CN115577290A - Distribution network fault classification and source positioning method based on deep learning - Google Patents
Distribution network fault classification and source positioning method based on deep learning Download PDFInfo
- Publication number
- CN115577290A CN115577290A CN202211192161.6A CN202211192161A CN115577290A CN 115577290 A CN115577290 A CN 115577290A CN 202211192161 A CN202211192161 A CN 202211192161A CN 115577290 A CN115577290 A CN 115577290A
- Authority
- CN
- China
- Prior art keywords
- data
- distribution network
- fault
- classification
- training 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.)
- Pending
Links
- 238000009826 distribution Methods 0.000 title claims abstract description 124
- 238000000034 method Methods 0.000 title claims abstract description 39
- 238000013135 deep learning Methods 0.000 title claims abstract description 18
- 238000012549 training Methods 0.000 claims abstract description 125
- 238000001514 detection method Methods 0.000 claims abstract description 20
- 230000008859 change Effects 0.000 claims abstract description 10
- 238000005259 measurement Methods 0.000 claims abstract description 5
- 238000000638 solvent extraction Methods 0.000 claims abstract description 4
- 230000006870 function Effects 0.000 claims description 35
- 238000004422 calculation algorithm Methods 0.000 claims description 30
- 239000013598 vector Substances 0.000 claims description 25
- 238000013434 data augmentation Methods 0.000 claims description 17
- 238000005192 partition Methods 0.000 claims description 17
- 238000013145 classification model Methods 0.000 claims description 16
- 238000013507 mapping Methods 0.000 claims description 12
- 238000009827 uniform distribution Methods 0.000 claims description 11
- 238000005516 engineering process Methods 0.000 claims description 7
- 238000005457 optimization Methods 0.000 claims description 6
- 238000012360 testing method Methods 0.000 claims description 5
- 238000010348 incorporation Methods 0.000 claims description 4
- 230000001186 cumulative effect Effects 0.000 claims description 3
- 238000009499 grossing Methods 0.000 claims description 3
- 238000012545 processing Methods 0.000 claims description 3
- 238000004088 simulation Methods 0.000 claims description 3
- 125000004122 cyclic group Chemical group 0.000 claims description 2
- 239000000126 substance Substances 0.000 claims description 2
- 230000002159 abnormal effect Effects 0.000 description 4
- 230000009286 beneficial effect Effects 0.000 description 3
- 241001522296 Erithacus rubecula Species 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 230000003416 augmentation Effects 0.000 description 1
- FFBHFFJDDLITSX-UHFFFAOYSA-N benzyl N-[2-hydroxy-4-(3-oxomorpholin-4-yl)phenyl]carbamate Chemical compound OC1=C(NC(=O)OCC2=CC=CC=C2)C=CC(=C1)N1CCOCC1=O FFBHFFJDDLITSX-UHFFFAOYSA-N 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000000354 decomposition reaction Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 230000035699 permeability Effects 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 238000010187 selection method Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/06—Energy or water supply
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
- Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
- Y04S10/00—Systems supporting electrical power generation, transmission or distribution
- Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications
- Y04S10/52—Outage or fault management, e.g. fault detection or location
Landscapes
- Business, Economics & Management (AREA)
- Health & Medical Sciences (AREA)
- Engineering & Computer Science (AREA)
- Economics (AREA)
- Public Health (AREA)
- Water Supply & Treatment (AREA)
- General Health & Medical Sciences (AREA)
- Human Resources & Organizations (AREA)
- Marketing (AREA)
- Primary Health Care (AREA)
- Strategic Management (AREA)
- Tourism & Hospitality (AREA)
- Physics & Mathematics (AREA)
- General Business, Economics & Management (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Complex Calculations (AREA)
Abstract
The invention discloses a distribution network fault classification and source positioning method based on deep learning, which comprises the following steps of: the method comprises the following steps that S1, a multi-classification SVDD data driving collection model is used for collecting power distribution network fault history data, and optimal parameters are obtained through multi-classification SVDD data driving collection model training data; s2, acquiring a DER dynamic change rate of the power distribution network according to the power distribution network offline data, acquiring a large amount of training data through the DER dynamic change rate of the power distribution network in combination with real-time data of the power distribution network, and effectively partitioning a plurality of branches of the power distribution network according to the training data; and S3, fault positioning detection is carried out on faults in the power distribution network according to the effective subareas of the power distribution network branches, fault sources are judged through the overall subarea fault probability measurement indication of the power distribution network, and the output probability of each subarea is compared, so that the subareas where the fault sources are positioned are effectively reduced, the fault range is gradually reduced, and operators of the power system are helped to find and remove fault nodes in time.
