KR20080074441A - Method for diagnosis and analysis of electric power quality using artificial intelligence - Google Patents

Abstract

A method for diagnosing and analyzing the quality of electric power by using artificial intelligence is provided to monitor a present state of an electrical signal in real time by identifying automatically power disturbance through a neural network. A method for diagnosing and analyzing the quality of input power is performed by using a power computer including a wavelet converter, a feature vector extractor, and a disturbance identifier. A feature vector of a voltage disturbance or a harmonic disturbance is extracted by wavelet-converting a waveform of an input voltage or a waveform of input current. A disturbance feature of the input voltage or the input current is identified by applying the feature vector of the voltage disturbance or the harmonic disturbance to an artificial intelligence neural network identifier. The waveform and the disturbance feature of the identified input voltage or current are displayed in a graphic manner.

Description

Method for diagnosis and analysis of electric power quality using artificial intelligence}

1 is a block diagram of an electrical quality diagnostic analysis system according to the present invention.

2 is a flowchart of a power disturbance automatic identification algorithm for electrical quality diagnostic analysis of the present invention.

Figure 3a shows an example of easy to discriminate on the linear discrimination analysis for the electrical quality diagnostic analysis of the present invention, Figure 3b shows an example that the determination is not easy.

Figure 4 shows an example of dimensional reduction by linear discrimination analysis for electrical quality diagnostic analysis of the present invention.

5 is a flowchart of a voltage disturbance detection algorithm for electrical quality diagnostic analysis of the present invention.

6 is a flowchart of a harmonic disturbance detection algorithm according to the present invention.

7 is a structural diagram of an artificial neural network of an identifier for an electrical quality diagnostic analysis of the present invention.

8 is an explanatory view of the basic structure and the forward path of the neural network applied to the present invention.

9 is a diagram illustrating an error back propagation process of a generalized delta rule of a neural network to which the present invention is applied.

10 is an overall flow chart of the forward and reverse directions of the neural network applied to the present invention.

11 shows one type of main graph for electrical quality diagnostic analysis displayed on the display of the present invention.

FIG. 12 shows a type of the main measurement information simulator screen for the non-expert for the electrical quality diagnosis result expressed through the display of the present invention.

FIG. 13 shows a type of power disturbance automatic identification screen through an artificial neural network algorithm displayed through the display of the present invention.

14 illustrates one type of event occurrence information expressed through the display of the present invention.

FIG. 15 is a sub-graph displayed through the display of the present invention and shows one type of event screen of one-phase signal among three-phase signals.

16 is a graph showing Fourier transform data for a harmonic signal of real-time three-phase power expressed through a display of the present invention.

* Description of the symbols for the main parts of the drawings *

10: power computer 11: wavelet converter

12: Feature Vector Extractor 13: Disturbance Identifier

14 memory 20 display

30: communication module

The present invention uses intelligent neural network to automatically identify the disturbance of the incoming power of the switchgear through a mathematical calculation to obtain the basic power parameter value and to analyze the intelligent electrical quality diagnosis based on the real-time monitoring of the current electrical signal state It is about a method.

As is well known, power quality refers to the degree to which the supply voltage and current deviate from the ideal sinusoidal wave, which affects instantaneous variations in voltage or current and voltage or current for a relatively long time. It is divided into that the fluctuation occurs in the waveform of or the noise of harmonic or low frequency (steady state variation).

Here, instantaneous voltage fluctuations (sag, swell, interruption) and transient voltage (transient) belong to the former, and harmonics, overvoltage and undervoltage belong to the latter.

The electrical quality diagnosis field, which analyzes electrical quality and determines and predicts its stability, is a very interesting and important field that has been in the spotlight since the late 80s. In particular, during this period, the power industry was deregulated and high-performance equipment and precision With the development of high-end systems, consumers using power energy have begun to demand high-quality power supplies that allow their equipment or equipment to continue to operate reliably.

Since then, high-efficiency power electronic devices, high-precision special motors, and application semiconductor devices have been developed in the 1990s, and microprocessor-based automation equipment, office equipment, information and communication equipment have been rapidly developed, In every field of use, the electrical quality problem has begun to attract new attention. Accordingly, efforts are being made in the industry to develop high performance and high accuracy diagnostic analysis systems that can easily and intuitively analyze electrical quality.

