阅读说明：本技术 一种基于narnn模型预测行波波头到达时刻的故障定位方法 (A kind of Fault Locating Method based on NARNN model prediction wavefront arrival time ) 是由 刘波 王丹 梁睿 滕松 郭振华 于 2019-10-12 设计创作，主要内容包括：本发明公开了一种基于NARNN模型预测行波波头到达时刻的故障定位方法,方法是,首先,利用导数法对所获得的初始电压行波信号进行处理,初步确定行波波头到达时刻。然后,选取行波波头到达时刻之前的一段数据作为训练集,利用NARNN模型对行波波形进行预测。最后,将NARNN模型预测的波形与实际波形作差,利用KPSS对所得差值进行校验,得到波头到达的精确时刻,并利用双端测距法进行故障定位。本发明避免了高频噪声对故障精确定位的影响,且不受故障电阻与故障初相角的影响,适用于含有高频噪声的电网故障定位,具有较高的精度和可靠性。(The invention discloses a kind of Fault Locating Method based on NARNN model prediction wavefront arrival time, method is, firstly, handling using derivative method initial voltage traveling wave signal obtained, to primarily determine wavefront arrival time.Then, the one piece of data before selection wavefront arrival time predicts traveling-wave waveform using NARNN model as training set.Finally, the waveform of NARNN model prediction and actual waveform work is poor, gained difference is verified using KPSS, obtains the precise moments of wave head arrival, and double-end distance measurement method is utilized to carry out fault location.The invention avoids high-frequency noises to be influenced on the pinpoint influence of failure, and not by fault resstance and failure initial phase angle, suitable for the electric network fault positioning containing high-frequency noise, precision and reliability with higher.)

1. a kind of Fault Locating Method based on NARNN model prediction wavefront arrival time, characterized in that including walking as follows It is rapid:

Measurement point is arranged in transmission line of electricity head and end in step 1), is indicated respectively with m and n, indicates fault point with f；Transmission line of electricity After breaking down, the initial voltage traveling wave signal of head end m point and end n point is extracted respectively, phase-model transformation then is carried out to it, is obtained Initial line mode voltage travelling wave signal is obtained, is denoted as respectivelyWithWherein subscript indicates that m point and n point, subscript (1) indicate line Mould；

It is step 2), right respectivelyWithWavefront arrival time, the head of note derivative method judgement are tentatively judged with derivative method End wavefront reaches time point for t '_m, wavefront arrival time point in end is t '_n；

Step 3), the initial wave head arrival time point t ' obtained according to step (2)_m、t′_n, respectively fromWithMiddle selection t '_m With t '_nPreceding length is the sequence of aWith

Step 4), withWithFor training set, establishes NARNN model and traveling-wave waveform is predicted, obtain forecasting sequence U_narm1And U_narn1；

Step 5), the difference for calculating forecasting sequence and corresponding actual sequence:

Step 6) examines Δ U using KPSS_narm1With Δ U_narm1To obtain accurate wavefront arrival time t_mAnd t_n；

Step 7), the wavefront arrival time t obtained according to step 6)_mAnd t_n, fault localization, which is carried out, using double-end distance measurement method obtains To accurate abort situation.

2. a kind of Fault Locating Method based on NARNN model prediction wavefront arrival time according to claim 1, It is characterized in that, identifying that the calculation formula of traveling wave Mintrop wave head is as follows by derivative method in the step (2):

Wherein, Y (i), Y (i-1) are traveling wave sampled value；Ts is the sampling period；K is the threshold value of setting, K=1000.

3. a kind of Fault Locating Method based on NARNN model prediction wavefront arrival time according to claim 1, It is characterized in that, choosing the sequence that length is a in the step (3), the value of a is 280, which eliminates wavefront arrival 70 sampled points before moment, taking preceding 71~350 sampled points is training sequence, is to guarantee not including row in high-frequency noise Wave wave head data.

4. a kind of Fault Locating Method based on NARNN model prediction wavefront arrival time according to claim 1, It is characterized in that, establishing the specific steps of nonlinear auto-companding neural network in the step (4) are as follows:

(a) structure of NARNN model is determined；

(b) parameter of NARNN model is determined；

(c) optimization of NARNN model parameter.

5. a kind of Fault Locating Method based on NARNN model prediction wavefront arrival time according to claim 4, It is characterized in that, step (a) specifically:

Nonlinear auto-companding neural network is to be trained and in advance based on non linear autoregressive model using neural network It surveys；

The time delay of output signal is fed back as time delay in the network in NARNN model because it be based on itself into Row returns, then by the time-delay signal of output as the output of network, acquires network using hidden layer and output layer calculating Output, and recycle the process.