Description
Technical Field
The invention relates to the technical field of power distribution networks, in particular to a distribution network fault classification and source positioning method based on deep learning.
Background
With the advance of the construction process of the smart power grid and the continuous expansion of the installed scale of renewable energy sources, the problem of renewable energy source grid connection is concerned, the energy sources highly dispersed in the power distribution network are called distributed energy source DER, with the incorporation of large-scale DER, a difficult challenge is brought to the reliability and stability of the power distribution network, and particularly in the aspect of system protection, therefore, the fault detection and positioning of the distributed energy source power distribution system become the problem to be solved urgently.
In recent years, the development of power grid and internet of things technology brings various and massive sensing monitoring data, so that the fault of a system can be detected and positioned by adopting a data driving method, and considering that the power grid works in a normal state for most of time and the possible fault condition exponentially increases along with the size of the power grid, the collected data of all fault conditions are unrealistic, and the problems that the fault of the power distribution system cannot be timely found and isolated and the like are caused after DER is incorporated into the power distribution network at present, the existing method for detecting and positioning the fault of the power distribution network mainly has the following three problems: the method has the advantages that firstly, although the existing method for positioning the faults of the power distribution system based on the neural network is simple to implement, the method needs enough and high-quality training data, parameters such as the number of hidden layers in the network, the number of nodes of each layer of the neural network and the like need to be subjectively preset, and the prediction result is relatively subjective; secondly, the existing power distribution network fault detection and positioning method monitors the unsuitably related faults when the dynamic response of the power distribution network contains DER during online self-adaptive updating, and the database created by the faults of all nodes is time-consuming.
Disclosure of Invention
The invention aims to provide a distribution network fault classification and source positioning method based on deep learning, and the method is used for solving the technical problems that the training data size is large, the parameter setting is relatively subjective, and the DER cannot be used for fault positioning under the online condition in the prior art.
In order to solve the technical problems, the invention specifically provides the following technical scheme:
a distribution network fault classification and source positioning method based on deep learning is characterized by comprising the following steps:
the method comprises the following steps that S1, a multi-classification SVDD data driving collection model is used for collecting power distribution network fault historical data, optimal parameters are obtained through multi-classification SVDD data driving collection model training data, and the optimal parameters are verified through a simulink simulation model to obtain power distribution network offline data;
s2, acquiring a DER dynamic change rate of the power distribution network according to the power distribution network offline data, acquiring a large amount of training data through the DER dynamic change rate of the power distribution network in combination with real-time data of the power distribution network, and effectively partitioning a plurality of branches of the power distribution network according to the training data;
and S3, fault positioning detection is carried out on faults in the power distribution network according to the effective subareas of the power distribution network branches, and the subareas where the fault sources are located are further judged through the overall subarea fault probability measurement indication of the power distribution network, so that the fault sources are positioned.
As a preferred embodiment of the present invention, in step S1, the multi-class SVDD data-driven acquisition model maps data to a high-dimensional space through a kernel function, quickly determines by describing data boundaries, and introduces the multi-class SVDD classifier and a penalty parameter epsilon into any given training data i The multi-classification SVDD classifier seeks the optimal hypersphere with the minimum volume in the hyperspace and judges the training data x i The Euclidean distance between the spherical center a of the optimal hyper-sphere and the radius R of the optimal hyper-sphere so as to judge whether the training data is in the envelope of the optimal hyper-sphere or not, and the punishment parameter epsilon i The penalty parameter epsilon is used in the optimization objective function of corresponding training data i The expression is as follows:
ε i =‖x i -a‖-R 2
when the penalty parameter ε i And when the result is more than or equal to 0, allowing a part of training data to be misjudged by the multi-classification SVDD classifier.
As a preferred scheme of the present invention, the optimization objective function of the training data specifically comprises the following steps:
first, a mapping function is adoptedTraining data set X = { X = 1 ,x 2 ,…,x i ,…,x n From the initial feature space R d Mapping to a higher dimensional space R d′ And d' is greater than or equal to d, the new data set is mapped as:
the training data containing the mapping function is processed by a kernel function k (x, y), and the expression is as follows:
wherein x and y respectively represent training data and mapping values of the power distribution network branch;
and verifying whether the training data x belongs to the class of the training set data by adopting a Gaussian kernel function, wherein the Gaussian kernel function expression is as follows:
wherein x is i Training data, x, representing the ith branch j Representing training data for the jth branch.
