CN105548807A - Single-phase fault line selection method of low current grounding system - Google Patents

Abstract

The invention relates to a single-phase fault line selection method of a low current grounding system. Wavelet decomposition is carried out on the zero-sequence current of feed lines to obtain wavelet energy of the feed lines, a sample set is established according to the obtained wavelet energy, and the sample set is normalized to obtain a normalized sample set X*. The Spiking neural network is established initially by utilizing an SRM neuron model, the weights of synaptic units among the neurons in the input layer H, neurons in the hidden layer I and neurons in the output layer J are calculated in the SpikeProp method, the connection weights are corrected to obtain the line section Spiking neural network, fault data is input to the line section Spiking neural network, and a line with the mark phi0 output by the output layer is determined to be a fault line. Operation is convenient, whether a grounding fault occurs in the electrical network is determined rapidly and effectively, the line with the grounding fault is determined after that the grounding fault is determined, and basis is provided for eliminating the fault timely.

Description

Small current neutral grounding system earth fault detection for power

Technical field

The present invention relates to a kind of fault-line selecting method, especially a kind of small current neutral grounding system earth fault detection for power, belongs to the technical field of electric power system fault line selection and location.

Background technology

Along with improving constantly of Living consumption, power industry constantly highlights as the importance of China's mainstay industry, is mainly reflected in the development having ensured nation's security and society.At present, small current neutral grounding system has been widely used in numerous electric power research both domestic and external, therefore, and the fault detect of the continuous exploratory development small current neutral grounding system of increasing scientific worker, and using it as one of primary problem.

When neutral grounding mode selects small current neutral grounding, in system operation, often there is singlephase earth fault.When system cloud gray model, the fault current produced due to singlephase earth fault is very little, smaller to the harm of the health of people, power equipment and communication, and meanwhile, the line voltage of three-phase keeps stablizing constant, can not have an impact to the power supply of load.Therefore, electric power system can continue when there is earth fault to run (but can not more than 2 hours), like this would not because of unexpected power failures to the normal work of user and life.The development of society improves the requirement of the reliability to electric power system, and what this made small current neutral grounding mode has had more and more wider using value.

Summary of the invention

The object of the invention is to overcome the deficiencies in the prior art, a kind of small current neutral grounding system earth fault detection for power is provided, it is easy to operate, can effectively judge whether electrical network earth fault occurs fast, and after determining earth fault, judge the circuit of earth fault, thus provide foundation for timely failture evacuation, safe and reliable.

According to technical scheme provided by the invention, a kind of small current neutral grounding system earth fault detection for power, described earth fault detection for power comprises the steps:

Step 1, foundation have the realistic model of the small current neutral grounding system of k bar feeder line, and arrange different circuit generation singlephase earth fault, the zero-sequence current of each feeder line under measuring each earth fault, to obtain training sample;

Step 2, choose each singlephase earth fault under, 3/4 cycle after front 1/4 cycle of fault of i-th feeder line zero-sequence current F and fault, and use db10 small echo to carry out 5 layers of WAVELET PACKET DECOMPOSITION, if F _{(5, s)}m () is coefficient under s sub-band of the 5th layer of WAVELET PACKET DECOMPOSITION, s=1,2, L, 31, there is m coefficient under each sub-band, then the zero-sequence current wavelet energy E of i-th feeder line _ifor:

$E_i=\underset{s=0}{\overset{31}{\mathrm{\Σ}}}\left(\underset{u=0}{\overset{m}{\mathrm{\Σ}}}{\[F_{(5,s)}\left(u\right)\]}^2\right),i=1,2,Lk;$

Step 3, zero-sequence current wavelet energy according to each feeder line obtained above, set up sample set X, then obtaining sample set X is

$X={\left[\begin{array}{cccccc}{x}_{11}& x_{12}& L& x_{1j}& L& x_{1k}\\ x_{21}& x_{22}& L& x_{2j}& L& x_{2k}\\ L& L& L& L& L& L\\ x_{n1}& x_{n2}& L& x_{nj}& L& x_{nk}\end{array}\right]}_{n\×k}$

