CN110909302A - Method and system for learning local disturbance characteristics of operating state parameters of alternating-current and direct-current power grid - Google Patents

Method and system for learning local disturbance characteristics of operating state parameters of alternating-current and direct-current power grid Download PDF

Info

Publication number
CN110909302A
CN110909302A CN201911064509.1A CN201911064509A CN110909302A CN 110909302 A CN110909302 A CN 110909302A CN 201911064509 A CN201911064509 A CN 201911064509A CN 110909302 A CN110909302 A CN 110909302A
Authority
CN
China
Prior art keywords
parameter
distribution
state
power grid
parameters
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
Application number
CN201911064509.1A
Other languages
Chinese (zh)
Inventor
杨晓楠
郎燕生
张印
李理
罗亚迪
李静
王少芳
王磊
王淼
宋旭日
吴奇
彭献永
林金星
张磊
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
State Grid Corp of China SGCC
China Electric Power Research Institute Co Ltd CEPRI
Electric Power Research Institute of State Grid Shandong Electric Power Co Ltd
Original Assignee
State Grid Corp of China SGCC
China Electric Power Research Institute Co Ltd CEPRI
Electric Power Research Institute of State Grid Shandong Electric Power Co Ltd
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by State Grid Corp of China SGCC, China Electric Power Research Institute Co Ltd CEPRI, Electric Power Research Institute of State Grid Shandong Electric Power Co Ltd filed Critical State Grid Corp of China SGCC
Priority to CN201911064509.1A priority Critical patent/CN110909302A/en
Publication of CN110909302A publication Critical patent/CN110909302A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10Complex mathematical operations
    • G06F17/15Correlation function computation including computation of convolution operations
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10Complex mathematical operations
    • G06F17/18Complex mathematical operations for evaluating statistical data, e.g. average values, frequency distributions, probability functions, regression analysis
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/045Combinations of networks
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06QINFORMATION 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
    • G06Q50/00Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
    • G06Q50/06Energy or water supply

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Mathematical Physics (AREA)
  • Computational Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Analysis (AREA)
  • Software Systems (AREA)
  • Mathematical Optimization (AREA)
  • Pure & Applied Mathematics (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Business, Economics & Management (AREA)
  • Computing Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biophysics (AREA)
  • Molecular Biology (AREA)
  • Evolutionary Computation (AREA)
  • Computational Linguistics (AREA)
  • Economics (AREA)
  • Databases & Information Systems (AREA)
  • Algebra (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • Public Health (AREA)
  • Primary Health Care (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Tourism & Hospitality (AREA)
  • Evolutionary Biology (AREA)
  • Strategic Management (AREA)
  • General Business, Economics & Management (AREA)
  • Operations Research (AREA)
  • Probability & Statistics with Applications (AREA)
  • Marketing (AREA)
  • Human Resources & Organizations (AREA)
  • Water Supply & Treatment (AREA)
  • Supply And Distribution Of Alternating Current (AREA)

Abstract

The invention discloses a method and a system for learning local disturbance characteristics of operating state parameters of an alternating current and direct current power grid, wherein the method comprises the following steps: carrying out noise reduction processing on the acquired historical operating parameters; generating a deep learning operation parameter feature vector for the operation parameters subjected to the noise reduction treatment by utilizing smooth pseudo-affine Wigner-Weiler distribution; learning the hidden variables of the operating parameter feature vector by using a Student-t distributed hybrid model to generate a state label corresponding to each operating parameter feature; and identifying the running state of the AC/DC power grid by adopting a deep shrinkage self-coding network-Gaussian process based on the state label corresponding to each running parameter characteristic. According to the invention, the corresponding state labels are generated for different operation parameter characteristics corresponding to the operation parameters, and the operation state of the AC/DC power grid is identified by adopting a deep shrinkage self-coding network-Gaussian process based on the state labels corresponding to the operation parameter characteristics, so that the identification accuracy is improved.