As a preferred scheme of the invention, the Gaussian kernel function of the training data is subjected to cyclic iteration processing, and the optimal parameters are selected in the parameter searching range.
As a preferred scheme of the present invention, a model is constructed for the optimal parameters by the multi-classification SVDD classifier, and the specific steps are as follows:
first, the optimal parameters are set as an initial training data set X 0 And introduce a new data set X 1 The initial training data set X is 0 The trained SVDD classification model is recorded as omega 0 At this time, the set of support vectors is denoted as SV 0 Set of unsupported vectors denoted as SV 1 ;
Second, the newly added data set x is validated in a round robin condition i ∈X 1 Whether the initial model omega is satisfied or not 0 KKT condition of (1), if satisfied, Ω 0 For the currently trained model, if not, x is set i Incorporation into collectionsIn (2), selecting a satisfied non-support vector set
Finally, willAnd SV 0 Merge into a new training set X 0 And retraining the SVDD classification model omega by using the new training set 1 When Ω is 0 =Ω 1 And then continuously participating in the updated output SVDD classification model omega of the new data vector 1 。
As a preferred embodiment of the present invention, in the step S2, the newly added data set X is subjected to a data updating process 1 Introducing a data augmentation algorithm which generates spherical uniform distribution by taking original data as the sphere center, and adding the new data set X 1 Carrying out data augmentation to improve the model training speed, generating over-solving uniform distribution data under Gaussian distribution, and acquiring effective data subareas, wherein the data augmentation algorithm specifically comprises the following steps:
first, the initial training data set X is divided into 0 Randomly introducing data to generate a mean value of 0, squareA Gaussian distribution with a range of 1, i.e., X to N (0, 1);
secondly, calculating the initial training data set X 0 The squared Euclidean distance R of each sample to the origin 2 The square distance is gamma 2 Distributed with d degrees of freedom, wherein R 2 =‖X 0 ‖ 2 ;
Finally, using γ 2 Cumulative distribution of distributionsR is to be 2 Is converted into a uniform distribution between 0 and 1, uniform distribution data generated by scaling and shifting in an arbitrary d-dimensional hyper-sphere is acquired.
As a preferred scheme of the present invention, an optimal classification model is trained for the uniformly distributed data through an NOF + QMS hybrid parameter-finding algorithm, and the specific steps are as follows:
first, according to the initial model Ω 0 Inputting training data x j ∈X 0 Judging whether the KKT condition is met, if yes, utilizing the data augmentation algorithm to augment the training data, and updating the SVDD model again to obtain omega 1 ;
Secondly, the support vector set is judged to be SV 0 Whether or not to equal the set of unsupported vectors SV 1 If the KKT condition is not met, adding data in the training data to judge whether the KKT condition is met again, and if the KKT condition is met, iterating again until omega 0 =Ω 1 Until now.
As a preferred embodiment of the present invention, according to said Ω 0 =Ω 1 The iteration times of the two models predict the deviation of the two models, and the optimal classification model is adjusted by adjusting the increment of training data so as to output a fault detection technology classification framework.
As a preferred scheme of the invention, the fault partition is obtained according to the fault detection technology classification framework, the distribution of one-dimensional distance decision values of training samples in each partition of the power distribution network in the SVDD model is estimated through a KDE algorithm by combining with the power grid topological structure of the power distribution network, the fault confidence of test data in each partition is calculated by using the distribution, fault points are compared and positioned, and the position of the fault source is obtained according to the fault confidence.
As a preferred scheme of the present invention, the KDE algorithm is a non-parametric method for estimating a random variable probability density function, and specifically includes:
will train data x 1 ,x 2 ,…,x n The kernel density estimation is made from some independent and identically distributed samples sampled with unknown probability density f, whose formula is:
wherein the content of the first and second substances,representing a symmetric but not necessarily positive kernel function, satisfying an in-domain integral of 1, h being a smoothing parameter or kernel bandwidth, n representing the amount of training data, x i Represents any one of the training data and,represents the mean of the training data.
