CN107992665B - Online fault diagnosis and analysis method for alternating current filter of extra-high voltage converter station - Google Patents

Online fault diagnosis and analysis method for alternating current filter of extra-high voltage converter station Download PDF

Info

Publication number
CN107992665B
CN107992665B CN201711207686.1A CN201711207686A CN107992665B CN 107992665 B CN107992665 B CN 107992665B CN 201711207686 A CN201711207686 A CN 201711207686A CN 107992665 B CN107992665 B CN 107992665B
Authority
CN
China
Prior art keywords
neural network
rbf neural
alternating current
extra
high voltage
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.)
Active
Application number
CN201711207686.1A
Other languages
Chinese (zh)
Other versions
CN107992665A (en
Inventor
刘志远
张沈习
韦鹏
史磊
李君宏
张志贤
高海洋
徐辉
王文刚
宁复茂
谢伟锋
韩慧麟
武嘉薇
尹琦云
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.)
Shanghai Jiaotong Electric Power Technology Co ltd
State Grid Corp of China SGCC
State Grid Ningxia Electric Power Co Ltd
Original Assignee
Shanghai Jiaotong Electric Power Technology Co ltd
State Grid Corp of China SGCC
State Grid Ningxia 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 Shanghai Jiaotong Electric Power Technology Co ltd, State Grid Corp of China SGCC, State Grid Ningxia Electric Power Co Ltd filed Critical Shanghai Jiaotong Electric Power Technology Co ltd
Priority to CN201711207686.1A priority Critical patent/CN107992665B/en
Publication of CN107992665A publication Critical patent/CN107992665A/en
Application granted granted Critical
Publication of CN107992665B publication Critical patent/CN107992665B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
    • Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
    • Y04S10/00—Systems supporting electrical power generation, transmission or distribution
    • Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Computational Linguistics (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computer Hardware Design (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Geometry (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Remote Monitoring And Control Of Power-Distribution Networks (AREA)
  • Emergency Protection Circuit Devices (AREA)

Abstract

The invention provides an on-line fault diagnosis and analysis method for an alternating current filter of an extra-high voltage converter station, belongs to the field of alternating current filter fault diagnosis, and aims to solve the problems that the existing fault diagnosis method is high in randomness, limited in application range and incapable of finding faults in time. The method comprises the following steps: training an RBF neural network by adopting a net receiving algorithm according to historical waveform data recorded by a fault recorder of the extra-high voltage converter station to determine model parameters of the RBF neural network, wherein the model parameters comprise the center, weight and width of the RBF neural network; acquiring current waveform data of the extra-high voltage converter station according to data recorded by a fault recorder of the extra-high voltage converter station, and extracting characteristic quantities of the current waveform data, wherein the extracted characteristic quantities comprise time domain characteristic quantities and frequency domain characteristic quantities; normalizing the extracted characteristic quantity; and inputting the characteristic quantity after the normalization processing into the trained RBF neural network, and diagnosing whether the alternating current filter has faults or not according to the output result of the trained RBF neural network.