Description

Method and system for learning local disturbance characteristics of operating state parameters of alternating-current and direct-current power grid
Technical Field
The invention relates to an alternating current-direct current power transmission network, in particular to a method and a system for learning local disturbance characteristics of running state parameters of an alternating current-direct current power network.
Background
Due to the fact that the alternating current and direct current hybrid power grid has the effect of an alternating current system on a direct current system, the influence of the direct current system on the alternating current system and the mutual influence among the direct current systems, huge impact is caused to the power grid. The traditional fault diagnosis mainly focuses on signal processing and physical model diagnosis, and how to perform transient time-frequency analysis on the operating parameters of the alternating current and direct current power grid is important for identifying fault reasons. The time-frequency characteristics of the alternating current and direct current power grid system are extracted manually, the limitation of the expert is limited, and the extracted information is incomplete because a lot of information of the alternating current and direct current power grid is not completely contained in the extracted characteristics. Therefore, the fault characteristics of the analog circuit are unsupervised and learned by using the noise reduction self-encoder in the morning and the like, and a good identification rate is obtained; shixin utilizes an automatic encoder to identify the fault of the transformer, and the identification rate is up to 90.96%. Aiming at the running state parameter state signals of the alternating current and direct current power grid, redundant information needs to be removed to increase the calculation efficiency, and the anti-interference capability is also needed to ensure high-efficiency and high-accuracy fault identification. However, the self-coding network disclosed in the above document only removes the signal redundancy to improve the calculation efficiency, but the classification between normal, weak and extreme fault signals with relatively small differences is not significant, but different protection measures can be adopted according to the normal, weak and extreme fault signals, which has an important influence on the selection of the fault processing mode.
Disclosure of Invention
In order to solve the above-mentioned deficiencies in the prior art, the present invention provides a learning method of local disturbance characteristics of an ac/dc power grid operating state parameter, comprising:
carrying out noise reduction processing on the acquired historical operating parameters;
generating a deep learning operation parameter feature vector for the operation parameters subjected to the noise reduction treatment by utilizing smooth pseudo-affine Wigner-Weiler distribution;
learning the hidden variables of the operating parameter feature vector by using a Student-t distributed hybrid model to generate a state label corresponding to each operating parameter feature;
and identifying the running state of the AC/DC power grid by adopting a deep shrinkage self-coding network-Gaussian process based on the state label corresponding to each running parameter characteristic.
Preferably, the performing noise reduction processing on the acquired historical operating parameters includes:
decomposing historical operating parameters of the alternating current and direct current power grid through DB10 wavelets to obtain corresponding wavelet coefficient amplitudes;
and removing the variation trend of the wavelet coefficient maximum value from the wavelet coefficient amplitude value into reduced operation parameters to obtain the operation parameters after the noise reduction treatment.
Preferably, the generating of the deep-learning operation parameter feature vector for the operation parameters after the noise reduction processing by using the smooth pseudo-affine wigner-wirler distribution includes:
performing time-frequency distribution transformation on the operation parameters subjected to noise reduction processing to obtain a Wigner distribution function;
adding an analysis signal into the Wigner distribution function to obtain a Wigner-Wirler distribution function;
adding a Kaiser window function to the Wigner-Wirler distribution function to obtain smooth pseudo-Wigner-Wirler distribution;
and performing energy spectrum density extraction based on the time-frequency distribution of the smooth pseudo-Wigner-Viller distribution to generate a running parameter feature vector for deep learning.
Preferably, the time-frequency distribution transformation is performed on the operating parameters after the noise reduction processing according to the following formula:
Figure BDA0002257191090000021
in the formula: s is an operation parameter; s*(u) is the complex conjugate of s (u);
Figure BDA0002257191090000022
is a kernel function; u is a time variable; theta is
Figure BDA0002257191090000023
Fixed parameters of the function; τ is the residence time; omega is frequency; t is time.
Preferably, the Kaiser window function is given by:
Figure BDA0002257191090000024
in the formula: g (t, w) is a window function; τ is the residence time of the Kaiser window; a is a non-negative real number that determines the shape of the window; i is0The Bessel function is modified for the zeroth order of the first class.
Preferably, the smooth pseudo-wigner-veller distribution is represented by the following formula:
Figure BDA0002257191090000025
in the formula: w′(l,m)A smooth pseudo-wigner-veller distribution; Δ ω is the frequency differential; Δ t is the time differential, l is the time; j is the time step; m is the frequency; k is the frequency length; p is a time parameter of the instantaneous frequency, and q is an angular frequency parameter of the instantaneous frequency; p-l is the time t of the window function G; q-m is the frequency omega of the window function G.
Preferably, the learning of the hidden variables of the operating parameter feature vector by using the Student-t distributed hybrid model to generate the state labels corresponding to the operating parameter features includes:
learning the hidden variables of the operating parameter feature vector according to the following formula to generate a state label corresponding to each operating parameter feature:
Figure BDA0002257191090000031
in the formula: p (Y | theta)s) Is the joint probability density of the hybrid model of the Student-t distribution; lambda [ alpha ]icIs the distribution of different students