Compared with the prior art, the invention has the following beneficial effects:
according to data attributes such as high data dimension, difference between training sample distribution and data real distribution in a power distribution network and the like, a support vector data description SVDD algorithm is adopted to perform offline training on a fault detection model to identify faults of the power distribution network, an NOF + QMS mixed parameter searching algorithm is adopted to establish an SVDD optimal model and shorten a parameter searching range, model identification precision is guaranteed, and training time overhead is reduced, so that the SVDD optimal model can be effectively applied to core bandwidth parameter selection of the SVDD model under power distribution network fault detection, the identification effect of newly added data is improved, an hypersphere data augmentation algorithm is adopted to analyze the performance influence of two parameters of data distribution radius and data distribution quantity on an online model, the online identification model is quickly and effectively established, a core estimation density algorithm KDE is adopted to convert the output of each partition classifier into a uniform probability P value form, the output probability of each partition is compared, the partition where a fault source is located is effectively reduced and positioned is beneficial to gradually reducing the fault range, and accordingly, operators of a power system are helped to find and remove fault nodes in time.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below. It should be apparent that the drawings in the following description are merely exemplary, and that other embodiments can be derived from the drawings provided by those of ordinary skill in the art without inventive effort.
Fig. 1 is a flowchart of a distribution network fault classification and source location method based on deep learning according to an embodiment of the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
As shown in fig. 1, the invention provides a distribution network fault classification and source positioning method based on deep learning, which comprises the following steps:
the method comprises the following steps that S1, a multi-classification SVDD data driving collection model is used for collecting power distribution network fault historical data, optimal parameters are obtained through multi-classification SVDD data driving collection model training data, and the optimal parameters are verified through a simulink simulation model to obtain power distribution network offline data;
in the embodiment, in order to solve the problem caused by the fact that DER with high permeability is merged into a power distribution network, a multi-classification SVDD data drive collection model is used for collecting historical fault data, a relevant model is built, optimal parameters are obtained to serve as a data base of subsequent fault location, and the reliability of a power system is improved.
In this embodiment, the multi-classification SVDD data-driven acquisition model can deal with high-dimensional data and can flexibly describe normal data and deal with data changes caused by a large number of branch topology changes in the power distribution network.
S2, acquiring a DER dynamic change rate of the power distribution network according to the power distribution network offline data, acquiring a large amount of training data through the DER dynamic change rate of the power distribution network in combination with real-time data of the power distribution network, and effectively partitioning a plurality of branches of the power distribution network according to the training data;
a large amount of training data are obtained according to the DER dynamic change rate real-time data of the power distribution network, and a new data mode in the power distribution network can be learned in real time with self-adaption capability aiming at a dynamic environment.
And S3, fault positioning detection is carried out on faults in the power distribution network according to the effective subareas of the power distribution network branches, and the subareas where the fault sources are located are further judged through the overall subarea fault probability measurement indication of the power distribution network, so that the fault sources are positioned.
The fault location detection model is trained by using enough global historical data, and the data has more power measurement values such as voltage, current, active power, reactive power and the like.
In the step S1, the multi-classification SVDD data-driven acquisition model maps data to a high-dimensional space through a kernel function, quickly determines a data boundary by describing a data boundary, and introduces the multi-classification SVDD classifier and a penalty parameter epsilon into any given training data i The multi-classification SVDD classifier seeks the optimal hypersphere with the minimum volume in the hyperspace and judges the training data x i The Euclidean distance from the sphere center a of the optimal hyper-sphere and the radius R of the optimal hyper-sphere to judge whether the training data is in the envelope of the optimal hyper-sphere or not, and the penalty parameter epsilon i The penalty parameter epsilon is used in an optimized objective function of corresponding training data i The expression is as follows:
ε i =‖x i -a‖-R 2
when the penalty parameter ε i And when the result is more than or equal to 0, allowing a part of training data to be misjudged by the multi-classification SVDD classifier.
Setting the space of the training data distribution as the envelope of the hypersphere, and introducing a penalty parameter epsilon for normal data falling outside the hypersphere by adopting the multi-classification SVDD classifier i Some of the training data is allowed to be misjudged, i.e. a mismatch between the training data and its correct label is allowed.
The optimization objective function of the training data comprises the following specific optimization steps:
first, a mapping function is adoptedTraining data set X = { X = 1 ,x 2 ,…,x i ,…,x n From the initial feature space R d Mapping to a higher dimensional space R d′ And d' is greater than or equal to d, the new data set is mapped as:
the training data containing the mapping function is processed by a kernel function k (x, y), and the expression is as follows:
wherein x and y respectively represent training data and mapping values of the power distribution network branch;
and verifying whether the training data x belongs to the class of the training set data by adopting a Gaussian kernel function, wherein the Gaussian kernel function expression is as follows:
wherein x is i Represents the training number of the ith branchAccording to x j Representing training data for the jth branch.
The Gaussian kernel function is adopted to model a training data set through a flexible description method, the introduction of the Gaussian kernel function method solves the problem of non-spherical distribution data set, and a more flexible data description mode is provided.
