineering.
Claims (4)
1. A power distribution network fault line selection method based on empirical wavelet transformation and graph annotation force network is characterized by comprising the following steps: comprising the following steps, and the following steps are carried out in sequence,
step one, power distribution network zero sequence current signal decomposition and characteristic component selection based on empirical wavelet transformation
Performing empirical wavelet decomposition on zero sequence current signals of all measuring points acquired by a power distribution network, arranging all MRA components obtained by decomposition from low to high according to frequency, acquiring the envelope entropy of all MRA components by a signal characteristic component selection method based on the envelope entropy, and selecting the MRA component corresponding to the minimum envelope entropy as the characteristic component of the zero sequence current signals;
step two, adopting a graph attention neural network model to learn the relationship among nodes by weighting attention of neighborhood nodes so as to realize fault line selection of the power distribution network
And finally, outputting a classification result through a Softmax classifier, wherein the output result is 0 for representing a sound feeder, the output result is 1 for representing a fault feeder, and the fault feeder of the distribution network containing distributed energy sources is selected.
2. The power distribution network fault line selection method based on the empirical wavelet transformation and graph annotation force network according to claim 1 is characterized in that: the specific method for performing empirical wavelet decomposition on the zero sequence current signals of each measuring point in the first step comprises the following steps: firstly, selecting a zero sequence current signal of a power distribution network as input data, carrying out Fourier transform on the data, then carrying out self-adaptive cutting on the frequency spectrum of the signal, constructing a series of filters in different frequency bands, filtering the signal and reconstructing the signal to correspondingly obtain a series of MRA components with tight support characteristics.
3. The power distribution network fault line selection method based on the empirical wavelet transformation and graph annotation force network according to claim 1 is characterized in that: the signal characteristic component selection method of the envelope entropy reflects the fluctuation degree of the envelope of different signal components, the envelope entropy is obtained by carrying out Hilbert demodulation on the component signals, and the envelope entropy value is calculated in a mode of combining the envelope signals with the information entropy and is used for representing the characteristic component with the largest fault characteristic information, so that the complexity of the signal is measured.
4. The power distribution network fault line selection method based on the empirical wavelet transformation and graph annotation force network according to claim 1 is characterized in that: the attention layer processes the relation among the nodes, calculates the attention weight of each node and the neighborhood nodes thereof through an attention mechanism, thereby obtaining the characteristics among the nodes and capturing the relevance among the data; and the image pooling layer performs edge contraction on the image structure, selects reserved or discarded nodes through fractional computation, selects a node subset with the highest characteristic information content, and generates a more representative new image.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310994871.9A CN117148034A (en) | 2023-08-09 | 2023-08-09 | Power distribution network fault line selection method based on empirical wavelet transformation and graph annotation force network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310994871.9A CN117148034A (en) | 2023-08-09 | 2023-08-09 | Power distribution network fault line selection method based on empirical wavelet transformation and graph annotation force network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN117148034A true CN117148034A (en) | 2023-12-01 |
Family
ID=88897812
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310994871.9A Pending CN117148034A (en) | 2023-08-09 | 2023-08-09 | Power distribution network fault line selection method based on empirical wavelet transformation and graph annotation force network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117148034A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117648634A (en) * | 2024-01-30 | 2024-03-05 | 合肥工业大学 | Method and system for predicting performance of connecting hardware fitting of power distribution network based on time domain and frequency domain information |
-
2023
- 2023-08-09 CN CN202310994871.9A patent/CN117148034A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117648634A (en) * | 2024-01-30 | 2024-03-05 | 合肥工业大学 | Method and system for predicting performance of connecting hardware fitting of power distribution network based on time domain and frequency domain information |
| CN117648634B (en) * | 2024-01-30 | 2024-04-16 | 合肥工业大学 | Method and system for predicting performance of connecting hardware fitting of power distribution network based on time domain and frequency domain information |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110263839B (en) | Power system load static characteristic online intelligent identification method based on big data | |
| CN112200694A (en) | Dominant instability mode identification model construction and application method based on graph neural network | |
| CN116245033B (en) | Artificial intelligent driven power system analysis method and intelligent software platform | |
| CN105354643A (en) | Risk prediction evaluation method for wind power grid integration | |
| CN117148034A (en) | Power distribution network fault line selection method based on empirical wavelet transformation and graph annotation force network | |
| CN110808580A (en) | Quick identification method for voltage sag source based on wavelet transformation and extreme learning machine | |
| CN111680875A (en) | Unmanned aerial vehicle state risk fuzzy comprehensive evaluation method based on probability baseline model | |
| CN116169675B (en) | Power system dynamic stability online evaluation method considering operation mode change | |
| CN116131313A (en) | Explanatory analysis method for association relation between characteristic quantity and transient power angle stability | |
| CN112633315A (en) | Electric power system disturbance classification method | |
| CN111401792A (en) | Dynamic safety assessment method based on extreme gradient lifting decision tree | |
| Hong et al. | Deep‐belief‐Networks based fault classification in power distribution networks | |
| CN114252739A (en) | Power distribution network single-phase earth fault distinguishing method, system, equipment and storage medium | |
| Bai et al. | Open circuit fault diagnosis of wind power converter based on VMD energy entropy and time domain feature analysis | |
| Chelaru et al. | Decision support system to determine the replacement ranking of the aged transformers in electric distribution networks | |
| Zhu et al. | Aiming to Complex Power Quality Disturbances: A Novel Decomposition and Detection Framework | |
| Wilson et al. | Uncertainty Quantification of Capacitor Switching Transient Location Using Machine Learning | |
| CN114689991A (en) | Method and system for determining fault time of high-voltage cable and related components | |
| CN110968073B (en) | Double-layer tracing identification method for commutation failure reasons of HVDC system | |
| CN114169395A (en) | Construction method and application of dominant instability mode recognition model of power system | |
| Patel et al. | Wavelet and Machine learning based approach for Fault classification in AC Micro-grid system | |
| Zhou et al. | Transient stability assessment of large-scale AC/DC hybrid power grid based on separation feature and deep belief networks | |
| Qiao et al. | Transient stability assessment for ac-dc hybrid systems based on bayesian optimization xgboost | |
| Ismail et al. | Classification of voltage sag using multi-resolution analysis and support vector machine | |
| Bi et al. | Explainable Artificial Intelligence for Power System Security Assessment: A Case Study on Short-Term Voltage Stability |
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 |