CN110968073A - Double-layer tracing identification method for commutation failure reasons of HVDC system - Google Patents

Double-layer tracing identification method for commutation failure reasons of HVDC system Download PDF

Info

Publication number
CN110968073A
CN110968073A CN201911158769.5A CN201911158769A CN110968073A CN 110968073 A CN110968073 A CN 110968073A CN 201911158769 A CN201911158769 A CN 201911158769A CN 110968073 A CN110968073 A CN 110968073A
Authority
CN
China
Prior art keywords
fault
commutation failure
degree
point
matrix
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN201911158769.5A
Other languages
Chinese (zh)
Other versions
CN110968073B (en
Inventor
王渝红
邰克强
宋雨妍
寇然
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.)
Sichuan University
Original Assignee
Sichuan University
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 Sichuan University filed Critical Sichuan University
Priority to CN201911158769.5A priority Critical patent/CN110968073B/en
Publication of CN110968073A publication Critical patent/CN110968073A/en
Application granted granted Critical
Publication of CN110968073B publication Critical patent/CN110968073B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B23/00—Testing or monitoring of control systems or parts thereof
    • G05B23/02—Electric testing or monitoring
    • G05B23/0205—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults
    • G05B23/0259—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults characterized by the response to fault detection
    • G05B23/0275—Fault isolation and identification, e.g. classify fault; estimate cause or root of failure

Landscapes

  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Engineering & Computer Science (AREA)
  • Automation & Control Theory (AREA)
  • Locating Faults (AREA)

Abstract

A double-layer tracing identification method for a commutation failure fault reason of an HVDC system is characterized in that wavelet entropy is used for carrying out feature extraction on a fault signal in shallow layer identification, and a fault feature space is constructed. An AP clustering algorithm is introduced, the data are divided only by using the characteristics of the data, the interference of human factors is avoided, and the tracking shallow layer identification of commutation failure can be realized without supervision so as to distinguish the converter valve fault from the AC side fault. In the deep layer identification, an auxiliary electrical signal is utilized to construct a fault space-time sample matrix, a convolutional neural network with excellent robustness is selected to realize the phase commutation failure traceability deep layer identification, and the specific reason causing the phase commutation failure is determined. Finally, the method is verified by taking a +/-500 kV high-voltage direct-current transmission system and a CIGRE direct-current transmission standard model as examples.