Wherein, x _nj=E _nj, E _njit is the zero-sequence current wavelet energy measuring jth bar feeder line n-th time;

Step 4, above-mentioned sample set X to be normalized, to obtain normalization sample set X ^*, then described normalization sample set X ^*for

$X^{*}={\left[\begin{array}{cccccc}\frac{x_{11}}{\max \left\{x_{1i}\right\}}& \frac{x_{12}}{\max \left\{x_{1i}\right\}}& L& \frac{x_{1i}}{\max \left\{x_{1i}\right\}}& L& \frac{x_{1k}}{\max \left\{x_{1i}\right\}}\\ \frac{x_{21}}{\max \left\{x_{2i}\right\}}& \frac{x_{22}}{\max \left\{x_{2i}\right\}}& L& \frac{x_{2i}}{\max \left\{x_{2i}\right\}}& L& \frac{x_{2k}}{\max \left\{x_{2i}\right\}}\\ L& L& L& L& L& L\\ \frac{x_{n1}}{\max \left\{x_{ni}\right\}}& \frac{x_{n2}}{\max \left\{x_{ni}\right\}}& L& \frac{x_{ni}}{\max \left\{x_{ni}\right\}}& L& \frac{x_{nk}}{\max \left\{x_{ni}\right\}}\end{array}\right]}_{n\×k}={\left[\begin{array}{cccccc}{x^{*}}_{11}& {x^{*}}_{12}& L& {x^{*}}_{1i}& L& {x^{*}}_{1k}\\ {x^{*}}_{21}& {x^{*}}_{22}& L& {x^{*}}_{2i}& L& {x^{*}}_{2k}\\ L& L& L& L& L& L\\ {x^{*}}_{n1}& {x^{*}}_{n2}& L& {x^{*}}_{ni}& L& {x^{*}}_{nk}\end{array}\right]}_{n\×k}$

Wherein, max{x _niit is element maximum in n-th line in sample set X;

Step 5, set up Spiking neural network and by above-mentioned normalization sample set X ^*in element convert corresponding pulse launch time to, wherein, the neuron of maximal value 1 correspondence is corresponded to first transponder pulse, and remembers that this pulse launch time is 0ms; And the neuron of minimum value 0 correspondence corresponds to last transponder pulse, and remember that this pulse launch time is T _maxms;

The corresponding neuron number of step 6, the input layer of Spiking neural network, output layer is all consistent with feeder line sum k; Determine Δ T time delay between above-mentioned input layer, output layer, and according to Δ T described time delay, mark the faulty line in output layer with Φ 0, and mark with Φ 1 and perfect circuit in output layer;

Step 7, by above-mentioned by normalization sample set X ^*the some pulse launch time obtained substitute into Spiking neural network, and utilize SpikeProp method to train Spiking neural network, and correction connects weights accordingly;

Step 8, connect weights according to after above-mentioned correction, obtain required route selection Spiking neural network; Gather the fault data under small area analysis singlephase earth fault, and by the fault data input value route selection Spiking neural network of collection, namely the circuit exporting Φ 0 mark in route selection Spiking neural network output layer is defined as faulty line.

In described step 5, adopt Time-to-first – Spike method by normalization sample set X ^*interior element converts corresponding burst length data to, and described transfer process is:

$T=T_{\max }(1-x_{ni}^{*})$

Wherein, for normalization sample set X ^*interior element, T _maxfor maximum impulse launch time, T is by normalization sample set X ^*the pulse launch time that corresponding element is converted to.

In described step 5, SRM neuron models are adopted to build Spiking neural network.