Figure BDA0002257191090000032
The composition ratio of (A); dir (lambda)ic| η) is the parameter λicA priori of (a); rcIs the precision of the distribution of each group t, d is the dimension of the operation parameter data, v is the degree of freedom parameter, η is the Dirichlet distribution parameter, c is the composition variable of the student distribution, y is the degree of freedom parameteriEach group of operating parameter characteristic data;
Figure BDA0002257191090000033
is the mean of the distribution of each group t; v. ofvIs an adjustable parameter for controlling the shape of the t-distribution; n is the operating parameter characteristic data dimension; c is the number of components distributed by the student; v. oficA degree of freedom parameter of group c which is a characteristic of group i; ricIs the accuracy parameter of the c-th group of the i-th group of operating parameter characteristics.
Preferably, the identifying the operating state of the ac/dc power grid based on the state label corresponding to each operating parameter feature by using a deep shrinkage self-coding network-gaussian process includes:
inputting each operation parameter characteristic with a state label into a depth contraction self-coding network-Gaussian process model, and learning abstract characteristics of an alternating current/direct current operation state;
and taking the abstract features as the input of a Gaussian process of a top-level classifier to identify the running state of alternating current and direct current.
Preferably, the operating parameters include: voltage, frequency and power.
Preferably, after the state tag corresponding to each operating parameter feature identifies the operating state of the ac/dc power grid by using a deep shrinkage self-coding network-gaussian process, the method further includes:
constructing an AC/DC power grid operation state identification model based on each operation parameter and the state label corresponding to each operation parameter characteristic;
and acquiring the running parameters of the AC/DC power grid in real time, and acquiring the running state of the current AC/DC power grid based on the AC/DC power grid running state identification model.
Based on the same invention concept, the invention provides an alternating current and direct current power grid operation state parameter local disturbance feature learning system, which comprises:
the noise reduction module is used for carrying out noise reduction processing on the acquired historical operating parameters;
the extraction module is used for generating a deep learning operation parameter feature vector for the operation parameters subjected to the noise reduction processing by utilizing smooth pseudo-affine Wigner-Weiler distribution;
the learning module is used for learning the hidden variables of the operating parameter feature vector by using a Student-t distributed hybrid model to generate state labels corresponding to the operating parameter features;
and the identification module is used for identifying the running state of the alternating current and direct current power grid by adopting a deep shrinkage self-coding network-Gaussian process based on the state label corresponding to each running parameter characteristic.
Preferably, the noise reduction module includes:
the decomposition unit is used for decomposing the historical operating parameters of the alternating current and direct current power grid through DB10 wavelet to obtain corresponding wavelet coefficient amplitude values;
and the acquisition unit is used for removing the variation trend of the wavelet coefficient maximum value from the wavelet coefficient amplitude value into reduced running parameters and acquiring the running parameters after the noise reduction processing.
Compared with the prior art, the invention has the beneficial effects that:
according to the technical scheme provided by the invention, the acquired historical operating parameters are subjected to noise reduction treatment; generating a deep learning operation parameter feature vector for the operation parameters subjected to the noise reduction treatment by utilizing smooth pseudo-affine Wigner-Weiler distribution; learning the hidden variables of the operating parameter feature vector by using a Student-t distributed hybrid model to generate a state label corresponding to each operating parameter feature; the running state of the alternating current and direct current power grid is identified by adopting a deep shrinkage self-coding network-Gaussian process based on the state label corresponding to each running parameter characteristic.
The technical scheme provided by the invention provides a novel time-frequency characteristic learning method based on smooth pseudo-affine Virgener-Ville distribution, which can provide three-dimensional dynamic change spectrum information of time-frequency-amplitude.
According to the technical scheme provided by the invention, the dynamic spectrum information is processed by utilizing the student mixed distribution model, so that the labels of normal, micro-normal and fault are attached to the characteristic data of the running state of the alternating current and direct current network.
According to the DC AEN-GP-based alternating current and direct current state identification model provided by the technical scheme, potential characteristics of an alternating current and direct current power grid in different states are automatically learned without supervision, abstract characteristics representing running postures of the alternating current and direct current power grid are extracted, and experimental results show that the model can identify running states of the power grid in different levels.
Compared with the experimental result of the traditional method, the alternating current and direct current state identification model based on DCAEN-GP provided by the technical scheme provided by the invention extracts the abstract characteristics of the alternating current and direct current running state to the maximum extent, and the accuracy of the experimental detection is up to 98.21%.
Drawings
FIG. 1 is a schematic diagram of four features obtained by WVD in an embodiment of the present invention;
FIG. 2 is a schematic diagram of four features obtained from a PWD in an embodiment of the present invention;
FIG. 3 is a schematic diagram of four features obtained by SPWVD in an embodiment of the present invention;
FIG. 4 is a schematic diagram of four features obtained by SPAWVD in an embodiment of the present invention;
FIG. 5 is a schematic diagram of the AC/DC fault state identification result based on the SPAWV spectral feature and the deep self-coding network-Gaussian process in the embodiment of the present invention;
