ANFIS based model for power loss estimation of metal oxide surge arrester

Abstract

This paper proposes a new method to estimate power loss characteristics of metal oxide surge arresters. The method was used to compute surge arrester power loss curve based on adaptive network based fuzzy inference system (ANFIS) and artificial neural network (ANN). Surge arrester operating history is an important factor that bears influence on its power loss. Degradation was, in this paper, introduced as a new index to represent operating history of metal oxide surge arrester. Therefore, applied voltage, temperature and degradation factor were considered as inputs in ANFIS system and ANN models to obtain accurate power loss which is a very important factor in surge arrester thermal stability. Degrading effect was undertaken by measurement voltage and current in varistors degraded by utilization in network. The results of the two artificial models were compared. Results show that ANFIS was more accurate than ANN. This study shows that the power loss characteristics of surge arrester are to a great extent accurately predictable using proposed artificial model.

Keywords

Adaptive network based fuzzy inference system artificial neural network degradation factor metal oxide surge arrester power loss

1 Introduction

Metal oxide surge arrester (MOSA) is a device used on electrical power and telecommunications systems to protect networks insulation and conductors from the damaging effects of high current and voltage strokes. MOSA is used to divert overvoltage transients damaging caused by external (lightning) or internal (switching) events safely to ground through property changes to its varistor in parallel arrangement to the conductor inside the unit [1, 2]. Metal oxide varistors, which are known to ZnO varistors, are the main components of the surge arrester. They provide the desired nonlinear voltage –current (V-I) characteristics and present a strong relation with temperature.

MOSAs are normally under maximum continuously operating voltage (MCOV) [3]. This leads to low power production in varistors of MOSA. This electrical power loss is important in evaluation of the arrester thermal stability [4 –6]. In order to analyze the thermal balance of MOSA, first the power loss model should be obtained. This model determines the operating reliability of MOSAs and their life-spans [7]. Voltages and currents of the MOSA are used as the main input for power loss calculation [8]. These values of ZnO resistors highly depend on temperature in low-current and voltage regions [9].

In low current range, power input to the varistor has complex function base on applied voltage, temperature and operating history [10]. Therefore, it is difficult to precisely describe power loss by conventional methods [11 –15]. Several methods [7 , 16–18] have been suggested for MOSA performance prediction. In [7], ANN is used to estimate power loss characteristics of ZnO varistors based on temperature and applied voltage. The structure of model and layer has been shown in Fig. 1. These methods are not accurate as all the aforementioned factors have not been considered. Using the previous methods, it is not possible to identify power loss equation of MOSA with different characteristics.

In this paper, the power loss characteristics of ZnO varistors have been modeled by an adaptive network based fuzzy inference system and artificial neural network. Applied voltage, temperature and operating history are the main parameters which have been considered as inputs in ANFIS and ANN. Therefore, degradation factor (DF) as a new index has been suggested to represent operating history of MOSAs. Based on proposed DF, electrical power loss of MOSAs can be modeled by considering degradation effect.

2 Power loss characteristic: Measurement and computation

In order to investigate the influence of temperature and degrading on MOSA V-I characteristic, experimental setup was arranged. The experimental setup was prepared as shown in Fig. 2 in order to perform the practical tests on MOSA varistor.

According to Fig. 2, the experimental setup consists of a high voltage transformer with adjustable voltage between 0–100 kV, capacitor divider, oven and data acquisition systems. Data acquisition system comprises a digital oscilloscope, back to back connected Zener diodes for overvoltage protection and a 470 Ω shunt resistor for leakage current measurement. R₁ is protective resistance which was used to limit short-circuit current. The resistive component of leakage current through the MOSA and applied voltage was captured using a two-channel digital oscilloscope. Oven was prepared to obtain the power loss of varistors at different temperatures. High voltage flange, heaters, heat sensor, and thermal controller are the main sections of electro-thermal setup. The temperature of the oven is measured and controlled by heat sensor and thermal controlling system. It is possible to increase temperature over 250±3°C, using heaters and thermal controller.

As already mentioned, varistors with different history should be examined. In order to achieve accurate measurements, four sets of varistors that used in 20 kV surge arrester were tested. The radius of ZnO varistors and their height are 20 mm and 30 mm, respectively. Four sets studied varistors are shown in Fig. 3.

The first set (A1,A2,A3) consists of new varistors with similar V-I characteristics. Varistors in second (B1,B2,B3,B4) and third (C1C2) groups extracted from old MOSA which had been used in distribution system. Finally, the last category (D1,D2,D3) consists of new varistors with different materials composition and V-I characteristic. It should be noted that in each group, the varistors’ names are classified according to degradation. For example B3 is degraded morethan B1.

Power loss characteristic is obtained in different temperatures for all studied varistors. Therefore, each varistor was placed in the oven and power loss was measured by the power loss measuring system. Oven temperature was fixed at a certain value firstly and then the power loss was measured in different applied voltage. In the next step, oven temperature was fixed at the new value and this process repeated.

Figure 4 Shows the Power loss characteristic for virgin and degraded varistors (shown in Fig. 3) as a function of temperature and applied voltage. As seen in Fig. 4, power loss increases with increment of applied voltage and temperature. In addition, comparing figures in left and right side of Fig. 4 illustrates power loss increment by varistor degradation, as well. Degraded varistors have higher power losses than virgin ones. This is owing to higher amplitudes of leakage currents in constant applied voltage.

According to power loss characteristics, increasing rate of power loss changes at a certain point which is called “critical voltage”. Therefore, power loss characteristics can be divided into two sections. When voltage is less than critical voltage, increasing rate of power loss is slow. But when voltage is higher than critical voltage, power loss increases quickly. The value of critical voltage depends on degradation of varistor. This issue has not been considered in analytical modeling of ZnO varistors power losses in references which are based on critical voltage [8].

3 Proposed method

Electrical power loss curve at different temperatures and applied voltages have been modeled by different equations in many published papers in the literature. But the effect of degrading on MOSA power loss curve has not been studied so far. Since performance history is an effective factor on MOSA leakage current, it must be considered in power loss modeling. Varistor degradation increases losses at a constant voltage. Also, defective varistor influence on losses increment at a constant temperature. Such conditions cause operating region of surge arrester and therefore, utilization conditions to be restricted.

3.1 Calculation of degradation factor

DF has been defined as a new index to model MOSA performance history. DF is computed by calculating non-linearity factors obtained from virgin and degraded V-I characteristics. As mentioned, material variation, temperature and operation condition contribute to a fall in rated and reference voltage amplitudes. DF is proposed as follows: $DF = \frac{{NF}_{virgin} - {NF}_{deg raded}}{{NF}_{virgin}}$ (1)

Where NF _virgin and NF _degraded are non-linearity factors (NFs) of virgin and degraded characteristics which are defined as follows: $NF = log (\frac{I_{ref}}{I_{rated}}) / log (\frac{V_{ref}}{V_{rated}})$ (2)

For each varistor, NF can be obtained from voltage and current information in rated and reference points. In this paper, NF _virgin is defined for varistor A1 which has greatest reference voltage among other studied varistors in 30°C. This is the reference characteristic for achieving DF of the other characteristics. In other words the DF for A1 is zero.

In order to calculate NF for studied varistors, the values of V _rated (rated voltage) and V _ref (reference voltage) are measured in constant I _rated-virgin (rated current of A1 varistor) and I _ref-virgin (reference current of A1 varistor as) values, respectively. Table 1 represents reference and rated voltages, their ratio, NF, and DF for studied varistors. The higher the value of DF shows the higher the deviation from the base V-I characteristic (A1). According to the DFs represented in Table 1, proposed index illustrates degradation of varistors appropriately.

3.2 ANFIS modeling of power loss

ANFIS was offered by Jang in 1993 [19]. In this system a hybrid learning method is used to create an input-output mapping. It works based on human knowledge and training data pairs. The fuzzy inference system is used in the framework of adaptive networks. ANFIS commonly includes a five-layer feed forward neural network to make the inference system [20]. In this structure, input data are not considered as a layer. Every layer includes several nodes which are expressed by nodes function. The nodes in previous layer supply input to the nodes in next layer.

Fuzzy inference system (FIS) is used to provide initial conditions for ANFIS training. Generation of FIS can be performed by three methods which are grid partitioning (GP), subtractive clustering (SC) and fuzzy c-means clustering (FCMC) [21, 22]. In this paper SC is used to provide FIS. This method is used to generate an initial FIS for ANFIS training by first applying subtractive clustering on the data. In other words, given separate sets of input and output data, a Sugeno-type fuzzy inference system can be achieved by this method. SC method accomplishes this act by extracting a set of rules that models the data behavior. The rule extraction method first uses the SUBCLUST function to determine the number of rules and antecedent membership functions and then uses linear least squares estimation to determine each rule’s consequent equations. It should be noted that SUBCLUST function Locates data cluster centers using subtractive clustering.

The ANFIS rules based contains fuzzy if-then rules of Sugeno type. The rules can be expressed as:

Rule 1: if V is A₁ and T is B₁ and DF is C₁ then f is f₁(V,T,DF)

Rule 2: if V is A₂ and T is B₂ and DF is C₂ then f is f₂(V,T,DF)

Here V, T and DF are the ANFIS inputs, A _i, B _i and C _i states the fuzzy sets and f _i (V, T, DF) represents the outputs of the first order Sugeno fuzzy inference system. According to (3), every node in first layer is denoted with a node function. $μ_{A_{i}} (x, σ, c) = e^{- \frac{(x - c)^{2}}{2 σ^{2}}}$ (3)

Where x is the input data to node i, A _i the linguistic tag associated with this node, c is central of membership function (MF), σ is standard deviation, and μ _{A
_i} is the MF of A _i. In the second layer, every node is a fixed node which computes the firing strength w _i of a fuzzy rule. The output of each node has been defined by product of all the incoming signals to node which is given by (4). This layer has been labeled as Π. $w_{i} = μ_{A_{i}} \times μ_{B_{i}} \times μ_{C_{i}} i = 1, 2, 3$ (4)

Nodes in third layer are shown by N are fixed. Ratio of w _i to the sum of the all w _i has been determined in ith node. The output of ith node is ${\bar{w}}_{i}$ which has been given as follows: $\bar{w} = \frac{w_{i}}{\sum_{i} w_{i}} i = 1, 2, 3$ (5)

Parameter function on output of third layer is computed for each node in 4th layer by (6) $f \bar{w} = \bar{w_{i}} f_{i}$ (6)

In 5th layer, the node f calculates the final output from all incoming signals. Overall output has been shown by (7). $f \bar{w}' = \sum_{i} \bar{w} i f_{i} = \frac{\sum_{i} w_{i} f_{i}}{\sum_{i} w_{i}} i = 1, 2, 3$ (7)

Table 2 has been shown Parameters setting of ANFIS. To demonstrate the process of an ANFIS (see Fig. 5), system with three inputs (V, T, DF) which are applied voltage, temperature, and degradation factor and one output which is power loss has been used. Numbers of MFs based on Gaussian have been used as ANFIS variables to get the optimum ANFIS structure. To train the model, a hybrid method included the least-squares method and the back propagation gradient descent is used as optimization technique to emulate a given training dataset. The linear output MFs have been used to create the electrical power loss values.

3.3 ANN modeling of power loss

The ANN is an adaptive nonlinear dynamic system that comprise of a lot of neurons [23]. ANN stores the information in the weight among neurons and has a good wrong-permitting capability. Moreover, it can remove the noise in the experimental data [24]. The ANN consists of input, hidden and output layers. The artificial neural network model adjusts the weights by a learning course, and reaches the desired precision [25]. In ANN, the inputs of neurons are: $u_{j} = \sum w_{ji} x_{i}$ (8)

Where w is a synaptic weight vector and x is an inputs vector. The output y _j of the neuron is $y_{j} = f (u_{j}) = \frac{1}{1 + e^{- (u_{j} + a_{j})}}$ (9)

Where f is a neuron exciting function, a _j is the threshold value of neuron. The square error E between target output {TO _d} and ANN output {y _d} is $y_{j} = f (u_{j}) = \frac{1}{1 + e^{- (u_{j} + a_{j})}}$ (10)

In this ANN model applied voltage, temperature and degradation factor are the input layers, and the power loss is the output layer. This structure has a hidden layer and six neurons. The experimental results are imported into ANN to study and the suitable weights and threshold values are obtained. The ANN model, which has been used to simulate the power loss characteristic of ZnO varistor, has been shown in Fig. 6. As shown in this model, similar to the ANFIS model, system with three inputs (V, T, DF) has been used to demonstrate the process of ANN model.

3.4 Power loss estimation algorithm

Proposed algorithm for prediction of power loss characteristic based on ANFIS and ANN models is shown in Fig. 7. In summary, the power loss estimation in the algorithm contains two steps:

In the first step, MOSA rated and reference voltages are measured in rated and reference resistive current of virgin varistor in ambient temperature. The DF can be calculated based on this information. Arbitrary temperature, arbitrary applied voltage based on measured reference voltage and the calculated DF which are the input variables of the ANFIS and ANN models will be given at the end of this step.

The input variables are given to the ANFIS and ANN models which have been trained by the previous experimental measurements (Section 2). The output of ANFIS and ANN models is tested for validation. If the error is minimized, the final output, which is the power loss, will be extracted.

4 Proposed model validation

In order to validate the proposed method results, A1 and D2 varistors were analyzed. Moreover, ANFIS results were compared with ANN. D2 information was used for ANFIS and ANN model training but A1 data was used as new input for these models.

The convergence of the ANFIS and ANN training after 100 epochs has been shown in Fig. 8. Comparing root mean square error (RMSE) of ANN, and ANFIS, it is obvious that ANFIS model is faster than the ANN and also has smaller final error.

Figure 9a, b shows a power loss comparison between 1000 experimental test data and estimated values from ANFIS and ANN models. In addition, error values are shown in Fig. 9c. According to Fig. 9c the errors of power loss values are less than 5% for ANFIS model. However, these values for ANN model are greater than 5% . Moreover, maximum errors happen in points with high power losses. In other words, the agreements between the results are higher in low applied voltages. Based on Figs. 8 and 9, it can be seen that there is a good agreement between ANFIS and experimentalresults.

In addition, Experimental and estimated power loss for A1 and D2 varistors is shown in Table 3. These results have been extracted from ANFIS model and experimental tests in several temperature and applied voltage values. Comparing the experimental and simulation results show that not only the proposed method based on ANFIS network can be modeled the influence of applied voltage and temperature variation on power loss characteristic but also it is able to model degradation effect properly.

5 Conclusions

In this paper, electrical properties of ZnO arrester blocks have been experimentally investigated. Power loss characteristics of metal oxide surge arrester have been studied using ANFIS and ANN models. The main findings are as follows:

Temperature, varistor material and operating history have changed varistors V-I characteristics which should be considered in power loss estimation.

Due to the effect of applied voltage and temperature on resistive component of leakage current, ZnO varistors power loss at ac stresses increases.

Operating history affects on power loss more than two above mentioned factors in same conditions.

DF as a new index has been introduced to represent operating history of MOSA.

ANFIS and ANN models has been used to estimate power loss characteristic based on three inputs which are applied voltage, temperature and DF.

The output of ANFIS and ANN had a good agreement with experimental results, which shows precision of proposed model.

Using ANFIS model, power loss estimation accuracy and speed of convergence increase and errors decrease.

References

Lee

2010

Analysis of thermal and electrical properties of ZnO arrester block

Elsevier Current Applied Physics 10 176 180

Fujiwara

Shibuya

Imataki

Nitta

1982

Evaluation of surge degradation of metal oxide surge arrester

IEEE Trans Power Apparatus and Systems PAS-101 978 985

Tominaga

Shibuya

Fujiwara

Imataki

Nitta

1980

Stability and long term degradation of metal oxide surge arresters

IEEE Transaction on Power Apparatus and Systems PAS-99 1548 1556

Lat

1985

Analytical method for performance prediction of metal oxide surge arresters

IEEE Transaction on Power Apparatus and Systems PAS-100 2665 2674

Bartkowiak

Comber

Mahan

1999

Failure modes and energy absorption capability of znovaristors

IEEE Transaction on Power Delivery 14 152 162

Imai

Udagawa

Ando

Tanno

Kayano

Kan

1998

Development of high gradient zinc oxide resistors and their application to surge arresters

IEEE Transaction on Power Delivery 13 1182 1187

Zeng

Chen

2003

Thermal characteristics of high voltage whole-solid-insulated polymeric zno surge arrester

IEEE Transaction on Power Delivery 18 1221 1227

Kan

Nishiwaki

Sato

Kojima

Yanabu

1983

Surge discharge capability and thermal stability of a metal oxide surge arrester

IEEE Transaction on Power Apparatus and Systems PAS-102 282 289

Miller

Woodworth

Daley

1999

Watts loss of polymer housed surge arresters in a simulated florida coastal climate

IEEE Transaction on Power Delivery 14 940 947

10.

Lat

1983

Thermal properties of metal oxide surge arresters

IEEE Transaction on Power Apparatus and Systems PAS-102 2194 2202

11.

Tominaga

Shibuya

Fujiwara

Imataki

Nitta

1980

Stability and long term degradation of metal oxide surge arresters

IEEE Transaction on Power Apparatus and Systems PAS-99 1548 1556

12.

Nishiwaki

Kimura

Satoh

Mizoguchi

Yanabu

1984

Study of thermal runaway/equivalent prorated model of a zno surge arrester

IEEE Transaction on Power Apparatus and Systems PAS-103 36 37

13.

Kobayashi

Mizuno

Hayashi

Sughita

1986

Metal oxide surge arrester

IEEE Transaction on Electrical Insulation EI-21 989 996

14.

Boggs

Kuang

Andoh

Nishiwaki

2000

Electro-thermal-mechanical computations in zno arrester elements

IEEE Transaction on Power Delivery 15 128 134

15.

Chrzan

2001

Degradation of metal oxide varistors caused by electrical stresses

11th Electromagnetic Disturbances (EMD) Conf

16.

Andrd Petit

Guy St-Jean

1990

Metal-oxide surge arrester operating limits defined by a temperature-margin concept

IEEE Transactions on Power Delivery 5 627 633

17.

Nigol

1992

Methods for analyzing the performance of gapless metal oxide surge arresters

IEEE Transaction on Power Delivery 3 1256 1264

18.

Zheng

Steven

Boggs

2010

Toshiya Imai and Susumu Nishiwaki, Computation of arrester thermal stability

IEEE Transaction on Power Delivery 3 1526 1529

19.

Jaya

ASM

Hashim

SZM

Haron

Ngah

Muhamad

Rahman

MNA

2013

Modeling of ANFIS in predicting tin coatings roughness

5th Computer Science and Information Technology (CSIT) Conf

20.

Jang

JSR

2002

ANFIS: Adaptive-network-based fuzzy inference systems

IEEE Transaction on Systems, Man and Cybernetics 23 665 685

21.

Rantala

Koivisto

2002

Optimized subtractive clustering for neuro-fuzzy models

3rd WSEAS International Conference on Fuzzy Sets and Fuzzy Systems Conf

22.

Buragohain

Mahanta

2006

ANFIS Modeling of Nonlinear System Based on Subtractive Clustering and V-fold Technique

India Conference of Annual IEEE 1 6

New Delhi

23.

Sikora

Baniukiewicz

Chady

Lopato

Psuj

Grzywacz

Misztal

2014

Artificial neural networks and fuzzy logic in nondestructive evaluation

Electromagnetic Nondestructive Evaluation 38 137 151

24.

Jain

Jianchang

Mohiuddin

1996

Artificial neural network: A tutorial

IEEE Computer Society 29 31 44

25.

H. Demuth and M. Beale, Neural Network Toolbox for Use with Matlab, The Mathwork, https://www.mathworks.com/, 1997-2002.