Description

Double-layer tracing identification method for commutation failure reasons of HVDC system
Technical Field
The invention belongs to the technical field of direct current transmission, and particularly relates to a double-layer tracing identification method for a commutation failure reason of an HVDC system.
Background
The phase commutation failure is one of typical faults of the traditional HVDC system, the voltage and current of the DC system are distorted after the phase commutation failure occurs, and if the reason causing the phase commutation failure continuously exists or is not properly processed, the subsequent phase commutation failure can be caused, so that the DC system is forced to be locked, and the safe and stable operation of the AC/DC system is seriously influenced. For the research of phase change failure, a certain mature theory is existed at present at home and abroad, the voltage drop of a current conversion bus at an inverter side and the rise of direct current are main reasons for causing the phase change failure, and the failure of a trigger pulse circuit in a converter valve can also cause the phase change failure. At present, research mostly focuses on the mechanism and influencing factors of commutation failure generation, commutation failure prediction and suppression, and the like. For phase commutation failure fault identification, the prior art mainly focuses on the aspects of relevant indexes for identifying fault reasons, threshold setting and the like, and the research idea is to generally extract the characteristics of fault information and realize fault diagnosis by combining relevant algorithms, but the process needs to be intervened by manually setting relevant parameters.
Disclosure of Invention
The invention aims to provide a double-layer tracing identification method for a commutation failure fault reason of an HVDC system, which is used for solving the problem of inaccurate prediction result caused by the fact that related parameters need to be manually set for intervention in the process of researching the commutation failure fault reason in the prior art.
In order to achieve the purpose, the technical scheme adopted by the invention is as follows:
a double-layer tracing identification method for a commutation failure fault reason of an HVDC system comprises the following steps:
s1, carrying out fault simulation, collecting wave recording signals, carrying out feature extraction on the fault signals by using wavelet entropy, and constructing a fault feature space T;
s2, calculating a similarity matrix S among samples to realize algorithm initialization, wherein diagonal elements of the matrix are called deviation parameters, and the numerical value is used as a standard for selecting the point to become a clustering center point; calculating the attraction degree and the attribution degree among the sample points and realizing iterative updating; introducing an oscillation limiting coefficient mu to avoid iterative oscillation, updating the iterative attraction degree and the attribution degree again, and introducing an iterative interval coefficient k, wherein k is eλ，
Figure BDA0002285496280000011
m is the number of input data to increase the discrimination between samples of different classes, when the set maximum iteration number or iteration immobility number is reached,stopping iterative updating of the algorithm, judging a clustering center point, and dividing the categories of the rest points to obtain a shallow layer identification result and determine the category of shallow layer fault reasons causing the commutation failure of the HVDC system;
and S3, selecting different corresponding convolutional neural networks according to the shallow layer identification result, and outputting the deep layer identification result to determine the specific fault reason.
Preferably, the step S1 further includes the following sub-steps:
s1.1, adopting wavelet energy spectrum entropy, wavelet singular entropy and wavelet distance entropy to carry out fault feature space composition, utilizing WEE to express energy distribution features of signals in time domains and frequency domains, using WSE to measure complexity and uncertainty degrees of the signals, adopting WDE to depict distance sizes among different coefficient matrixes, reflecting internal connection features of the fault signals, processing the fault signals and extracting feature information contained in the fault signals;
s1.2, a fault feature space T is constructed by adopting three different wavelet entropies, feature information contained in complex and disordered fault signals is shown from different angles, the complexity of the signals is reduced, and the operation speed of a subsequent algorithm is improved.
3. The double-layer tracing identification method for the commutation failure reason of the HVDC system according to claim 2, wherein the expressions of the inter-sample similarity matrix S, the attraction degree and the attribution degree are respectively as follows:
Figure BDA0002285496280000021
Figure BDA0002285496280000022
Figure BDA0002285496280000023
wherein the attraction degree R (i, j) represents the attraction degree of the sample point i, and the larger the numerical value, the more suitable the point j is to be the center of the point i; the attribution degree A (i, j) represents the support degree of the point i to the point j to become the center of the point i, the numerical values of the attribution degree A (i, j) and the attribution degree A (i, j) are larger, the probability that the two points are classified into one type is higher, S (i, j) and S (i, j ') are the points i and j, the similarity degree of the points i and j' is higher, subscripts t and t +1 represent the iteration times, algorithm initialization is achieved by calculating a similarity matrix S between samples, and iterative updating is achieved by calculating the attraction degree R (i, j) and the attribution degree A (i, j) between the sample points;
the formula for updating the iterative attraction degree and the attribution degree is as follows:
Rt+1(i,j)←eλ[μRt(i,j)+(1-μ)Rt+1(i,j)]
At+1(i,j)←eλ[μAt(i,j)+(1-μ)At+1(i,j)]
wherein the content of the first and second substances,
Figure BDA0002285496280000024
the formula for judging the position basis of the clustering center point is as follows:
C＝argmax{Q}＝argmax{A(i,j)+R(i,j)}
