CN105759167A - Wavelet neural network-based distribution network single-phase short circuit line selection method - Google Patents

Abstract

The invention discloses a wavelet neural network-based distribution network single-phase short circuit line selection method, which belongs to the technical field of distribution network protection. The method comprises the following steps: 1) db4 wavelets are selected to serve as a wavelet packet basis function to decompose transient zero-sequence current in each feeder line of the distribution network, and the sampling frequency is 10KHz; 2) the modulus maximum of the transient zero-sequence current is calculated; 3) the modulus maximum obtained in the second step and the polarity of the modulus maximum are used for training the neural network; and 4) the BP neural network after the training in the third step is applied to the distribution network with single-phase short circuit, and a fault line is determined according to a different output result. The method of fault line selection by using the wavelet neural network disclosed by the invention has good reliability and practicability, a single-phase grounding fault line can be effectively eliminated, stable operation of the distribution network system is facilitated, and an important role is played in planning and applications of a distributed power supply.

Description

A kind of power distribution network single-phase short circuit selection method based on wavelet neural network

Technical field

The present invention relates to a kind of power distribution network single-phase short circuit selection method, belong to distribution protection technical field.

Background technology

Distributed power source refers generally to, for meeting power system and the specific requirement of user, be connected on the compact electrical generating systems near user side.Distributed power generation adopts clean energy resource or regenerative resource mostly, not only inexhaustible, nexhaustible, the impact of environment is also preferably minimized, thus obtains and widely popularize.Mostly concentrating on user side due to distributed power source, do not pass through upper level transformator directly to customer power supply, the installation scale of distributed power source is all smaller, builds and also disperses very much.

At present, the earthing mode of China's distribution network system is mainly neutral non-effective grounding, namely isolated neutral, through grounding through arc or high resistance ground, under this earthing mode, low-impedance short-circuit loop can not be constituted when power system is broken down, therefore earth current is only small, so system with non effectively earth ed neutral is also called small current neutral grounding system.In all faults of power distribution network, the probability that singlephase earth fault occurs is the highest, constitutes about about the 80% of total failare.

Compared with conventional electrical distribution net, after distributed power source is connected to the grid, bigger uncertain factor will be brought to electrical network.On the one hand, distributed power source has the reliability of better motility and Geng Gao, it is possible to be effectively ensured the power demands of user side, improves the quality of power supply；On the other hand, after accessing distributed power source, distribution net work structure there occurs the change of essence, single source network be progressively changed into multi-source network.When, after power distribution network single phase ground fault fault, big system can provide short circuit current to trouble point as power supply, meanwhile, trouble point also will be provided short circuit current by distributed power source, change the distribution of power distribution network short circuit currents.

In existing selection method, zero-sequence current amplitude method selects the feeder line of zero-sequence current amplitude maximum to be faulty line；Zero-sequence current real component method extracts zero-sequence current real component and the polarity of each bar feeder line respectively, and the active power of fault feeder is bigger, and direction is with to perfect feeder line contrary；Negative-sequence current method utilizes the negative-sequence current of fault feeder substantially opposite with the negative-sequence current phase place of normal feeder line, but basically identical with the phase place of faulted phase voltage selects faulty line；First half-wave route selection method utilizes transient state capacitance current contrary with polarity of voltage at initial time, and the feature that normal feeder line is all identical, construct route selection criterion；Fault signature component is carried out exponential term matching by Prony method, and then analyze its frequency spectrum, because fault signature reflects with the amplitude of voltage, frequency and phase place typically via fault current, and Prony method can therefrom extract high frequency and DC component, realizes route selection by calculating.But after addition distributed power source, faulty line all changes with the amplitude of electric current and phase angle in regular link, owing to distributed power source provides certain fault current, the zero-sequence current of faulty line is not necessarily maximum, simultaneously because the effect of arc suppression coil, its direction is likely consistent with the zero-sequence current in regular link, thus such as the earth fault detection for power of zero-sequence current amplitude method, real component method etc be not suitable for.Simultaneously because the change of phase angle, first half-wave method can be influenced by impact in actual applications, and fault occurs when phase angle is only small, there will not be first half-wave, and the interference such as passage drift, out-of-balance current also can affect the polarity of first half-wave.And Prony method is by the impact of amplitude with phase angle, it is possible to produce erroneous judgement.