Description

Online fault diagnosis and analysis method for alternating current filter of extra-high voltage converter station
Technical Field
The invention relates to the technical field of fault diagnosis of an alternating current filter of an extra-high voltage converter station, in particular to an on-line fault diagnosis and analysis method of the alternating current filter of the extra-high voltage converter station.
Background
The extra-high voltage converter station is a system for mutual energy conversion between direct current and alternating current in a direct current transmission system. The transmission principle of the converter station is different from that of a conventional alternating-current substation. In a dc transmission system, a large amount of reactive power needs to be consumed, and a large amount of harmonics exist in a converted waveform, so that a plurality of ac filter banks need to be arranged in a converter station. The alternating current filter bank is usually arranged in an outdoor place due to the characteristics of large heat productivity, large volume and the like, and is prone to defects because of being saturated by wind and being exposed to the sun. When the defect reaches a certain degree, the protection device in the station can detect the fault and further send out an alarm signal, and when the fault is serious, the AC filter bank is directly tripped off.
Common fault conditions of ac filter banks can be classified into the following three categories. (1) And (3) element failure: the capacitor unit leaks oil, and the occurrence of the conditions of damage of a resistor, a reactor and the like can cause unbalance of three-phase current and increase of harmonic waves; (2) mechanical failure of the circuit breaker body: the mechanical fault of the breaker body can influence the synchronism of three-phase closing; (3) joint aging: joints of metal materials are susceptible to oxidation, resulting in increased contact resistance and abnormal heat generation.
At present, the fault of the alternating current filter bank is mainly discovered by means of manual inspection, infrared temperature measurement, protection alarm and the like. The manual patrol frequency is low, and the found fault has higher randomness; the infrared temperature measurement has better capability of finding heating points, but only can detect the parts such as equipment shells, joints and the like, and has limited application range; when the fault of the equipment is found in a protection alarm mode, the equipment affects the normal operation of the direct current transmission system, and early small voltage and current differences cannot be detected at the first time.
Disclosure of Invention
The invention aims to solve the technical problems that the existing fault diagnosis method for the alternating current filter of the extra-high voltage converter station has high randomness and limited application range and cannot find faults in time, and provides an on-line fault diagnosis analysis method for the alternating current filter of the extra-high voltage converter station.
In order to solve the technical problems, the invention adopts the technical scheme that:
an on-line fault diagnosis and analysis method for an alternating current filter of an extra-high voltage converter station comprises the following steps:
s1, training an RBF neural network by adopting a net receiving algorithm according to historical waveform data recorded by a fault recorder of the extra-high voltage converter station to determine model parameters of the RBF neural network, wherein the model parameters comprise the center, weight and width of the RBF neural network;
s2, acquiring current waveform data of the extra-high voltage converter station according to data recorded by a fault recorder of the extra-high voltage converter station, and extracting characteristic quantities of the current waveform data, wherein the extracted characteristic quantities comprise time domain characteristic quantities and frequency domain characteristic quantities;
s3, normalizing the extracted characteristic quantity;
and S4, inputting the characteristic quantity after the normalization processing into the trained RBF neural network, and diagnosing whether the alternating current filter has faults or not according to the output result of the trained RBF neural network.
Optionally, the time-domain characteristic quantity includes an opening and closing time T of the circuit breaker, and a current maximum value I recorded in a waveform when the alternating-current filter opens and closesmaxAnd minimum value of current IminAnd a three-phase matching degree gamma; wherein:
T＝max(TA,TB,TC) (1)
in the formula (1), TA、TBAnd TCThe switching-on and switching-off time of three phases of the alternating current filter A, B, C is respectively;
Figure BDA0001484018260000021
in the formula (2), ia、ibAnd icRespectively representing the instantaneous values of the three-phase-shifted currents, I, of the AC filter A, B, Ca、IbAnd IcRespectively, the current amplitudes of the three phases of the ac filter A, B, C, and d represents the waveform of the d-th cycle in the waveform data.
Optionally, the frequency domain characteristic quantity comprises a zero sequence current harmonic maximum value IH,maxHarmonic content delta and total harmonic distortion rate K of zero sequence current of alternating current filter circuit breakerTHD(ii) a Wherein:
Figure BDA0001484018260000031
in equation (3): i isH,eIs the effective value of the e-harmonic wave, IrIs the rated current of the corresponding alternating current filter bank;
Figure BDA0001484018260000032
Figure BDA0001484018260000033
optionally, the step S1, when determining the model parameters of the RBF neural network, includes:
s11, extracting characteristic quantity of M groups of historical waveform data recorded by a fault recorder when an alternating current filter of the extra-high voltage converter station is switched to obtain seven characteristic quantities including time domain characteristic quantity and frequency domain characteristic quantity, marking abnormal states of the M groups of historical waveform data to obtain state values, and combining the seven characteristic quantities and one state value into an initial training data set SM×8Wherein, the time domain characteristic quantity comprises the opening and closing time T of the circuit breaker and the current maximum value I recorded in the waveform when the alternating current filter opens and closesmaxAnd minimum value of current IminAnd three-phase matching degree gamma, frequency domain characteristic quantity including zero sequence current harmonic maximum value IH,maxHarmonic content delta and total harmonic distortion rate K of zero sequence current of alternating current filter circuit breakerTHD；
S12, for the initial training data set SM×8The seven characteristic quantities in the sequence are normalized to form a standard training data set S'M×8；
S13, determining the number z of hidden layers of the RBF neural network and the development coefficient C of the network receiving algorithm, and determining the node number A of the network receiving algorithm according to the number z of the hidden layers and the characteristic quantity, wherein:
Figure BDA0001484018260000034
A＝2(7×z+2×z)y (7)；
in formula (6), K represents the total number of iterations, and K represents the kth iteration; in the formula (7), y represents the number of nodes selected by each hyperplane;
s14, determining an objective function of the RBF neural network as the following formula (8), iteratively updating the formula of the net collecting algorithm as the following formula (9), and selecting a standard training data set S'M×8The RBF neural network is subjected to parameter identification through H data by a net collecting algorithm with preset development coefficients and iteration times to obtain a center T of the RBF neural networkiWeight WiAnd width σi；
Figure BDA0001484018260000041
Figure BDA0001484018260000042
In formula (8), y (j) represents an actual output value of the jth iteration, and S' (j, m) represents an input value of the jth iteration; in formula (9), X(k)Indicates the node position, X, at the k-th iteration(k+1)Represents the node position, X, at the k +1 th iterationbestIndicating the optimal node location, ξ, occurring in the search history(k)And ζ(k)Respectively represent [0, 1]A random number in between;
s15, using the state value in the H pieces of data as reference, determining the center T using the step S14iWeight WiAnd width σiThe accuracy of the RBF neural network is calculated for 100 values by stepping 0.01 from 0 to 1 each time, and the value with the highest accuracy is selected as a standard threshold.
Optionally, the step S4, when diagnosing whether the ac filter is faulty according to the output result of the trained RBF neural network, includes: when the output result of the trained RBF neural network is greater than a standard threshold value, determining that the alternating current filter has a fault; and when the output result of the trained RBF neural network is not greater than the standard threshold, determining that the alternating current filter has no fault.
The invention has the beneficial effects that:
the RBF neural network is trained by adopting a network receiving algorithm according to historical waveform data recorded by a fault recorder of the extra-high voltage converter station, so that when the alternating current filter is diagnosed and analyzed subsequently, the characteristic quantity in the waveform data is extracted, the characteristic quantity is input into the trained RBF neural network after normalization processing, and whether the alternating current filter has a fault or not can be determined according to the output result of the trained RBF neural network. Therefore, compared with the background art, the method and the device can diagnose and analyze whether the alternating current filter has faults or not according to the waveform data acquired by the fault recorder in time, so that unsafe hidden dangers to a power grid can be avoided; the diagnosis and analysis method is suitable for all the extra-high voltage converter stations and has a wide application range; the method can overcome the randomness of artificially discovering faults and the like.
Drawings
FIG. 1 is a flow chart of the present invention.
FIG. 2 is a schematic diagram of the RBF neural network according to the present invention.
Fig. 3 is a schematic diagram of a convergence curve of a network receiving algorithm when parameter identification is performed on the RBF neural network.
Fig. 4 is a diagram illustrating the result of verifying the trained RBF neural network.
Detailed Description
The present invention will be described in further detail with reference to the accompanying drawings and examples.
As shown in fig. 1, the method for diagnosing and analyzing the online fault of the ac filter of the extra-high voltage converter station in the embodiment includes the following steps:
s1, training an RBF neural network by adopting a net receiving algorithm according to historical waveform data recorded by a fault recorder of the extra-high voltage converter station to determine model parameters of the RBF neural network, wherein the model parameters comprise the center, weight and width of the RBF neural network;
s2, acquiring current waveform data of the extra-high voltage converter station according to data recorded by a fault recorder of the extra-high voltage converter station, and extracting characteristic quantities of the current waveform data, wherein the extracted characteristic quantities comprise time domain characteristic quantities and frequency domain characteristic quantities;
s3, normalizing the extracted characteristic quantity;
and S4, inputting the characteristic quantity after the normalization processing into the trained RBF neural network, and diagnosing whether the alternating current filter has faults or not according to the output result of the trained RBF neural network.