the result matrix Q is the sum of the membership matrix and the membership matrix, the position of the maximum value of each row is sequentially searched, and if i is j, the point j is determined as a clustering center; if i ≠ j, the point i is classified into the category represented by the point j.
Preferably, the step S3 includes the following sub-steps:
s3.1, selecting an Adam optimization algorithm as a learning rate adjustment optimization algorithm to improve the accuracy of deep layer identification, and adopting a Dropout mechanism to improve the network generalization capability and improve the accuracy of deep layer identification;
s3.2, selecting the voltage and current values of the direct-current line and the voltage and current effective values of the alternating-current side of the converter valve during the fault period, and constructing a fault space-time sample matrix FiForming a fault space-time sample matrix set F as a sample set;
s3.3, reconstructing the fault space-time sample matrix set F to adapt to the input format of the CNN network;
s3.4, realizing F pair by adopting a 'overlapped slicing' methodiExpanding the data in (1);
s3.5, expanding FiAnd (3) processing fault data into a convolutional neural network test set, entering into a correspondingly trained convolutional neural network, obtaining a deep identification result of a commutation failure fault reason, and determining the type of a deep fault causing the commutation failure of the HVDC system.
Preferably, said step 4.3 further comprises the following sub-steps:
s4.3.1, matrix FiEach element value in the (1) is divided into an integer and a decimal part;
s4.3.2, converting the two parts of decimal value into fourteen digit binary number, combining into a twenty eight digit binary number according to the principle that the integer part is before and after the decimal part, and getting the sample set F under binary expression.
Preferably, the "overlapped slicing" refers to dividing the original fault electrical signal by using an equal-length time window, and when the length of the time window is smaller than the length of the signal and the time window moves by a fixed step length smaller than the length of the signal, overlapping may exist between the collected samples, so as to realize the expansion of the sample data.
The beneficial technical effects of the invention are as follows:
1. the shallow layer identification method based on the wavelet entropy fault feature space and the AP algorithm does not need human intervention, can quickly and effectively judge whether a fault signal belongs to a converter valve fault or an alternating current side fault, and completes shallow layer identification.
2. On the basis of shallow layer identification, direct current, alternating current voltage and current signals are supplemented, the phase change failure traceable deep layer identification is realized by utilizing a convolutional neural network, and the real reason of the phase change failure caused by locking is locked.
Drawings
FIG. 1 is a flow chart of the method of the present invention;
FIG. 2 is a diagram showing DC voltage waveforms during a fault in accordance with the present invention;
FIG. 3 is a schematic diagram of a fault signature space according to the present invention;
FIG. 4 is a diagram of a dual-level tracing identification result according to 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 fig. 1 to 4 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 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, a double-layer tracing identification method for a commutation failure cause of an HVDC system includes the following steps:
s1, carrying out fault simulation, collecting wave recording signals, carrying out feature extraction on the fault signals by using wavelet entropy, and constructing a fault feature space T;
s2, calculating a similarity matrix S among samples to realize algorithm initialization, wherein diagonal elements of the matrix are called deviation parameters, and the numerical value is used as a standard for selecting the point to become a clustering center point; calculating the attraction degree and the attribution degree among the sample points and realizing iterative updating; introducing an oscillation limiting coefficient mu to avoid iterative oscillation, updating the iterative attraction degree and the attribution degree again, and introducing an iterative interval coefficient k, wherein k is eλ，
Figure BDA0002285496280000041
m is the number of input data to increase the discrimination among samples of different types, when the set maximum iteration times or iteration immobility times are reached, the iterative updating of the algorithm is stopped, the clustering center point is judged, the categories of the rest points are divided, so that a shallow layer identification result is obtained, and the category of a shallow layer fault cause causing the commutation failure of the HVDC system is determined;
and S3, selecting different corresponding convolutional neural networks according to the shallow layer identification result, and outputting the deep layer identification result to determine the specific fault reason.
Further, the step S1 further includes the following sub-steps:
s1.1, adopting wavelet energy spectrum entropy, wavelet singular entropy and wavelet distance entropy to carry out fault feature space composition, and utilizing WEEPresenting a signalFeatures of energy distribution in time and frequency domains, using WSEMeasuring the complexity and uncertainty degree of the signal by adopting WDEDepicting the distance between different coefficient matrixes to reflect the internal connection characteristics of the fault signal, so as to process the fault signal and extract the characteristic information contained in the fault signal;
are expressed as follows:
Figure BDA0002285496280000042
Figure BDA0002285496280000043
Figure BDA0002285496280000044
where the subscript l indicates the number of wavelet decomposition layers,
Figure BDA0002285496280000045
the sum of energy of each layer is decomposed by wavelet;
Figure BDA0002285496280000046
representing the sum of the singular values of the coefficient matrix,
Figure BDA0002285496280000047
and the distance between the wavelet coefficients of the j-th layer and the i-th layer is represented.