For the power distribution network containing distributed power source; instantly the Main way studied is the grid-connected regulating strategy of distributed power source; and the impact that grid-connected rear distributed power source is on aspects such as distribution power flow, relay protection and the qualities of power supply, and about the power distribution network single-phase earthing fault route selection containing distributed power source research almost without.

Summary of the invention

The technical problem to be solved in the present invention is, not enough for prior art, propose a kind of based on wavelet neural network containing distributed power source power distribution network single-phase short circuit selection method, faulty line can be detected in time, being conducive to the stable operation of distribution network system, the planning simultaneously for distributed power source is also significant with application.

The present invention solves that the technical scheme that above-mentioned technical problem proposes is: a kind of electrical power distribution net single-phase short circuit selection method based on wavelet neural network, perform step as follows:

1) choose db4 small echo as wavelet packet basis functions, the transient zero-sequence current of each bar feeder line of described distribution network to be decomposed, sample frequency is 10KHz, and the resolution of described wavelet packet basis functions is 4, decomposes 16 frequency bands, the bandwidth of each frequency band is 312.5Hz

2) calculating the modulus maximum of described transient zero-sequence current, computing formula is as follows,In formula(j, k) coefficient under sub-band, total n coefficient under each sub-band for WAVELET PACKET DECOMPOSITION；

3) utilize step 2) in the polarity of the modulus maximum that obtains and described modulus maximum train described neutral net, the relatively desired output of described neutral net and actual output, to the weight of synapse in described neutral net and in described neutral net neuronic threshold value be modified；

Described neutral net is BP neutral net；

4) by step 3) trained after BP Application of Neural Network in the power distribution network that single-phase short circuit occurs, with according to output result difference determine faulty line.

The improvement of technique scheme is, described BP neutral net composition is as follows:

Input vector P_k=(a₁,a₂,…,a_n)；

Object vector T_k=(y₁,y₂,…,y_q)；

Hidden layer input vector S_k=(s₁,s₂,…,s_p), output vector B_k=(b₁,b₂,…,b_p)；

Output layer input vector L_k=(l₁,l₂,…,l_q), output vector C_k=(c₁,c₂,…,c_q)；

Output layer is to the weight w of hidden layer_ij, i=1,2 ..., n, j=1,2 ..., p；

Hidden layer is to the weight v of output layer_jt, j=1,2 ..., p, t=1,2 ..., p；

Hidden layer output threshold θ_j, j=1,2 ..., p；

Output layer output threshold gamma_j, j=1,2 ..., p；

Parameter k=1,2 ..., m.

The improvement of technique scheme is, the training process of described BP neutral net is as follows:

(1) initial value is composed, to w_ij、v_jt、θ_jAnd γ_jRandom assignment in interval (-1,1)；

(2) one group of training sample is arbitrarily chosenAs the input of network, output vector；

(3) according to formula (3-12a), utilizew_ijAnd θ_jCalculate the input vector s of hidden layer neuron_j；According to formula (3-12b), utilize s_jCalculate the output vector b of hidden layer neuron_j；

$s_j=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{w}_{ij}{a}_i-{\mathrm{\θ}}_j,(j=1,2,...,p)---(3-12a)$

b_j=f (s_j) (j=1,2 ..., p)

(3-12b)

(4) according to formula (3-13a), b is utilized_j、w_ijAnd γ_jCalculate the neuronic input vector L of output layer_t；According to formula (3-13b), utilize L_tCalculate the neuronic output vector C of output layer_t；

