KR20020076413A - Fault Classification of Power System using Neuro-Fuzzy Network and method for judgement thereof - Google Patents

Abstract

PURPOSE: A method for detecting and discriminating errors of a power system using a neuro-fuzzy network is provided to detect and process rapidly errors by detecting and calculating each phase current to obtain an image current and inputting effective values of each phase current and the image current into 4 neuro-fuzzy networks. CONSTITUTION: A phase current is detected from a line supplying three phase power source. An image current is calculated by using the detected phase current. Each effective value of the detected phase current and the calculated image current are produced. The produced effective values of the detected phase current and the calculated image current are inputted into 4 neuro-fuzzy networks and the neuro-fuzzy networks are operated. A kind of error is determined and the error is reported by generating an alarm sound if the neuro-fuzzy networks are completed.

Description

Fault Classification of Power System using Neuro-Fuzzy Network and method for judgmentment}

The present invention detects each phase current, calculates the image current, calculates the image current, and then the effective values of each phase current and the image current are inputted to four neurofuge networks for each operation so as to operate a short circuit and a power supply line or a distribution line in a power system. The present invention relates to a failure detection of a power system and a method of determining the failure of a power system using a Neuro-Fuzzy Network that can detect, discriminate, and handle the failure quickly and accurately when a failure such as a ground fault occurs.

In general, as the industry develops and the standard of living improves, the demand for electric power gradually increases, and the facilities constituting the electric power system are increasing in capacity and diversification.

In other words, the power system is growing on a very large scale, and its configuration is also very complicated. As a result, reliable operation of the power system and the transmission of high-quality power to customers are becoming increasingly difficult. If a failure occurs in any part of the complex power system, it is difficult to detect the failure in a short time. Protection coordination is fragile and the spread of accidents becomes very large.

Therefore, effective protection system is important for supplying high quality power. Among them, as the power system becomes very high voltage, the protection of the transmission line is increasing.

Such transmission line protection can be mainly cut off and repair of ground fault, short circuit, etc. For effective protection, it is very important to accurately identify the type of accidents occurring on the track and take appropriate measures quickly. This helps to reduce unnecessary power outages, thereby greatly improving the reliability of the power system. In order to protect such transmission lines, various research and development is being carried out at home and abroad, and many developments have been made.

Most power system failures are represented by unbalanced faults such as single-phase and two-phase ground faults, and in the case of unbalanced faults, circuits are often asymmetrical. For unbalanced fault analysis, FORTESCUE uses the symmetric coordinate method ( Symmetrical Component Method) has been proposed, and there have been many studies for applying the distance relay method using the symmetric coordinate method to the actual power system.

According to this symmetric coordinate method, assuming that a transmission terminal is viewed from a virtual terminal when a failure occurs at a specific point of a power system, the power system includes a generator, an ancestor, a transformer, and the like, and there is a line connecting them. Although complex, the condition that each phase voltage and each constant are balanced is generally satisfied, so the power system can be thought of as a symmetric three-phase generator.

Therefore, knowing the voltage just before the failure and the impedance at that point, various failures occurring in the power system can be treated as a failure at the terminal of the three-phase generator which makes the voltage no-load and the impedance at that point as the impedance. By using the basic formula of the three-phase AC generator, it is possible to calculate the voltage, current and symmetry of the failure point.

As a technique to apply the symmetric coordinate method to the distance relay protection method of the real power system, in the case of a failure of the power system, the phase voltage is processed in real time to obtain the symmetry. AG Phadke, JS Thorp, and MG Adamiak jointly Presented paper "A new measurement technique for tracking voltage phasors, local system frequency, and rate of change off requency" (IEEE Transactions on power apparatus and systems, Vol.PAS 102, No.5, pp.1025-1038, May 1983 Is a distance relay that calculates the distance of failure after determining the occurrence of failure by setting the threshold in advance regardless of the type of failure. Co-published by Phadke, M. Ibrahim and T. Hibka, "Fundamental basis for distance relaying with symmetrical components" (IEEE transactions on power apparatus and systems, Vol. PAS 96, No.2, pp.635-646, March / April 1977).

However, these papers only define partial algorithms for fault determination of power systems, and there are many difficulties in applying them to actual power systems.

In addition, the protection relay against ground fault in the current power system uses a ground fault relay as a transmission line protection method and a directional ground current over relay (DOCGR) to protect the main transformer of a substation.

However, such a protective relay system has a ground fault distance relay for transmission line protection of a substation due to a very small ground fault current caused by a high resistance ground fault caused by a foreign body such as a crane or heavy equipment such as a tree in the transmission line. There are cases of detecting and not detecting an accident.

As a result, a ground fault distance relay (Zone-3) installed in the substation at the rear stage is operated or a directional ground overcurrent relay (DOCGR), which is a main transformer protection relay, is operated so that the substation becomes pressureless. The situation is causing a ripple.

Therefore, many studies have been attempted to detect such high resistance ground faults at home and abroad, but are mainly limited to 22.9kV class (distribution line), and the countermeasures against transmission line have no obvious problem.

On the other hand, the method used to detect such a high resistance ground fault generally uses a Fourier transform which analyzes the signal as a sinusoidal function composed of several frequencies, among the methods of analyzing voltage and current signals generated when a transmission line breakdown occurs. have.

However, these Fourier transforms do not know when a specific accident occurs when analyzing a power system fault having transient and nonstationary characteristics, such as a high resistance ground fault, and the nonstationary signal analysis. As it requires a lot of sampling and long time to improve the efficiency, it is not an efficient analysis method for transient signal analysis such as high resistance ground fault.

In addition, some countries have applied PCM transmission current differential meter electric protective relays for transmission line protection, which are somewhat advantageous in detecting high resistance ground faults. However, it is impossible to replace these PCM type protective relays with existing relays in all transmission lines of all substations currently installed.

Therefore, the development of a high resistance ground fault protection relay system that detects a high resistance ground fault and selectively operates a breaker (CB) of the corresponding line, while suppressing the operation of other relays including the main transformer side. It became necessary.

SUMMARY OF THE INVENTION The present invention has been made to solve the above problems, and an object of the present invention is to detect each phase current, calculate the image current and calculate the effective values of each phase current and the image current for each four neurofuge networks. By inputting to the power supply, it detects and discriminates quickly and accurately in case of failure of 1 line ground, 2 line ground, line short circuit, 3 phase fault, high resistance ground fault in power transmission line or distribution line of power system. The present invention provides a method for detecting a failure of a power system using a neuro-fuzzy network and a method of determining the same.

1 is a circuit diagram illustrating the overall configuration of a power system to which the fault detection and determination method of the present invention is applied;

2 is a flowchart illustrating a failure detection of the power system and a method of determining the same using a neurofuge network according to the present invention;

3 is a configuration diagram for explaining the structure of a neurofuzzy network used in the failure determination method of the present invention;

4 is a view for explaining a membership function used in the first layer of FIG.

5 is a view for explaining the membership function of the effective value of each phase current (I _A , I _B , I _C ) according to the present invention,

6 is a view for explaining a membership function for the effective value of the image current (Io) according to the present invention,

7 is a view for explaining a model of the error back-propagation algorithm according to the present invention,

8A to 8C are graphs showing the output of the neurofuge network for a single phase ground fault according to the present invention;

9A to 9C are graphs showing the output of the neurofuge network for a two-phase ground fault according to the present invention;

10A to 10C are graphs showing the output of the neurofuge network for line short circuit failure according to the present invention;

11 is a graph showing the output of the neurofuzzy network for the three-phase failure according to the present invention,

12A to 12C are graphs showing the output of the neurofuge network for the high resistance ground fault according to the present invention;

13 is a diagram showing an output value of the neurofuzzy network for each fault according to the present invention.

Explanation of symbols on the main parts of the drawings

1: microcomputer 2: current signal output unit

3: A / D converter 4: Input unit

5: display part 6: ROM

7: RAM CT ₁ ~ CT ₃ : Current transformer

In order to achieve the above object, the fault detection and the method of detecting the present invention detect and determine a fault of a power system using a neurofuge network. The learning of the neurofuge network is performed by a back propagation algorithm, the first step of detecting a phase current of a line supplying three-phase power, the second step of calculating an image current using each phase current, and A third step of calculating an effective value of each phase current and an image current, a fourth step of inputting the effective values of each phase current and an image current to four neurofuge networks, and a fourth step of the neurofuge network of the fourth step. When the execution is completed, it characterized in that it comprises a fifth step of determining the type of failure and alert it.

In addition, the neurofuge network of the fourth step may include a sixth step of outputting a membership function value by inputting the effective values of each phase current and an image current, and a seventh step of calculating the membership function value and outputting a fuzzy operation value. And an eighth step of normalizing the fuzzy operation value according to the active intensity of the corresponding rule and outputting a normalized value of the active intensity, and a ninth step of inferring the normalized value of the active intensity and outputting a non-fuzzy value; And a tenth step of outputting an inference value by performing an addition operation on each non-fuzzy value.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. This embodiment is not intended to limit the scope of the invention, but is presented by way of example only.

1 is a circuit configuration diagram for explaining the overall configuration of a power system to which the fault detection and determination method of the present invention is applied. The microcomputer 1 controls the overall operation of detecting and discriminating a failure of a power system, and a load path. A current transformer CT _{1 to} CT ₃ for detecting phase currents of each of the supplied A, B, and C phases, a current signal output unit 2 for outputting a phase current value detected at the current transformers CT _{1 to} CT ₃ , An A / D converter 3 for converting an output signal of the current signal output unit 2 into a digital signal and inputting it to the microcomputer 1, an input unit 4 for operating commands to the microcomputer 1, and a microcomputer 1 Display unit 5 for displaying the operation status and alarm under the control of the control panel, a ROM 6 storing the operation program of the microcomputer 1, and a RAM 7 storing data necessary for fault determination. It was.

Power system configured in this way is, if supplying power to the load A, B on the phase current and the C (I _A, I _B, I _C), a current transformer (CT ₁ ~CT ₃₎ is detected by a current signal output unit (2) A virtual current value is output, and each phase current value of the current signal output unit 2 is converted into a digital signal by the A / D converter 3 and input to the microcomputer 1, so that the microcomputer 1 is an A / D converter. 3 of a, B and C phase current (I _a, I _B, I _C), the calculated after obtaining the sequence current (Io), these phase currents (I _a, I _B, I _C) and the image current (Io on the ) The effective values are input to four neurofuge networks for each to determine the type of failure of the power system, and when the type of failure is determined, the alarm is alerted.

Figure 2 is a phase current (I _A, I _B on the A, B and C when supplied to, first, power supply to the load a flowchart illustrating a fault detection and a determination method of a power system using a neuro-fuzzy network according to the present invention , I _C ) is detected by the current transformers CT _{1 to} CT ₃ (S10), and the microcomputer ₁ records the input phase currents I _A , I _B and I _C as shown in Equation 1 below. The current Io value is calculated (S20).

The formula for calculating the image current (Io) by the value of each phase current (I _A , I _B , I _C ) is

Becomes

Here, I _A , I _B , and I _C are the phase current values, and Io is the image current value.

When the image current Io value is calculated in this manner, the calculated image current Io and the effective value of each phase current I _A , I _B , I _C are calculated (S30), and the respective phase currents I _A. The effective values of, I _B , I _C ) and the image current Io are inputted into four neurofuge networks of the present invention that have been learned in advance (S40).

Here, the four neurofuge networks used in the failure determination method of the present invention will be described in detail.

FIG. 3 is a block diagram illustrating the structure of a neurofuzzy network used in the failure determination method of the present invention, and has three fuzzy rules in the form of one input and one output, and three phase currents I _A ,. The effective values of I _B , I _C ) and the image current Io are used for four neurofuge networks for each, and each of the four neurofuge networks has the following structure.

First, the three fuzzy rule bases have the following rules.

As shown in Equation 2 and Equation 3, a first layer is formed that receives each input variable (ie, I _A , I _B , I _C , Io) and outputs a fuzzy variable for the membership function.

here, Is the output of the first layer, i is the output at each node (i.e. i = 1,2,3), and the membership degree of the input value x for the language term A _i is calculated. Value. As shown in FIG. 4, the membership function is a value of the Schlumberm function defined as Small, Medium, and Big, and is represented by the input / output relationship as shown in Equation 4 below.

Here, a, b, c, and d are the x-axis positions of each vertex of the trapezoid shape, respectively.

As shown in Fig. 5 or 6, in the embodiment of the present invention is a membership function used for the effective value of each phase current (I _A , I _B , I _C ) and the image current (Io), its a, b, c and d values are as follows.

As shown in FIG. 5, it is a membership function for the effective value of each phase current I _A , I _B , I _C , and the a, b, c, d values of the membership function are as follows.

SM: Small Membership Function

[a, b, c, d] = [0.95, 1.00, 1.00, 1.05]

ME: Medium membership function

[a, b, c, d] = [1.00, 1.05, 1.70, 1.95]

BI: big membership function

[a, b, c, d] = [1.75, 1.95, 4.90, 5.00]

As shown in FIG. 6, it is a membership function for the effective value of the image current Io, and a, b, c, and d values of the membership function are as follows.

SM: Small Membership Function

[a, b, c, d] = [-0.05, 0.00, 0.00, 0.05]

ME: Medium membership function

[a, b, c, d] = [0.00, 0.05, 0.50, 0.75]

BI: big membership function

[a, b, c, d] = [0.50, 0.75, 4.90, 5.00]

Next, a second layer for outputting a fuzzy operation value by calculating the fuzzy variable output from the first layer is formed. Here, when there are several inputs, the second layer is calculated by a product operator or the like. However, in the present invention, since the input is one, the output of the first layer is used as it is.

Next, a third layer for outputting the normalized value of the active intensity is formed by normalizing the fuzzy operation value output from the second layer according to the active intensity of the corresponding rule. If this is expressed as an input-output relation, it is as follows.

Next, a fourth layer is formed which infers the normalized value of the active intensity output from the third layer and outputs a non-fuzzy value, and the second half is represented by a linear function. If this is expressed as an input-output relation, it is as follows.

Where P _i is a set of latter parameters.

Finally, a fifth layer representing the sum of the respective unpurged values output from the fourth layer is formed, and this is represented by the input / output relational expression (Equation 8).

The output of the fifth layer corresponds to the inference value of the neurofuge network of the present invention, which indicates the propensity of the type of failure in the power system.

In the present invention, the learning of the neurofuzzy network is performed by a back propagation learning algorithm method. The backpropagation algorithm extends the Widrow-Hoff learning rule to multilayer networks and nonlinear differential transition functions. As a result, neural network research became active and backpropagation learning algorithm became the main area of neural network.

As shown in FIG. 7, the backpropagation learning algorithm is also referred to as a backpropagation technique because it performs a recursive calculation in a reverse direction to the signal flow during the learning.

Multilayer perceptron refers to a neural network consisting of an input layer, an output layer, and one or more hidden layers. The hidden layer has hidden units that are not directly connected by input and output. Multilayer perceptron has a structure similar to single layer perceptron having only input layer and output layer, but improves its ability by using nonlinear functions such as sigmoid function for hidden layer and input unit function to overcome the shortcomings of single layer perceptron. It was.

The equation for obtaining the delta from the error of the nerve cells is as shown in (Equations 9) to (Equation 11).

Where δ _{j is the} delta of the output layer neuron j, f '(net _j ) is the derivative of the activity function of the output layer neuron _j , e j is the error of the output layer neuron _j , t _{j is the} output layer neuron j Is a component of the desired pattern corresponding to a _j : the activity value of the output layer neuron j.

When this delta is obtained, the delta is multiplied from the top to the bottom by multiplying the connection weights associated with the output stratum neurons, which are then transferred to the stratum neurons. The combined value of these processes is the error of the corresponding neuron in the hidden layer. When the error of the hidden layer neurons is found, the delta of the hidden layer neurons can be obtained again as shown in Equations 12 to 14 below.

Where δ _{i is the} delta of the hidden layer neuron i, f '(net _i ): the derivative of the activity function of the hidden layer neuron _i , e _i : the error of the hidden layer neuron _i , a _i : the hidden layer neuron i Is the active value of.

The improvement of the back propagation algorithm to improve the learning speed and learning ability includes the addition of momentum, the improvement of the initial condition, and the adaptive learning speed method. Among these, the adaptive learning speed method makes learning as stable as possible and the size of the learning stage. In order to increase the step size, the adaptive learning speed can be used to reduce the learning time. If the learning is stable at very large learning speeds, the learning speed will increase, and if the learning speed is too large to reduce the error, it will decrease until the learning is stable again.

In other words, the back propagation algorithm is a generalized Least Mean Square (LMS) algorithm, and uses the gradient search method to minimize the objective function corresponding to the mean square distance between the teacher data output and the network output.

The network is taught by iteratively applying all the teacher data in the state with an initial random random coefficient of coupling and a threshold. When all teacher data is applied, the coupling coefficient is changed, and the learning is terminated when the coupling coefficient converges, the value of the objective function falls below a certain value, or the learning is performed for the specified number of learning times.

In the present invention, the learning count is used as the end-of-learning condition, the probability of local convergence of the backpropagation algorithm is reduced, the learning rate used to adjust the coupling coefficient is small, and the new initial value of the coupling coefficient is used to improve the performance of the algorithm. I learned by.

As described above, when the neurofuge network is completed, the type of failure is determined, and if normal (S51), the next operation is performed (S70), single phase ground (S52), two phase ground (S53), line short ( If it is determined as S54, three-phase fault (S55), high resistance ground (S56), the fault for each fault is displayed and alarmed (S60) and the next operation is performed (S70).

In the neurofuge network of the present invention, fault data are calculated as shown in FIGS. 8A to 8C according to the type of the fault using the EMTP (Electrical Magnetic Transient Program), that is, in case of a phase A fault, shown in FIG. 8A. As shown, the inference value of the neuro purge network was obtained for the effective values of the phase currents and the image currents, and in the case of the B phase and C phase ground faults, as shown in FIGS. 8B and 8C, Inference values of the neurofuge network for the effective values were obtained.

In case of a two-phase ground fault, the fault data as shown in FIGS. 9A to 9C are calculated, that is, in the case of an AB phase ground fault, as shown in FIG. 9A, the effective values of the respective phase currents and the image current are shown. Inference values of the neurofuge network were obtained, and in case of the BC and CA phase ground faults, as shown in FIGS. 9B and 9C, the inference values of the neurofuge network were obtained for the effective values of the phase currents and the image currents.

In addition, in the case of a line short circuit fault, the fault data as shown in FIGS. 10A to 10C are calculated, that is, in the case of an AB phase short circuit fault, as shown in FIG. 10A, a neuron for the effective value of each phase current and the image current is shown. Inferred values of the fuzzy network were obtained, and in case of short-circuit failures of the BC and CA phases, the inferred values of the neurofuge network were obtained for the effective values of the phase currents and the image currents, as shown in FIGS. 10B and 10C.

In addition, in the case of a three-phase failure, as shown in Fig. 11, the inferred value of the neurofuge network with respect to the effective value of each phase current and image current was obtained.

Also, in the case of a high resistance ground fault, the fault data as shown in FIGS. 12A to 12C are calculated. In other words, in the case of an A phase high resistance ground fault, as shown in FIG. 12A, the effective value of each phase current and an image current is shown. Inferred values of the neurofuge network are obtained, and in case of B and C phase high resistance ground faults, as shown in FIGS. 12B and 12C, the neurofuge network inferred values for the effective values of the phase currents and the image currents are shown. Got.

As such, the inference values of the neurofuge network for each type of failure are shown in the diagram of FIG. 13.

Therefore, as described above, the present invention detects each phase current, calculates the image current, calculates the image current, and then inputs the effective values of each phase current and the image current to four neurofuge networks for each, thereby operating the power system. In addition to the rapid detection of the fault, it is possible to accurately determine the type of failure, thereby improving the supply reliability of the transmission system.

In addition, by detecting each phase current in real time and combining the algorithm of the neurofuge network with an existing relay or the like, there is an advantage in that the hardware can be easily implemented and the fault detection device can be manufactured at low cost.

Claims (7)

A first step of detecting phase current of a line supplying three-phase power; A second step of calculating an image current using each phase current; A third step of calculating an effective value of each phase current and an image current; A fourth step of performing the input of the effective values of the phase current and the image current to four neurofuge networks; And a fifth step of determining a type of a failure and an alarm when the execution of the neurofuge network of the fourth step is completed, and a method for detecting the failure of the power system using the neurofuge network. The neurofuge network of claim 4, wherein A sixth step of outputting a membership function value by inputting an effective value of each phase current and an image current; A seventh step of calculating a membership function value and outputting a fuzzy operation value; An eighth step of normalizing the fuzzy operation value according to the active intensity of the corresponding rule and outputting a normalized value of the active intensity; A ninth step of inferring a normalized value of the active intensity and outputting a non-fuzzy value; And a tenth step of outputting an inference value by performing an addition operation on each of the non-fuzzy values, and detecting a failure of the power system using the neurofuge network. The membership function of claim 6, wherein It consists of three small, medium, and big membership functions defined by the following Equation 2, and the effective values of each of the phase currents I _A , I _B , I _C and the image current Io are respectively small, medium, Inputted to a big membership function, wherein the sixth step is a formula (3) characterized in that the fault detection of the power system using a neurofuge network and its method. here, Is the output of the sixth step, i is the output at each node (i.e. i = 1,2,3) The method of claim 2, wherein the seventh step comprises the following equation (5). The method of claim 2, wherein the eighth step comprises the following Equation (6). The method of claim 2, wherein the ninth step is performed by the following equation (7). Where P _i is a later set of parameters. The method of claim 2, wherein the tenth step consists of the following equation (8).