S1.2, considering a plurality of fault reasons causing commutation failure, constructing a fault characteristic space T by adopting three different wavelet entropies according to the theory, representing characteristic information contained in complex and disordered fault signals from different angles, reducing the complexity of the signals and improving the operation speed of a subsequent algorithm;
Ti＝[WSE,WEE,WDE]
T＝[T1,T2,…,Ti]T,i＝1,2,…,m
wherein i represents a specific cause corresponding to the phase commutation failure, and m represents the kind of the fault causing the phase commutation failure
Further, in the inter-sample similarity matrix S, expressions of the attraction degree and the attribution degree are respectively:
Figure BDA0002285496280000051
Figure BDA0002285496280000052
Figure BDA0002285496280000053
wherein the attraction degree R (i, j) represents the attraction degree of the sample point i, and the larger the numerical value, the more suitable the point j is to be the center of the point i; the attribution degree A (i, j) represents the support degree of the point i to the point j to become the center of the point i, the numerical values of the attribution degree A (i, j) and the attribution degree A (i, j) are larger, the probability that the two points are classified into one type is higher, S (i, j) and S (i, j ') are the points i and j, the similarity degree of the points i and j' is higher, subscripts t and t +1 represent the iteration times, algorithm initialization is achieved by calculating a similarity matrix S between samples, and iterative updating is achieved by calculating the attraction degree R (i, j) and the attribution degree A (i, j) between the sample points;
the formula for updating the iterative attraction degree and the attribution degree is as follows:
Rt+1(i,j)←eλ[μRt(i,j)+(1-μ)Rt+1(i,j)]
At+1(i,j)←eλ[μAt(i,j)+(1-μ)At+1(i,j)]
wherein the content of the first and second substances,
Figure BDA0002285496280000054
the formula for judging the position basis of the clustering center point is as follows:
C＝argmax{Q}＝argmax{A(i,j)+R(i,j)}
the result matrix Q is the sum of the membership matrix and the membership matrix, the position of the maximum value of each row is sequentially searched, and if i is j, the point j is determined as a clustering center; if i ≠ j, the point i is classified into the category represented by the point j.
Convolutional neural networks are one of the deep learning modes. The core idea is to alternately utilize the convolutional layer and the pooling layer to perform feature extraction and pooling on input data layer by layer. Compared with other artificial neural networks, the CNN has the characteristics of local connection, region weight sharing, feature refining transmission and the like, can reduce model training parameters, improves training cost and algorithm accuracy, and is suitable for processing large-scale data. In order to construct a network with excellent structure, excellent training and testing results and strong generalization capability, the CNN network structure is optimized, and the optimized structure adopted in the present application is explained below:
further, the step S3 includes the following sub-steps:
s3.1, selecting an Adam optimization algorithm as a learning rate adjustment optimization algorithm to improve the accuracy of deep layer identification, and adopting a Dropout mechanism to improve the network generalization capability and improve the accuracy of deep layer identification;
during training of the neural network, the setting of the learning rate is related to the updating speed of the parameters, and the method plays a vital role in network training. The Adam optimization algorithm has gradient diagonal scaling invariance and self-adaptive learning rate, has unique advantages in processing large-scale and high-latitude electrical signal data, and has good performance for noise-containing electrical fault data and sparse gradient unsteady-state problems. The overshoot parameters are simple and clear, the application range of default parameters is wide, manual adjustment is not needed, the Adam optimization algorithm is selected as the learning rate adjustment optimization algorithm, the deep-layer identification accuracy is improved, the convolutional neural network has unique advantages in processing mass data, when training data contain noise and the sample capacity is insufficient, an overfitting phenomenon often occurs, the phenomenon is shown in that the training set has good memory capacity, characteristics contained in the training set are not learned, and the phenomenon is not ideal in identifying unknown data. Considering that the fault electric signal has noise and interference of an uncertain signal, a Dropout mechanism with wider application is adopted to improve the network generalization capability and improve the deep identification accuracy.
S3.2, selecting the voltage and current values of the direct-current line and the voltage and current effective values of the alternating-current side of the converter valve during the fault period, and constructing a fault space-time sample matrix FiForming a fault space-time sample matrix set F as a sample set;
aiming at the learning characteristics of the convolutional neural network, a high-dimensional space-time sample matrix needs to be constructed, so that the CNN network can fully and effectively learn the state characteristics of the system during the fault period. The single electrical signal of the HVDC system is usually a one-dimensional time signal which is a time function varying with time, and the operation state of the HVDC system under the condition of commutation failure cannot be well represented by selecting the single electrical signal. Considering the change rule of the electric quantity of the alternating current side and the direct current side during the fault, expressing the characteristics of the running states of the alternating current side and the direct current side and the like, selecting the voltage and the current value of the direct current line and the effective voltage and the effective current value of the alternating current side of the converter valve during the fault, and constructing a fault space-time sample matrix FiAnd forming a fault space-time sample matrix set F as a sample set:
Fi＝[Ud,Id,Uac,Iac]
F＝{F1,F2,…,Fi},i＝1,2,…,m
s3.3, reconstructing the fault space-time sample matrix set F to adapt to the input format of the CNN network;