$L_t=\underset{j=1}{\overset{p}{\mathrm{\Σ}}}{v}_{jt}{b}_j-{\mathrm{\γ}}_t,(t=1,2,...,q)---(3-13a)$

C_t=f (L_t) (t=1,2 ..., q)

(3-13b)

(5) by object vectorC is exported with reality_tSubstitution formula (3-14), calculates the learning error of output layer

$d_t^k=(y_t^k-C_t)C_t(1-C_t),(t=1,2,...,q)---(3-14)$

(6) by v_jt、d_tAnd b_jSubstitution formula (3-15), calculates the learning error in intermediate layer

$e_j^k=\[\underset{t=1}{\overset{q}{\mathrm{\Σ}}}{d}_t\·v_{jt}\]b_j(1-b_j)---(3-15)$

(7) willWith and b_jSubstitution formula (3-16), to v_jtAnd γ_jIt is modified；

$v_{jt}(N+1)=v_{jt}\left(N\right)+{\mathrm{\αd}}_t^k{b}_j---(3-16a)$

${\mathrm{\γ}}_t(N+1)={\mathrm{\γ}}_t\left(N\right)+{\mathrm{\αd}}_t^k---(3-16b)$

T=1,2 ..., q；J=1,2 ..., p；0 < α < 1

(8) willAnd P_k=(a₁,a₂,…,a_n) substitute into formula (3-16), to w_ijAnd θ_jIt is modified；

$w_{ij}(N+1)=w_{ij}\left(N\right)+{\mathrm{\&beta;e}}_j^k{a}_i---(3-17a)$

${\mathrm{\&theta;}}_j(N+1)={\mathrm{\&theta;}}_j\left(N\right)+{\mathrm{\&beta;e}}_j^k---(3-17b)$

I=1,2 ..., n；J=1,2 ..., p；0 < β < 1

(9) all samples are continued training according to above-mentioned steps, require until all inputs and output meet；

(10) again choosing one group of sample, after BP neural network learning, the error of comparing cell output and target sample, if error is in preset range, then network convergence, study terminates；Otherwise network is not restrained, and learns unsuccessfully.

This machine of technique scheme is: when BP neural metwork training, and the state of each feed line of described power distribution network is all as the input of input layer, input vector P=(x₁,x₂,x₃,x₄,…,x_n), wherein x_i(i=1,2,3,4 ..., n) represent the modulus maximum of zero-sequence current wavelet package transforms on i-th feeder line, therefore the nodes of input layer is n；The nodes of output layer is also n, output vector T=(y₁,y₂,y₃,y₄,…,x_n), y_i=1 (i=1,2,3,4 ..., n) represent i-th feeder line single-phase short circuit, y_i=0 (i=1,2,3,4 ..., n) represent that i-th is regular link.

The present invention adopts technique scheme to provide the benefit that: 1) signal that the present invention gathers is transient zero-sequence current, than other fault-signals, initial instant in fault, the amplitude of zero-sequence current and frequency are mainly determined by transient state capacitance current, not by the impact of neutral grounding mode.So extracting transient zero-sequence current as the fault-signal of the method, it is effectively improved the route selection success rate of the method.

2) wavelet package transforms is the effective tool of time frequency analysis, by zero-sequence current is carried out wavelet packet analysis, it is possible to accurately extract effective information therein, such as feature band, modulus maximum and polarity etc..BP neutral net has powerful learning capacity, pass through error back propagation, realizing the nonlinear mapping of input and output, characteristic quantity wavelet packet extracted hence with BP neutral net is trained, not by the impact of neutral grounding mode, fault moment and earth resistance.

3) when distributed power source is popularized gradually, also little for the research of the Single-phase Earth-fault Selection in Distribution Systems containing distributed power source both at home and abroad, and utilize the method that wavelet neural network carries out failure line selection to have good reliability and practicality, and singlephase earth fault circuit can be got rid of timely and effectively, being conducive to the stable operation of distribution network system, the planning simultaneously for distributed power source is also significant with application.