Advantage of the present invention: the zero-sequence current information utilizing each feeder line during the realistic model acquisition fault of the small current neutral grounding system set up, and carry out wavelet decomposition and obtain its wavelet energy, and sample set can be set up, to obtaining normalization sample set X after sample set normalization according to the wavelet energy obtained ^*.Adopt SRM neuron models Primary Construction Spiking neural network, SpikeProp method is adopted to calculate the neuron in input layer H and the neuron in hidden layer I and the neuron in hidden layer I and the weights of each cynapse between the neuron in output layer J, route selection Spiking neural network is obtained after carrying out connection modified weight, fault data input route selection Spiking neural network, namely the circuit exporting Φ 0 mark in output layer is defined as faulty line, easy to operate, can effectively judge whether electrical network earth fault occurs fast, and after determining earth fault, judge the circuit of earth fault, thus provide foundation for timely failture evacuation, safe and reliable.

Accompanying drawing explanation

Fig. 1 is process flow diagram of the present invention.

Fig. 2 is SRM neuron models schematic diagram of the present invention.

Fig. 3 is the schematic diagram of the present invention's three layers of forward direction Spiking neural network.

Embodiment

Below in conjunction with concrete drawings and Examples, the invention will be further described.

As shown in Figure 1: in order to effectively judge whether electrical network earth fault occurs fast, and after determining earth fault, judge the circuit of earth fault, thus provide foundation for timely failture evacuation, earth fault detection for power of the present invention comprises the steps:

Step 1, foundation have the realistic model of the small current neutral grounding system of k bar feeder line, and arrange different circuit generation singlephase earth fault, the zero-sequence current of each feeder line under measuring each earth fault, to obtain training sample;

The present invention sets up the realistic model of small current neutral grounding system by conventional technological means, set up the process of realistic model known by the art personnel, in realistic model, the quantity k of feeder line can carry out selection as required and determines, be specially known by the art personnel, repeat no more herein.After setting up the realistic model of low current grounding, by arranging different circuit generation singlephase earth fault, thus the zero-sequence current of each feeder line under each earth fault can be measured, namely the above-mentioned zero-sequence current obtaining each feeder line is training sample.

Step 2, choose each singlephase earth fault under, 3/4 cycle after front 1/4 cycle of fault of i-th feeder line zero-sequence current F and fault, and use db10 small echo to carry out 5 layers of WAVELET PACKET DECOMPOSITION, if F _{(5, s)}m () is coefficient under s sub-band of the 5th layer of WAVELET PACKET DECOMPOSITION, s=1,2, L, 31, there is m coefficient under each sub-band, then the zero-sequence current wavelet energy E of i-th feeder line _ifor:

$E_i=\underset{s=0}{\overset{31}{\mathrm{\Σ}}}\left(\underset{u=0}{\overset{m}{\mathrm{\Σ}}}{\[F_{(5,s)}\left(u\right)\]}^2\right),i=1,2,Lk;$

In the embodiment of the present invention, utilize db10 small echo to carry out the process of 5 layers of WAVELET PACKET DECOMPOSITION known by the art personnel, specifically repeat no more.

Step 3, zero-sequence current wavelet energy according to each feeder line obtained above, set up sample set X, then obtaining sample set X is

$X={\left[\begin{array}{cccccc}{x}_{11}& x_{12}& L& x_{1j}& L& x_{1k}\\ x_{21}& x_{22}& L& x_{2j}& L& x_{2k}\\ L& L& L& L& L& L\\ x_{n1}& x_{n2}& L& x_{nj}& L& x_{nk}\end{array}\right]}_{n\×k}$

Wherein, x _nj=E _nj, E _njit is the zero-sequence current wavelet energy measuring jth bar feeder line n-th time;

In the embodiment of the present invention, n is the number of times measured, and namely obtain the number of times of training sample in step 1, concrete pendulous frequency n can carry out selection as required and determine.

Step 4, above-mentioned sample set X to be normalized, to obtain normalization sample set X ^*, then described normalization sample set X ^*for

Wherein, max{x _niit is element maximum in n-th line in sample set X;

In the embodiment of the present invention, the element x in sample set X ₁₁obtain after normalization max{x _nibe element maximum in n-th line in sample set X, as max{x _1ibe the maximal value in the first row all elements in sample set X, the rest may be inferred by analogy for it, specifically will not enumerate.