And performing loop iteration processing on the Gaussian kernel function of the training data, and selecting an optimal parameter in a parameter searching range.
The parameter searching range is a subjective task, and when the set searching range is too large, a large amount of time and calculation amount are consumed; when the set search range is too small, it may result in an optimization in the suboptimal hyper-parameter space, resulting in the sought hyper-parameter being not globally optimal.
And (3) acquiring an optimal parameter searching range by adopting a k nearest neighbor algorithm, firstly sorting the distances from all data to the center of the sphere by utilizing the kNN, rejecting data of a plurality of suspected abnormal values, and selecting the European distance between the kth nearest data and the training data as the optimal parameter searching range, wherein the data contained in the training set is considered to be all normal data, and any data in the training set is not rejected.
Constructing a model for the optimal parameters through the multi-classification SVDD classifier, which comprises the following specific steps:
first, the optimal parameters are set as an initial training data set X 0 And introducing a new data set X 1 The initial training data set X is 0 The trained SVDD classification model is recorded as omega 0 At this time, the set of support vectors is denoted as SV 0 Set of unsupported vectors denoted as SV 1 ;
Training data in the multi-classification SVDD classifier is input in a mode of adding sequences one by one or a mode of a small number of samples, and a model is automatically updated to deal with newly added data under the attribute of a new mode.
Second, the newly added data set x is validated in a round robin condition i ∈X 1 Whether the initial model omega is satisfied or not 0 KKT condition of (1), if satisfied, Ω 0 For the currently trained model, if not, x is set i Incorporation into collectionsIn (2), selecting a satisfied non-support vector set
Finally, willAnd SV 0 Merge into a new training set X 0 And retraining the SVDD classification model omega by using the new training set 1 When Ω is 0 =Ω 1 And then continuously participating in updating of new data vectors and outputting the SVDD classification model omega 1 。
The newly added data set X 1 KTT condition for the original model, then x i Will not alter the previous support vector set SV 0 If the samples in the new training set violate the KTT condition, at least some of the data will become the new support vector SV 0 If the samples in the new training set violate the KTT condition, then the non-support vector data in the previous training setPossibly converted into a new support vector SV 0 。
In the step S2, the newly added data set X is processed 1 Introducing a data augmentation algorithm for generating spherical uniform distribution by taking original data as the sphere center, and adding the new data set X 1 Carrying out data augmentation to improve the model training speed, generating over-solving uniformly distributed data under Gaussian distribution, and acquiring effective data subareas, wherein the data augmentation algorithm specifically comprises the following steps:
first, the initial training data set X is set 0 Randomly introducing data to generate a Gaussian distribution with a mean value of 0 and an equation of 1, namely X-N (0, 1);
secondly, calculating the initial training data set X 0 The squared Euclidean distance R of each sample to the origin 2 This squared distance is γ 2 Distributed with d degrees of freedom, wherein R 2 =‖X 0 ‖ 2 ;
Finally, using γ 2 Cumulative distribution of distributionR is to be 2 The distribution of (2) is converted into a uniform distribution between 0 and 1, and uniform distribution data generated by scaling and shifting in an arbitrary d-dimensional hypersphere is acquired.
The hypersphere data augmentation algorithm increases the training data volume, and can ensure that newly added data are not completely discarded in SVDD (singular value decomposition) updating iteration, thereby effectively contributing key support vector data in a super-dimensional space to construct hypersphere under a new mode, performing data augmentation on the data is equivalent to oversampling on the data, accelerating the updating of a detection model under the new mode, and improving the identification precision in a shorter time.
By applying to the newly added data set X 1 And data augmentation is carried out, the trained model is updated, the concept of new data learning is accelerated, and the performance of the model under the new mode data is improved, so that the dynamic change of the online environment is effectively coped with.
Training the uniformly distributed data to obtain an optimal classification model through an NOF + QMS hybrid parameter searching algorithm, and specifically comprising the following steps of:
firstly, according to the initial model omega 0 Inputting training data x j ∈X 0 Judging whether KKT conditions are met, if yes, utilizing the data augmentation algorithm to augment the training data, and updating the SVDD model again to obtain omega 1 ;
Secondly, the support vector set is judged to be SV 0 Whether or not to equal the set of unsupported vectors SV 1 If the KKT condition is not met, adding data in the training data to judge whether the KKT condition is met again, and if the KKT condition is met, iterating again until omega 0 =Ω 1 Until now.