6. a kind of Fault Locating Method based on NARNN model prediction wavefront arrival time according to claim 4, It is characterized in that, step (b) specifically:

NARNN model is constructed, neural network parameter P, Q, E are given；Wherein, P is the autoregression number of plies, and Q is to hide the number of plies, and E is repeatedly Generation number；

Only include the traveling wave data of HF noise signal to use NARNN model to obtain, adds two special parameters: when input Between sequence length I and input time series position L；

The length of time series I of input is the length for importing the travelling wave signal in NARNN model before wave head arrival, and time series is long Degree I needs the high-frequency signal comprising sufficient length；Input time sequence location L is the traveling wave imported in NARNN before wave head arrival The position of signal, the time series position L of input cannot include the data of wavefront；

P, the value of Q is respectively 10,10, it is contemplated that time cost, test setting the number of iterations E is 20 times, defeated by repeatedly testing Angle of incidence sequence length I is that 280 sampled points can satisfy ranging requirement.

7. a kind of Fault Locating Method based on NARNN model prediction wavefront arrival time according to claim 4, It is characterized in that, step (c) specifically:

In the training of NARNN model, in order to measure the quality of selected neural network parameter, RMSE root mean square is selected to miss Evaluation index of the difference as NARNN model in test set；RMSE value is smaller, and representative model predicts that precision is higher.

8. a kind of Fault Locating Method based on NARNN model prediction wavefront arrival time according to claim 1, It is characterized in that, the specific steps that KPSS is verified in the step (6) are as follows:

To residual sequence Δ U_narThe KPSS that a circulation is arranged is examined:

ΔU_nar=[Δ u₁,Δu₂,…,Δu₁₀]

Since KPSS checking sequence shortest length is 10, so most short sequence is [Δ u₁,Δu₂,…,Δu₁₀], k=[1, n- 10], wherein k is integer, and according to sequentially successively Check-Out Time sequence [Δ u₁,Δu₂,…,Δu_10+k], when each checking sequence The test value of current sequence can be returned to, return value composition one contains only 0 and 1 sequence.

9. a kind of Fault Locating Method based on NARNN model prediction wavefront arrival time according to claim 8, It is characterized in that,

When the waveform of NARNN output sufficiently restores true waveform, and any trend is not included in residual sequence, only comprising white When noise, residual sequence is stationarity sequence；

When the time series position of input does not include mutation, NARNN model output signal waveform and wavefront arrival time Waveform before still conforms to the characteristic that residual error surrounds about 0 minor swing, and residual sequence is also stationary sequence at this time；

When wavefront reaches, which will deviate from 0, Δ U_narNumerical value mutates, and residual sequence is steady at this time Property change, can obtain wavefront arrival exact time.

10. a kind of fault location side based on NARNN model prediction wavefront arrival time according to claim 1 Method, which is characterized in that the formula of both-end method fault localization in the step (7) are as follows:

If the initial travelling wave signal of failure is t with the time that spread speed v reaches route both ends m, n_mAnd t_n, using double-end distance measurement method Formula are as follows:

Wherein, d_mAnd d_nIt is distance between fault point and head end m, end n respectively；L is the length of transmission line of electricity.

Both-end distance measuring formula model is solved, is obtained:

It is d that fault point, which is finally obtained, at a distance from head end_m。

Technical field

The invention belongs to electric network fault field of locating technology, in particular to are arrived based on NARNN model prediction type wave wave head Up to the Fault Locating Method at moment.

Background technique

With the propulsion of intelligent power distribution network construction, fault location technology is for searching rapidly failure, when fast recovery of power supply Between, reduce each side's economic loss important in inhibiting.The topology of power distribution network is more complicated, and line connection is also a variety of more Sample, HF noise signal are largely present in various power equipments, these factors bring many challenges to failure accurate positionin.

It is respectively surveyed currently, the traveling wave method of electric network fault positioning reaches power grid generally by detection initial transient traveling wave Mintrop wave head At the time of amount point, fault location is carried out using the internal logical relationship between these moment and fault distance, relative to impedance Method fault location has the advantages that precision is high, is not influenced by factors such as electric arcs, therefore traveling wave method fault location technology application is more next It is wider.But due to containing a large amount of high-frequency noise in most of travelling wave signal, this directly affects wavefront arrival time Identification and determination, and one of the major influence factors of levels of precision of fault location based on both-end method are exactly the accurate of wave head Identification.