An object of the present invention is to provide an electrical quality diagnostic analysis method to automatically identify the power disturbance through the artificial neural network in accordance with the demands of the times and to monitor the current electrical signal status in real time through the display. have.

In particular, the present invention is widely used, unlike the conventional method using a special neural network, and using a three-layer neural network, by limiting the power parameter values obtained through mathematical operations to increase the precision, as a back-propagation learning algorithm In order to detect the disturbance feature using the back propagation algorithm, we perform the fast Fourier transform and discrete wavelet transform, which is excellent for the multi-resolution analysis of the signal. There is.

Electrical quality diagnostic analysis method of the present invention for achieving the above object is a waveform of the input voltage or input current to diagnose the quality of the incoming power using a power computer including a wavelet converter, a feature vector extractor and a disturbance identifier Extracting the feature vector of the voltage waveform or the current waveform by wavelet transforming them, and substituting the extracted characteristic vector of the voltage disturbance or harmonic disturbance into the artificial intelligence neural network identifier, respectively, Identifying from the learning data of the identifier and displaying the waveform and disturbance characteristics of the identified input voltage or current in a graphic form.

Hereinafter, the present invention will be described with reference to the accompanying drawings.

1 is a block diagram of an electrical quality diagnosis and analysis system according to the present invention, wherein the voltage value and current value picked up from an incoming power line such as a switchgear are included in a wavelet converter 11, a feature vector extractor 12, and a disturbance identifier ( 13) and a result value and various signals which are configured to be applied to the power computer 10 having the memory 14, and the feature vector extracted by the wavelet transform in the power computer 10 is identified by the artificial intelligence of the neural network. The waveform and event information are configured to be displayed on the display 20, and the power computer 10 is configured to be connected to a control panel or a control computer at a remote location through the communication module 30.

Basic power parameters required for mathematical operation processing for power quality diagnostic analysis in the above system block are as follows.

<RMS>

The first value to monitor for power quality measurements is the RMS value. This RMS value is obtained by calculating the rms values of voltage and current, which can be used to detect RMS variation events such as Sag and Swell.

The RMS value is calculated using original data of digitally sampled voltage and current, and the formula is as follows.

(식 1)

(Equation 1)

In the present invention, the effective value obtained by Equation 1 was applied to the judgment of Sag and Swell.

Sag has a duration of about 1 minute at 0.5 cycle, and it is applied to judge Sag when it is 0.9 p.u. at 0.1 p.u. (per / unit) based on effective value. In the case of swell, it has a duration of about 1 minute at 0.5 cycle, and it is applied to judge the swell when the effective value is 1.1 p.u. to 1.8 p.u.

<Unbalance>

Unbalance refers to a condition in which the three phases of voltage or current are not constant but are unbalanced, and the difference between the voltage and the average voltage of each phase is calculated by dividing it by the average voltage to express the degree of unbalance. 2) and the voltage or current can be decomposed into three-phase symmetric components to calculate the ratio between the normal and reverse phase components (Equation 3).

(식 2)

(Equation 2)

(식 3)

(Equation 3)

<Active power, reactive power, apparent power>

In the case of an ideal sinusoidal voltage and current, the equations of power and energy are undisputed worldwide, but there are many differences in the equations that define power and energy when the waveform of voltage or current is distorted. Reactive power Q and apparent power S are defined as follows.

(식 4)

(Equation 4)

(식 5)

(Eq. 5)

(식 6)

(Equation 6)

<Power factor, displacement rate>

The displacement of a waveform is the degree to which phases of two periodic waveforms having the same period are out of phase. If, as if the I differ from each other both in the shape of a wave period equal to one another but only the phase Φ (rad) The Φ determines the size of the displacement. However, if two waveforms have only the same period and have an arbitrary shape, it is impossible to determine the posterior relationship between the two waveforms. In this case, the magnitude of the displacement is determined by the phase difference of the fundamental waves of the two waveforms, and in terms of power transmission, when the two waveforms considering the displacement are voltage and current, the degree of agreement between the waveforms is closely related to the efficiency of power transmission, that is, the power factor. do.

Displacement Factor (DPF) for periodic current and voltage can be defined by the following equation.

(식 7)

(Eq. 7)

Here, Φ ₁ represents the phase of the fundamental wave voltage based on the phase of the fundamental wave current, and when the displacement ratio DPF is compared with the power factor PF, PF can be expressed as follows.

(식 8)

(Eq. 8)

Harmonics

Harmonics obtain the magnitude and phase of each harmonic component from the coefficient X [k] of each frequency component obtained through the FFT. From the output of the FFT, the magnitude and phase of each harmonic component can be obtained as follows.

(식 9)

(Eq. 9)

(식 10)

(Eq. 10)

Where a [k] and b [k] mean real and imaginary values of the FFT output. THD (total harmonic distortion), distortion ratio (DF), total demand distortion (TDD), and K factor can be obtained from the magnitudes of the harmonics obtained in Equations 9 and 10 above.

Harmonic Distortion

(식 11)

(Eq. 11)

(식 12)

(Eq. 12)

(식 13)

(Eq. 13)

<Distortion Factor>

(식 14)

(Eq. 14)

(식 15)

(Eq. 15)

(식 16)

(Eq. 16)

<Demand Distortion>

(식 17)

(Eq. 17)

(식 18)

(Eq. 18)

<K factor>

(식 19)

(Eq. 19)

The following describes power disturbance detection and automatic identification.

Power disturbance refers to the fact that power cannot maintain the rated voltage (current) or rated frequency (60 Hz) due to the electromagnetism disturbance that actually occurs in the power system. Harmonic current from the interruption of the power supply due to the ground fault There are various causes and types such as distortion of supply voltage due to the inflow of water.

In the detection of power disturbance, the time and end point of the disturbance are measured to determine the existence period of the disturbance, and there are methods of comparing adjacent cycle values, using an expert system, and using a wavelet transform.

In the present invention, a method of extracting / identifying a feature of power disturbance using wavelets is used. By applying linear discrimination, only a main component is extracted from data of a disturbance signal to automatically identify power disturbance through an artificial neural network algorithm.

The entire flow of the process of wavelet transforming power signals to detect disturbances, extracting feature vectors, and automatically identifying them through an artificial neural network algorithm is shown in FIG. 2.

<Wavelet transform>

For the disturbance detection and feature vector extraction according to the present invention, the wavelet transform, which is widely used for the analysis of the transition signal, is applied.

Wavelet transform is an analysis method that represents a signal in the time-frequency plane and refers to an algorithm capable of multi-resolution analysis by moving and expanding a mother wavelet.

In particular, when the discrete wavelet transform is used among the wavelet transforms in the wavelet converter 11, the power disturbance signal without information duplication can be detected.

Wavelets are generated by expanding or contracting or translating a single function, and provide an extension of any function belonging to L ² (R). These wavelets are used to decompose blocked signals, and their feature is that they can exhibit good resolution in both time and frequency domains.

In the present invention, the discrete wavelet transform (DWT) is used in consideration of the calculation amount of the power computer 10, which can be defined as Equation 20 below.

(식 20)

(Eq. 20)

In particular, in the present invention, the wavelet transform is performed based on the Dovesh 4 function, and the Dovesh 4 function can be defined using the following transformation matrix.

Where the empty input represents zero and the numbers _{C0 , C1 , C2} And _C3 satisfies the following orthogonal properties:

Meet with the only solution of:

The wavelet Dovesh 4 transform of the array X is defined as in Equation 21 below.

Wavelet Dovesh 4 {X} = C * X (Equation 21)

<Linear Discrimination Analysis>

Linear discriminant analysis is one of the representative feature vector dimension reduction techniques, which reduces the dimension of feature vectors for data by maximizing the ratio of between-class scatter and with-class scatter. Say how.

3A and 3B are used to express the concept of class dispersion and class dispersion in a two-dimensional space. From this, FIG. 3A shows a wider space between classes and better discrimination between classes than in FIG. 3B. It can be seen that.

If there are two-dimensional data forming three classes as shown in FIG. 4, it is possible to map to one-dimensional subspace through various transformation matrices. Among them, in terms of class classification, there may be an idea of making a distribution that is easy to discriminate and an idea of making a distribution that is difficult to distinguish. Linear discriminant analysis is a method of reducing dimensions to linear subspaces by mapping them to the major axis that maximizes class separation in the feature space.

After all, the linear discriminant analysis method can be said to reduce the data in view of the optimal classification of the data.

Based on this principle, the optimized conversion matrix can be obtained from Eq.

(식 22)

(Eq. 22)

In the present invention, the performance is improved by reducing the amount of data input to the neural network algorithm by extracting only the principal component of the detected disturbance data by applying Equation (20).

<Feature Vector Extraction of Power Disturbance>

In order to determine the type of detected disturbance, appropriate information on the disturbance should be extracted and applied to the disturbance identifier 13. In particular, the input of the identifier should be as small as possible and contain the maximum characteristics of each signal. The feature vector extraction by the feature vector extractor 12 refers to a process of extracting values that can represent the characteristics of each disturbance detected in this way, and by using the feature vector as an input to the identifier, the identifier of the identifier can be further improved. Can be.

In the present invention, since the characteristics of voltage disturbance and current harmonic disturbance are different, the feature vectors are extracted by different methods, and the method is as follows.

Feature Vector Extraction Process of Voltage Disturbance

Step 1: From the detected data, the disturbance data having a predetermined length is obtained to sufficiently include the disturbance characteristics. At this time, if the length of the extracted data is too short, the characteristics of disturbance do not appear well. If the length of the data is too long, the input of the identifier becomes large.

In the present invention, one cycle of disturbance data is used.

Step 2: Remove the fundamental wave (60Hz) using a fast Fourier transformer (FFT) to make the frequency characteristic of power disturbance stand out.

Fast Fourier transform is an algorithm that converts the signal expressed on the time axis into the frequency axis and can be used for signal processing and analysis because it can recognize the frequency information of the signal, but it cannot represent all the time-frequency information of the signal. In the present invention, a wavelet transform of a signal from which fundamental waves are removed by fast Fourier transform is used to extract a feature vector of voltage disturbance.

Step 3: Wavelet transform the signal from which the fundamental wave is removed. In the present invention, after analyzing various signals by applying various wavelet filters, the wavelet transform observes the calculation amount and the degree of feature extraction of the signal and applies the wavelet filter showing the best performance.

Step 4: Combine the power value of the signal with the feature vector to improve the identification of low frequency disturbances. Since the feature vectors obtained through the first to third steps have been removed from the 60hz fundamental component, it is difficult to identify sag and swell in which the power of the fundamental wave greatly affects the identification. It allows for better identification performance.

Step 5: Apply linear identification to extract only the components that contain the most characteristic of each power disturbance. Criterion of the linear identification can be expressed as follows.

(식 23)

(Eq. 23)

The detection algorithm of the voltage disturbance is shown in FIG. 5.

First, the power computer performs wavelet conversion on a sampled incoming signal in one cycle unit. Wavelet transforming the sampling signal yields a high frequency region and a low frequency region. The coefficient of the high frequency region stores the disturbance occurrence time and waveform when the threshold is satisfied by comparing with the threshold value that determines the disturbance signal. If the threshold is not met, check if there is any stored data. Here, if there is stored data, the power value of the stored data is compared with the power value of the normal waveform, and if there is no stored data, it is input again.

The reason for checking whether there is stored data and comparing it with the power value of the normal signal is due to wavelet transient. The Dovesh 4 wavelet exhibits a high coefficient value when a transient transient occurs, and thus an instantaneous voltage increase and an instantaneous voltage drop. In the case of momentary power failure, disturbance detection is possible through wavelets only at the start and end of disturbance. This is because it is difficult to determine whether the disturbance lasts more than one cycle.

Feature Vector Extraction Process of Current Harmonic Disturbance

Step 1: Extract the disturbance data of one cycle and obtain the harmonic spectrum up to 50th order (3000Hz) normalized by fast Fourier transform.

Stage 2: Harmonic Disturbance Due to the nature of the spectrum, the presence or magnitude of the third harmonic (180 Hz) harmonics has a great influence on the identification of disturbances. Therefore, harmonic disturbances are divided into the following two types through appropriate threshold values.

i) when the magnitude of the third harmonic is greater than a certain value

ii) the third harmonic is smaller than a certain value;

Step 3: Calculate Total Harmonic Distortion (THD) and add it to the feature vector. The THD value is obtained from Equation 11 described above.

Step 4: Apply Fisher's Criterion to extract only the components that contain the most characteristic of each harmonic disturbance.

Harmonic disturbance detection algorithm using the above harmonic feature vector is shown in FIG.

In this case, the analog input signal is sampled and stored in memory, and then subjected to wavelet transformation to apply the threshold value that determines the disturbance, such as a voltage disturbance detection algorithm. If not, determine whether there is any saved data.

At this time, if there is no stored data, the data of the memory is loaded and the wavelet transformation process described above is performed again. If the stored data exists, the processed data is regarded as disturbance, and the disturbance is identified and stored in the database.

<Identifier Design>

In the present invention, a neural network is used for identification of power disturbance. The neural network has an advantage of being able to be learned by an algorithm modeled using a human brain structure, and thus, the neural network has been widely used in the field of pattern recognition.

7 is an artificial neural network that mimics human neurons, and is composed of an input layer, a hidden layer, and an output layer. In this case, the input layer and the output layer must exist one by one, and in the case of the hidden layer, there may be no or several pieces depending on the type of work.

The biggest advantage of neural network is learning ability. This learning ability uses the data to learn the weights of the network so that when a real signal comes in, it can identify the signal without prior statistical knowledge. In particular, multi-layer perceptrons with multiple hidden layers are capable of nonlinear separation.

The multilayer neural network identifier is the most widely used identifier of the supervised learning method that learns while the output value of the input is known, and adds one or more hidden-layers to the network. It is designed to enable nonlinear separation.

Therefore, in the present invention, an identifier is implemented by using a multi-layered neural network having a three-layer hidden layer, and a forward propagation obtained by propagating the output value from the input end of the neural network to the output end as a learning method and the output value thus obtained. The backpropagation algorithm was used to reduce the error by changing the weight by propagating the error between the output and the target from the output to the input.

FIG. 8 is a schematic diagram showing the structure of a three-layer neural network to be described through the present invention and the omnidirectional path of data input to the input layer.

Here, the connection weights between the input layer and the hidden layer and between the hidden layer and the output layer are initialized through a random function.

The calculation of forward propagation on the input pattern is performed by outputting the value input to the input layer as it is in the input layer, and multiplying the value output from the input layer by the connection weight between the input layer and the hidden layer as shown in Eq. It is an input of the hidden layer, and is added if a threshold exists.

(식 24)

(Eq. 24)

In the above formula 24 are connected between the neuron i wji and neurons j weight, netj an input of the neuron j, the threshold value θj of the neuron j, oi is an activity value of the neuron i.

The input of the hidden layer is taken as an output function by taking an active function. The activity function was used as sigmoid function and is shown in the following equation.

(식 25)

(Eq. 25)

In the output layer, the value output from the hidden layer and the connection weight between the hidden layer and the output layer are multiplied as inputs of the output layer, and are added when a threshold exists. Thus, the input of the output layer is taken into an active function and the final output value is calculated. Compute the error by comparing the target pattern with the final output value. If the error is appropriate, stop it. Otherwise, back propagate.

9 illustrates an error backpropagation algorithm of the generalized delta rule used in the present invention.

In FIG. 9, the output layer obtains the delta value by taking the above-described error into the derivative of the active function. Based on this value, the new connection weight between the output layer and the hidden layer is calculated. Multiplying the delta values obtained at the output layer by the connection weights between the output layer and the hidden layer yields a new delta value for the hidden layer. Based on this value, the new connection weight between the hidden layer and the input layer is calculated.

Redefined Eq 24 and Eq 25 as

(Sigmoid 함수) (식 26)

(Sigmoid function) (Equation 26)

(식 27)

(Eq. 27)

The amount of adjustment of the link weights using the generalized delta rules can be obtained as shown in Eqs. 28 and 29.

(식 28)

(Eq. 28)

(식 29)

(Eq. 29)

The error between the output of the output layer neuron and the target pattern for the input / output pattern p in the neural network is defined as shown in Eq.

(식 30)

(Eq. 30)

(식 31)

(Eq. 31)

Equation 31 above shows the total error of m input / output patterns.

The overall flow of the above process is represented as shown in FIG.

In FIG. 10, inpat is an input pattern, hidunit is an output value of a hidden layer neuron, outunit is an output value of an output layer neuron, wh is a connection weight value between the hidden layer and the input layer, wo is a connection weight value between the output layer and the hidden layer, and net is an input of each layer neuron. Where h is a hidden layer and o is an output layer.

The electrical quality diagnostic analysis data by the power computer 10 is displayed in various forms through the display 20.

FIG. 11 is a main graph mode screen of an electrical quality diagnostic analysis system. Here, voltage and current quality can be monitored through voltage and current waveform graphs of a three-phase input signal shown in a main graph of a display screen. .

12 is a display screen of the simulator mode, in order to help the non-experts to understand the electrical signal, it is possible to intuitively grasp the change of the waveform according to the change of the voltage value, the current value, and the frequency through the simulation function.

FIG. 13 is a display screen of a power disturbance automatic identification mode, and shows a function of automatically identifying Sag, Swell, Interruption, and Harmonic by computer learning by applying an artificial neural network algorithm.

FIG. 14 is an information mode. When an event occurs, FIG. 14 displays information on an event which is input / stored in advance so that an administrator can quickly and appropriately respond.

FIG. 15 shows a subgraph mode. Unlike a main graph that monitors all three-phase signals, the subgraph allows monitoring of a phase signal in which an event has occurred.

FIG. 16 is a harmonic graph that monitors harmonic values of voltage and current of a real-time three-phase power signal in real time through a fast Fourier transform (FFT).

Such electrical quality diagnostic analysis method of the present invention is to automatically identify the power disturbance through the artificial neural network and to monitor the current electrical signal state in real time through the display, even the unskilled person can intuitively grasp the electrical quality. You can get the distinctive effect.

In addition, the present invention is widely used, unlike the conventional method using a special neural network, and using a three-layer neural network, by limiting the power parameter values obtained through mathematical operations to increase the precision.

In addition, the present invention uses a back propagation algorithm as a learning algorithm, and extracts only the principal component of the feature data by linear discrimination through a fast Fourier transform and discrete wavelet transform, which is excellent for multi-resolution analysis of a signal for detecting disturbance features. This results in an accurate and simple diagnostic analysis of the electrical quality.

Claims (4)

A method for diagnosing the quality of incoming power using a power computer including a wavelet converter, a feature vector extractor, and a disturbance identifier, wherein the waveform of the input voltage or the input current is wavelet-converted, Extracting the feature vector, and assigning the extracted feature vector of the voltage disturbance or harmonic disturbance to the artificial intelligence neural network identifier, respectively, and identifying the disturbance characteristic of the current input voltage or input current in the learning data of the identifier; Intelligent electrical quality diagnostic analysis method comprising the step of displaying in a graphical form the waveform and disturbance characteristics of the input voltage or current. The method of claim 1, wherein the extraction of the voltage disturbance feature vector comprises: obtaining disturbance data having a predetermined length so that the characteristic of the disturbance can be sufficiently included from the detected data. Making the frequency characteristic of power disturbance stand out by removing the fundamental wave (60Hz) by using the fast Fourier transformer (FFT), and wavelet transforming the disturbance data signal from which the fundamental wave is removed; Combining the power value of the signal with the feature vector to improve the identification performance of the low frequency disturbance, and extracting only the components that contain the most characteristic of each power disturbance by linear discrimination. Quality diagnostic analysis method. The intelligent electrical quality diagnostic analysis method of claim 2, wherein the length of the disturbance data for extracting the feature vector is one period of data. The method of claim 1, wherein the extracting of the harmonic disturbance feature vector comprises extracting disturbance data having a predetermined length to obtain harmonic spectra of several orders of magnitude normalized by fast Fourier transform, and performing a third order (180 Hz) in the harmonic disturbance spectrum. Classifying the harmonic disturbance identification threshold value using the presence or absence of harmonics into a case where the third harmonic is larger than a predetermined value and the third harmonic is smaller than a predetermined value, and the harmonics Intelligent harmonic quality comprising the step of calculating the total harmonic distortion based on the size and phase of the component and adding it to the feature vector, and extracting only the component containing the most characteristic of each harmonic disturbance from the feature vector. Diagnostic analysis method.

Images