s3.4, realizing F pair by adopting a 'overlapped slicing' methodiExpanding the data in (1);
s3.5, expanding FiAnd (3) processing fault data into a convolutional neural network test set, entering into a correspondingly trained convolutional neural network, obtaining a deep identification result of a commutation failure fault reason, and determining the type of a deep fault causing the commutation failure of the HVDC system.
Further, the step 4.3 further comprises the following substeps:
s4.3.1, matrix FiEach element value in the (1) is divided into an integer and a decimal part;
s4.3.2, converting the two parts of numerical values expressed in decimal system into binary numbers with fourteen bits, combining the binary numbers into a binary number with twenty eight bits according to the principle that the integer part is before and after the decimal part, and obtaining a sample set F under binary expression:
FDecimal→FBinary。
furthermore, the overlapping slicing is to divide the original fault electrical signal by using an equal-length time window, when the length of the time window is smaller than the signal length and the time window moves by a fixed step length smaller than the signal length, overlapping exists between the collected samples, so as to realize the expansion of the sample data, and assuming that a fault space-time sample matrix F in the textiIs L, according to the above method, FiExpandable to a number N ═ Lsignal-Lw)/LstepA subset of samples.
The influence factors causing the commutation failure are complex, and at present, the commonly adopted expression of the turn-off angle of the inverter side converter of the high-voltage direct-current transmission system is
Figure BDA0002285496280000071
Wherein: i isdIs direct current; xcFor commutation reactance, Xc ═ ω Lc；Uacβ is trigger lead angle;
Figure BDA0002285496280000072
the asymmetric ground fault voltage zero crossing is offset by an angle.
Typically, when the converter off-angle γ is less than its minimum off-angle γminWhen the current converter fails to change the phase. As can be seen from the expression, the failure of the inverter-side ac system, which results in the decrease of the ac side line voltage and the increase of the dc current, is the main cause of the decrease of γ. Meanwhile, internal faults of the converter, such as loss of trigger pulses, short-circuit faults of bridge arms and the like, can also cause the converter valve to be incapable of being normally conducted, and cause phase conversion failure.
The expression can theoretically analyze which factors can cause the reduction of the gamma angle, but in practical engineering application, it is very difficult to judge which fault causes the generation of commutation failure.
The causes of faults causing commutation failure in HVDC systems can be roughly divided into two broad categories, which can be expressed as two sets V and W, where V is the converter valve fault set and W is the ac side fault set. Each fault set can be respectively and specifically refined into different fault types, namely V ═ V1，V2,…,Vn]，W＝[W1,w2,…,Wn]. The details are shown in Table 1.
Figure BDA0002285496280000073
The simulation verification adopts a typical direct current engineering model and a CIGRE standard model for double verification, and the typical direct current engineering simulation experiment and result analysis are as follows:
a direct current power transmission system with the rated power of 3000MW is built in PSCAD/EMTDC software, and a digital simulation calculation result is used as a numerical value source verified by the method. Faults listed in the figure 1 are applied to the inversion bus of the inversion side converter station, simulation is carried out, and fault data are recorded. By fault V3,W1,W2For example, the dc voltage waveform during the fault is observed, as shown in fig. 2, during the fault duration, the dc voltage waveform in all three fault situations falls to different degrees, and as the fault duration, the dc voltage oscillates to different amplitudes, wherein the dc voltage has substantially the same trend from 10ms to 40ms after the fault occurs, and the fluctuation range is from 100kV to 300 kV. Observing the waveform change between different DC voltages at the later stage of the fault, i.e. the fault V3And fault W2The direct-current voltage waveform oscillates near 100kV, different voltage waveforms of the direct-current voltage waveform slightly differ in the fault period but are not distinguished, and the specific reason for causing the commutation failure cannot be judged only from the direct-current voltage waveform.
Selecting two-phase short circuit grounding faults of an alternating current side converter bus with grounding resistance of 10 omega and a converter valve 6 valve multi-pulse loss fault as test faults, numbering S1 and S2, constructing a fault characteristic space together with fault simulation in a table 1, and performing shallow layer identification, wherein the fault characteristic space is as shown in a table 2:
Figure 1
setting the maximum iteration times of the AP algorithm as 500 times, the iteration immobility times as 50 times, the damping coefficient as 0.5 and the deviation parameter P as a data median. And calculating a similarity matrix and starting shallow layer identification. After the iteration of the algorithm stops, the results are shown in table 3:
Figure 2
observing the table 3, according to the formula for judging the position basis of the cluster center point and the judgment condition of the improved AP algorithm, the position of the maximum value of each row of elements is found, and when i is equal to j (namely, the diagonal elements of the table), the [ V ] can be determined1,W1]Respectively as the category [ V, W ] of the fault cause of this shallow layer identification]And the center is taken as a representative of each shallow fault cause category. And when i ≠ j (namely other elements of the table), sequentially judging the specific category of each data according to the maximum value position of each row. For example, observing the second row and the third row of elements of the result matrix Q, the maximum value appears in column 1, and V2 and V3 belong to the shallow fault cause category represented by V1, and similarly, it can be known that the attributions of other data points are consistent with the expected design. Then observing the iteration result of the test signals in the table 3, judging the test signal S1 as the converter valve AC side fault W, judging the test signal S2 as the converter valve pulse loss fault V, and accurately identifying the fault reason,