FIG. 6 is a schematic diagram of the AC/DC fault state identification result based on the SPAWV spectral characteristics and the deep noise reduction self-coding network-Gaussian process in the embodiment of the present invention;
FIG. 7 is a schematic diagram of the AC/DC fault state identification result based on the SPAWV spectral feature and the deep sparse self-coding network-Gaussian process in the embodiment of the present invention;
FIG. 8 is a schematic diagram of the AC/DC fault state identification result based on the SPAWV spectral characteristics and the depth shrinkage self-coding network-Gaussian process in the embodiment of the present invention;
fig. 9 is a flowchart of a method for learning local disturbance characteristics of an ac/dc power grid operating state parameter according to the present invention.
Detailed Description
For a better understanding of the present invention, reference is made to the following description taken in conjunction with the accompanying drawings and examples.
Example 1
Liuxing et al in 2016 published a document entitled "high-power rectifying device fault online diagnosis method using current fault characteristics", which only determined the high-power rectifying device fault through the time domain variation of the current state variable and failed to learn the multimode multi-domain characteristics of the current state variable; the Zhang Jing 2013 published a document named 'research on a fault diagnosis method of a power distribution network by using multi-source information', and the document mentions that fault diagnosis at a characteristic level or a data level is selected based on manual experience, so that time and labor are consumed; liu Biao et al published a document entitled "transformer fault diagnosis based on deep belief network" in 2018, used a conventional deep belief network, failed to improve the theory of the deep belief network, and resulted in low accuracy.
Aiming at the defects in the prior art, the invention provides a method for learning the local disturbance characteristics of the running state parameters of the alternating current and direct current power grid, which solves the following technical problems:
(1) the characteristic data is too simple to avoid noise interference.
(2) A better dimensionality reduction algorithm is needed so that the data dimensionality can be reduced as much as possible under the condition of ensuring that the original signal energy is sufficient.
(3) A rigorous classification algorithm is required to make it adapt to the newer, not completely parameter dependent. And the classification algorithm should be able to handle non-linear data.
As shown in fig. 9, the technical solution of the present invention includes the following steps:
step S1: carrying out noise reduction processing on the acquired historical operating parameters;
step S2: generating a deep learning operation parameter feature vector for the operation parameters subjected to the noise reduction treatment by utilizing smooth pseudo-affine Wigner-Weiler distribution;
step S3: learning the hidden variables of the operating parameter feature vector by using a Student-t distributed hybrid model to generate a state label corresponding to each operating parameter feature;
step S4: and identifying the running state of the AC/DC power grid by adopting a deep shrinkage self-coding network-Gaussian process based on the state label corresponding to each running parameter characteristic.
Step S1, carrying out noise reduction processing on the typical running state parameters of the AC/DC power grid, such as voltage, frequency and power, by utilizing wavelet packet conversion;
the method specifically comprises the following steps: the original signals of the operation parameters of the AC/DC power grid are decomposed through DB10 wavelets to obtain corresponding wavelet coefficient amplitudes, and the amplitudes can reflect the amount of signal energy carried by the signals through the amplitudes. During the decomposition of the wavelet packet, the maximum value of the wavelet coefficient of the subspace where the target signal is located slightly increases with the increase of the decomposition level, and the maximum value of the wavelet coefficient of the subspace containing noise decreases. According to the characteristic, selection is carried out according to the variation trend of the wavelet coefficient, and the noise elimination of the original alternating current and direct current parameter signal can be completed.
Step S2, namely, carrying out transient time-frequency analysis on the typical parameter signals by utilizing smooth pseudo-affine Wigner-Ville distribution (transformation is carried out to obtain the transient spectrum characteristics of the typical parameters and form the running parameter characteristic vector of deep learning;
the method specifically comprises the following steps: in order to obtain the instantaneous frequency spectrum information of the running parameter signals of the alternating current and direct current power grid, an improved Wigner-Ville Distribution (WVD) based on a Kaiser window is provided. The Wigner Distribution Function (WDF) is an important method for analyzing non-stationary signals. The spectral information may be negative due to cross term interference. The interference of cross terms can be effectively inhibited by adding a sliding index window to carry out affine smoothing on the Wigner-Viller distribution, and the improved method is called smooth Pseudo-affine Wigner-Viller distribution (SPAWVD).
The time-frequency distribution of the signal s (u) can be expressed as:
Figure BDA0002257191090000071
wherein s is*(u) is the complex conjugate of s (u),
Figure BDA0002257191090000072
is a kernel function.
Different distributions may be generated depending on the choice of kernel function. When kernel function
Figure BDA0002257191090000073
Then, the wigner distribution can be obtained:
Figure BDA0002257191090000074
the energy spectral density function p (ω) can be expressed as:
Figure BDA0002257191090000075
wherein R ist(τ) is a time-varying autocorrelation function, whereby a time-varying energy spectral density function can be obtained.
For one continuous WDF:
w(t,ω)=∫Rt(τ)e-jwτdτ (4)
wherein R ist(τ) can be expressed in a symmetric form as:
Figure BDA0002257191090000076
additionally, the WDF of discrete time may be expressed as:
Figure BDA0002257191090000077
when sampling signal s (t), which results in aliasing of the signal in the WDF, an effective way to avoid aliasing is to use an analysis signal before calculating the WDF, also known as the wigner-verler distribution, which can be expressed as:
Figure BDA0002257191090000078
wherein H { s }r(n) is a Hilbert transform, generated by convolution of a 90-degree phase-shifted impulse response h (t).
In order to avoid negative values caused by the influence of the interference term, a Kaiser window function G (t, w) is added to WVD:
Figure BDA0002257191090000081
where τ is the dwell time of the Kaiser window and a is a non-negative real number that determines the shape of the window. I is0Is a first class of zero-order modified Bessel functions.
Accordingly, WVD can be expressed as:
Figure BDA0002257191090000082
selecting proper w and t to obtain a sampled Kaiser window function:
Figure BDA0002257191090000083
wherein the values of p and q vary between + -2 j and + -2 k, respectively. From the convolution of the sampled WDF and the Kaiser window function, a smooth Pseudo-Wigner-Ville Distribution (SPWVD) can be obtained:
Figure BDA0002257191090000084
then, the affine SPWVD can be expressed as:
Figure BDA0002257191090000085
where Ψ (t, m) is a smoothing function.
The invention alternating current and direct current power grid operation parameter s is subjected to time-frequency distribution transformation of formula (1) to obtain a time-frequency-amplitude three-dimensional map of the power grid parameter s through formula (12), and further subjected to SPAWVD transformation to obtain a spectral density ps(ω). The invention uses the transient frequency spectrum of SPAWVD to replace the frequency spectrum of FFT, because FFT can only provide average frequency spectrum information in a given time period, and SPAWVD can provide the time frequency amplitude information of AC/DC power grid running state signals in real time. The operation parameter feature vector in the invention is a frequency domain feature vector of the operation parameter, the forms of the state labels are 1, 2 and 3, which respectively correspond to normal, micro-normal and fault, and the state label is the quantitative display representation of the operation state.
Further, in step S3, the hybrid student distribution model is used to learn the hidden variables of the initial frequency operating state, so as to realize the label of the characteristics of the ac/dc operating state (normal \ micro-normal \ fault), that is, the student hybrid distribution model is used to perform attribute calibration on the typical parameter characteristics of the ac/dc power grid, and a student hybrid distribution model is provided in order to calibrate the instantaneous frequency spectrum information of the ac/dc power grid operating parameter signal. The student mixed model is a mixed model which is composed of a plurality of student t distributions, is similar to a Gaussian mixed model, and one student t distribution is infinite approximation of the Gaussian distribution, so that the student mixed distribution model has stronger hidden variable learning capacity.
The student mixed model is as follows:
Figure BDA0002257191090000091
wherein P (Y | theta)s) Is the joint probability density of the student's distributed mixture model. Lambda [ alpha ]icIs the distribution of different students
Figure BDA0002257191090000092
The composition ratio of (A); dir (lambda)ic| η) is the parameter λicA priori of (a); rcIs the precision of the distribution of each set t, d is the dimension of the operating parameter data, v is the degree of freedom parameter, η isDirichlet distribution parameters; c is a compositional variable of the student distribution; y isiEach set of operating parameter signature data corresponds to T1, T2, T3 and T4;
Figure BDA0002257191090000093
is the mean of the distribution of each group t; v. ofvIs an adjustable parameter for controlling the shape of the t-distribution; n is the operating parameter characteristic data dimension; c is the number of components distributed by the student; v. oficA degree of freedom parameter of group c which is a characteristic of group i; ricIs the accuracy parameter of the c-th group of the i-th group of operating parameter characteristics.
Preferably, in step S4, the modeling method for the local disturbance feature learning problem of the operation state parameters of the ac/dc power grid is characterized in that the preliminary features with the labels are input into the DCAEN-GP model, and abstract features of the operation state of the ac/dc power grid are learned, and the abstract features are used as input of a gaussian process of a top-level classifier, so as to realize the identification of the operation state of the ac/dc power grid.
The AC/DC power grid operation state fault identification is divided into three categories, namely normal, micro-normal and fault, the Gaussian process classification is a two-classification method, the method is a generation model, a support vector machine is a discrimination model, and the Gaussian process has the advantage that the probability explanation of a solution can be provided. Compared with a Softmax multi-classifier, the Gaussian process can provide the nuclear space learning performance of sample learning, and Softmax only carries out multi-classification on input samples, so the learning capability of the Gaussian process is stronger than that of Softmax, and the two-classification method needs to be popularized to three-classification for realizing multi-state identification of an alternating current-direct current power grid.
The specific three classification processes are as follows:
(1) the training process of the first gaussian process classifier is as follows: the first gaussian process classifier mainly recognizes normal and abnormal states. We first label the abnormal state (micro normal and abnormal) data points as-1 and the normal state data points as + 1. Training the Gaussian process classification by using the two types of training data to obtain a likelihood function of a covariance hyperparameter hyp1 related to the first Gaussian process classification:
Figure BDA0002257191090000101
(2) data points for abnormal states (micro-normal and fault) are selected as training samples for the second gaussian process classifier. The data points for the second type of micro normal state are labeled-1 and the data points for the third type of abnormal state are labeled + 1. And training the Gaussian process classification by using the two types of training data to obtain a covariance hyperparameter hyp1 of the second Gaussian process classification.
(3) And classifying the working states of the three alternating current and direct current power grids through two Gaussian process classifications.
The invention performs a spectral density p of the time-frequency distribution from the Wigner-Viller distribution (WVD)sAnd (omega) extracting to form a typical AC/DC power grid operation parameter feature vector, learning the multi-parameter fusion feature of AC/DC power grid operation by using Treelet transformation, and acquiring four main features of T1, T2, T3 and T4.
This example illustrates the spectral density extraction of four methods, respectively, and as shown in FIG. 1, the spectral density p is performed for the time-frequency distribution from the Wigner-Viller distribution (WVD)sAnd (omega) extracting to form a typical AC/DC power grid operation parameter feature vector, learning the multi-parameter fusion feature of the AC/DC power grid operation by using Treelet transformation to obtain four main features of T1, T2, T3 and T4, and finding that the features obtained by WVD are almost fused together.
As shown in FIG. 2, the spectral density p is performed for the time-frequency Distribution from a Pseudo Wigner Ville Distribution (PWVD)sAnd (omega) extracting to form a typical AC/DC power grid operation parameter feature vector, learning the multi-parameter fusion feature of the AC/DC power grid operation by using Treelet transformation to obtain four main features T1, T2, T3 and T4, and finding that the four features obtained by PWD are obviously divided into two categories.
As shown in FIG. 3, the spectral density p is performed for the time-frequency information from a smooth Pseudo-Wigner-VilleDisposition (SPWVD)s(omega) extracting to form a typical AC/DC power grid operation parameter characteristic vector, and using Treelet transformation to carry out AC/DC conversionThe characteristics of multi-parameter fusion during the operation of the flow power grid are learned, four main characteristics T1, T2, T3 and T4 are obtained, and the four characteristics obtained by SPWVD can be found to be mixed together.
As shown in FIG. 4, spectral density p is performed for the time-frequency information from a smooth Pseudo-affine Weiner-Weiler distribution (SPAWVD)sAnd (omega) extracting to form a typical AC/DC power grid operation parameter feature vector, learning the multi-parameter fusion feature of the AC/DC power grid operation by using Treelet transformation, and acquiring four main features of T1, T2, T3 and T4.
In the embodiment, the abstract characteristics of the AC/DC state are extracted by comparing different modes, and then the Gaussian process model is used for identification, and the result shows that the accuracy of the depth shrinkage self-coding network-Gaussian process provided by the invention is highest.
As shown in fig. 5, the Deep Auto-Encoder Network-gaussian process (DAE-GP) extracts abstract features of ac and dc states, and then performs recognition using a gaussian process model. The method adopts 5-fold cross validation, and the identification accuracy of the model is as follows: DAE-GP: 75.6 percent.
As shown in fig. 6, the Deep noise reduction self-encoding network-Gaussian Process (Deep noise Auto-encoder network-Gaussian Process, DDAE-GP) extracts the abstract features of the ac/dc state, and then identifies by using a Gaussian Process model, and adopts 5-fold cross validation, where the identification accuracy of the DDAE-GP model is: 89.29 percent.
As shown in fig. 7, the Deep Sparse Auto-encodings network-Gaussian Process (DSAE-GP) extracts abstract features of the ac/dc state, and then identifies the abstract features by using a Gaussian Process model. The method adopts 5-fold cross validation, and the identification accuracy of the DSAE-GP model is as follows: 94.05 percent.
As shown in fig. 8, a Deep contracting self-encoding network-Gaussian Process (Deep continuous Auto-encoder network-Gaussian Process, DCAEN-GP) extracts abstract features of an ac/dc state, and then performs recognition using a Gaussian Process model. The method adopts 5-fold cross validation, and the identification accuracy of the DCAEN-GP model is as follows: 98.21 percent.
Overall DCAEN-GP has the highest recognition accuracy. Because the contraction automatic encoder adds the punishment of the Jacobian item, has stronger disturbance capture capability on the periphery of the sample and realizes more fine potential characteristic learning capability of the sample, the performance of the AC/DC operation fault state identification method provided by the invention is better than that of the traditional DSAE, DDAE and DAE methods.
Example 2
Based on the same invention concept, the invention also provides an alternating current and direct current power grid operation state parameter local disturbance feature learning system, which comprises:
the noise reduction module is used for carrying out noise reduction processing on the acquired historical operating parameters;
the extraction module is used for generating a deep learning operation parameter feature vector for the operation parameters subjected to the noise reduction processing by utilizing smooth pseudo-affine Wigner-Weiler distribution;
the learning module is used for learning the hidden variables of the operating parameter feature vector by using a Student-t distributed hybrid model to generate state labels corresponding to the operating parameter features;
and the identification module is used for identifying the running state of the alternating current and direct current power grid by adopting a deep shrinkage self-coding network-Gaussian process based on the state label corresponding to each running parameter characteristic.
In an embodiment, the noise reduction module includes:
the decomposition unit is used for decomposing the historical operating parameters of the alternating current and direct current power grid through DB10 wavelet to obtain corresponding wavelet coefficient amplitude values;
and the acquisition unit is used for removing the variation trend of the wavelet coefficient maximum value from the wavelet coefficient amplitude value into reduced running parameters and acquiring the running parameters after the noise reduction processing.
As will be appreciated by one skilled in the art, embodiments of the present application may be provided as a method, system, or computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.
The present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the application. It will be understood that each flow and/or block of the flow diagrams and/or block diagrams, and combinations of flows and/or blocks in the flow diagrams and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
The present invention is not limited to the above embodiments, and any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention are included in the scope of the claims of the present invention which are filed as the application.

Claims (12)

1. The utility model provides an alternating current-direct current network operation state parameter local disturbance feature study which characterized in that includes:
carrying out noise reduction processing on the acquired historical operating parameters;
generating a deep learning operation parameter feature vector for the operation parameters subjected to the noise reduction treatment by utilizing smooth pseudo-affine Wigner-Weiler distribution;
learning the hidden variables of the operating parameter feature vector by using a Student-t distributed hybrid model to generate a state label corresponding to each operating parameter feature;
and identifying the running state of the AC/DC power grid by adopting a deep shrinkage self-coding network-Gaussian process based on the state label corresponding to each running parameter characteristic.
2. The method of claim 1, wherein the denoising the obtained historical operating parameters comprises:
decomposing historical operating parameters of the alternating current and direct current power grid through DB10 wavelets to obtain corresponding wavelet coefficient amplitudes;
and removing the variation trend of the wavelet coefficient maximum value from the wavelet coefficient amplitude value into reduced operation parameters to obtain the operation parameters after the noise reduction treatment.
3. The method of claim 1, wherein the generating a deep-learning operating parameter feature vector for the noise-reduced operating parameters using a smooth pseudo-affine wiener-wirler distribution comprises:
performing time-frequency distribution transformation on the operation parameters subjected to noise reduction processing to obtain a Wigner distribution function;
adding an analysis signal into the Wigner distribution function to obtain a Wigner-Wirler distribution function;
adding a Kaiser window function to the Wigner-Wirler distribution function to obtain smooth pseudo-Wigner-Wirler distribution;
and performing energy spectrum density extraction based on the time-frequency distribution of the smooth pseudo-Wigner-Viller distribution to generate a running parameter feature vector for deep learning.
4. The method of claim 3, wherein the noise-reduced operating parameters are transformed in a time-frequency distribution according to the following equation:
Figure FDA0002257191080000011
in the formula: s is an operation parameter; s*(u) is the complex conjugate of s (u);
Figure FDA0002257191080000012
is a kernel function; u is a time variable; theta is
Figure FDA0002257191080000013
Fixed parameters of the function; τ is the residence time; omega is frequency; t is time.
5. The method of claim 3, wherein said Kaiser window function is given by:
Figure FDA0002257191080000021
in the formula: g (t, w) is a window function; τ is the residence time of the Kaiser window; a is a non-negative real number that determines the shape of the window; i is0The Bessel function is modified for the zeroth order of the first class.
6. The method of claim 3, wherein the smooth pseudo-Virger-Viller distribution is represented by the following equation:
Figure FDA0002257191080000022
in the formula: w'(l,m)A smooth pseudo-wigner-veller distribution; Δ ω is the frequency differential; Δ t is the time differential, l is the time; j is the time step; m is the frequency; k is the frequencyLength; p is a time parameter of the instantaneous frequency, and q is an angular frequency parameter of the instantaneous frequency; p-l is the time t of the window function G; q-m is the frequency omega of the window function G.
7. The method of claim 1, wherein learning the hidden variables of the operating parameter feature vector by using a Student-t distributed hybrid model to generate the state labels corresponding to the operating parameter features comprises:
learning the hidden variables of the operating parameter feature vector according to the following formula to generate a state label corresponding to each operating parameter feature:
Figure FDA0002257191080000023
in the formula: p (Y | theta)s) Is the joint probability density of the hybrid model of the Student-t distribution; lambda [ alpha ]icIs the distribution of different students
Figure FDA0002257191080000024
The composition ratio of (A); dir (lambda)ic| η) is the parameter λicA priori of (a); rcIs the precision of the distribution of each group t, d is the dimension of the operation parameter data, v is the degree of freedom parameter, η is the Dirichlet distribution parameter, c is the composition variable of the student distribution, y is the degree of freedom parameteriEach group of operating parameter characteristic data;
Figure FDA0002257191080000025
is the mean of the distribution of each group t; v. ofvIs an adjustable parameter for controlling the shape of the t-distribution; n is the operating parameter characteristic data dimension; c is the number of components of the student distribution; v. oficA degree of freedom parameter of group c which is a characteristic of group i; ricIs the accuracy parameter of the c-th group of the i-th group of operating parameter characteristics.
8. The method of claim 1, wherein identifying the operating state of the ac/dc power grid based on the state labels corresponding to the operating parameter features by a deep-shrinkage self-coding network-gaussian process comprises:
inputting each operation parameter characteristic with a state label into a depth contraction self-coding network-Gaussian process model, and learning abstract characteristics of an alternating current/direct current operation state;
and taking the abstract features as the input of a Gaussian process of a top-level classifier to identify the running state of alternating current and direct current.
9. The method of claim 1, wherein the operating parameters comprise: voltage, frequency and power.
10. The method of claim 1, wherein after identifying the operating state of the ac/dc power grid based on the state labels corresponding to the operating parameter features by using a deep-shrinkage self-coding network-gaussian process, the method further comprises:
constructing an AC/DC power grid operation state identification model based on each operation parameter and the state label corresponding to each operation parameter characteristic;
and acquiring the running parameters of the AC/DC power grid in real time, and acquiring the running state of the current AC/DC power grid based on the AC/DC power grid running state identification model.
11. The utility model provides an alternating current-direct current network operation state parameter local disturbance characteristic learning system which characterized in that includes:
the noise reduction module is used for carrying out noise reduction processing on the acquired historical operating parameters;
the extraction module is used for generating a deep learning operation parameter feature vector for the operation parameters subjected to the noise reduction processing by utilizing smooth pseudo-affine Wigner-Weiler distribution;
the learning module is used for learning the hidden variables of the operating parameter feature vector by using a Student-t distributed hybrid model to generate state labels corresponding to the operating parameter features;
and the identification module is used for identifying the running state of the alternating current and direct current power grid by adopting a deep shrinkage self-coding network-Gaussian process based on the state label corresponding to each running parameter characteristic.
12. The system of claim 11, wherein the noise reduction module comprises:
the decomposition unit is used for decomposing the historical operating parameters of the alternating current and direct current power grid through DB10 wavelet to obtain corresponding wavelet coefficient amplitude values;
and the acquisition unit is used for removing the variation trend of the wavelet coefficient maximum value from the wavelet coefficient amplitude value into reduced running parameters and acquiring the running parameters after the noise reduction processing.
CN201911064509.1A 2019-11-01 2019-11-01 Method and system for learning local disturbance characteristics of operating state parameters of alternating-current and direct-current power grid Pending CN110909302A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201911064509.1A CN110909302A (en) 2019-11-01 2019-11-01 Method and system for learning local disturbance characteristics of operating state parameters of alternating-current and direct-current power grid

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201911064509.1A CN110909302A (en) 2019-11-01 2019-11-01 Method and system for learning local disturbance characteristics of operating state parameters of alternating-current and direct-current power grid

Publications (1)

Publication Number Publication Date
CN110909302A true CN110909302A (en) 2020-03-24

Family

ID=69816388

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201911064509.1A Pending CN110909302A (en) 2019-11-01 2019-11-01 Method and system for learning local disturbance characteristics of operating state parameters of alternating-current and direct-current power grid

Country Status (1)

Country Link
CN (1) CN110909302A (en)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111966455A (en) * 2020-08-05 2020-11-20 中国建设银行股份有限公司 Method, device, equipment and medium for generating operation and maintenance component of stateful application instance
CN112348158A (en) * 2020-11-04 2021-02-09 重庆大学 Industrial equipment state evaluation method based on multi-parameter deep distribution learning
CN113688537A (en) * 2021-09-09 2021-11-23 云南电网有限责任公司迪庆供电局 Power distribution network relay protection parameter extraction method based on ENTSILDIg software
CN117853827A (en) * 2024-03-07 2024-04-09 安徽省大气探测技术保障中心 Sampling pump working state operation monitoring system and method for atmospheric chamber gas monitoring

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111966455A (en) * 2020-08-05 2020-11-20 中国建设银行股份有限公司 Method, device, equipment and medium for generating operation and maintenance component of stateful application instance
CN112348158A (en) * 2020-11-04 2021-02-09 重庆大学 Industrial equipment state evaluation method based on multi-parameter deep distribution learning
CN112348158B (en) * 2020-11-04 2024-02-13 重庆大学 Industrial equipment state evaluation method based on multi-parameter deep distribution learning
CN113688537A (en) * 2021-09-09 2021-11-23 云南电网有限责任公司迪庆供电局 Power distribution network relay protection parameter extraction method based on ENTSILDIg software
CN113688537B (en) * 2021-09-09 2023-05-23 云南电网有限责任公司迪庆供电局 Distribution network relay protection parameter extraction method based on ENTSILDig software
CN117853827A (en) * 2024-03-07 2024-04-09 安徽省大气探测技术保障中心 Sampling pump working state operation monitoring system and method for atmospheric chamber gas monitoring
CN117853827B (en) * 2024-03-07 2024-05-14 安徽省大气探测技术保障中心 Sampling pump working state operation monitoring system and method for atmospheric chamber gas monitoring

Similar Documents

Publication Publication Date Title
CN110909302A (en) Method and system for learning local disturbance characteristics of operating state parameters of alternating-current and direct-current power grid
Veerasamy et al. LSTM recurrent neural network classifier for high impedance fault detection in solar PV integrated power system
Cai et al. Classification of power quality disturbances using Wigner-Ville distribution and deep convolutional neural networks
WO2022261805A1 (en) Diesel engine gearbox fault diagnosis method
Radil et al. PQ monitoring system for real-time detection and classification of disturbances in a single-phase power system
Zhang et al. Morphological undecimated wavelet decomposition for fault location on power transmission lines
CN110648088B (en) Electric energy quality disturbance source judgment method based on bird swarm algorithm and SVM
CN110726898B (en) Power distribution network fault type identification method
Taskovski et al. Classification of power quality disturbances using wavelets and support vector machine
CN111275108A (en) Method for performing sample expansion on partial discharge data based on generation countermeasure network
Koley et al. Wavelet-aided SVM tool for impulse fault identification in transformers
CN109283389B (en) Wind farm harmonic emission level assessment method and system
CN112881942A (en) Abnormal current diagnosis method and system based on wavelet decomposition and empirical mode decomposition
Asman et al. Identification of transient overvoltage using discrete wavelet transform with minimised border distortion effect and support vector machine
CN112327104A (en) Fault detection and positioning method for power distribution network with distributed power supply
CN114386452B (en) Nuclear power circulating water pump sun gear fault detection method
CN112151067B (en) Digital audio tampering passive detection method based on convolutional neural network
CN117116290B (en) Method and related equipment for positioning defects of numerical control machine tool parts based on multidimensional characteristics
CN111896890B (en) Micro-grid line fault diagnosis method and system based on Hilbert-Huang transform
CN113076892A (en) Sag type identification method, device and system based on deep learning SDAE-BP
CN111027587A (en) Method and system for calibrating attributes of AC/DC network operation data
CN117572161A (en) Arc fault detection method and device and related components
CN114755529A (en) Multi-feature fusion single-phase earth fault type identification method based on deep learning
CN115662456A (en) Adaptive filtering method, device and equipment for environmental noise and storage medium
CN115563480A (en) Gear fault identification method for screening octave geometric modal decomposition based on kurtosis ratio coefficient

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