Accompanying drawing explanation

Below in conjunction with accompanying drawing, the invention will be further described:

Fig. 1 is the schematic flow sheet of a kind of power distribution network single-phase short circuit selection method based on wavelet neural network of the embodiment of the present invention.

Detailed description of the invention

Embodiment

A kind of electrical power distribution net single-phase short circuit selection method based on wavelet neural network of the present embodiment, as it is shown in figure 1, it is as follows to perform step:

1) choosing db4 small echo as wavelet packet basis functions, the transient zero-sequence current of each bar feeder line of distribution network to be decomposed, sample frequency is 10KHz, and the resolution of wavelet packet basis functions is 4, decomposes 16 frequency bands, and the bandwidth of each frequency band is 312.5Hz,

2) calculating the modulus maximum of transient zero-sequence current, computing formula is as follows,In formula(j, k) coefficient under sub-band, total n coefficient under each sub-band for WAVELET PACKET DECOMPOSITION；

3) utilize step 2) in the polarity training neutral net of the modulus maximum that obtains and modulus maximum, the desired output of comparative neural network and actual output, to the weight of synapse in neutral net and in neutral net neuronic threshold value be modified；

Neutral net is BP neutral net；

4) by step 3) trained after BP Application of Neural Network in the power distribution network that single-phase short circuit occurs, with according to output result difference determine faulty line.

BP neutral net composition is as follows:

Input vector P_k=(a₁,a₂,…,a_n)；

Object vector T_k=(y₁,y₂,…,y_q)；

Hidden layer input vector S_k=(s₁,s₂,…,s_p), output vector B_k=(b₁,b₂,…,b_p)；

Output layer input vector L_k=(l₁,l₂,…,l_q), output vector C_k=(c₁,c₂,…,c_q)；

Output layer is to the weight w of hidden layer_ij, i=1,2 ..., n, j=1,2 ..., p；

Hidden layer is to the weight v of output layer_jt, j=1,2 ..., p, t=1,2 ..., p；

Hidden layer output threshold θ_j, j=1,2 ..., p；

Output layer output threshold gamma_j, j=1,2 ..., p；

Parameter k=1,2 ..., m.

The training process of BP neutral net is as follows:

(1) initial value is composed, to w_ij、v_jt、θ_jAnd γ_jRandom assignment in interval (-1,1)；

(2) one group of training sample is arbitrarily chosenAs the input of network, output vector；

(3) according to formula (3-12a), utilizew_ijAnd θ_jCalculate the input vector s of hidden layer neuron_j；According to formula (3-12b), utilize s_jCalculate the output vector b of hidden layer neuron_j；

$s_j=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{w}_{ij}{a}_i-{\mathrm{\&theta;}}_j,(j=1,2,...,p)---(3-12a)$

b_j=f (s_j) (j=1,2 ..., p)

(3-12b)

(4) according to formula (3-13a), b is utilized_j、w_ijAnd γ_jCalculate the neuronic input vector L of output layer_t；According to formula (3-13b), utilize L_tCalculate the neuronic output vector C of output layer_t；

$L_t=\underset{j=1}{\overset{p}{\mathrm{\&Sigma;}}}{v}_{jt}{b}_j-{\mathrm{\&gamma;}}_t,(t=1,2,...,q)---(3-13a)$

C_t=f (L_t) (t=1,2 ..., q)

(3-13b)

(5) by object vectorC is exported with reality_tSubstitution formula (3-14), calculates the learning error of output layer

$d_t^k=(y_t^k-C_t)C_t(1-C_t),(t=1,2,...,q)---(3-14)$

(6) by v_jt、d_tAnd b_jSubstitution formula (3-15), calculates the learning error in intermediate layer

$e_j^k=\&lsqb;\underset{t=1}{\overset{q}{\mathrm{\&Sigma;}}}{d}_t\&CenterDot;v_{jt}\&rsqb;b_j(1-b_j)---(3-15)$

(7) willWith and b_jSubstitution formula (3-16), to v_jtAnd γ_jIt is modified；

$v_{jt}(N+1)=v_{jt}\left(N\right)+{\mathrm{\&alpha;d}}_t^k{b}_j---(3-16a)$

${\mathrm{\&gamma;}}_t(N+1)={\mathrm{\&gamma;}}_t\left(N\right)+{\mathrm{\&alpha;d}}_t^k---(3-16b)$

T=1,2 ..., q；J=1,2 ..., p；0 < α < 1

(8) willAnd P_k=(a₁,a₂,…,a_n) substitute into formula (3-16), to w_ijAnd θ_jIt is modified；

$w_{ij}(N+1)=w_{ij}\left(N\right)+{\mathrm{\&beta;e}}_j^k{a}_i---(3-17a)$

${\mathrm{\&theta;}}_j(N+1)={\mathrm{\&theta;}}_j\left(N\right)+{\mathrm{\&beta;e}}_j^k---(3-17b)$

I=1,2 ..., n；J=1,2 ..., p；0 < β < 1

(9) all samples are continued training according to above-mentioned steps, require until all inputs and output meet；

(10) again choosing one group of sample, after BP neural network learning, the error of comparing cell output and target sample, if error is in preset range, then network convergence, study terminates；Otherwise network is not restrained, and learns unsuccessfully.

When BP neural metwork training, the state of each feed line of power distribution network is all as the input of input layer, input vector P=(x₁,x₂,x₃,x₄,…,x_n), wherein x_i(i=1,2,3,4 ..., n) represent the modulus maximum of zero-sequence current wavelet package transforms on i-th feeder line, therefore the nodes of input layer is n；The nodes of output layer is also n, output vector T=(y₁,y₂,y₃,y₄,…,x_n), y_i=1 (i=1,2,3,4 ..., n) represent i-th feeder line single-phase short circuit, y_i=0 (i=1,2,3,4 ..., n) represent that i-th is regular link.

Utilize MATLAB/Simulink to set up a 10kV distribution network system containing distributed power source, have four feeder lines, respectively L1, L2, L3 and L4.Infinite bus system three phase mains equivalence replaces, U_s=10.5kV, f=50Hz.Article four, the line parameter circuit value of feeder line, length and load are in Table 1 and table 2.

	Positive order parameter	Zero sequence parameter
			R(Ω/km)	0.01273	0.3864
L(mH/km)	0.9337	4.1264
			C(nF/km)	12.74	7.751

Table 1

Table 2

Such as neutral by arc extinction coil grounding, the overcompensation degree of arc suppression coil takes 10%, is computed, and the three-phase ground capacitance of system is C₀=6.2008 × 10^-7F.AgainThen U₀/ j ω L=1.1U₀J ω C, soGenerally, the resistance of arc suppression coil takes the 10% of reactance value, i.e. R_L=10%X_L=155.56 Ω.

In phantom, DG three phase mains represents, and sets active power and the reactive power of its output, the DG in Fig. 1 in advance.Access its capacity of distributed power source in distribution and be generally not very big, be generally 6MW and following, take 2MW at this.

1, isolated neutral

Breaking down during voltage zero-cross, the A phase ground connection of feeder line L4, from bus 4.5km place, earth resistance 50 Ω, fault moment is 0.03s, and distributed power source is connected on L2, from bus 13.3km.

The feeder line L4 amplitude maximum broken down, and the sense of current and other three-lines be contrary.After wavelet package transforms, at the energy respectively [1.333,0.5876,2.032,0.001340,0.007455] of (4,1) to (4,5) five frequency bands, so feature band is (4,3).On (4,3), the wavelet coefficient of each feeder line is similar with zero-sequence current waveform, the zero-sequence current I of L4₄Wavelet coefficient on (4,3) is maximum, and opposite polarity.Each feeder line wavelet modulus maxima is [0.3594,0.5141,0.3561 ,-1.229] respectively, so faulty line is L4.

2, neutral by arc extinction coil grounding

The A phase ground connection of feeder line L1, from bus 7.7km place, earth resistance 120, fault moment is 0.027s, and distributed power source is connected on L4, from bus 15km.

After accessing arc suppression coil, when overcompensation, the zero-sequence current steady-state value phase place of each feeder line is identical.After wavelet package transforms, at the energy respectively [57.87,86.57,91.91,4.074,19.52] of (4,1) to (4,5) five frequency bands, so feature band is (4,3).On (4,3), the zero-sequence current I of the wavelet coefficient L1 of each feeder line₁Wavelet coefficient on (4,3) is maximum, and polarity is contrary with other feeder lines.Each feeder line wavelet modulus maxima is [8.181 ,-3.921 ,-2.580 ,-1.717] respectively, so faulty line is L1.

Based on the route selection flow process of wavelet neural network, first have to through substantial amounts of simulated extraction training sample and object vector as data base.This power distribution network one has four feeder lines, therefore the nodes of input layer is 4, input vector P=(x₁,x₂,x₃,x₄), wherein x_i(i=1,2,3,4) modulus maximum of zero-sequence current wavelet package transforms on i-th feeder line is represented；The nodes of output layer is also 5, output vector T=(y₁,y₂,y₃,y₄), when i-th feeder line generation singlephase earth fault, y_i(i=1,2,3,4) are 1, otherwise take 0.

Extracting 50 groups as training sample from emulation data, another extraction 30 groups is as test data.Table 3 is training sample and object vector, and table 4 is test data and route selection result.

Table 3

Table 4

The present invention is not limited to above-described embodiment.All employings are equal to replaces the technical scheme formed, and all falls within the protection domain of application claims.

Claims (4)

1. the electrical power distribution net single-phase short circuit selection method based on wavelet neural network, it is characterised in that perform step as follows:

1) choose db4 small echo as wavelet packet basis functions, the transient zero-sequence current of each bar feeder line of described distribution network to be decomposed, sample frequency is 10KHz, the resolution of described wavelet packet basis functions is 4, decomposes 16 frequency bands, and the bandwidth of each frequency band is 312.5Hz；

2) calculating the modulus maximum of described transient zero-sequence current, computing formula is as follows,In formula(j, k) coefficient under sub-band, total n coefficient under each sub-band for WAVELET PACKET DECOMPOSITION；

3) utilize step 2) in the polarity of the modulus maximum that obtains and described modulus maximum train described neutral net, the relatively desired output of described neutral net and actual output, to the weight of synapse in described neutral net and in described neutral net neuronic threshold value be modified；

Described neutral net is BP neutral net；

4) by step 3) trained after BP Application of Neural Network in the power distribution network that single-phase short circuit occurs, with according to output result difference determine faulty line.

2. the electrical power distribution net single-phase short circuit selection method based on wavelet neural network as claimed in claim 1, it is characterised in that described BP neutral net composition is as follows:

Input vector P_k=(a₁,a₂,…,a_n)；

Object vector T_k=(y₁,y₂,…,y_q)；

Hidden layer input vector S_k=(s₁,s₂,…,s_p), output vector B_k=(b₁,b₂,…,b_p)；

Output layer input vector L_k=(l₁,l₂,…,l_q), output vector C_k=(c₁,c₂,…,c_q)；

Output layer is to the weight w of hidden layer_ij, i=1,2 ..., n, j=1,2 ..., p；

Hidden layer is to the weight v of output layer_jt, j=1,2 ..., p, t=1,2 ..., p；

Hidden layer output threshold θ_j, j=1,2 ..., p；

Output layer output threshold gamma_j, j=1,2 ..., p；

Parameter k=1,2 ..., m.