According to the omega 0 =Ω 1 Predicting the deviation of the two models by the iteration number, and adjusting the optimal classification model by adjusting the increment of the training data to outputFault detection technology classification framework.
The non-support vector set of the optimal classification model and subsequent new data are periodically retrained, so that the parameters of the online model can be adjusted, in addition, the prediction results output by all the models can be compared with each other, and after the deviation is found, the models can be adjusted, so that the performance of a fault detection framework is ensured in time.
According to the augmentation data, the previous training results are effectively utilized to obtain better classification results, and the same training sample set is prevented from being repeatedly trained.
And acquiring fault partitions according to the fault detection technology classification framework, estimating the distribution of one-dimensional distance decision values of training samples in each partition of the power distribution network in the SVDD model by a KDE algorithm in combination with the power grid topological structure of the power distribution network, calculating the fault confidence of test data in each partition by using the distribution, comparing and positioning fault points, and acquiring the position of a fault source according to the fault confidence.
The KDE algorithm is adopted to estimate the distribution of the one-dimensional distance decision values of the training samples in each partition of the power distribution network in the SVDD model, a simple and visual kernel estimation is provided for the partition of the power distribution network, the asymptotic deviation and the mean square error are obviously reduced, the kernel estimation algorithm has better boundary deviation performance, the optimal kernel bandwidth selection method is improved, and the common experience rule based on Gaussian distribution hypothesis is avoided.
The KDE algorithm is a nonparametric method for estimating a random variable probability density function, and specifically comprises the following steps:
will train data x 1 ,x 2 ,…,x n The kernel density estimation is made from some independent and identically distributed samples sampled with unknown probability density f, whose formula is:
wherein, the first and the second end of the pipe are connected with each other,represents aA symmetric but not necessarily positive kernel function satisfying an intra-domain integral of 1, h being a smoothing parameter or kernel bandwidth, n representing the number of training data, x i Represents any one of the training data and,represents the mean of the training data.
The kernel density estimation is used as a fault confidence coefficient, the accuracy of fault detection is improved through a training data decision branch value, when an abnormal value and a normal value have the same probability distribution, the abnormal value is assumed to have extremely low occurrence probability and is positioned at the tail of the sample data decision value distribution, therefore, good estimation of the probability of the abnormal value can be obtained, the estimation quantity must allow the tail of unknown distribution f to be well estimated, and the fault confidence coefficient of the test data can be measured through the kernel density estimation for any test data.
According to data attributes such as high data dimension, difference between training sample distribution and data real distribution in a power distribution network and the like, a support vector data description SVDD algorithm is adopted to perform offline training on a fault detection model to identify faults of the power distribution network, an NOF + QMS mixed parameter searching algorithm is adopted to establish an SVDD optimal model and shorten a parameter searching range, model identification precision is guaranteed, and training time overhead is reduced, so that the SVDD optimal model can be effectively applied to core bandwidth parameter selection of the SVDD model under power distribution network fault detection, the identification effect of newly added data is improved, an hypersphere data augmentation algorithm is adopted to analyze the performance influence of two parameters of data distribution radius and data distribution quantity on an online model, the online identification model is quickly and effectively established, a core estimation density algorithm KDE is adopted to convert the output of each partition classifier into a uniform probability P value form, the output probability of each partition is compared, the partition where a fault source is located is effectively reduced and positioned is beneficial to gradually reducing the fault range, and accordingly, operators of a power system are helped to find and remove fault nodes in time.
The above embodiments are only exemplary embodiments of the present application, and are not intended to limit the present application, and the protection scope of the present application is defined by the claims. Various modifications and equivalents may be made by those skilled in the art within the spirit and scope of the present application and such modifications and equivalents should also be considered to be within the scope of the present application.
Claims (10)
1. A distribution network fault classification and source positioning method based on deep learning is characterized by comprising the following steps:
the method comprises the following steps that S1, a multi-classification SVDD data driving collection model is used for collecting power distribution network fault historical data, optimal parameters are obtained through multi-classification SVDD data driving collection model training data, and the optimal parameters are verified through a simulink simulation model to obtain power distribution network offline data;
s2, acquiring a DER dynamic change rate of the power distribution network according to the power distribution network offline data, acquiring a large amount of training data through the DER dynamic change rate of the power distribution network in combination with real-time data of the power distribution network, and effectively partitioning a plurality of branches of the power distribution network according to the training data;
and S3, fault positioning detection is carried out on faults in the power distribution network according to the effective subareas of the power distribution network branches, and the subareas where the fault sources are located are further judged through the overall subarea fault probability measurement indication of the power distribution network, so that the fault sources are positioned.