Step 5, set up Spiking neural network and by above-mentioned normalization sample set X ^*in element convert corresponding pulse launch time to, wherein, the neuron of maximal value 1 correspondence is corresponded to first transponder pulse, and remembers that this pulse launch time is 0ms; And the neuron of minimum value 0 correspondence corresponds to last transponder pulse, and remember that this pulse launch time is T _maxms;

In the embodiment of the present invention, SRM neuron models are adopted to build Spiking neural network.Adopt Time-to-first – Spike method by normalization sample set X ^*interior element converts corresponding burst length data to, and described transfer process is:

$T=T_{\max }(1-x_{ni}^{*})$

Wherein, for normalization sample set X ^*interior element, T _maxfor maximum impulse launch time, T is by normalization sample set X ^*the pulse launch time that corresponding element is converted to.

Particularly, the SRM neuron models in Spiking neural network adopt accurate pulse launch time coding method to carry out transmission and the calculating of information, therefore, it is possible to the neuron in the true biological nervous system of more close description.

Similar to traditional BP-ANN, Spiking neural network also can form the various network such as feedforward network, feedback network with Spiking neuron.In the embodiment of the present invention, the SNN structure of employing is 3 layers of feedforward network structure, and any one neuron in every layer is only connected with the neuron in adjacent layer, and with other neurons in same layer without being connected.Wherein H represents input layer, and I represents hidden layer, and J represents output layer; The neuron number of input layer H, hidden layer I neuron number, the neuronic number of output layer J are respectively w, p, q, as shown in Figures 2 and 3.

Because Spiking neural network carries out the transmission of information and calculating based on accurate pulse launch time, therefore must analog data be converted to the burst length.The present invention adopts Time-to-first-Spike method to the input of Spiking neural network, exports and encode, and the data value after exomonental moment and normalization is proportional, and the value of general pulse launch time is larger, and the corresponding exomonental time more early.In the data through normalizing in interval [0,1], the neuron of maximal value 1 correspondence corresponds to first transponder pulse, and remembers that this pulse launch time is 0ms.And the neuron of minimum value 0 correspondence is last transponder pulse in data, and remember that this pulse launch time is T _maxms.In the embodiment of the present invention, choose maximum burst length T _max=5ms.

The corresponding neuron number of step 6, the input layer of Spiking neural network, output layer is all consistent with feeder line sum k; Determine Δ T time delay between above-mentioned input layer, output layer, and according to Δ T described time delay, mark the faulty line in output layer with Φ 0, and mark with Φ 1 and perfect circuit in output layer;

In the embodiment of the present invention, in Spiking neural network, pulse is at 0ms and T _maxtwo moment are delivered to output layer J from hidden layer I from input layer H again to hidden layer I, have the individual cynapse sub-connection of s (s gets 16), synaptic delay d between often two-layer ^kminimum 1ms, maximum 16ms.Therefore, the input layer of Spiking neural network, export between postpone Δ T if having time, scope is from 0+2*1=2ms to 5+2*16=37ms.In the embodiment of the present invention, choosing time delay Δ T is 19ms, and namely [0,5] ms of input layer corresponds to [19,24] ms of output layer.In output layer, 19ms correspond to circuit and is faulty line, marks with Φ 0, and being of other perfects circuit, marks with Φ 1.

In the embodiment of the present invention, the neuron number of input layer, output layer is feeder line sum k.The experimental formula of the basis for selecting BP-ANN hidden layer neuron number of hidden layer neuron number to determine through test of many times, is specially known by the art personnel, specifically repeats no more.