3. the electrical power distribution net single-phase short circuit selection method based on wavelet neural network as claimed in claim 2, it is characterised in that the training process of described BP neutral net is as follows:

(1) initial value is composed, to w_ij、v_jt、θ_jAnd γ_jRandom assignment in interval (-1,1)；

(2) one group of training sample is arbitrarily chosenAs the input of network, output vector；

(3) according to formula (3-12a), utilizew_ijAnd θ_jCalculate the input vector s of hidden layer neuron_j；According to formula (3-12b), utilize s_jCalculate the output vector b of hidden layer neuron_j；

$s_j=\underset{i=1}{\overset{n}{\&Sigma;}}{w}_{ij}{a}_i-{\mathrm{\&theta;}}_j(j=1,2,\mathrm{...},p)---(3-12a)$

b_j=f (s_j) (j=1,2 ..., p)

(3-12b)

(4) according to formula (3-13a), b is utilized_j、w_ijAnd γ_jCalculate the neuronic input vector L of output layer_t；According to formula (3-13b), utilize L_tCalculate the neuronic output vector C of output layer_t；

$L_t=\underset{j=1}{\overset{p}{\&Sigma;}}{v}_{jt}{b}_j-{\mathrm{\&gamma;}}_t(t=1,2,\mathrm{...},q)---(3-13a)$

C_t=f (L_t) (t=1,2 ..., q)

(3-13b)

(5) by object vectorC is exported with reality_tSubstitution formula (3-14), calculates the learning error of output layer

$d_t^k=(y_t^k-C_t)C_t(1-C_t)(t=1,2,\mathrm{...},q)---(3-14)$

(6) by v_jt、d_tAnd b_jSubstitution formula (3-15), calculates the learning error in intermediate layer

$e_j^k=\&lsqb;\underset{t=1}{\overset{q}{\&Sigma;}}{d}_t\&CenterDot;v_{jt}\&rsqb;b_j(1-b_j)---(3-15)$

(7) willWith and b_jSubstitution formula (3-16), to v_jtAnd γ_jIt is modified；

$v_{jt}(N+1)=v_{jt}\left(N\right)+{\mathrm{\&alpha;d}}_t^k{b}_j---(3-16a)$

${\mathrm{\&gamma;}}_t(N+1)={\mathrm{\&gamma;}}_t\left(N\right)+{\mathrm{\&alpha;d}}_t^k---(3-16b)$

T=1,2 ..., q；J=1,2 ..., p；0 < α < 1

(8) willAnd P_k=(a₁,a₂,…,a_n) substitute into formula (3-16), to w_ijAnd θ_jIt is modified；

$w_{ij}(N+1)=w_{ij}\left(N\right)+{\mathrm{\&beta;e}}_j^k{a}_i---(3-17a)$

${\mathrm{\&theta;}}_j(N+1)={\mathrm{\&theta;}}_j\left(N\right)+{\mathrm{\&beta;e}}_j^k---(3-17b)$

I=1,2 ..., n；J=1,2 ..., p；0 < β < 1

(9) all samples are continued training according to above-mentioned steps, require until all inputs and output meet；

(10) again choosing one group of sample, after BP e-learning, the error of comparing cell output and target sample, if error is in preset range, then network convergence, study terminates；Otherwise network is not restrained, and learns unsuccessfully.

4. as claimed in claim 2 or claim 3 based on the electrical power distribution net single-phase short circuit selection method of wavelet neural network, it is characterised in that: during training, the state of each feed line of described power distribution network is all as the input of input layer, input vector P=(x₁,x₂,x₃,x₄,…,x_n), wherein x_i(i=1,2,3,4 ..., n) represent the modulus maximum of zero-sequence current wavelet package transforms on i-th feeder line, therefore the nodes of input layer is n；The nodes of output layer is also n, output vector T=(y₁,y₂,y₃,y₄,…,x_n), y_i=1 (i=1,2,3,4 ..., n) represent i-th feeder line single-phase short circuit, y_i=0 (i=1,2,3,4 ..., n) represent that i-th is regular link.