the fuzzy C clustering algorithm is selected as a comparison algorithm in the simulation, the same data are recalculated, the clustering number is set to be 2, the membership degree is used as a fault type division basis, an Euclidean distance basic distance formula is selected, and the results after the algorithm is stopped are shown in Table 4:
Figure 3
for the classification of the test signals, the fuzzy clustering algorithm is in accordance with the pre-design, and the judgment result and the AP calculation are carried outThe method is identical in judgment, but the converter valve multi-phase one-shot pulse loses the fault signal V2Misjudging as a valve AC side fault W, and sending a converter valve AC side single-phase short circuit grounding fault signal W1The misjudgment is a pulse loss fault V. Therefore, the accuracy of the AP algorithm is superior to that of the fuzzy C clustering algorithm, and the shallow layer identification of unknown fault reasons can be accurately realized. In order to better show the shallow layer identification result, three kinds of wavelet entropy values are respectively used as the coordinates of the fault signal, the identification result of the improved AP algorithm is visualized, a three-dimensional stereogram is established, and the result is shown in FIG. 3
The method selects a TensorFlow architecture based on Python language, builds a deep recognition method convolutional neural network model, and the network structure parameters are shown in Table 5:
Figure BDA0002285496280000092
aiming at the characteristic that a large number of samples are required for the training of the convolutional neural network as supports, the data samples are respectively expanded under different fault types, and different fault states of the system are supplemented under the condition of the same fault. For the converter valve faults, different bridges of the converter valve and different combined bridges of the converter valve are set, and the faults of single-trigger pulse loss and multi-trigger pulse loss occur; for alternating current side faults, the alternating current system adopts a Thevenin equivalent model, under the condition of considering each fault, different grounding resistances of a transmission line are modified between 0 and 50 ohms, different fault states under the fault type are simulated, data are preprocessed, a random sub-sampling verification method is adopted, 21600 sample data are generated in total, 2400 samples are randomly selected as a test set, 9600 converter valve fault training sets are selected, and 1200 test sets are selected; the alternating current side fault training sets are 9600, and the testing sets are 1200. The convolutional neural network is adopted to train different types of faults in sequence, and the process is shown in fig. 4.
In order to verify the deep layer identification method provided by the text, the text adopts a BP neural network as a CNN comparison algorithm, sets related parameters of the BP neural network, and trains and tests the same sample respectively, wherein the results are shown in Table 6:
Figure BDA0002285496280000101
the result shows that for +/-500 kV direct current engineering, the accuracy of the identification results of the converter valve fault and the alternating current side fault is 94.51 percent and 98.08 percent respectively; for the CIGRE standard model, the accuracy is 96.75% and 98.96% respectively. Compared with the traditional BP neural network, the convolutional neural network has high result accuracy, and can accurately identify the commutation failure fault reason for a HVDC system.
The working principle of the invention is as follows: fault simulation is carried out, a recording signal is collected, a fault feature space is formed by adopting wavelet energy spectrum entropy, wavelet singular entropy and wavelet distance entropy, WEE is utilized to represent energy distribution features of the signal in time domain and frequency domain, WSE is utilized to measure complexity and uncertainty degree of the signal, WDE is utilized to depict distance between different coefficient matrixes, internal connection features of the fault signal are reflected, the fault signal is processed, and feature information contained in the fault signal is extracted; three different wavelet entropies are adopted to construct a fault characteristic space T, and the characteristic information contained in complex and disordered fault signals is presented from different angles, so that the complexity of the signals is reduced, and the operation speed of a subsequent algorithm is improved; and calculating an inter-sample similarity matrix S to realize algorithm initialization, wherein diagonal elements of the matrix are called deviation parameters, and the numerical value is used as a standard for selecting the point to become a clustering center point. And calculating the attraction degree and the attribution degree among the sample points and realizing iterative updating. And introducing a damping coefficient lambda to avoid iterative oscillation, updating the iterative attraction degree and the attribution degree again, stopping the iterative updating of the algorithm when the set maximum iterative times or the set iteration immobility times are reached, judging the clustering center point, and dividing the categories of the rest points to obtain a shallow layer identification result, determining the category of shallow layer fault reasons causing the commutation failure of the HVDC system, selecting different corresponding convolutional neural networks according to the shallow layer identification result, and outputting a deep layer identification result to determine the specific fault reasons.
In the description of the present invention, it is to be understood that the terms "counterclockwise", "clockwise", "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate orientations or positional relationships based on those shown in the drawings, and are used for convenience of description only, and do not indicate or imply that the referenced devices or elements must have a particular orientation, be constructed and operated in a particular orientation, and thus, are not to be considered as limiting.