Step 7, by above-mentioned by normalization sample set X ^*the some pulse launch time obtained substitute into Spiking neural network, and utilize SpikeProp method to train Spiking neural network, and correction connects weights accordingly;

In the embodiment of the present invention, SpikeProp method is similar to the error back propagation in second generation artificial neural network, and target minimizes network error function, trains, obtain best initial weights to Spiking neural network.Neuron in neuron in input layer H and hidden layer I and the neuron in hidden layer I and between the neuron in output layer J the initial value of the weights of each cynapse get the random value in interval [0,1].

By above-mentioned by normalization sample set X ^*the pulse launch time transformed substitutes into Spiking neural network, and the process revising connection value is:

$e=\frac{1}{2}\underset{j\∈J}{\Σ}{(t_j^a-t_j^d)}^2$

Wherein: for the time of neuron actual transmission pulse in output layer J, for the pulse launch time expected.According to revised connection weights, calculate each layer and export and network training error e, if e< is ε (the error upper limit of setting), then train termination, otherwise modified weight is carried out in continuation calculating, until meet the demands, obtains best initial weights.The size of error upper limit ε can be carried out selection according to specific needs and be determined, is specially known by the art personnel, repeats no more herein.

Step 8, connect weights according to after above-mentioned correction, obtain required route selection Spiking neural network; Gather the fault data under small area analysis singlephase earth fault, and by the fault data input value route selection Spiking neural network of collection, namely the circuit exporting Φ 0 mark in route selection Spiking neural network output layer is defined as faulty line.

In the embodiment of the present invention, after carrying out connection modified weight, required Spiking neural network can be obtained, be route selection Spiking neural network.Be input to by the fault data of the low current grounding of collection in route selection Spiking neural network, namely the circuit exporting Φ 0 mark in output layer is defined as faulty line.

The present invention utilize the realistic model of the small current neutral grounding system of foundation obtain fault time each feeder line zero-sequence current information, and carry out wavelet decomposition and obtain its wavelet energy, and sample set can be set up, to obtaining normalization sample set X after sample set normalization according to the wavelet energy obtained ^*.Adopt SRM neuron models Primary Construction Spiking neural network, SpikeProp method is adopted to calculate the neuron in input layer H and the neuron in hidden layer I and the neuron in hidden layer I and the weights of each cynapse between the neuron in output layer J, route selection Spiking neural network is obtained after carrying out connection modified weight, fault data input route selection Spiking neural network, namely the circuit exporting Φ 0 mark in output layer is defined as faulty line, easy to operate, can effectively judge whether electrical network earth fault occurs fast, and after determining earth fault, judge the circuit of earth fault, thus provide foundation for timely failture evacuation, safe and reliable.

Claims (3)

1. a small current neutral grounding system earth fault detection for power, is characterized in that, described earth fault detection for power comprises the steps:

Step 1, foundation have the realistic model of the small current neutral grounding system of k bar feeder line, and arrange different circuit generation singlephase earth fault, the zero-sequence current of each feeder line under measuring each earth fault, to obtain training sample;

Step 2, choose each singlephase earth fault under, 3/4 cycle after front 1/4 cycle of fault of i-th feeder line zero-sequence current F and fault, and use db10 small echo to carry out 5 layers of WAVELET PACKET DECOMPOSITION, if F _{(5, s)}m () is coefficient under s sub-band of the 5th layer of WAVELET PACKET DECOMPOSITION, s=1,2, L, 31, there is m coefficient under each sub-band, then the zero-sequence current wavelet energy E of i-th feeder line _ifor:

$E_i=\underset{s=0}{\overset{31}{\mathrm{\&Sigma;}}}\left(\underset{u=0}{\overset{m}{\mathrm{\&Sigma;}}}{\&lsqb;F_{(5,s)}\left(u\right)\&rsqb;}^2\right),i=1,2,Lk;$

Step 3, zero-sequence current wavelet energy according to each feeder line obtained above, set up sample set X, then obtaining sample set X is

$X={\left[\begin{array}{cccccc}{x}_{11}& x_{12}& L& x_{1j}& L& x_{1k}\\ x_{21}& x_{22}& L& x_{2j}& L& x_{2k}\\ L& L& L& L& L& L\\ x_{n1}& x_{n2}& L& x_{nj}& L& x_{nk}\end{array}\right]}_{n\&times;k}$

Wherein, x _nj=E _nj, E _njit is the zero-sequence current wavelet energy measuring jth bar feeder line n-th time;

Step 4, above-mentioned sample set X to be normalized, to obtain normalization sample set X ^*, then described normalization sample set X ^*for

$x^{*}={\left[\begin{array}{cccccc}\frac{x_{11}}{\max \left\{x_{1i}\right\}}& \frac{x_{12}}{\max \left\{x_{1i}\right\}}& L& \frac{x_{1i}}{\max \left\{x_{1i}\right\}}& L& \frac{x_{1k}}{\max \left\{x_{1i}\right\}}\\ \frac{x_{21}}{\max \left\{x_{2i}\right\}}& \frac{x_{22}}{\max \left\{x_{2i}\right\}}& L& \frac{x_{2i}}{\max \left\{x_{2i}\right\}}& L& \frac{x_{2k}}{\max \left\{x_{2i}\right\}}\\ L& L& L& L& L& L\\ \frac{x_{n1}}{\max \left\{x_{ni}\right\}}& \frac{x_{n2}}{\max \left\{x_{ni}\right\}}& L& \frac{x_{ni}}{\max \left\{x_{ni}\right\}}& L& \frac{x_{nk}}{\max \left\{x_{ni}\right\}}\end{array}\right]}_{n\&times;k}={\left[\begin{array}{cccccc}{x^{*}}_{11}& {x^{*}}_{12}& L& {x^{*}}_{1i}& L& {x^{*}}_{1k}\\ {x^{*}}_{21}& {x^{*}}_{22}& L& {x^{*}}_{2i}& L& {x^{*}}_{2k}\\ L& L& L& L& L& L\\ {x^{*}}_{n1}& {x^{*}}_{n2}& L& {x^{*}}_{ni}& L& {x^{*}}_{nk}\end{array}\right]}_{n\&times;k}$

Wherein, max{x _niit is element maximum in n-th line in sample set X;

Step 5, set up Spiking neural network and by above-mentioned normalization sample set X ^*in element convert corresponding pulse launch time to, wherein, the neuron of maximal value 1 correspondence is corresponded to first transponder pulse, and remembers that this pulse launch time is 0ms; And the neuron of minimum value 0 correspondence corresponds to last transponder pulse, and remember that this pulse launch time is T _maxms;

The corresponding neuron number of step 6, the input layer of Spiking neural network, output layer is all consistent with feeder line sum k; Determine Δ T time delay between above-mentioned input layer, output layer, and according to Δ T described time delay, mark the faulty line in output layer with Φ 0, and mark with Φ 1 and perfect circuit in output layer;

Step 7, by above-mentioned by normalization sample set X ^*the some pulse launch time obtained substitute into Spiking neural network, and utilize SpikeProp method to train Spiking neural network, and correction connects weights accordingly;

Step 8, connect weights according to after above-mentioned correction, obtain required route selection Spiking neural network; Gather the fault data under small area analysis singlephase earth fault, and by the fault data input value route selection Spiking neural network of collection, namely the circuit exporting Φ 0 mark in route selection Spiking neural network output layer is defined as faulty line.

2. small current neutral grounding system earth fault detection for power according to claim 1, is characterized in that, in described step 5, adopts Time-to-first – Spike method by normalization sample set X ^*interior element converts corresponding burst length data to, and described transfer process is:

$T=T_{\max }(1-x_{ni}^{*})$

Wherein, for normalization sample set X ^*interior element, T _maxfor maximum impulse launch time, T is by normalization sample set X ^*the pulse launch time that corresponding element is converted to.

3. small current neutral grounding system earth fault detection for power according to claim 1, is characterized in that, in described step 5, adopts SRM neuron models to build Spiking neural network.