2. The distribution network fault classification and source positioning method based on deep learning of claim 1, wherein: in the step S1, the multi-classification SVDD data-driven acquisition model maps data to a high-dimensional space through a kernel function, quickly determines a data boundary by describing a data boundary, and introduces the multi-classification SVDD classifier and a penalty parameter epsilon into any given training data i The multi-classification SVDD classifier seeks the optimal hypersphere with the minimum volume in the hyperspace and judges the training data x i The Euclidean distance from the sphere center a of the optimal hyper-sphere and the radius R of the optimal hyper-sphere to judge whether the training data is in the envelope of the optimal hyper-sphere or not, and the penalty parameter epsilon i In an optimized objective function for corresponding training data, the penaltyParameter epsilon i The expression is as follows:
ε i =‖x i -a‖-R 2
when the penalty parameter ε i And when the result is more than or equal to 0, allowing a part of training data to be misjudged by the multi-classification SVDD classifier.
3. The distribution network fault classification and source positioning method based on deep learning of claim 2, wherein the optimization objective function of the training data specifically comprises the following steps:
first, a mapping function is adoptedTraining data set X = { X = { X 1 ,x 2 ,…,x i ,…,x n From the initial feature space R d Mapping to a higher dimensional space R d′ And d' is greater than or equal to d, the new data set is mapped as:
the training data containing the mapping function is processed by a kernel function k (x, y), and the expression is as follows:
wherein x and y respectively represent training data and mapping values of the branch of the power distribution network;
and verifying whether the training data x belongs to the class of the training set data by adopting a Gaussian kernel function, wherein the Gaussian kernel function expression is as follows:
wherein x is i Training representing the ith branchData, x j Representing training data for the jth branch.
4. The distribution network fault classification and source positioning method based on deep learning of claim 3, wherein the Gaussian kernel function of the training data is subjected to cyclic iteration processing, and optimal parameters are selected in a parameter searching range.
5. The distribution network fault classification and source positioning method based on deep learning of claim 4, wherein a model is constructed for the optimal parameters through the multi-classification SVDD classifier, and the method comprises the following specific steps:
first, the optimal parameters are set as an initial training data set X 0 And introducing a new data set X 1 The initial training data set X 0 The trained SVDD classification model is marked as omega 0 At this time, the set of support vectors is denoted as SV 0 Set of unsupported vectors denoted as SV 1 ;
Second, the new data set x is validated in a loop condition i ∈X 1 Whether the initial model omega is satisfied or not 0 KKT condition of (1), if satisfied, Ω 0 For the currently trained model, if not, x is set i Incorporation into collectionsIn the method, a set of satisfied non-support vectors is selected
6. The distribution network fault classification and source location method based on deep learning of claim 5, wherein in the step S2, the newly added data set X is subjected to 1 Introducing a data augmentation algorithm which generates spherical uniform distribution by taking original data as the sphere center, and adding the new data set X 1 Carrying out data augmentation to improve the model training speed, generating over-solving uniform distribution data under Gaussian distribution, and acquiring effective data subareas, wherein the data augmentation algorithm specifically comprises the following steps:
first, the initial training data set X is set 0 Randomly introducing data to generate Gaussian distribution with the mean value of 0 and the equation of 1, namely X to N (0, 1);
secondly, calculating the initial training data set X 0 The squared Euclidean distance R of each sample to the origin 2 This squared distance is γ 2 Distributed with d degrees of freedom, where R 2 =‖X 0 ‖ 2 ;
7. The distribution network fault classification and source positioning method based on deep learning of claim 6, wherein an optimal classification model is trained for the uniformly distributed data through an NOF + QMS hybrid parameter searching algorithm, and the method comprises the following specific steps:
first, according to the initial model Ω 0 Inputting training data x j ∈X 0 Judging whether the KKT condition is met, if yes, utilizing the data augmentation algorithm to augment the training data, and updating the SVDD model again to obtain omega 1 ;
Secondly, judging that the set of support vectors is SV 0 Whether or not to equal the set of unsupported vectorsHesv 1 If the KKT condition is not met, adding data in the training data to judge whether the KKT condition is met again, and if the KKT condition is met, iterating again until omega 0 =Ω 1 Until now.
8. The distribution network fault classification and source positioning method based on deep learning of claim 7, wherein the distribution network fault classification and source positioning method is based on the omega 0 =Ω 1 The iteration times of the two models predict the deviation of the two models, and the optimal classification model is adjusted by adjusting the increment of training data so as to output a fault detection technology classification framework.
9. The distribution network fault classification and source location method based on deep learning of claim 8, wherein fault partitions are obtained according to the fault detection technology classification framework, distribution of one-dimensional distance decision values of training samples in each partition of the distribution network in an SVDD model is estimated through a KDE algorithm by combining with a power grid topological structure of the distribution network, fault confidence degrees of test data in each partition are calculated by using the distribution, fault points are compared and located, and a fault source location is obtained according to the fault confidence degrees.
10. The distribution network fault classification and source positioning method based on deep learning of claim 9, wherein the KDE algorithm is a non-parametric method for estimating a random variable probability density function, and specifically comprises:
will train data x 1 ,x 2 ,…,x n The kernel density estimation is made from some independent and identically distributed samples sampled with unknown probability density f, whose formula is:
wherein the content of the first and second substances,representing a symmetric but not necessarily positive kernel function, satisfying an in-domain integral of 1, h as a smoothing parameter or kernel bandwidth, n representing the number of training data, x i Represents any one of the training data and,represents the mean of the training data.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202211192161.6A CN115577290A (en) | 2022-09-28 | 2022-09-28 | Distribution network fault classification and source positioning method based on deep learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202211192161.6A CN115577290A (en) | 2022-09-28 | 2022-09-28 | Distribution network fault classification and source positioning method based on deep learning |
Publications (1)
Publication Number | Publication Date |
---|---|
CN115577290A true CN115577290A (en) | 2023-01-06 |
Family
ID=84583184
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202211192161.6A Pending CN115577290A (en) | 2022-09-28 | 2022-09-28 | Distribution network fault classification and source positioning method based on deep learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN115577290A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116070151A (en) * | 2023-03-17 | 2023-05-05 | 国网安徽省电力有限公司超高压分公司 | Ultra-high voltage direct current transmission line fault detection method based on generalized regression neural network |
-
2022
- 2022-09-28 CN CN202211192161.6A patent/CN115577290A/en active Pending
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116070151A (en) * | 2023-03-17 | 2023-05-05 | 国网安徽省电力有限公司超高压分公司 | Ultra-high voltage direct current transmission line fault detection method based on generalized regression neural network |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111783953B (en) | 24-point power load value 7-day prediction method based on optimized LSTM network | |
CN112508442B (en) | Transient stability assessment method and system based on automatic and interpretable machine learning | |
CN111523785A (en) | Power system dynamic security assessment method based on generation countermeasure network | |
CN112200694B (en) | Dominant instability mode identification model construction and application method based on graph neural network | |
CN112069727B (en) | Intelligent transient stability evaluation system and method with high reliability for power system | |
CN116448419A (en) | Zero sample bearing fault diagnosis method based on depth model high-dimensional parameter multi-target efficient optimization | |
CN112800682A (en) | Feedback optimization fan blade fault monitoring method | |
Xingrong | Research on time series data mining algorithm based on Bayesian node incremental decision tree | |
CN113612237A (en) | Method for positioning resonance-induced subsynchronous oscillation source in offshore wind farm | |
CN115688563A (en) | Multi-agent evolution reinforcement learning-based processing parameter optimization method | |
CN113139570A (en) | Dam safety monitoring data completion method based on optimal hybrid valuation | |
CN115577290A (en) | Distribution network fault classification and source positioning method based on deep learning | |
CN115329669A (en) | Power system transient stability evaluation method based on deep Bayes active learning | |
Park et al. | Source term estimation using deep reinforcement learning with Gaussian mixture model feature extraction for mobile sensors | |
CN111797935B (en) | Semi-supervised depth network picture classification method based on group intelligence | |
Li et al. | Online course learning outcome evaluation method based on big data analysis | |
CN116485049B (en) | Electric energy metering error prediction and optimization system based on artificial intelligence | |
CN117151488A (en) | Method, system, storage medium and equipment for expanding cold tide and strong wind weather sample | |
CN111242266A (en) | Operation data management system | |
CN113836818B (en) | Ocean current motion prediction algorithm based on BP neural network prediction model | |
CN113128130B (en) | Real-time monitoring method and device for judging stability of direct-current power distribution system | |
CN112465253B (en) | Method and device for predicting links in urban road network | |
Kuang et al. | An Association Rules-Based Method for Outliers Cleaning of Measurement Data in the Distribution Network | |
CN114169395A (en) | Construction method and application of dominant instability mode recognition model of power system | |
Chen et al. | Distributed hierarchical deep reinforcement learning for large-scale grid emergency control |
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 |