Optionally, the time-domain characteristic quantity includes an opening and closing time T of the circuit breaker, and a current maximum value I recorded in a waveform when the alternating-current filter opens and closesmaxAnd minimum value of current IminAnd a three-phase matching degree gamma; wherein:
T＝max(TA,TB,TC) (1)
in the formula (1), TA、TBAnd TCThe switching-on and switching-off time of three phases of the alternating current filter A, B, C is respectively;
Figure BDA0001484018260000051
in the formula (2), ia、ibAnd icRespectively representing the instantaneous values of the three-phase-shifted currents, I, of the AC filter A, B, Ca、IbAnd IcRespectively, the current amplitudes of the three phases of the ac filter A, B, C, and d represents the waveform of the d-th cycle in the waveform data.
Optionally, the frequency domain characteristic quantity comprises a zero sequence current harmonic maximum value IH,maxHarmonic content delta and total harmonic distortion rate K of zero sequence current of alternating current filter circuit breakerTHD(ii) a Wherein:
Figure BDA0001484018260000061
in equation (3): i isH,eIs the effective value of the e-harmonic wave, IrIs the rated current of the corresponding alternating current filter bank;
Figure BDA0001484018260000062
Figure BDA0001484018260000063
in the formula (5), IH,nRepresenting the effective value of the nth harmonic.
Optionally, the step S1, when determining the model parameters of the RBF neural network, includes:
s11, extracting characteristic quantity of M groups of historical waveform data recorded by a fault recorder when the AC filter of the ultra-high voltage converter station is switched to obtain seven characteristic quantities including time domain characteristic quantity and frequency domain characteristic quantity, and marking abnormal states of the M groups of historical waveform data to obtain the characteristic quantityCombining seven characteristic quantities and one state value into an initial training data set SM×8Wherein, the time domain characteristic quantity comprises the opening and closing time T of the circuit breaker and the current maximum value I recorded in the waveform when the alternating current filter opens and closesmaxAnd minimum value of current IminAnd three-phase matching degree gamma, frequency domain characteristic quantity including zero sequence current harmonic maximum value IH,maxHarmonic content delta and total harmonic distortion rate K of zero sequence current of alternating current filter circuit breakerTHD；
S12, for the initial training data set SM×8The seven characteristic quantities in the sequence are normalized to form a standard training data set S'M×8；
S13, determining the number z of hidden layers of the RBF neural network and the development coefficient C of the network receiving algorithm, and determining the node number A of the network receiving algorithm according to the number z of the hidden layers and the characteristic quantity, wherein:
Figure BDA0001484018260000064
A＝2(7×z+2×z)y (7)；
in formula (6), K represents the total number of iterations, and K represents the kth iteration; in the formula (7), y represents the number of nodes selected by each hyperplane;
s14, determining an objective function of the RBF neural network as the following formula (8), iteratively updating the formula of the net collecting algorithm as the following formula (9), and selecting a standard training data set S'M×8The RBF neural network is subjected to parameter identification through H data by a net collecting algorithm with preset development coefficients and iteration times to obtain a center T of the RBF neural networkiWeight WiAnd width σi；
Figure BDA0001484018260000071
Figure BDA0001484018260000072
Figure BDA0001484018260000073
In formula (8), y (j) represents an actual output value of the jth iteration, and S' (j, m) represents an input value of the jth iteration; in formula (9), X(k)Indicates the node position, X, at the k-th iteration(k+1)Represents the node position, X, at the k +1 th iterationbestIndicating the optimal node location, ξ, occurring in the search history(k)And ζ(k)Respectively represent [0, 1]A random number in between;
s15, using the state value in the H pieces of data as reference, determining the center T using the step S14iWeight WiAnd width σiThe accuracy of the RBF neural network is calculated for 100 values by stepping 0.01 from 0 to 1 each time, and the value with the highest accuracy is selected as a standard threshold.
Optionally, when the step S4 diagnoses whether the ac filter has a fault according to the output result of the trained RBF neural network, the specific manner is: when the output result of the trained RBF neural network is greater than a standard threshold value, determining that the alternating current filter has a fault; and when the output result of the trained RBF neural network is not greater than the standard threshold, determining that the alternating current filter has no fault.
In order to facilitate understanding of the online fault diagnosis and analysis method for the alternating current filter of the extra-high voltage converter station, a training process and a subsequent diagnosis and analysis process of the online fault diagnosis and analysis method for the alternating current filter of the extra-high voltage converter station are described in detail below with reference to an example.
For example, the on-line fault diagnosis and analysis method for the alternating current filter of the extra-high voltage converter station provided by the embodiment is applied to fault early warning of the alternating current filter of a converter station in northwest region. The alternating current filters of the converter station are divided into 4 groups, 16 groups are provided, each group has a capacitive reactive capacity of 295Mvar, and the converter station comprises: 4 groups of BP11/BP13 alternating current filters, 4 groups of HP24/36 alternating current filters, 3 groups of HP3 alternating current filters and 5 groups of SC parallel capacitor banks.
Specifically, for the extra-high voltage converter station, in order to perform fault diagnosis and analysis on the ac filter of the extra-high voltage converter station, an RBF neural network for the extra-high voltage converter station needs to be established first, and the process is specifically realized by the following steps 1 to 5:
1. establishing an initial training data set S of a RBF neural networkM×8。
Specifically, extracting characteristic quantity of 600(M) groups of historical waveform data recorded during switching of the alternating-current filter of the extra-high voltage converter station, extracting seven characteristic quantities including time domain characteristic quantity and frequency domain characteristic quantity, marking abnormal states of the 600 groups of historical waveform data to obtain state values, and combining the seven characteristic quantities and one state value into an initial training data set S600×8. Wherein an initial training data set S600×8Part of the data in (1) is shown in table one. The time domain characteristic quantity comprises the opening and closing time T of the circuit breaker and the current maximum value I recorded in the waveform when the alternating current filter opens and closesmaxAnd minimum value of current IminAnd three-phase matching degree gamma, frequency domain characteristic quantity including zero sequence current harmonic maximum value IH,maxHarmonic content delta and total harmonic distortion rate K of zero sequence current of alternating current filter circuit breakerTHD. In table one, S is "0" to indicate that the ac filter is not in failure, and S is "1" to indicate that the ac filter is not in failure.
Watch 1
Figure BDA0001484018260000081
2. Standard training data set S 'for establishing RBF neural network'M×8。
In order to determine the upper and lower limits of the parameters of the net receiving algorithm in the optimization process and avoid the input quantity of the RBF neural network from exceeding the limits of the hidden layer and the output layer, the embodiment applies the initial training data set SM×8All the feature quantities in (1) are subjected to normalization processing. The specific implementation method of the normalization processing comprises the following steps: initial training data set S600×8Dividing the data of each column by the maximum value of the corresponding column to limit all the characteristic quantities to [0, 1]]Form a standard training data set S'600×8。
3. Determining the number z of hidden layers of the RBF neural network and the development coefficient C of the network receiving algorithm, and determining the node number A of the network receiving algorithm according to the number z of the hidden layers and the characteristic quantity.
As shown in fig. 2, the RBF neural network includes an input layer, a hidden layer, and an output layer. Input layer X ═ X1,X2,...,Xn]There are n inputs (7 in this embodiment); hidden layer total z nodes, hidden center Ti＝[Ti1,,Ti2,,...,Tin,]Each center width is σi，i∈[1,z](ii) a The output is Y. "G" represents a radial basis function, and a commonly used Gaussian function can be selected; "Wi"represents the weight between the ith hidden unit and the output, and the relationship between the actual output Y and the input X is shown in the following formula (10).
Figure BDA0001484018260000091
Since the number z of hidden layers of the RBF neural network determines the number of centers thereof, the larger this number is, the more complex the neural network model is, the more precisely the quantity to be represented can be represented expropriately, however, the more complex the operation of training the RBF neural network is. In this example, assume that the RBF neural network includes five hidden layers. Generally, when the RBF neural network includes five hidden layers (i.e., z is 5), the fault diagnosis analysis of the ac filter can be completed.
The net collecting algorithm is a group intelligent optimization algorithm which utilizes the idea of fishermen who net-casting and fish catching to carry out optimization solution, and has the advantages of strong global search capability, few adjustable parameters, fast convergence and the like. The fishing net in the net collecting algorithm is compiled by information shared by all nodes and the nodes, the target to be captured is wrapped in the fishing net, the search space is reduced through one-time collecting operation, the distance between each node on the fishing net and the captured target is shortened until the whole fishing net is gathered, and the captured target is obtained. The nodes on the network surface represent a solution of the optimization problem, each node obtains respective quality degree through calculation of an objective function, shared information between the nodes forms a whole network, the process of closing each node to the historical optimal node is called closing, and the position updating formula of each node in the embodiment is as shown in the formula (9).
The net collecting algorithm is to start searching from the boundary of the search space, and when initializing, for the x-dimensional search space, the whole net is 2 consisting of upper and lower limits of each variablexThe vertex of (2) and a plurality of nodes on 2x hyperplanes are formed together, and the node on each hyperplane meets the requirement that a certain one-dimensional variable is in an upper limit or a lower limit. If each hyperplane randomly selects y nodes, the whole net has 2x+2xy nodes. For an RBF neural network with n-dimensional input z hidden layers, the parameter to be identified has a center Ti(total of zn variables), width σi(z variables in total), weight Wi(z variables in total), for a total of x z n + z + z variables. And when the value of x is large, reserving 2xy nodes on the hyperplane for initialization.
In this example, the parameters that the RBF neural network needs to recognize include 7 eigenvalues of 5 hidden layers, 5 widths σi5 weights WiThe total number of the parameters is 45, so that the searching space required by the net collecting algorithm is extremely large, in this example, 2 nodes are selected in each hyperplane of the net collecting algorithm, and therefore, the whole net surface of the net collecting algorithm is composed of 180 nodes.
The development coefficient C determines the global optimization capability and the optimization accuracy of the net-receiving algorithm, the global optimization capability is reflected in the early iteration process, and the search accuracy is reflected in the later iteration process. In order to enable the net collecting algorithm to have strong searching capability and higher searching precision, in the embodiment, the exploitation coefficient C is set to be a variable which is reduced along with the increase of the furling times k, the exploitation coefficient C is larger than 1 in the early stage of furling so as to ensure the global searching performance of the net collecting algorithm, and the exploitation coefficient C is set to be between [0, 1] in the later stage of furling so as to improve the searching precision of the net collecting algorithm. The specific development coefficient is shown in formula (6).
4. Determining an objective function of the RBF neural network as a formula (8), iteratively updating the formula of the net collecting algorithm as a formula (9), and then selecting a standard training data set S'600×8Front of (5)500(H) pieces of data are used for parameter identification of the RBF neural network.
The closer the actual output value Y after calculation of the RBF neural network is to the actual value R, the more the RBF neural network conforms to the actual situation, so that when M pieces of data are shared in the learning library, the formula (8) is used as the criterion for judgment of the RBF neural network. Therefore, the network receiving algorithm is used for optimization, so that the target function takes the f minimum value min (f), and the center T needing to be identified can be obtainediWidth σiAnd weight Wi。
The specific parameter identification process is as follows: step 1: generating an initialization network surface according to the upper limit and the lower limit of each dimension variable; step 2: and (4) enabling each parameter information contained in each node to be matched with a standard training data set S'600×8Calculating an actual output value using formula (8); and step 3: calculating the evaluation value min (f) of each node by using the actual output value and the real value obtained by RBF network calculation through an equation (8); and 4, step 4: updating the historical optimal nodes, and updating the positions of all nodes according to a formula (9); and 5: returning to the step 2, preparing for the next round of folding the net surface until all the nodes are gathered together; step 6: and finishing the calculation and outputting each parameter of the RBF neural network.
In this example, the maximum iteration number of the network receiving algorithm is set to 100, the RBF neural network is subjected to parameter identification, the convergence curve of the network receiving algorithm is shown in fig. 3, and the calculated model parameters of the RBF neural network are shown in table two.
Watch two
(a) Width parameter
σ1 σ2 σ3 σ4 σ5
0.386 0.421 0.415 0.246 0.315
(b) Weight parameter
W1 W3 W4 W5
0.543 0.681 0.409 0.587
(c) Center parameter
T’ I’max I’min γ’ I’H，max δ' K’THD
T1 0.66 0.77 0.50 0.32 0.62 0.65 0.64
T2 0.52 0.57 0.43 0.44 0.41 0.53 0.53
T3 0.69 0.81 0.71 0.27 0.56 0.75 0.65
T4 0.58 0.53 0.61 0.68 0.57 0.55 0.59
T5 0.59 0.72 0.77 0.43 0.70 0.56 0.64
5. A standard threshold is determined.
For the selection of the abnormal output state judgment standard threshold, this embodiment determines by an enumeration method, steps from 0 to 1 by 0.01, calculates the accuracy for 100 values between 0 and 1, and selects the value with the highest accuracy as the standard threshold. In combination with the above values, the standard threshold value is finally determined to be 0.31 in this example. Under the standard threshold, the accuracy of judging the abnormal state of the training data reaches 98 percent.
To this end, the RBF neural network has been trained. To verify the accuracy of the trained RBF in the analysis and diagnosis of AC filter faults, the trained RBF neural network may be used to calculate S'600×8The final calculation result obtained from the last 100 groups of test data in the process is shown in fig. 4, and the data accuracy of the abnormal state judgment of the alternating current filter group reaches 98%. As can be seen in fig. 4: the RBF neural network is used for diagnosing and analyzing the internal abnormality of the alternating current filter, the accuracy rate is high, only 1 time of tests with abnormality is found, and 1 time of tests without abnormality but the alarm is output (points in circles in fig. 4), and the rest tests are correct, and the accuracy rate reaches 98%.
In conclusion, the method provided by the embodiment can be used for diagnosing and analyzing the fault of the alternating current filter of the extra-high voltage converter station. Specifically, during analysis and diagnosis, when a waveform file of the converter station alternating current filter generated in real time is detected, the waveform data features in the waveform file are extracted and normalized, the set of feature values are used for calculating an abnormal state value by using an RBF neural network, and when the value exceeds a standard threshold value, the alternating current filter can be determined to be in fault. At the moment, an alarm can be sent out, so that the fault of the alternating current filter of the extra-high voltage converter station can be found in time.

Claims (4)

1. An on-line fault diagnosis and analysis method for an alternating current filter of an extra-high voltage converter station is characterized by comprising the following steps:
s1, training an RBF neural network by adopting a net receiving algorithm according to historical waveform data recorded by a fault recorder of the extra-high voltage converter station to determine model parameters of the RBF neural network, wherein the model parameters comprise the center, weight and width of the RBF neural network;
s2, acquiring current waveform data of the extra-high voltage converter station according to data recorded by a fault recorder of the extra-high voltage converter station, and extracting characteristic quantities of the current waveform data, wherein the extracted characteristic quantities comprise time domain characteristic quantities and frequency domain characteristic quantities;
s3, normalizing the extracted characteristic quantity;
s4, inputting the characteristic quantity after normalization processing into the trained RBF neural network, and diagnosing whether the alternating current filter has faults or not according to the output result of the trained RBF neural network;
the time domain characteristic quantity comprises the opening and closing time T of the circuit breaker and the current maximum value I recorded in the waveform when the alternating current filter opens and closesmaxAnd minimum value of current IminAnd a three-phase matching degree gamma; wherein:
T＝max(TA,TB,TC) (1)
in the formula (1), TA、TBAnd TCThe switching-on and switching-off time of three phases of the alternating current filter A, B, C is respectively;
Figure FDA0002982298600000011
in the formula (2), ia、ibAnd icRespectively representing the instantaneous values of the three-phase-shifted currents, I, of the AC filter A, B, Ca、IbAnd IcRespectively, the current amplitudes of the three phases of the ac filter A, B, C, and d represents the waveform of the d-th cycle in the waveform data.
2. The on-line fault diagnosis and analysis method for the alternating-current filter of the extra-high voltage converter station according to claim 1, wherein the frequency domain characteristic quantity comprises a zero-sequence current harmonic maximum value IH,maxHarmonic content delta and total harmonic distortion rate K of zero sequence current of alternating current filter circuit breakerTHD(ii) a Wherein:
Figure FDA0002982298600000021
in equation (3): i isH,eIs the effective value of the e-harmonic wave, IrIs the rated current of the corresponding alternating current filter bank;
Figure FDA0002982298600000022
Figure FDA0002982298600000023
3. the method for analyzing the on-line fault diagnosis of the alternating current filter of the extra-high voltage converter station as claimed in claim 1, wherein the step S1 when determining the model parameters of the RBF neural network comprises:
s11, extracting characteristic quantity of M groups of historical waveform data recorded by a fault recorder when an alternating current filter of the extra-high voltage converter station is switched to obtain seven characteristic quantities including time domain characteristic quantity and frequency domain characteristic quantity, marking abnormal states of the M groups of historical waveform data to obtain state values, and combining the seven characteristic quantities and one state value into an initial training data set SM×8Wherein, the time domain characteristic quantity comprises the opening and closing time T of the circuit breaker and the current maximum value I recorded in the waveform when the alternating current filter opens and closesmaxAnd minimum value of current IminAnd three-phase matching degree gamma, frequency domain characteristic quantity including zero sequence current harmonic maximum value IH,maxHarmonic content delta and total harmonic distortion rate K of zero sequence current of alternating current filter circuit breakerTHD；
S12, for the initial training data set SM×8The seven characteristic quantities in the sequence are normalized to form a standard training data set S'M×8；
S13, determining the number z of hidden layers of the RBF neural network and the development coefficient C of the network receiving algorithm, and determining the node number A of the network receiving algorithm according to the number z of the hidden layers and the characteristic quantity, wherein:
Figure FDA0002982298600000024
A＝2(7×z+2×z)y (7)；
in formula (6), K represents the total number of iterations, and K represents the kth iteration; in the formula (7), y represents the number of nodes selected by each hyperplane;
s14, determining the target function of the RBF neural network as the following formula (8), iteratively updating the formula of the net receiving algorithm as the following formula (9), and selecting the standard trainingRefining data set S'M×8The RBF neural network is subjected to parameter identification through H data by a net collecting algorithm with preset development coefficients and iteration times to obtain a center T of the RBF neural networkiWeight WiAnd width σi；
Figure FDA0002982298600000031
Figure FDA0002982298600000032
Figure FDA0002982298600000033
In formula (8), y (j) represents an actual output value of the jth iteration, and S' (j, m) represents an input value of the jth iteration; in formula (9), X(k)Indicates the node position, X, at the k-th iteration(k+1)Represents the node position, X, at the k +1 th iterationbestIndicating the optimal node location, ξ, occurring in the search history(k)And ζ(k)Respectively represent [0, 1]A random number in between;
s15, using the state value in the H pieces of data as reference, determining the center T using the step S14iWeight WiAnd width σiThe accuracy of the RBF neural network is calculated for 100 values by stepping 0.01 from 0 to 1 each time, and the value with the highest accuracy is selected as a standard threshold.
4. The on-line fault diagnosis and analysis method for the extra-high voltage converter station alternating current filter according to claim 3, wherein the step S4 includes the following steps when diagnosing whether the alternating current filter has a fault according to the output result of the trained RBF neural network:
when the output result of the trained RBF neural network is greater than a standard threshold value, determining that the alternating current filter has a fault; and when the output result of the trained RBF neural network is not greater than the standard threshold, determining that the alternating current filter has no fault.
CN201711207686.1A 2017-11-27 2017-11-27 Online fault diagnosis and analysis method for alternating current filter of extra-high voltage converter station Active CN107992665B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201711207686.1A CN107992665B (en) 2017-11-27 2017-11-27 Online fault diagnosis and analysis method for alternating current filter of extra-high voltage converter station

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201711207686.1A CN107992665B (en) 2017-11-27 2017-11-27 Online fault diagnosis and analysis method for alternating current filter of extra-high voltage converter station

Publications (2)

Publication Number Publication Date
CN107992665A CN107992665A (en) 2018-05-04
CN107992665B true CN107992665B (en) 2021-05-07

Family

ID=62032104

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201711207686.1A Active CN107992665B (en) 2017-11-27 2017-11-27 Online fault diagnosis and analysis method for alternating current filter of extra-high voltage converter station

Country Status (1)

Country Link
CN (1) CN107992665B (en)

Families Citing this family (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109358255B (en) * 2018-12-18 2024-03-05 北京交通大学 Traction converter fault analysis method and device
CN111076927A (en) * 2019-12-19 2020-04-28 佛山科学技术学院 Gear fault diagnosis method, device and system based on deep neural network
CN111551819B (en) * 2020-04-16 2022-04-29 国网湖南省电力有限公司 Micro-grid fault detection method and device and storage medium
CN113128111A (en) * 2021-04-12 2021-07-16 中国南方电网有限责任公司超高压输电公司昆明局 Diagnostic method and diagnostic device for operating parameters in ultra-high voltage circuit
CN113360555A (en) * 2021-04-30 2021-09-07 国网浙江省电力有限公司丽水供电公司 Fault diagnosis and analysis method and system based on big data of power grid secondary equipment
CN113315099B (en) * 2021-05-20 2022-10-21 国网宁夏电力有限公司检修公司 Fault early warning method, medium and system for converter station alternating current filter circuit breaker
CN113702780B (en) * 2021-08-20 2024-01-19 中国南方电网有限责任公司超高压输电公司大理局 BP neural network-based high-voltage power supply online monitoring method and device
CN113792636B (en) * 2021-09-07 2024-05-14 国网安徽省电力有限公司 Extra-high voltage direct current transmission converter transformer recessive fault prediction method based on fault wave recording

Family Cites Families (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101975910B (en) * 2010-09-07 2012-07-04 昆明理工大学 Intelligent fault classification and location method for ultra-high voltage direct current transmission line
CN103048041B (en) * 2012-12-20 2014-11-26 北京信息科技大学 Fault diagnosis method of electromechanical system based on local tangent space and support vector machine
US9461535B2 (en) * 2013-12-30 2016-10-04 King Fahd University Of Petroleum And Minerals Photovoltaic systems with maximum power point tracking controller
CN106199412B (en) * 2016-07-01 2019-01-29 太原理工大学 A kind of permanent magnet mechanism high-pressure vacuum breaker method of fault pattern recognition
CN106707153A (en) * 2016-12-27 2017-05-24 北京合纵科技股份有限公司 FOA-RBF based high-voltage circuit breaker fault diagnosis method
CN107132443B (en) * 2017-03-01 2019-07-12 大连理工大学 A kind of three-level STATCOM intelligent failure diagnosis method
CN107909035B (en) * 2017-11-16 2021-05-28 国家电网公司 Quasi-real-time fault recording file reading and analyzing system

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
基于小波包熵和概率神经网络的微特电机轴承故障监控;李珂; 邰能灵; 张沈习;《微特电机》;20160430;第44卷(第4期);第37-39页 *

Also Published As

Publication number Publication date
CN107992665A (en) 2018-05-04

Similar Documents

Publication Publication Date Title
CN107992665B (en) Online fault diagnosis and analysis method for alternating current filter of extra-high voltage converter station
CN108681633B (en) Condensate pump fault early warning method based on state parameters
Nurmanova et al. A new transformer FRA measurement technique to reach smart interpretation for inter-disk faults
CN106443316B (en) Multi-information detection method and device for deformation state of power transformer winding
CN113159077B (en) Three-phase inverter fault identification method based on hybrid convolutional neural network
CN106841949B (en) Method and device for monitoring stator insulation of three-phase asynchronous alternating current motor on line
CN111797365B (en) Converter transformer temperature abnormity judgment method and system
CN109443717B (en) On-load tap-changer mechanical fault on-line monitoring method
CN109376778B (en) Fault classification diagnosis method based on characteristic variable weighting
CN115187832A (en) Energy system fault diagnosis method based on deep learning and gram angular field image
CN111160241A (en) Power distribution network fault classification method, system and medium based on deep learning
CN107909035B (en) Quasi-real-time fault recording file reading and analyzing system
CN113358993A (en) Online fault diagnosis method and system for multi-level converter IGBT
CN115980513A (en) Power distribution network single-phase earth fault distinguishing method and device based on particle swarm optimization
CN114358116A (en) Oil-immersed transformer fault diagnosis method and system and readable storage medium
CN114330486A (en) Power system bad data identification method based on improved Wasserstein GAN
CN117630800A (en) Fault diagnosis method and system for automatic calibrating device of electric energy meter
CN116699400A (en) Generator rotor short-circuit fault monitoring system, method and readable storage medium
CN109389313B (en) Fault classification diagnosis method based on weighted neighbor decision
CN116975389A (en) Transformer substation fault analysis method and device
CN116990633A (en) Fault studying and judging method based on multiple characteristic quantities
CN116256402A (en) Valve side sleeve defect detection method based on frequency domain dielectric spectrum data
CN115130550A (en) Distribution transformer fault identification method based on gradient lifting decision tree
CN114091593A (en) Network-level arc fault diagnosis method based on multi-scale feature fusion
CN110135021B (en) ATRU system fault grading diagnosis method based on multi-source signals and RBF neural network

Legal Events

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