Claims (6)

1. A double-layer tracing identification method for a commutation failure fault reason of an HVDC system is characterized by comprising the following steps:
s1, carrying out fault simulation, collecting wave recording signals, carrying out feature extraction on the fault signals by using wavelet entropy, and constructing a fault feature space T;
s2, calculating a similarity matrix S among samples to realize algorithm initialization, wherein diagonal elements of the matrix are called deviation parameters, and the numerical value is used as a standard for selecting the point to become a clustering center point; calculating the attraction degree and the attribution degree among the sample points and realizing iterative updating; introducing an oscillation limiting coefficient mu to avoid iterative oscillation, updating the iterative attraction degree and the attribution degree again, and introducing an iterative interval coefficient k, wherein k is eλ，
Figure FDA0002285496270000011
m is the number of input data to increase the discrimination among samples of different types, when the set maximum iteration times or iteration immobility times are reached, the iterative updating of the algorithm is stopped, the clustering center point is judged, the categories of the rest points are divided, so that a shallow layer identification result is obtained, and the category of a shallow layer fault cause causing the commutation failure of the HVDC system is determined;
and S3, selecting different corresponding convolutional neural networks according to the shallow layer identification result, and outputting the deep layer identification result to determine the specific fault reason.
2. The double-layer tracing identification method for the commutation failure reason of the HVDC system according to claim 1, wherein the step S1 further includes the following sub-steps:
s1.1, adopting wavelet energy spectrum entropy, wavelet singular entropy and wavelet distance entropy to carry out fault feature space composition, utilizing WEE to express energy distribution features of signals in time domains and frequency domains, using WSE to measure complexity and uncertainty degrees of the signals, adopting WDE to depict distance sizes among different coefficient matrixes, reflecting internal connection features of the fault signals, processing the fault signals and extracting feature information contained in the fault signals;
s1.2, a fault feature space T is constructed by adopting three different wavelet entropies, feature information contained in complex and disordered fault signals is shown from different angles, the complexity of the signals is reduced, and the operation speed of a subsequent algorithm is improved.
3. The double-layer tracing identification method for the commutation failure reason of the HVDC system according to claim 2, wherein the expressions of the inter-sample similarity matrix S, the attraction degree and the attribution degree are respectively as follows:
Figure FDA0002285496270000012
Figure FDA0002285496270000013
Figure FDA0002285496270000021
wherein the attraction degree R (i, j) represents the attraction degree of the sample point i, and the larger the numerical value, the more suitable the point j is to be the center of the point i; the attribution degree A (i, j) represents the support degree of the point i to the point j to become the center of the point i, the numerical values of the attribution degree A (i, j) and the attribution degree A (i, j) are larger, the probability that the two points are classified into one type is higher, S (i, j) and S (i, j ') are the points i and j, the similarity degree of the points i and j' is higher, subscripts t and t +1 represent the iteration times, algorithm initialization is achieved by calculating a similarity matrix S between samples, and iterative updating is achieved by calculating the attraction degree R (i, j) and the attribution degree A (i, j) between the sample points;
the formula for updating the iterative attraction degree and the attribution degree is as follows:
Rt+1(i,j)←eλ[μRt(i,j)+(1-μ)Rt+1(i,j)]
At+1(i,j)←eλ[μAt(i,j)+(1-μ)At+1(i,j)]
wherein the content of the first and second substances,
Figure FDA0002285496270000022
the formula for judging the position basis of the clustering center point is as follows:
C＝argmax{Q}＝argmax{A(i,j)+R(i,j)}
the result matrix Q is the sum of the membership matrix and the membership matrix, the position of the maximum value of each row is sequentially searched, and if i is j, the point j is determined as a clustering center; if i ≠ j, the point i is classified into the category represented by the point j.
4. The double-layer tracing identification method for the commutation failure reason of the HVDC system according to claim 1, wherein the step S3 includes the following sub-steps:
s3.1, selecting an Adam optimization algorithm as a learning rate adjustment optimization algorithm to improve the accuracy of deep layer identification, and adopting a Dropout mechanism to improve the network generalization capability and improve the accuracy of deep layer identification;
s3.2, selecting the voltage and current values of the direct-current line and the voltage and current effective values of the alternating-current side of the converter valve during the fault period, and constructing a fault space-time sample matrix FiForming a fault space-time sample matrix set F as a sample set;
s3.3, reconstructing the fault space-time sample matrix set F to adapt to the input format of the CNN network;
s3.4, realizing F pair by adopting a 'overlapped slicing' methodiExpanding the data in (1);
s3.5, expanding FiThe fault data is processed into a test set of the convolutional neural network and enters the convolutional neural network which is trained correspondingly, and deep discrimination of the fault reason of commutation failure can be obtainedAnd identifying the type of deep fault causing the commutation failure of the HVDC system.
5. The HVDC system commutation failure cause double-layer tracing identification method according to claim 4, wherein the step 4.3 further comprises the following substeps:
s4.3.1, matrix FiEach element value in the (1) is divided into an integer and a decimal part;
s4.3.2, converting the two parts of decimal value into fourteen digit binary number, combining into a twenty eight digit binary number according to the principle that the integer part is before and after the decimal part, and getting the sample set F under binary expression.
6. The double-layer tracing identification method for the commutation failure reason of the HVDC system according to claim 5, wherein the "overlapped slicing" is to divide the original failure electrical signal by using an equal-length time window, and when the length of the time window is smaller than the length of the signal and the time window is moved by a fixed step length smaller than the length of the signal, there is an overlap between the collected samples, so as to expand the sample data.
CN201911158769.5A 2019-11-22 2019-11-22 Double-layer tracing identification method for commutation failure reasons of HVDC system Active CN110968073B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201911158769.5A CN110968073B (en) 2019-11-22 2019-11-22 Double-layer tracing identification method for commutation failure reasons of HVDC system