Summary of the invention

In order to solve the technical issues of above-mentioned background technique proposes, the present invention is intended to provide (non-linear to return certainly based on NARNN Return neural network) Fault Locating Method of model prediction wavefront arrival time, overcome traditional traveling wave method wave head identification inaccurate Really, by noise jamming the defects of.

To achieve the above object, the technical solution adopted by the present invention are as follows:

One kind being based on the fault location of NARNN (nonlinear auto-companding neural network) model prediction wavefront arrival time Method comprising following steps:

(1) measurement point is set in transmission line of electricity head and end, is indicated respectively with m and n, fault point is indicated with f.Transmission line of electricity After breaking down, the primary voltage travelling wave signal of head end m point and end n point is extracted respectively, phase-model transformation then is carried out to it, is obtained Original line mode voltage travelling wave signal is obtained, is denoted as respectivelyWithWherein subscript indicates that m point and n point, subscript (1) indicate line Mould；

(2) right respectivelyWithTraveling wave Mintrop wave head arrival time point, derivative method identifying rows are substantially judged with derivative method The calculation formula of wave Mintrop wave head is as follows:

Wherein, Y (i), Y (i-1) are traveling wave sampled value；Ts is the sampling period；K is the threshold value of setting, K=1000.Note is led The head end traveling wave Mintrop wave head that number method judges reaches time point as t '_m, end traveling wave Mintrop wave head arrival time point is t '_n；

(3) the initial wave head arrival time point t ' obtained according to (2)_m、t′_n, respectively fromWithMiddle selection t '_mWith t '_n Preceding length is the sequence of aWithThe value of a is 280, which eliminates 70 before wavefront arrival time Sampled point, taking preceding 71~350 sampled points is training sequence, is to guarantee not including wavefront data in high-frequency noise.

(4) withWithFor training set, establishes NARNN model and traveling-wave waveform is predicted, obtain forecasting sequence U_nam1And U_narn1.The specific steps of nonlinear auto-companding neural network are as follows:

(a) structure of NARNN model is determined

Nonlinear auto-companding neural network be based on non linear autoregressive model, be trained using neural network and Prediction.The time delay of output signal is fed back as time delay in the network in NARNN model because it be based on itself into Row returns, then by the time-delay signal of output as the output of network, acquires network using hidden layer and output layer calculating Output, and recycle the process.The present invention obtains the rule of traveling wave high-frequency noises using nonlinear auto-companding neural metwork training Rule and feature, and by neural network test come the high-frequency noise in traveling-wave waveform before wavefront.

(b) parameter of NARNN model is determined

NARNN model is constructed, neural network parameter P, Q, E are given；Wherein, P is the autoregression number of plies, and Q is to hide the number of plies, E For the number of iterations.Only include the traveling wave data of HF noise signal to use NARNN model to obtain, adds two special ginsengs Number: the time series position L of input time sequence length I and input.The length of time series I of input is to import NARNN model The length of travelling wave signal before middle wave head arrival, length of time series I need the high-frequency signal comprising sufficient length；Input time Sequence location L is the position for importing the travelling wave signal in NARNN before wave head arrival, and the time series position L of input cannot include The data of wavefront.P, the value of Q is respectively 10,10, it is contemplated that time cost, test setting the number of iterations E is 20 times, is passed through Repeatedly test, input time sequence length I are that 280 sampled points can satisfy ranging requirement.

(c) optimization of NARNN model parameter

In the training of NARNN model, in order to measure the quality of selected neural network parameter, RMSE has been selected herein Evaluation index of the root-mean-square error as NARNN model in test set.RMSE value is smaller, and representative model prediction precision is got over It is high.

With head end line mode voltage signalFor, after selecting optimized parameter, NARNN model output waveform obtains prediction wave Graphic data: U_nar.It seeksAnd U_narDifference DELTA U_nar。

(5) difference of forecasting sequence and corresponding actual sequence is calculated:

(6) Δ U is examined using KPSS_narm1With Δ U_narm1To obtain accurate wavefront arrival time t_mAnd t_n。KPSS The specific steps of verification are as follows:

To residual sequence Δ U_narThe KPSS that a circulation is arranged is examined:

ΔU_nar=[Δ u₁, Δ u₂..., Δ u₁₀]

Since KPSS checking sequence shortest length is 10, so most short sequence is [Δ u₁, Δ u₂..., Δ u₁₀], k= [1, n-10], wherein k is integer, and according to sequentially successively Check-Out Time sequence [Δ u₁, Δ u₂..., Δ u_10+k], it examines every time The test value of current sequence can be returned to when sequence, return value composition one contains only 0 and 1 sequence.

When NARNN output waveform sufficiently restore true waveform, and in residual sequence do not include any trend, only wrap When containing white noise, residual sequence are stationarity sequences；When the time series position of input does not include mutation, NARNN model is defeated Signal waveform and the waveform before wavefront arrival time still conform to the characteristic that residual error surrounds about 0 minor swing out, residual at this time Difference sequence is also stationary sequence.When wavefront reaches, which will deviate from 0, Δ U_narNumerical value mutates, this When residual sequence stationarity change, can obtain wavefront arrival exact time.

(7) the traveling wave Mintrop wave head t obtained according to (6)_mAnd t_n, fault localization is carried out using double-end distance measurement method and is obtained accurately Abort situation.The formula of both-end method fault localization are as follows:

If the initial travelling wave signal of failure is t with the time that spread speed v reaches route both ends m, n_mAnd t_n, surveyed using both-end Away from method formula are as follows:

Wherein, d_mAnd d_nIt is distance between fault point and head end m, end n respectively；L is the length of transmission line of electricity；

Both-end distance measuring formula model is solved, is obtained:

It is d that fault point, which is finally obtained, at a distance from head end_m。

The invention has the advantages that:

The invention avoids high-frequency noises on the pinpoint influence of failure, and not by fault resstance and failure initial phase angle It influences, suitable for the electric network fault positioning containing high-frequency noise, precision and reliability with higher.

Detailed description of the invention

Fig. 1 is flow chart of the present invention；

Fig. 2 is the topology diagram of two ends of electric transmission line ranging；

Specific embodiment

The present invention will be further explained with reference to the accompanying drawing.

The topological structure of the transmission line of electricity of one both-end distance measuring is as shown in Fig. 2, the invention proposes one kind to be based on NARNN The Fault Locating Method of (nonlinear auto-companding neural network) model prediction wavefront arrival time, overall flow such as Fig. 1 It is shown, include the following steps:

(1) measurement point is set in transmission line of electricity head and end, is indicated respectively with m and n, fault point is indicated with f.Transmission line of electricity After breaking down, the primary voltage travelling wave signal of head end m point and end n point is extracted respectively, phase-model transformation then is carried out to it, is obtained Original line mode voltage travelling wave signal is obtained, is denoted as respectivelyWithWherein subscript indicates that m point and n point, subscript (1) indicate line Mould；

(2) right respectivelyWithTraveling wave Mintrop wave head arrival time point, derivative method identifying rows are substantially judged with derivative method The calculation formula of wave Mintrop wave head is as follows:

Wherein, Y (i), Y (i-1) are traveling wave sampled value；Ts is the sampling period；K is the threshold value of setting, K=1000.Note is led The head end traveling wave Mintrop wave head that number method judges reaches time point as t '_m, end traveling wave Mintrop wave head arrival time point is t '_n；

(3) the initial wave head arrival time point t ' obtained according to (2)_m、t′_n, respectively fromWithMiddle selection t '_mWith t '_n Preceding length is the sequence of aWithThe value of a is 280, which eliminates 70 before wavefront arrival time Sampled point, taking preceding 71~350 sampled points is training sequence, is to guarantee not including wavefront data in high-frequency noise.

(4) withWithFor training set, establishes NARNN model and traveling-wave waveform is predicted, obtain forecasting sequence U_narm1And U_narn1.The specific steps of nonlinear auto-companding neural network are as follows:

(a) structure of NARNN model is determined

Nonlinear auto-companding neural network be based on non linear autoregressive model, be trained using neural network and Prediction.The time delay of output signal is fed back as time delay in the network in NARNN model because it be based on itself into Row returns, then by the time-delay signal of output as the output of network, acquires network using hidden layer and output layer calculating Output, and recycle the process.The present invention obtains the rule of traveling wave high-frequency noises using nonlinear auto-companding neural metwork training Rule and feature, and by neural network test come the high-frequency noise in traveling-wave waveform before wavefront.

(b) parameter of NARNN model is determined

NARNN model is constructed, neural network parameter P, Q, E are given；Wherein, P is the autoregression number of plies, and Q is to hide the number of plies, E For the number of iterations.Only include the traveling wave data of HF noise signal to use NARNN model to obtain, adds two special ginsengs Number: the time series position L of input time sequence length I and input.The length of time series I of input is to import NARNN model The length of travelling wave signal before middle wave head arrival, length of time series I need the high-frequency signal comprising sufficient length；Input time Sequence location L is the position for importing the travelling wave signal in NARNN before wave head arrival, and the time series position L of input cannot include The data of wavefront.P, the value of Q is respectively 10,10, it is contemplated that time cost, test setting the number of iterations E is 20 times, is passed through Repeatedly test, input time sequence length I are that 280 sampled points can satisfy ranging requirement.

(c) optimization of NARNN model parameter

In the training of NARNN model, in order to measure the quality of selected neural network parameter, RMSE has been selected herein Evaluation index of the root-mean-square error as NARNN model in test set.RMSE value is smaller, and representative model prediction precision is got over It is high.

With head end line mode voltage signalFor, after selecting optimized parameter, NARNN model output waveform obtains prediction wave Graphic data: U_nar.It seeksAnd U_narDifference DELTA U_nar。

(5) difference of forecasting sequence and corresponding actual sequence is calculated:

(6) Δ U is examined using KPSS_narm1With Δ U_narm1To obtain accurate wavefront arrival time t_mAnd t_n。KPSS The specific steps of verification are as follows:

To residual sequence Δ U_narThe KPSS that a circulation is arranged is examined:

ΔU_nar=[Δ u₁, Δ u₂..., Δ u₁₀]

Since KPSS checking sequence shortest length is 10, so most short sequence is [Δ u₁, Δ u₂..., Δ u₁₀], k= [1, n-10], wherein k is integer, and according to sequentially successively Check-Out Time sequence [Δ u₁, Δ u₂..., Δ u_10+k], it examines every time The test value of current sequence can be returned to when sequence, return value composition one contains only 0 and 1 sequence.

When NARNN output waveform sufficiently restore true waveform, and in residual sequence do not include any trend, only wrap When containing white noise, residual sequence are stationarity sequences；When the time series position of input does not include mutation, NARNN model is defeated Signal waveform and the waveform before wavefront arrival time still conform to the characteristic that residual error surrounds about 0 minor swing out, residual at this time Difference sequence is also stationary sequence.When wavefront reaches, which will deviate from 0, Δ U_narNumerical value mutates, this When residual sequence stationarity change, can obtain wavefront arrival exact time.

(7) the traveling wave Mintrop wave head t obtained according to (6)_mAnd t_n, fault localization is carried out using double-end distance measurement method and is obtained accurately Abort situation.The formula of both-end method fault localization are as follows:

If the initial travelling wave signal of failure is t with the time that spread speed v reaches route both ends m, n_mAnd t_n, surveyed using both-end Away from method formula are as follows:

Wherein, d_mAnd d_nIt is distance between fault point and head end m, end n respectively；L is the length of transmission line of electricity；

Both-end distance measuring formula model is solved, is obtained:

It is d that fault point, which is finally obtained, at a distance from head end_m。

Simulating, verifying

In order to verify reliability and validity of the invention, 500KV transmission line of electricity simulation model is built on RSCAD, it is double Hold the transmission line travelling wave propagation path of ranging as shown in Figure 2.Transmission line of electricity overall length is 37.715km, and system fundamental frequency is 60Hz, Voltage class is 500kV.Transmission line of electricity, which uses, meets actual frequency dependent character model.The model of conducting wire and lightning conducter is respectively as follows: 4 × LGJ-400/35 and GJ-80.In route head end and end, two voltage traveling wave measurement points (being expressed as m and n) are set, Sample frequency is 1MHz.Failure is with respect to range error e_rIt is defined as follows:

Wherein, d_testSurvey fault distance, d_realFor physical fault distance, l_m-nFor transmission line of electricity total length, value is 37.715km。

In order to verify the validity of proposition method in the field, this method is added to the noise of 60dB on the line, and It joined error in identification first time, point wavefront arrival time.Respectively the distance apart from head end m be 5km, 10km, Singlephase earth fault is simulated at 14km and 23km, fault resstance is 10 Ω and 200 Ω, and failure initial phase angle is 30 ° and 90 °.Table 1 is After above 2 positions are broken down, fault distance definitive result.Proposed method can for the failure of different location Fault distance is accurately calculated, is not influenced by fault resstance and failure initial phase angle.

1 fault localization result of table

To sum up, the invention proposes the Fault Locating Methods of wave head arrival time based on NARNN model a kind of.This method It is possible to prevente effectively from the influence that HF noise signal identifies wave head, and not by transition resistance, the influence of initial phase angle of failure etc., Improve the precision of fault location.

The above is only a preferred embodiment of the present invention, it should be pointed out that: for the ordinary skill people of the art For member, various improvements and modifications may be made without departing from the principle of the present invention, these improvements and modifications are also answered It is considered as protection scope of the present invention.