international journal of scientific & technology … · copies all the training observations...

INTERNATIONAL JOURNAL OF SCIENTIFIC & TECHNOLOGY RESEARCH VOLUME 9, ISSUE 06, JUNE 2020 ISSN 2277-8616

20 IJSTR©2020 www.ijstr.org

Neural Network Prediction Parameters Quality of Electrical Energy

Huthaifa A. Al_Issa, Mohammad Qawaqzeh, O. Oleksandr Miroshnyk, Oleksandr Savchenko, Irina Trunova

Abstract— A method for predicting physical parameters is proposed. The apparatus is considered and an analysis is made of the need to use a

neural network for the problem of predicting the quality of electrical energy. The analysis and the structure of neural networks, which are expedient

for using for the estimation and forecasting of the quality of electric energy, are chosen. Neural network models are constructed to calculate

additional indicators of the quality of electrical energy. Also mathematical expressions for the description of neural networks and their work are given.

Index Terms— neural network, forecasting, quality electrical energy.

—————————— ——————————

1 INTRODUCTION

Electric energy supplied by energy supplying organizations to consumers under contracts, acts as a special kind of product, characterized by a coincidence of production, transportation and consumption in time, as well as inability to store and return it. Accordingly, as a commodity of any kind, the concept of "quality" is applicable to electricity. The deviation of power quality from the standards set by the standards worsen the operating conditions of electrical installations both of the network and of consumers. [1].

In connection with the random nature of the change in electrical loads, the requirement of observance of the Quality Standards of Electric Energy (QSEE) norms throughout this time is practically unrealistic, therefore, the standard establishes the probability of exceeding the QSEE norms. The measured Indicators of the Quality of Electric Energy (IQEE) should not go beyond the normal allowable values with a probability of 0.95 for the standard time period specified by the standard (this means that it is possible not to reckon with individual excesses of the standardized values if the expected total duration is less than 5% for a specified period of time).

One of the main and most important regime parameters determining the quality of electricity are asymmetry and nonsinusoidal stresses in three-phase networks [2], which lead to additional voltage deviations at the terminals of consumers, increasing losses in network elements and electric receivers, deterioration of electrical equipment operating conditions of electrical equipment, etc.

2 ANALYSIS OF RECENT RESEARCH

In industry, the quality of electrical energy is estimated by technical and economic indicators, taking into account the damage caused by damage to materials and equipment,

technological process deterioration, deterioration in the quality of products, and loss of productivity – so-called technological damage [3, 4]. In addition, there is also electromagnetic damage from low-quality electricity, which is characterized by increased power losses, failure of electrical equipment, disruption of automation, telemechanics, communications, electronic equipment, etc. In this connection, the problem arises of estimating the damage from low-quality electric power, which is expressed in additional losses of electric energy and in reducing the service life of the equipment. Therefore, to solve this problem, it is necessary to measure the IQEE at the substation and assess how much the network losses will increase in comparison with the network operation mode, where the IQEE do not exceed the permissible limits, and how long the equipment will be used. For this we use a mathematical apparatus based on neural networks.

3 STATEMENT OF THE MAIN RESEARCH MATERIAL

The main feature of the neural network is the parallel processing of information by all links, which greatly speeds up the processing of information. Also, with a large number of connections, the network becomes more reliable even if the connections between the neurons are damaged.

Also, neural networks are capable of learning and summarizing the accumulated knowledge. The neural network has the features of artificial intelligence. A network trained on a limited number of data is able to generalize the information obtained and to show good results on data that were not used in its training [5, 6, 13].

To date, neural networks are used to solve a number of problems, one of which is the prediction task.

The purpose of the study is development of a method for estimating and predicting the quality of electrical energy using artificial neural networks.

Forecasting – these are predictions of future events.