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201911158769.5A CN110968073B (en) 2019-11-22 2019-11-22 Double-layer tracing identification method for commutation failure reasons of HVDC system

Publications (2)

Publication Number Publication Date
CN110968073A true CN110968073A (en) 2020-04-07
CN110968073B CN110968073B (en) 2021-04-02

Family

ID=70031326

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201911158769.5A Active CN110968073B (en) 2019-11-22 2019-11-22 Double-layer tracing identification method for commutation failure reasons of HVDC system

Country Status (1)

Country Link
CN (1) CN110968073B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113791294A (en) * 2021-09-07 2021-12-14 国网安徽省电力有限公司 Extreme high voltage direct current transmission system commutation failure tracing method based on edge cloud coordination

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102520317A (en) * 2011-12-15 2012-06-27 四川大学 Commutation failure fault detection method based on wavelet Shannon entropy standardization characteristic vector
CN103837784A (en) * 2014-03-27 2014-06-04 国家电网公司 Artificial intelligence identification method of external faults of conversion current bus of HVDC system
KR20150142763A (en) * 2014-06-11 2015-12-23 한국전력공사 Apparatus for predicting and controlling commutation failure of high voltage direct current system
KR20160033362A (en) * 2014-09-17 2016-03-28 한국전력공사 Apparatus controlling commutation failure of hvdc system
CN106291239A (en) * 2016-08-02 2017-01-04 昆明理工大学 A kind of direct current transmission line fault recognition methods using filter branches electric current and principal component analytical method
CN107077625A (en) * 2014-10-27 2017-08-18 电子湾有限公司 The deep convolutional neural networks of layering
CN109635732A (en) * 2018-12-12 2019-04-16 国网山东省电力公司电力科学研究院 A kind of continuous commutation failure fault pre-alarming network training method of LCC-HVDC
CN110470937A (en) * 2019-07-15 2019-11-19 昆明理工大学 Based on FEEMD Sample Entropy+neural network HVDC transmission system line fault and commutation failure method for diagnosing faults

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102520317A (en) * 2011-12-15 2012-06-27 四川大学 Commutation failure fault detection method based on wavelet Shannon entropy standardization characteristic vector
CN103837784A (en) * 2014-03-27 2014-06-04 国家电网公司 Artificial intelligence identification method of external faults of conversion current bus of HVDC system
KR20150142763A (en) * 2014-06-11 2015-12-23 한국전력공사 Apparatus for predicting and controlling commutation failure of high voltage direct current system
KR20160033362A (en) * 2014-09-17 2016-03-28 한국전력공사 Apparatus controlling commutation failure of hvdc system
CN107077625A (en) * 2014-10-27 2017-08-18 电子湾有限公司 The deep convolutional neural networks of layering
CN106291239A (en) * 2016-08-02 2017-01-04 昆明理工大学 A kind of direct current transmission line fault recognition methods using filter branches electric current and principal component analytical method
CN109635732A (en) * 2018-12-12 2019-04-16 国网山东省电力公司电力科学研究院 A kind of continuous commutation failure fault pre-alarming network training method of LCC-HVDC
CN110470937A (en) * 2019-07-15 2019-11-19 昆明理工大学 Based on FEEMD Sample Entropy+neural network HVDC transmission system line fault and commutation failure method for diagnosing faults

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113791294A (en) * 2021-09-07 2021-12-14 国网安徽省电力有限公司 Extreme high voltage direct current transmission system commutation failure tracing method based on edge cloud coordination
CN113791294B (en) * 2021-09-07 2023-10-31 国网安徽省电力有限公司 Extra-high voltage direct current transmission system commutation failure tracing method based on edge cloud cooperation

Also Published As

Publication number Publication date
CN110968073B (en) 2021-04-02

Similar Documents

Publication Publication Date Title
Veerasamy et al. LSTM recurrent neural network classifier for high impedance fault detection in solar PV integrated power system
CN110398663B (en) Flexible direct current power grid fault identification method based on convolutional neural network
CN109635928B (en) Voltage sag reason identification method based on deep learning model fusion
US20180238951A1 (en) Decision Tree SVM Fault Diagnosis Method of Photovoltaic Diode-Clamped Three-Level Inverter
CN109975673B (en) Method for identifying fault arc at direct current side of photovoltaic microgrid
CN110596530B (en) Low-current ground fault line selection method
CN110726898B (en) Power distribution network fault type identification method
CN110687393B (en) Valve short-circuit protection fault positioning method based on VMD-SVD-FCM
CN110247420B (en) Intelligent fault identification method for HVDC transmission line
CN105938578A (en) Large-scale photovoltaic power station equivalent modeling method based on clustering analysis
Song et al. A negative selection algorithm-based identification framework for distribution network faults with high resistance
CN113093058A (en) NPC three-level inverter open-circuit fault diagnosis method
CN110968073B (en) Double-layer tracing identification method for commutation failure reasons of HVDC system
Ye et al. Single pole‐to‐ground fault location method for mmc‐hvdc system using wavelet decomposition and dbn
CN111654392A (en) Low-voltage distribution network topology identification method and system based on mutual information
CN115469184A (en) New energy transmission line fault identification method based on convolutional network
CN115267428A (en) LCC-MMC single-pole grounding fault positioning method based on VMD-ET feature selection
CN109342885B (en) Method and system for positioning line fault of direct-current power distribution network
CN114266396A (en) Transient stability discrimination method based on intelligent screening of power grid characteristics
Li et al. Incipient Fault Detection in Power Distribution System: A Time–Frequency Embedded Deep-Learning-Based Approach
Cui et al. T-type inverter fault diagnosis based on GASF and improved AlexNet
CN108108659B (en) Island detection key feature extraction method based on empirical mode decomposition
CN117148034A (en) Power distribution network fault line selection method based on empirical wavelet transformation and graph annotation force network
Luo et al. Stacked auto-encoder-based transients recognition in VSC-HVDC
CN116131313A (en) Explanatory analysis method for association relation between characteristic quantity and transient power angle stability

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant