JP3304452B2 - Restoration method of current and phase distribution in case of transmission line current sensor failure - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for restoring a current / phase distribution in a transmission line fault section detection system when a fault occurs when a ground wire current sensor fails or is abnormal. The purpose of restoring the current / phase distribution is to accurately locate the fault section.

[0002]

2. Description of the Related Art Many current detection sensors are provided on a transmission line. If an accident occurs in the transmission line, the accident section can be specified from the distribution of the current value of the current detection sensor. Here, the section means a line segment between adjacent current detection sensors. When there is an accident section between two adjacent current detection sensors, the section sandwiched by these sensors is called an accident section. Identifying the accident section is called accident section location.

If all the towers are provided with current detection sensors and all the current detection sensors are normal, the fault section can be located without error. Regarding this, there is a technique for obtaining an accident section from a current / phase distribution of a current detection sensor of all towers. However, if some of the current detection sensors are faulty or abnormal, the values of these current detection sensors cannot be used to locate an accident section. Conventionally, there has been no system capable of locating an accident section when a current detection sensor has failed or is abnormal. Therefore, there is no prior art corresponding to the present invention.

[0004]

Conventionally, current detection sensors are provided in all towers, or current detection sensors are provided in the number of towers required for orientation, and when all the current detection sensors are normal, In addition, the accident section was accurately located. However, if the current detection sensor is abnormal or the transmission device is abnormal and the central device cannot collect accurate current / phase information at the time of the accident, accurate accident section location cannot be performed. The present invention solves such a problem, and relates to a method for restoring a current / phase distribution of a transmission line, which can accurately perform fault section location even when some current detection sensors are abnormal.

[0005]

According to the present invention, information including missing data collected by a central unit is calculated in advance by simulation, and all types of expected accidents and accidents at all accident occurrence locations are calculated. By using a neural network in which the current / phase distribution at the time has been learned, the data of the missing portion is restored. First, the current and phase values, which are the outputs of the current detection sensors, are used as the values of each unit in the input and output layers, and the current and phase distribution at the time of an accident are input to the units in the input layer. Construct a neural network to be output to the output layer. A news that the output layer becomes equal to the input layer under the condition of failure
Lal network. However, even though the output layer and the input layer are equal, they are not exactly equal but within a certain allowable range.

Further, if some of the current detection sensors are faulty, it is known which of the current detection sensors is faulty. Therefore, it is also possible to know which current detection sensor is normal. In the event of an accident, the correct current and phase of the tower to which it is attached are required from a normal current detection sensor.
Therefore, the current / phase information of these normal sensors is
Input to the input layer of the ral network.

[0007] The information of the failed sensor is missing, but this is tentatively determined by interpolating the current / phase values of the closest normal sensor and input to the corresponding unit of the input layer. Performs neural network calculations. This is the first calculation. The output of the output layer corresponding to the normal sensor is used for determining the convergence. The output layer corresponding to the fault sensor is returned to the input layer and input to the corresponding unit. The unit in the input layer corresponding to the normal sensor inputs the measurement value of the normal sensor again. Thus the second
Make the second calculation. Then, the calculation of the neural network is repeated for the third and fourth times. At this time,
The current / phase information of the normal sensor is used as it is. It uses the same value no matter how many times it is calculated. This is a true value because it is a measured value.

On the other hand, a fault sensor is corrected for each calculation. The calculation is performed by returning the output result of the previous output layer to the corresponding input unit of the next input layer. As described above, the technique of the present invention is to obtain the current / phase distribution of the fault sensor asymptotically by calculation. If the process is repeated several times, the value of the unit in the output layer converges to a constant value. This means that the current and phase distributions corresponding to all the sensors have been obtained. Based on this, the accident section can be located.

The above is the outline of the present invention. Next, the neural network will be described in more detail.
As shown in FIGS. 5, 6, and 7, the neural network
The block has a plurality of input layers, intermediate layers, and output layers, which are connected by weight coefficients. The input layer has n units, which can take one value. The output layer also has n units. This also has one value. There are also some units in the middle layer. This is often n or less. When the number of intermediate layers is large, the number of parameters is large, so the calculation takes time. However, on the other hand, a lot of accident information can be included.

A unit is a unit of calculation and is sometimes called a neuron. It receives the sum of the output from the unit of the immediately preceding layer multiplied by the weighting factor as input, binarizes it, and outputs one output value. This output is the input of several neurons in the next layer. Simply outputting the product multiplied by the weighting factor merely creates a linear combination. However, this is further binarized and non-linearly output in Newlon.

Information propagates in a complicated manner, but a neural network in which information is propagated unidirectionally and in layers is called a multilayer type. There are other types of neural networks. A linear combination of the values of the several input layers in the previous stage enters the unit of the hidden layer. Also, a linear combination of several units of the middle layer enters the output layer. Since the unit of the intermediate layer and the output layer is one neuron, this input is binarized into 0 and 1 as performed by the neuron. However, in the neural network here, a step is taken from the meaning of the initial neuron, and the conversion function is not necessarily binarized, but can be an intermediate value between 0 and 1. This is called a sigmoid function. The input information is nonlinearized by a sigmoid function. This will be described later.

In a transmission line system, it is assumed that a number N of current detection sensors required for locating an accident section are mounted on a steel tower. The neural network used in the present invention uses the current / phase information of the current detection sensors of all the towers required for fault section location as an input layer and an output layer. The number of the intermediate layers may be smaller or the same. Since one current detection sensor has two parameters, current and phase, assuming that the number of towers providing the current and phase distribution required for fault location is N, the observable parameters of these towers are as follows. Is 2N.

First, a neural network of the present invention is constructed. For example, assume that the number of input layers and the number of output layers of the neural network are 2N (n = 2N). This corresponds to the current and the phase value of the current detection sensors of all the towers. That is, the information of the transmission line can be represented by the set {I _j , P _j } of the current I _j of the sensor S _j of the J-th tower and the phase P _j , but all of them are stored in the input layer and the output layer. Pretend Then, for all aspects of the accident (the type and location of the accident), the weight coefficients of the intermediate layer, the input layer, and the output layer are determined by trial and error so that the input layer and the output layer have almost the same value. I do. Determining the weighting factors is to construct a neural network.

What has been described above is a mixed neural network having both current and phase. This is a method of constructing and using one neural network. Instead, other methods are possible. One current detection sensor gives two measured values of current and phase (both are collectively called parameters), but there is a case where there is almost no correlation between current and phase. In such a case, separate neural networks can be constructed for current and phase information. In this case, the number of current detection sensors can be N, and the number of units in the input layer and output layer of the neural network can be N (n = N).

Assuming that the parameters of the input layer are Q _i and the parameters of the output layer are R _i , the parameter group {Q _i } takes a certain value in the event of an accident. On the other hand, the weight coefficient is determined so that the parameter group {R _i } of the output layer has the same value. A number of calculations are performed so that this holds for all accident modes O _k (type, location). The weighting factor is changed each time it is calculated, and approaches an appropriate value. Finally, for any accident,

{Q _i } {R _i } (arbitrary O _k ) (1)

The following holds. It is important to note that this equality only holds in the event of an accident. Otherwise it does not hold. If it is also true in a non-accident situation, simply the relation that the individual components are always equal (Q _i =
R _i ). The linear transformation matrix becomes the unit matrix. In this case, no neural network is required. Nor does it require that equal signs be exactly equal. It is good if they are almost equal.

[0018]

When several of the current detection sensors installed at a plurality of points on the transmission line fail, accurate current / phase information at the time of the accident is not transmitted to the central unit. Therefore, it was necessary to locate the accident section from the incomplete information collected. However, in the present invention, missing and abnormal data are restored to normal data from a known current / phase using a neural network. For this reason, even for an accident that occurs when the current detection sensor is abnormal or faulty, the fault section can be located with the same accuracy as when the sensor is normal. This will be described in more detail below.

A current detection sensor (CT) is provided on a steel tower and constantly measures a current value I and a phase P of a current flowing through an electric wire. However, under normal circumstances, these values are unnecessary.
The instantaneous current and phase values when an accident occurs are required. This is one value for the current and one value for the phase. Since the current is an alternating current, a current from a peak to a peak or an effective value is used as the current value.
In the event of an accident, the transmission will be interrupted because the circuit breaker will work. In other words, in a short time from the occurrence of an accident until the circuit breaker operates, the current detection sensor in each tower measures the above-described accident current,
Measure the phase.

Changes in current and phase depend on the type of accident. For example, in the case of a ground fault, the current immediately before the fault section increases. The current decreases before the accident section. The phase changes by almost 180 degrees before and after the accident section. In such an accident O _k , the current / phase information I
_j and P _j are obtained. This news - Rarunettowa - parameters of click input layer - entered as data Q _i. Then, although the value to the output layer R _i through the intermediate layer comes, this R _i is Q
_Make the value close to _i .

This is a unique point of the present invention. For all the accident modes {O _k }, R _i → Q _i . In all situations, this is not the case where the input layer is equal to the output layer, but only during an accident.

However, if the output and the input are simply connected by a linear transformation, the matrix relating the output and the input is simply an identity matrix, regardless of whether the number of accidents is large or small. If the number of accidents is small, there may be a possibility other than the identity matrix, but even in that case, it is not possible to select which is better. In other words, in a linear transformation such as a linear transformation, if the constraint that the output layer is equal to the input layer is imposed, the relation matrix becomes a unit matrix and is useless. This is because the effect of other input parameters cannot be taken into the value of each output. Therefore, we should not simply combine the input and output with a linear transformation.

One of the reasons the neural network is used in the present invention is its nonlinearity. If the relation between input and output is non-linear, even if the limitation of input = output is imposed in the event of an accident as described above, the relation connecting them is not affected by other parameters like a unit matrix. It doesn't matter. Even if the mode of the accident is larger than the number of parameters, the effects of many inputs can be included in one output if the relationship is nonlinear.

A multilayer network using an input layer, an intermediate layer, and an output layer will be described. The components of these layers may be units or neurons. FIG. 8 shows a conceptual diagram of the Neuron. The neurons are interconnected and exchange signals. In the present invention, a multilayer type is used. As described above, this is one-sided because the signal flow is performed in a layered manner. It flows from the input layer to the hidden layer and from the hidden layer to the output layer.

[0025] One of the Niyu - Ron N _j, a plurality of input X
₁ · · X _n receives, this sum is multiplied by a weight coefficient W _ij Y
_{Find j} . Outputting this to the next layer as it is does not differ from linear combination. Instead, the sum Y _j is made non-linear through the sigmoid function F (Y). In other words, what Newlon does is

[0026] _{Y j = ΣX i W ij (} 2) F (Y j) = (1 + e -Yj) -1 (3)

The operation is as follows. The former linear combination formula indicates that the neuron is affected by several preceding neurons. Neural network
Constructing a _matrix means determining a coefficient W _ij of a linear combination. In the present invention, since one intermediate layer is used, a neural network can be created by determining the coefficients between the input layer and the intermediate layer and the coefficients between the intermediate layer and the output layer. Thus, the determination of the coefficients is equivalent to creating a neural network, but a linear transformation is performed no matter how many linear combining stages are stacked. This does not allow for non-linear effects. A binarization function for nonlinearization is a sigmoid function.

FIG. 9 shows a schematic form of the sigmoid function. The sigmoid function associates an input with a value between 0 and 1 by a non-linear relationship, and is a simple binarization function in the original sense of a neuron. However, a simple binarization function is inconvenient because an intermediate value does not come out. Therefore, the binarization function is a symmetric function in which a tail is subtracted near 0 and 1. There may be several other than those shown in FIG. The value of the binarization function with respect to the input can be changed by selecting the weight coefficient W _ij , and any sigmoid function may be selected.

This F (Y _j ) is sent as an output of this neuron N _{j to} the next layer of neurons, and the input X _j = F (Y _j ) for these is obtained.

However, the reason why the sigmoid function has a form as shown in (3) is that the sigmoid function receives an input having an average value of 0,
In this case, an output in the range of 1 is generated. If the average value is not 0, the function falling within the shoulder of e in (3) is (Y _j -av)
Becomes Here, av is an average value.

To set the output range to a value other than 0 to 1,
What is necessary is just to multiply (3) by a coefficient and add a constant. The range of the input is not limited. This is because the input range can be effectively limited by selecting the coefficient. The reason for subtracting the average value is to make the output 0.5 here.

The neural network used in the present invention handles parameters twice as many as the total number N of towers required for fault location (since one sensor obtains current and phase). Therefore, the neural network includes the current and the phase (2N = n), and makes the number of input layers and output layers equal to the total number of observation values n. Alternatively, a current neural network and a phase neural network may be separately constructed so that n = N. It is easier to handle input and output values by defining ranges and normalizing them.
1] or -1 to +1 [-1, 1].

For example, since the phase is from -180 ° to 180 °, this is divided by 180 to set the variation to [-1, 1], or to add 0.5 over 360 to add [0, 1]. , 1]
To In the case of the current, for example, if the maximum value is 1000 V, the actual current value is divided by 1000 to set the variation to [0,
1]. These are only procedures to facilitate handling.

The output of the i-th neuron of the input layer is x _i
(I = 1,... N). Let s be the number of intermediate layers. The input of the h-th neuron of the intermediate layer is represented by Y _h , and the output is represented by y _h . Here, the intermediate layer is described as one layer, but there may be two or more layers. The input of the j-th neuron of the output layer is denoted by Z _j , and the output is denoted by z _j . There is the following relationship between them.

Y _h = Σ _{i = 1} ⁿ x _i W _ih (h = 1, 2,... S) (4)

Y _h = (1 + e ^−Yh ) ⁻¹ (5)

[0037] ^{_{Z j = Σ h = 1 s}} y h U hj (j = 1,2, ·· n) (6)

Z _j = (1 + e ^−Zj ) ⁻¹ (7)

Here, the subscripts i = 1 and n of Σ mean that i = 1 to n are added. In JIS, subscripts cannot be added at the top and bottom, so they are added at the end.

Although there is such a relationship, the weighting factors W _ih and U _hj are constants, but are determined so as to provide a desired result. According to the present invention, in all faults, when all parameters of fault current and phase are given to the input layer, these weighting factors W _ih , U _i , U Determine _hj .

This is not something that can be obtained by solving the equation. That is, they cannot be uniquely determined, and even if input = output, they are not exactly equal. Calculate repeatedly using a calculator. The learning is stopped when the double error between the desired output result and the output from the network falls within a satisfactory range. Learning means obtaining an optimal weighting factor.

In the above (4) to (7), if the input and output of each neuron are assumed to be equal without using the sigmoid functions (5) and (7), even if there is an intermediate layer, it is simple. With the limitation of output = input, which is only a linear transformation, the weighting factor can be determined accurately by solving the equation, but this is often meaningless and becomes meaningless.

[1. Determination of Weight Coefficient] First, a neural network is constructed. In all accidents, input =
Determine the weighting factor so that it becomes an output. For example, assuming various weighting factors, the above calculation may be performed for one accident parameter, the output may be returned to the input, and further calculation may be performed to obtain the output. . By performing such repetition, the weight coefficient is corrected so that the output solution converges. Since this is performed for the parameter input at the time of all accidents, the amount of calculation is enormous.

Since a certain degree of magnitude relationship between the weighting factors can be estimated from physical considerations and experience, it is possible to start from here. Starting from the calculation, the difference between the output and the input is checked, the weight coefficient is corrected, and the output gradually approaches the input.

When the fault current I _i and the phase P _j of the i-th current detection sensor for all the fault _modes are input to the input layer, almost the same current and phase appear in the output layer. Thus, the weighting factors _Wih and _Uhj are determined. The above is the preparation stage of the neural network of the present invention.

The description so far is based on the assumption that one sensor is attached to all necessary towers. However, if sensors are provided in all the towers, the above neural network is of course unnecessary. Since some of the sensors have failed and there are not enough normal sensors, the above-described neural network is required.

In the above description, the sensor attached to the tower is
Since two pieces of information, current and phase, are measured, the number of towers is N
Then, the number n of neurons in the input layer becomes 2N. However, the current and the phase information of the sensor do not actually correlate with each other. Then, what connects the current and the phase neuron with the weighting coefficient can be set to 0 a priori. If so, the current and the phase can be handled from separate neural networks from the beginning. Of course, a neural network may be constructed in which current and phase are mixed and handled.
The present invention can be realized by any of the methods.

[2. Calculation at the time of an accident] The neural network obtained above gives the current and phase information at the time of the accident to the corresponding neuron of the input layer, and the corresponding neuron of the output layer has the same There is a property that current / phase information appears. There is doubt as to whether such a thing would be useful, but it is significant because it involves the tendency of the parameters at the time of the accident. This means that the output is equal to the input only in the event of an accident, rather than simply passing the input to the output.

Now, let N be the number of all towers required for orientation, and install a current detection sensor in these towers. However, (N
-M) Assume that the sensors are faulty or abnormal. It is assumed that the sensor malfunction is already known by other means. This can be seen by monitoring the sensor signals at the central unit. Let M be the number of normal sensors. The current detection sensor measures the current and phase information of the transmission line,
This is stored in the memory for a certain period of time. However, since the current and phase information in the normal state is unnecessary, the previously stored information is discarded. In the event of an accident, current / phase information is left in the memory. This can be done automatically if the memory is stored on a first-in first-out basis.

When an accident occurs, the circuit breaker works immediately, so that no current flows through the transmission line. The current and phase in a short time from the occurrence of an accident until the circuit breaker operates are referred to as the current and the phase at the time of the accident. This is a clear value I, P given for each sensor. Remains in memory. This is taken out and used as an input value of the neural network. Collectively write {K _h }. The values obtained directly from these sensors are placed directly in the corresponding neurons of the input layer {X _i }.

However, the current of the tower where the sensor has failed
Phase information cannot be obtained. The problem is how to do this. The current phase information of the tower that is not measured by the sensor due to a failure is represented by _Hg . This is also a parameter belonging to the input {X _i } of the input layer. I mean

{X _i } = {K _h } + {H _g } (8)

Can be written as Here, the former is a known parameter measured by a sensor. The latter is an unknown parameter which is not measured due to a failure. At the time of an accident, current / phase information is given from the current detection sensor, but this is only data corresponding to the former {K _h }. The latter {H _g } has no data. Unless inputs are provided for all input layers, a neural network cannot be operated. Therefore, the neural network at the time of failure
For the first time in the calculation of the peaks, {H _g } starts from an appropriate number. A certain initial value is made to correspond to these {H _g } to obtain input parameters {X _i } of all input layers. This is put into the neural network. The calculation is automatically performed by the computer, the parameters of the intermediate layer and the output layer are obtained, and the output {z _i } is obtained in the output layer. This completes the first calculation.

The second calculation is performed immediately, but the parameters of the input layer are slightly different from those of the first calculation. For the known parameter {K _h }, the same measured value as the previous one is used. This is immutable. However, the parameter that could not be measured ｛H
_{For g} ｝, substitute the output in the corresponding order in the output layer. That is, {z _g } → {H _g }. Thus, the second neural network is calculated.

The second calculation is performed to obtain the output {z _i } of the second output layer. Next, the third calculation is performed. For the known input layer input {K _h }, the measured value is substituted as in the previous time. However, the unknown input {H _g } is the corresponding output {z _g } of the second output layer. That is, {z _g } → {H _g }. Similarly, unknown parameters are substituted with the previous calculation results of the output layer. The known parameter always uses the measured value as it is. That is, the output of the output layer of the known parameter is unnecessary except for the determination of the convergence.

The calculation is performed several times in this manner. Each parameter of the output layer gradually converges to a certain value. In the case of an accident, only such calculations are performed, but the number of calculations can be reduced by selecting a neural network so that it does not take much time.

This is where the calculation is stopped. This is because the output of the output layer converges to a certain finite definite value. , The calculation may be repeatedly stopped. Alternatively, the number of calculations may be determined, and the calculation may be stopped when the number of calculations is reached. If the solution of the output layer converges uniformly, it will be easy to determine the number of calculations. Since it is often not easy to determine the uniform convergence, the output value of the normal sensor converges to the input value (constant value of the measured value). Abort.

Of course, the speed of convergence depends on how the initial value is selected. Of course, all unknown input layer parameters may start from 0 or 1, but this is not always the optimal choice. More preferably, the unknown parameter
The first value (initial value) of the parameter is
It would be better to take an interpolated value of the data (measured value). For example the unknown parameters - data H _i is known next parameter - data K
When we have _i-1 and K _{i + 1,} as a simple average of

H _i = (K _i−1 + K _{i + 1} ) / 2 (9)

It is assumed that: Alternatively, if the number of sensors is smaller and there are no adjacent known parameters, the values of the nearest known parameters on both sides are interpolated to obtain the initial values of the unknown parameters. This is just the first initial value story.
From the second time on, the unknown parameters use the output of the previous output layer as it is ({z _g } → {H _g }).
There is no such problem.

As a result of repeating such neural network calculation several times, an output of the output layer is obtained. This gives the current and phase information of all towers at the time of an accident. For known parameters, the values as measured by the sensors are given. The unknown parameters relating to the current and phase of the tower in which the sensor has failed can be obtained by this calculation. The purpose of the present invention is here. In this way, the current and phase information of all the towers required for fault location can be obtained.

If the current and phase information of all the towers required for the location is known, a technique for accurately performing the fault section location has already been established. This was mentioned earlier. The unknown parameter according to the present invention
Data, you can find all parameters,
Accident section location can be performed. Accident section location must be performed in a short time, but the neural network of the present invention
Since the repetitive calculation of the clock can be performed instantaneously by a computer, it is well suited for the purpose.

The neural network of the present invention takes in all the current and phase information at the time of an accident and selects a weighting factor so that the input and output of the input layer coincide with the output at the time of the accident. It can be said that the accident coefficient information is included in the weight coefficient. In the case of linear transformation, if the condition of equalizing the output and the input is set, the transformation matrix will inevitably become an identity matrix.However, since nonlinear transformation is used here, it is necessary to incorporate the accident information into the weighting factor. You can.

[0064]

FIG. 1 shows a basic configuration of a system according to an embodiment of the present invention. The system consists of a current detection sensor installed at multiple points on the transmission line, a sensor signal transmission device that transmits the signal to a central device installed at the substation, and a central device that determines the accident section from the collected sensor signals. Be composed. The current detection sensors are installed on the tower as required for fault location. If necessary, they may be installed on all towers. It is assumed that there are N current detection sensors, some of which are out of order, and which one is out of order. There are only M normal current detection sensors. It is assumed that an accident section cannot be located with M sensor signals.

The normal current detection sensor is always operating and transmits detected current information to the central unit. The central unit does not use those sensor signals when no accident occurs, but for a certain period of time (for example, 1 second: 50 seconds).
(Equivalent to 60 cycles or 60 cycles) Discard oldest ones while storing them in memory. When an accident occurs in the transmission line (that is, when the circuit breaker operates), the central device knows that the accident has occurred by, for example, detecting a sudden change in the waveform by setting a certain threshold, and for a certain period of time (for example, 10 cycles). After that, the writing to the memory is stopped.

Therefore, in the memory, as shown in FIG. 2, 10 cycles before the occurrence of the accident until the occurrence of the accident, several cycles from the occurrence of the accident to the interruption (3 cycles in the figure), and several tens of cycles after the interruption.
This means that all sensor information for a total of 50 cycles remains. The signal before the accident and the signal after the circuit breaker has been activated are unnecessary and are discarded. Current / phase information is obtained from the remaining three cycle signals at the time of the accident. The signal at the time of the accident stored in the memory is an AC signal, but the current value is obtained by taking an effective value or a peak value. Although the phase can be specified because of the AC signal, the phase of this sensor can be obtained by comparing with the phase of the reference AC signal.
For each sensor, two parameters, current and phase, are required.

When all sensors are normal, the central unit extracts the current value and phase distribution of the entire transmission line from the sensor information stored in these memories, and it is possible that the most accident has occurred. It is possible to find a section with a high probability.
However, if some of the sensors in the current / phase distribution of the entire transmission line are faulty and their data show abnormal values or are missing, the section where the accident occurred can be accurately identified. Cannot be oriented. In such a case, by applying the method of the present invention, an unknown current / phase distribution can be obtained, and the entire current / phase distribution can be reproduced. As a result, the accident section can be accurately located.

A description will now be given of a network for restoring the current / phase distribution. It is assumed that n current detection sensors are installed on a target transmission line. It is assumed that some of them are out of order. It is assumed that which one has a failure is known in advance from a current / phase signal distribution in a normal state. In the event of an accident, as described above, the circuit breaker works and power transmission is stopped. A current / phase distribution signal at the time of an accident can be obtained from the signal left in the memory retroactively from this. In this way, the current distribution and the phase distribution at the time of an accident can be calculated for the n current detection sensors installed on the transmission line. FIG. 3 is a current distribution diagram. The horizontal axis represents the number of the current detection sensor, and the vertical axis represents the current value. The current increases around the mth. It is highly probable that an accident occurred in the vicinity. FIG. 4 shows the phase distribution at the time of the accident. The horizontal axis is the number of the sensor. The vertical axis is the phase. Since it is a phase, it is in the range of -180 to 180 or 0 to 360. Some sensors have failed and there are no signals from them. The current / phase distribution information obtained by the current detection sensor is used as the input of the neural network.

Here, the current distribution shown in FIG.
[1, 1] or [0, 1]. As already mentioned, normalization is generally performed on the input of the neural network. For example, input data
The maximum value in the data is set to 1 (in the case of a phase, it is a positive / negative value, so each is set to +1, -1), or a predetermined value (for example, 1000 [A] is set to 1 or ± 180 of the phase). °
Is set to ± 1), and the input value is set to 0 to 1 or -1.
Normalize to a value of ~ 1. The current and phase information at the time of an accident at a tower where the current detection sensor has failed by inputting to the constructed network and performing repeated calculations with the following configuration and procedure based on each normalized information Is estimated.

[Structure of Network] Here, a multilayer perceptron type network was used (FIG. 5). It consists of an input layer, a middle layer, and an output layer, and the flow of operation is unilateral. Each has several units. The number of units in the input layer and the output layer was the same (n) and matched with the number of current detection sensors installed on the transmission line. In other words, this constructs a current neural network and a phase neural network separately. An input layer unit is connected to a plurality of intermediate layer units. A plurality of units in the intermediate layer are connected to a plurality of units in the output layer. The thin line may be considered as a weighting factor between these outputs and inputs. In each unit which is a neuron, a linear combination of the parameters from the preceding stage is calculated and binarized through a sigmoid function. This is sent to the next unit as an output. The output layer is also linearly combined and passed through a sigmoid function. About the number of units of the intermediate layer,
The network learning was set to such an extent that the learning progressed smoothly.

[Acquisition of Weight Coefficients by Learning] If possible, simulations are performed for all accident types and accident occurrence locations if possible, and the current and phase in each accident case in n towers required for orientation. Find the distribution. Using such information, a network is constructed to faithfully output the input distribution from the output layer in the event of an accident.
This network is such that the output becomes the input in the event of an accident.

That is, teacher data (consisting of a pair of an input (input data) and a desired output result (output data) given the input data) used for network learning is as follows. The input data and the output data are completely identical. FIG. 6 illustrates this. When the current / phase distribution in each fault section is input to the input layer by adjusting the repetition weight coefficient, the current / phase distribution at the time of the fault appears in the corresponding output of the output layer as it is. This is to determine the weight coefficients W _ij and U _jk . The weight coefficient is obtained by using, for example, a back propagation algorithm.

[Method of Estimating Current Information at the Time of Fault When Current Detection Sensor Fails (Abnormal)] When some of the sensors are faulty, at the time of a fault, unknown currents / phases can be obtained from known current / phase distributions. It needs to be estimated. Therefore, the neural network of the present invention is used. In the event of an accident, the current / phase distribution from the normal sensor is input to the input layer. Since some of the sensors are out of order, the data corresponding to the current and the phase is missing or has an abnormal value. As an input to the unit corresponding to the faulty current detection sensor, for example, an average value of data from both adjacent sensors sandwiching the faulty current detection sensor is used. FIG. 7 shows this. ○ indicates a normal sensor, and ● indicates a failed sensor. The current and phase for the failed sensor are input to the input layer of the network by substituting the average value of the current and phase value of the adjacent sensor.

The first operation of the neural network is performed. Outputs corresponding to the individual sensors are taken from the output layer. The output corresponding to a normal sensor is approximately equal to the input value. This is done so that the input and output layers are equal.
Naturally, we are building a ral network. However, because the input to the fault sensor is incorrect,
The input and output of the normal sensor do not fall within the predetermined range. The output of the normal sensor is used to determine the convergence of the calculation result, but is not returned to the input layer repeatedly. That is, the output of the output layer for the normal sensor is discarded. In the second and subsequent calculations, the output value from the unit corresponding to the failed sensor is the same as that of the same parameter of the same sensor.
Back to the input layer corresponding to the The values of the current and phase information of the normal sensor are directly input to the corresponding unit as in the previous case. Such calculations are repeated many times.

That is, for the input of the normal sensor, the actual measurement value is input to the corresponding unit of the input layer, and for the input corresponding to the failure sensor, the result output to the previous output layer is returned and input. When such repetitive calculations are performed, the output value corresponding to the current / phase value of the fault sensor gradually changes, as shown in FIG. In FIG. 7, the output value of ● is indicated by a broken line, but as the calculation proceeds for the first, second, and third times, the current and phase values approach the values at the time of the accident.

What should be noted here is the value of the output layer for the normal sensor. From the beginning, in this neural network, when the current and phase values are input to the input layer at the time of an accident, almost the same values appear in the output layer. Therefore, when the input value to the failed sensor is far from the actual value, the output value of the normal sensor and the input value should greatly differ. Then, as the output value for the failed sensor approaches the true value, the output value and the input value of the normal sensor approach. The output value of the normal sensor is not simply discarded, but if it converges to the input value (measured value), it means that the output value for the failed sensor is close to the true value. The measured value of the known input parameter {K _h } obtained from the normal sensor is defined as K _hs . In the neural network calculation,
The n-th input is K _h ⁱⁿ and the n-th output is K _h ^on .
For normal sensors, regardless of the number of times,

K _h ⁱⁿ = K _hs (10)

Assume that This output will not be used next time,
Use in the determination of convergence S _n.

S _n = Σh _{= 1} ^m (K _h ^on −K _hs ) ² (11)

Here, the integration is performed for all normal sensors (h =
1, 2,... M). This is a sum of squares, but may be a sum of absolute values.

S _n = Σh _{= 1} ^m | K _h ^on −K _hs | (12)

Since the error defined in this way decreases by repeating the calculation (n increases), a certain appropriate positive number ε is determined.

S _n <ε (13)

The calculation is stopped when As for the parameter {H _g } for the faulty sensor, the first input value is the average value of the adjacent normal sensors as shown in equation (9) (of course, other values may be used). From the second time, the previous output value is used. The input value of the n-th time is H _g ⁱⁿ ,
If the output value is H _g ^on ,

H _g ⁱ¹ = appropriate constant (14) H _g ^{in + 1} = H _g ^on (15)

Assume that: If this converges uniformly,
If an appropriate N is selected, for any positive number ε,

When n, q> N, | H _g ^on −H _g ^oq | <ε (16)

Is always satisfied. This seems to be able to determine the convergence of the solution. Although the convergence can be determined by this, it is not easy. (11),
The convergence of the known parameter (12) is easier to calculate. When the known parameters converge, the unknown parameters
Should also converge. That is, regarding the known parameters, if (11) and (12) can be said, the unknown parameters are also close to the true values. Conversely, if the output value of the normal sensor is far from the input value, it means that the output value of the failed sensor is still incorrect. As described above, the values of the input layer and the output layer of the normal sensor are important because they are a measure of the approximate correctness.

That is, what is to be said here is whether to determine the convergence of the data of the normal sensor or the data of the failed sensor. Since we want to find the parameters for the fault sensor,
If this converges to a certain value, I think that is fine, but this is not easy. Since the true value of the normal sensor is known in advance, the difference from the output can be calculated immediately. The fact that this differs from the input value means that the neural network
Means that the correct value for the failed sensor has not yet been obtained by the calculation of the fault. Such error complementarity between two different data sets is unique to neural networks.

The calculation of the neural network is repeated, and the calculation is stopped when the output of the normal sensor converges on the input. At this time, the current and phase for the fault sensor appear in the corresponding output layer of the neural network. When these are combined, the current and phase distributions of all sensors required for fault section location are obtained. From this, it is possible to accurately determine the accident section.

[0091]

As described above, according to the present invention, it is possible to restore the information of the abnormal sensor even in the event of an abnormality or failure of the ground wire current sensor attached to the power transmission line. The fault section of the transmission line can be located with almost the same accuracy as when the detection sensor is normal. The reliability of the transmission line accident section location system can be greatly improved.

[Brief description of the drawings]

FIG. 1 is a schematic diagram of a transmission line accident section locating system.

FIG. 2 is a schematic diagram of all sensor information at the time of an accident.

FIG. 3 is a diagram showing a current distribution at the time of an accident extracted by a central device.

FIG. 4 is a diagram showing a phase distribution at the time of an accident extracted by a central device.

FIG. 5 is a configuration diagram of a neural network used in the present invention.

FIG. 6 is a conceptual diagram of a neural network construction.

FIG. 7 is a diagram showing a process of estimating missing data using normal sensor information and restoring a current distribution at the time of an accident.

FIG. 8 is a conceptual diagram simply showing the operation of a neuron.

FIG. 9 is a schematic diagram of a sigmoid function.

Claims (1)

(57) [Claims] 1. An accidental current / phase flowing through an overhead ground line at the time of an accident of an overhead power transmission line is detected by current detection sensors attached to a plurality of locations of the overhead ground line, and the detected data is transmitted to a central device via a transmission device. It has a means for measuring the current and phase distribution to be transmitted to the equipment.
This is an accident section locating method for finding the section of the accident from the phase data.
When the phase information is given to the input layer, it has a neural network that outputs the same value as the input layer to the output layer. When the current and phase information at the time of an accident at all accident locations are input to the input layer, it has been learned to output the same value as that of the input layer to the output layer. If some of the current and phase information transmitted to the central unit is missing or abnormal due to a failure or transmission device failure, if the current and phase information is not missing, the current The detected current / phase information is supplied as an initial value to the corresponding input layer, and the current / phase information is missing or the current detection sensor that has output abnormal data. As the initial value, an arbitrary value in the variation range or, for example, the current / phase information of the current detection sensor in which the current / phase information of the nearest current / phase information is not missing and which is not abnormal, or the average value, The current and phase information is obtained in the output layer, and then the input layer corresponding to the current detection sensor in which the current and phase information is not missing is the same as the previous calculation. The current / phase information actually detected is given to the corresponding input layer, and for the current detection sensor with missing current / phase information and the current detection sensor that output abnormal data, the current obtained in the output layer in the previous calculation Given the phase information, perform the neural network operation again, and repeat the neural network operation in the same manner until the output layer value becomes a constant value or a predetermined operation. When the number is exceeded, the calculation is terminated, the current / phase information of the current detection sensor that lacks current / phase information is obtained, and the current that should have been generated in the normal case of the current detection sensor that generated abnormal data A method for restoring the current / phase distribution in the event of a transmission line current sensor failure, which obtains phase information and restores the distribution of current / phase information in the event of a transmission line fault.