Suppose given n discrete samples {y(t1), y(t2)..., y(tn)} at

successive instants of time t1, t2,..., tn. Then the prediction

problem is to predict the value y(tn+1) at some future time

tn+1. The purpose of forecasting is to reduce the risk in decision-making. The forecast usually goes erroneous, but the error depends on the forecasting system used. By providing more resources for the forecast, you can increase its accuracy

————————————————

Huthaifa A. Al_Issa, Mohammad Qawaqzeh, are currently assistant Professors, Electrical and Electronics Engineering Department, Al Balqa Applied University, Jordan. E-mail: [email protected], [email protected]

O. Oleksandr Miroshnyk, Oleksandr Savchenko, Irina Trunova are in Kharkiv Petro Vasylenko National University of Agriculture.

mailto:[email protected]



and reduce losses associated with uncertainty in decision-making. In this connection, it seems expedient to use a neural network to solve the problem of forecasting time series. The user selects an arbitrary time series that contains N samples, and breaks it into three sets: the training, testing and control samples, which are then fed to the network input. The result of the prediction is the values of the time series at the required time. To improve the quality of the forecast, it is necessary to make preliminary processing of information since the time series is a sequence of numerical samples, the preliminary processing is reduced to scaling the sample values in order to bring them into a single range. Each sample is a discrete function defined at points on the interval [0, N] in increments of 1, where N - the maximum value of the argument of this function.

When solving prediction tasks, the role of the neural network is to predict the future response of the system to its previous behavior. Possessing information on the meaning of

the replaceable x in the moments preceding the prediction

x(k-1), x(k-2), …, x(k-n). The network produces a solution in which the most probable value of the sequence x(k) at present

k. To adapt the weights of the network coefficients, the actual

prediction error ε = x(k) − x(k) and the value of this error at the previous instants of time [7, 8].

When choosing a network architecture, several configurations with different number of elements are usually tested. Proceeding from the fact that the prediction problem is a separate case of the regression problem, it turns out that it can be solved by the following types of neural networks: a multilayer perceptron, a radial-basic network, a generalized regression network, a Volterra network and an Elman network.

When solving the problem of forecasting time series, a generalized regression network was chosen as the neural network, which realizes methods of nuclear approximation. In regression problems, the network output can be considered as the expected value of the model at a given point in the input space. This expected value is related to the probability density of the general distribution of input and input data. At the location of each training observation, there is a Gaussian nuclear function. It is assumed that each observation indicates some certainty that the response surface at a given point has a certain height, and this confidence decreases as you move away from the point. The generalized regression network copies all the training observations inside itself and uses them to evaluate the response at an arbitrary point. The final initial estimate of the network comes out as a weighted average of the outputs for all the learning observations, where the value of the weights represent the distance from these observations to the point at which the assessment is performed. The structure of the generalized regression neural network is shown in Fig 1.

Fig 1. Structure of the generalized regression neural network.

The generalized regression network has two hidden layers: a layer of radial elements and a layer of elements that form a weighted sum for the corresponding element of the source layer. The weighted average is determined in the source layer by distributing the weighted sum to the sum of the weights. The Gaussian function is used as the radial function. The input layer transmits signals to the first intermediate layer of neurons, which are radially symmetric. They carry information about these learning cases or their clusters and transfer it to the second intermediate layer. It forms the weighted sums for all elements of the original layer and the sum of the weights, which is calculated by a special element. If we denote the output i- th the neuron of a radially-basic network layer as vi, then the original signal i- th the neuron of the second intermediate layer is calculated by expression.

𝑢 = ∑ 𝑣

, (1)

where k - number of neurons in the radial-basis network

layer. Having now designated the weight coefficients of the i-th

neuron in the radial-basis network layer as ωi . We obtain the expression for the sum of the weights:

𝑣 = ∑ 𝜔 . (2)

So, the source layer divides the weighted sums by the sum

of weights and issues the final forecast. Denoting it yl. We

obtain the expression:

𝑦 =

(3)

Let us now consider the principles of the functioning of the first intermediate layer, the structure of which is shown in Fig 2.



+

+

+

+

+ .

.

.

.

x1

x2

xn

(1)

1C

S1

vi

.

.

.

.

x1

.

.

.

.

.

.

.

.

+

+

+

+

+

+

(1)

nC

( )

1

kC

( )k

nC

(1)

1S

(1)

nS

( )

1

kS

( )k

nS

Sk

1

Fig 2. The structure of the radial-base layer of the

generalized regression neural network. The input of radial elements from the input layer is fed by a

vector x. Basis functions of the radial-base layer are given by the matrix 𝑄, but in practical terms it is more convenient to use for the description of elements the correlation matrix C, which leaves the matrix 𝑄 in this way:

TC Q Q . (4)

The center of the i-th neuron of the radial layer is denoted

by ci.

The residual result of the processing of the input signals Sj is calculated by the expressions 5, 6 and 7:

2

1

1

2

nt t

j t i

t

S x c

, (5)

1

nt

t j

j

S S

, (6)

21

1exp

2

kt

i t

t t

Sv

. (7)

Then the vector of the original signals v is transmitted to

the input of the second intermediate layer of the network. Advantage of the generalized regression neural network

can be considered the certainty of the structure: the network actually contains all the training data. On the other hand, with a large volume of training data, the speed of the network decreases through the increase in the complexity of the architecture.

The initial value of the network has a probabilistic form, so it is easier to interpret. With a small amount of input data, the network learns very quickly. Network training needs to be done separately for each time series, because trying to predict a line on which the network was not taught will lead to an erroneous result [9, 10].

Considering the above, we will build a neural network for forecasting the economic indicators of the electric network

and the compatibility of consumers with it as shown in Fig 3, which will take into account the parameters:

Increasing the power consumption of the motor and lighting load in comparison with the symmetrical mode;

Coefficients of asymmetry of currents, voltages and additional losses;

Coefficient of uneven load phases;

Additional losses of active power in asynchronous motors, transformers and lines due to the asymmetry of the regime;

Additional losses of active power in asynchronous motors, transformers and lines, caused by non-sinusoidality of the regime;

Service life of insulation TD .

Fig 3 . Neural network input and output signals.

To determine the symmetrical components of the voltages

and currents of the three-phase system, we use the known formulas and construct the neurons as shown in Fig 4:

2 4

3 31

1

3A A B CI I e I e I

;

�̱� =

(�̱� + 𝑒

�̱� + 𝑒

�̱� );

4 2

3 32

1

3A A B CU U e U e U

; 4 2

3 32

1

3A A B CI I e I e I

; (8)

0

1

3A A B CU U U U

; 0

1

3A A B CI I I I

.



Fig 4. Neurons that determine the symmetrical components

of a three-phase network. Similar procedures can be performed for:

The coefficient of increase in the power consumption of the motor load in comparison with the symmetric mode.

1 1 2 2 0 0

1 1engine

asym

P

sym

Р I U I U I UК

Р I U

D + += =

D, (9)

The coefficient of increase in the power consumption of the lighting load in comparison with the symmetrical mode

1 1

1

asym

P

sym

lighti

m

ng

no

Р I UК

Р I U

D= =

D, or

2 4

3 3

220lighting

A B C

asym

P

sym

U e U e UP

KP

, (10)

Unbalance coefficients of currents and voltages:

2

2

1

I

IK

I

,

0

0

1

I

IK

I

,

2

2

1

U

UK

U

,

0

0

1

U

UK

U

. (11) 4 2

3 3

2 2 4

3 3

A B C

I

A B C

I e I e I

K

I e I e I

;

Additional loss factor

2 2

2 0

31 1

asym N

add I I

sym P

Р rК К K

Р r

æ öD ÷ç ÷= = + + +ç ÷ç ÷çD è ø, (12)

where rN and rP – resistance of zero and phase conductors (rN = rP).

Factor of uneven load of phases

N

un

A B C

IК

I I I

. (13)



Additional losses of active power in the asynchronous motor AD caused by the asymmetry of the regime

2 2

2 22,41engine nom st UP P I KD = ×D × ×, (14)

where nomPD - losses in stator copper at rated load, and

stI- frequency of the starting current.

Additional losses of active power in transformers due to the asymmetry of the regime

2. .

2 22

. .

s c

tr U

s c

PP K

u

DD = ×

, (15)

where . .s cPD - short circuit losses, and . .s cu - short-circuit voltage.

Additional losses of active power in the line, due to the asymmetry of the regime 𝛥𝑃 = 𝐼

𝑅 + 𝐼 𝑅 + 𝐼

𝑅, (16)

Additional losses of active power in the engine, caused by non-sinusoidality of the regime

( ) ( )2

. . .

2

1n

U v

engine m nom o f

v

KP P k v v

vnS

=

æ ö÷ç ÷çD = D + ±÷ç ÷ç ÷è øå

, (17)

. .o fk- overload factor,

v - harmonic number

Additional losses of active power in the transformer, caused by non-sinusoidality of the regime

2

. .

2

3n

vt s c vt

v

P I r knS

=

D = å, (18)

where

тIn - harmonic n current

. .s cr - short-circuit resistance of the transformer

0,47тkn n=

Ne

f(2

Ne

Ne

f(2)

ІA

ІB

ІС



Additional losses of active power in the line due to the non-sinusoidal nature of the regime

𝛥𝑃 = 3𝑟 𝑙 ∑ 𝐼 (𝑘 + 𝑘 )

, (19)

Where 2

1,18

0,27

Pv

бv

Pv

k dk

k a

,

d - diameter of core conductor, mm; а - distance between centers of cores, mm

0 0v Pv pr vr r k k , where r0 – resistivity of conductor to

direct current;

Pvk - coefficient, taking into account the phenomenon of

surface effect for 𝑣-th harmonics, and equal

0,021Pvk f - for copper and 0,01635Pvk f

- for aluminum. pr vk – coefficient taking into account the

proximity effect for 𝑣-th harmonics; 𝑣 - harmonic number; п - number of harmonics;

Iv - current of the 𝑣-th harmonic.

Service life of insulation

( )0,0860,086

2T

tt

DD = D +

(20)

where 2

. . . . ( )

2

nstat v

s l s l I v

vstat

PК

Pt t t

=

DD = =

Då

The ultimate goal of measuring electricity quality indicators

is the adoption of appropriate measures to improve them. At present, in accordance with the technical regulations, the equalization of electric loads of transformers and trunk lines of 0.38 kV is made by the construction crews on the basis of the results of measuring the currents in phases during the



period of the greatest loads. If in the wires of a trunk or the winding of a transformer

𝐼 ≠ 𝐼 ≠ 𝐼 , (21) then after measuring the load of single-phase inputs to

consumers, the redistribution of the latter along the line phases is carried out in such a way that the load of the sections becomes more or less uniform at the time of measurement:

𝐼 ≈ 𝐼 ≈ 𝐼 . (22) As a rule, the equalization of electrical loads is limited by

this [11]. However, since the inclusion or deenergization of any receiver is equally probable and the total load is subject to statistical regularities, the unsuitable symmetry over instantaneous currents even in the period of the greatest loads is obvious. In this regard, it is proposed to use a neural network that takes into account the change in loads of consumers in time and makes it possible to perform an even distribution of loads across phases on other principles. At the same time recommendations on the reconnection of a particular consumer from the most loaded phase to the least loaded will be provided by the neural network on the basis of the history of the accumulated data [12]. In Fig 5 shows the neural network for creating a strategy for uniform distribution of loads in the 0.38 / 0.22 kV network.

Fig 5. Neural network for uniform distribution of loads The proposed neural network based on the removal and

analysis of daily load graphs for six months gives recommendations on how much current must be redistributed to a particular phase.

4 CONCLUSIONS

The developed method has the ability to evaluate and predict the indications of the main and additional indicators of the quality of electric power, and the network can provide a reduction in costs in the prevention of emergencies. The results of these studies can be used in various areas for forecasting the parameters of technical systems and aggregates, and used to prevent the emergence of emergency and emergency situations. The constructed neural network

allows using information about the 0.38 / 0.22 kV network mode in real time and, if necessary, to use it to control the mode in order to reduce the asymmetry of the currents and to reduce the additional losses of electric energy.

REFERENCES

[1] Naumov I.V. Reducing losses and improving the quality of electrical energy in rural distribution networks 0.38 kV with the help of balancing devices / I.V. Naumov. - Diss. Doct. those. Sciences, 05.20.02 - Irkutsk, 2002. - 387 p.

[2] O. Mіroshnik. Statistical analysis of the basic parameters of the SIL'Sh least 0.38 / 0.22 kV / O. Mіroshnik // Scientific herald of the National University of Bioresources and nature conservation of Ukraine. Series "Engineering and Power Engineering of Agroindustrial Complex" - Kyiv: NUBiPU, 2011. - № 166. ч. 4. - С. 203-211.

[3] Levin MS Analysis of asymmetric modes of rural networks 0.38 kV / M.S. Levin, TB Leshchinskaya // Electricity, 1999, №5. - P. 18 - 22.

[4] Angelov D.D. Investigation of nonsymmetry in aerial power distribution networks 0.38 / 0.22 kV and ways of its limitation: Abstract of thesis. Dis. Cand. Tech. D.D. The angels. - Sofia, 1980- 24 p.

[5] Anil K. Jain, Jianchang Mao, K.M. Mohiuddin. Artificial Neural Networks: A Tutorial, Computer, Vol.29, No.3, March / 1996, p. 31-44.

[6] Gorban AN A generalized approximation theorem and computational possibilities of neural networks / A.N. Gorban / Siberian Journal of Computational Mathematics, 1998. Vol. 1, No. 1, p.12-24.

[7] Wassermen F. Neurocomputer Technology: Theory and Practice / F. Wasserman / Transl. From the English Yu.A. Zuyev. - Moscow: Mir, 1992. - 378 c.

[8] Bodyansky EV, Rudenko O.G. Artificial neural networks: architecture, training and application / Е.В. Bodiansky, OG Rudenko. - Kharkiv: TELETEH, 2004. - 372 p.

[9] Bilas O. Є. A cloak of chapters of lines for the possible neuromereg / O. Є. Bilas // Proceedings of the Odessa Polytechnic University. - 2001. - 3 (15). - P. 127-131.

[10] Handbook of Neural Computation / Ed. By E. Fiesler, R. Beale - Bristol: IOP Publishing, 1997. - 988 p.

[11] Derrick V.G. Simulation of unbalance of load of phase lines of 0.38 kV in calculations of electric energy losses during its transmission under conditions of uncertainty / VG Derzky, VF Skiba // Energy Saving Energy Energy audit. - 2007. – No. 6. - pp. 9 - 22.

[12] O Mіroshnik Uniform distribution of loads in the network 0,38 / 0,22 kV using the neural network / O Mіroshnik // News of NTU "KhPI" - Khar'kov: NTU "KhPI", 2013. - №17 . - P. 107-114.

[13] Al Smadi, T., Al_Issa, Huthaifa A., Trad, E. and Al Smadi, K. (2015), ―Artificial intelligence techniques for speech recognition based on neural networks’’, Journal of Signal and Information Processing, 6 (02), 66-72. http:// dx.doi.org/10.4236/jsip.2015.62006. Scientific Research, USA.

international journal of scientific & technology … · copies all the training observations...

Documents