On-Line Trained Adaptive Neuro-Fuzzy Inference System for Distance

Document Sample
On-Line Trained Adaptive Neuro-Fuzzy Inference System for Distance Powered By Docstoc
					The Online Journal on Electronics and Electrical Engineering (OJEEE)                                           Vol. (1) – No. (2)

           On-Line Trained Adaptive Neuro-Fuzzy
             Inference System for Distance Relay
               of Transmission line Protection
                                 Tamer S. Kamel M. A. Moustafa Hassan
            Electrical Power and Machines Department, Faculty of Engineering, Cairo University.
                              Corresponding E-Mail: mmustafa@eng.cu.edu.eg

Abstract-This paper presents a new distance relay                 commercial developments. However, both these approaches
technique for transmission line protection by using well          are based on deterministic computations on a well defined
known control technique; Adaptive Neuro-Fuzzy                     model of the system to be protected. This results in difficulty
Inference System (ANFIS). The ANFIS can be viewed                 in taking system variation into account as the rules are fixed.
either as a fuzzy system, a neural network or fuzzy neural        They do not have the ability to adapt dynamically to the
network FNN. The structure is seen as a neural network            system operating conditions, and to make correct decisions if
for training and a fuzzy viewpoint is utilized to gain            the signals are uncertain. Recently, intelligent soft
insight into the system and to simplify the model. The            computational techniques such as Artificial Neural Network
integration with neural network technology enhances               (ANN), Fuzzy Inference System (FIS) and (ANFIS) can
fuzzy logic systems on learning capabilities. The                 model superiority of human knowledge features. They also
integration with neural network technology enhances               re-establish the process without plenty of analysis. Thus these
fuzzy logic systems on learning capabilities. It also             techniques are attracting great attention in an environment
provides a natural framework for combining both                   that is obvious with the absence of a simple and well-defined
numerical information in the form of input/output pairs           mathematical model. Besides, these models are characterized
and linguistic information in the form of IF–THEN rules           by nonrandom uncertainties which associated with
in a uniform fashion. The proposed technique is                   imprecision and elusiveness in real-time systems [3-4]. Many
accomplished by one ANFIS that achieves accurate and              researchers have studied the application of neural networks to
fast estimation of distance to fault from the relay point.        overcome most of the problems above outlined. The fuzzy set
The normalized positive sequence impedance of the three           theory is also used to solve uncertainty problems [5-11]. The
phases are considered as inputs to the network The input          use of neural nets in applications is very sparse due to its
data of the ANFIS were firstly derived from the                   implicit knowledge representation, the prohibitive
fundamental values of the voltage and current                     computational effort and so on. The key benefit of fuzzy logic
measurements after making Fourier transform. Computer             is that its knowledge representation is explicit, using simple
simulation results are shown in this paper and they               IF-THEM relations. However, it is at the same time its major
indicate this approach can be used as an effective tool for       limitation. The power system operation in transient period
location of faults for different fault conditions in fault        cannot be easily described by artificial explicit knowledge,
inception time, fault impedance, fault distance and fault         because it is affected by many unknown parameters. The
types                                                             integration of neural network into the fuzzy logic system
                                                                  makes it possible to learn from the prior obtained data sets.
Keywords: Fuzzy Neural Networks FNN, Adaptive Neuro
                                                                     This research employs adaptive-network-based fuzzy
Fuzzy Inference System ANFIS, fault location,
                                                                  inference system for fault location in the transmission lines.
transmission line protection.                                     This novel approach overcomes the difficulties associated
                                                                  with conventional voltage and current based measurements
                    I. INRODUCTION                                for transmission line protection algorithms. These difficulties
                                                                  are due to effect of factors such as fault inception time, fault
   Protection of transmission lines is very important for         impedance and fault distance. This research is integrating the
preserving of the power system. With the advent of                learning capabilities of neural network to the robustness of
microprocessors and digital electronics, digital-based relaying   fuzzy logic systems in the sense that fuzzy logic concepts are
has been developed since the late 1960s. Research activity has    embedded in the network structure. It also provides a natural
covered virtually every protection technique. Furthermore,        framework for combining both numerical information in the
many novel algorithms and associated hardware                     form of input/output pairs and linguistic information in the
implementations have emerged. The fundamentals of                 form of IF–THEN rules in a uniform fashion.
transmission lines protection were developed many years ago
[1-2]. Some of them such as representing transmission lines
by either first- or second order differential equations and
traveling-wave techniques have resulted in several

Reference Number: W09-0013                                                                                                     73
The Online Journal on Electronics and Electrical Engineering (OJEEE)                                             Vol. (1) – No. (2)

     II. ADAPTIVE NEURO FUZZY INFERENCE                                   It has no rule sharing. Different rules do not share
                 SYSTEM (ANFIS)                                            the same output membership function, namely the
                                                                           number of output membership functions must be
   A neural network which can perform pattern matching task
                                                                           equal to the number of rules.
has a large number of highly interconnected processing
                                                                          It has unity weight for each rule.
elements (nodes). These elements demonstrate the ability to
learn and generalize from training patterns. Distributed
                                                                 Figure (1) shows the architecture of the ANFIS, comprising
representation and strong learning capabilities are the major
                                                                 by input, fuzzification, inference and defuzzification layers.
features of neural network. On the other hand, decisions using
                                                                 The network can be visualized as consisting of inputs, with N
fuzzy logic systems are based on inputs in the form of
                                                                 neurons in the input layer and F input membership functions
linguistic variables. These linguistic variables are derived
                                                                 for each input, with F*N neurons in the fuzzification layer.
from membership functions which are formulas used to
                                                                 There are F^N rules with F^N neurons in the inference and
determine the fuzzy set to which a value belongs and the
                                                                 defuzzification layers and one neuron in the output layer.
degree of membership in that set. The variables are then
matched with the specific linguistic IF-THEN rules and the
                                                                 Input    inputmf              rules     outputmf
response of each rule is obtained through fuzzy implication.
To perform compositional rule of inference, the response of
each rule is weighted according to the values or degree of
membership of its inputs and the centroid of response is
calculated to generate the appropriate output. Neural network
has the shortcoming of implicit knowledge representation.
However, fuzzy logic systems are subjective and heuristic.
The determination of fuzzy rules, input and output scaling
factors and choice of membership functions depend on trial
and error that makes the design of fuzzy logic system a time
consuming task. These drawbacks of neural network and
fuzzy logic systems are overcome by the integration between
the neural network technology and the fuzzy logic systems.
The ANFIS could be viewed as a fuzzy system, a neural
network or fuzzy neural network. The structure is seen as a
neural network for training and a fuzzy viewpoint is utilized
to gain insight into the system and to simplify the model. The            Figure (1): The architecture of the ANFIS
neuro-adaptive learning method works similarly to that of        For simplicity, it is assumed that the fuzzy inference system
neural networks. Neuro-adaptive learning techniques provide      under consideration has two inputs x and y and one output z
a method for the fuzzy modeling procedure to learn               as shown in Figure (1). For a zero-order Sugeno fuzzy model,
information about a data set. It computes the membership         a common rule set with two fuzzy if-then rules is the
function parameters that best allow the associated fuzzy         following:
inference system to track the given input/output data. A         Rule 1: If x is A1 and y is B1, Then f1=r1                     (1)
network-type structure similar to that of a neural network can   Rule 2: If x is A2 and y is B2, Then f2 = r2                   (2)
be used to interpret the input/output map. So it maps inputs     Here the output of the ith node in layer n is denoted as On,i:
through input membership functions and associated
parameters. This will be translated through output               Layer 1
membership functions and associated parameters to outputs.       Every node i in this layer is an adaptive node with a node
The parameters associated with the membership functions          function:
changes through the learning process. The computation of                     O1,i=μAi(x)         for      i=1,2,3 or             (3)
these parameters (or their adjustment) is facilitated by a                   O1,i =μBi-3(y)       for    i=4,5,6                 (4)
gradient vector. This gradient vector provides a measure of
how well the fuzzy inference system is modeling the              Where x (or y) is the input to node i and Ai (or Bi) is a
input/output data for a given set of parameters. When the        linguistic label associated with this node. In other words, O1,i
gradient vector is obtained, any of several optimization         is the membership grade of a fuzzy set A1, A2 and A3 (or B1,
routines can be applied in order to adjust the parameters to     B2 and B3) and it specifies the degree to which the given
reduce some error measure. This error measure (performance       input x (or y) satisfies the quantifier A (or B). Here the
index) is usually defined by the sum of the squared difference   membership function for A (or B) is triangular membership
between actual and desired outputs. ANFIS uses a                 function and is given as:
combination of least squares estimation and back propagation
                                                                                    1                              if u  c L
for membership function parameter estimation.                                       
The suggested ANFIS has several attributes:                      Left :      u \ 
                                                                             L            0 ,1  c L  u                       (5)
      The output is zero th order Sugeno-type system.                               Max                         otherwise
                                                                                         0.5 w L 
      It has a single output, obtained using weighted                                                  
          average defuzzification. All output membership
          functions are constant.

Reference Number: W09-0013                                                                                                       74
The Online Journal on Electronics and Electrical Engineering (OJEEE)                                                            Vol. (1) – No. (2)

                      0 ,1  c  u                                                       wi. fi
                  Max                               otherwise                                                   w1
                      0 .5 w                                                   O4 i    i
Centers :  u \ 
                                                                                             wi
                                                                    (6)                                       w1  w2  w3
                  Max 0 , 1  u  c 
                                                     if u  c                                i
                      0 .5 w                                                                             w2                w3
                                                                                                                   f 2              f3     (11)
                                                                                                       w1  w2  w3      w1  w2  w3
                                          
Right :  u \ Max 0 , 1  u  c
         R                          R
                                                                   (7)   As w1, w2 and w3 are assumed to be constant. Therefore,
                                                      if u  c R
                    0 .5 w L                                           equation (2) can be rewritten as follows:
                                       
Notice that for Equation (5) c specifies the “saturation point”                   O4,i =c1.r1+c2.r2+c3.r3                                    (12)
and wL specifies the slope of the nonunity and nonzero part of            where
μL as shown in Figure (2) Similarly, for μR. For μC notice that                             w1
c is the center of the triangle and w is the base-width. cL, cR, c,               c1                                                        (13)
                                                                                       w1  w2  w3
wL, wR, and w are the parameters set. As the values of these
parameters change, the triangular function varies accordingly,                              w2
                                                                                  c2                                                        (14)
thus exhibiting various forms of membership functions for                              w1  w2  w3
fuzzy set A. Parameters in this layer are referred to as premise                            w3
parameters.                                                                       c3                                                        (15)
                                                                                       w1  w2  w3
                                   u 
                                                                          This is linear in the consequent parameters r1, r2, and r3.
                                                                          From this observation, It can be concluded that:
                                                                               S = set of total parameters,
                                                                               S1 = set of premise (nonlinear) parameters,
                                                                               S2 = set of consequent (linear) parameters
                                                                          Therefore the overall output will be:
           Figure (2) : Input triangular membership
                       functions Layer 2                                          O4,i = F(i, S)                                             (16)
Every node in this layer is a fixed node whose output is the              Where i is the vector of input variables, F is the overall
product of all the incoming signals:                                      function implemented by the adaptive network, and S is the
                                                                          set of all parameters which can be divided into two sets:
       O2,i= wi = μAi (x) μBi (y)                i=1,2,3            (8)
Each node output represents the firing strength of a rule.                        S = S1  S2                                                (17)
                                                                          Where  represents direct sum.
Layer 3
Every node i in this layer is an adaptive node with a node                Therefore, the hybrid learning algorithm can be applied
function:                                                                 directly. More specifically, the error signals propagate
                                                                          backward and the premise parameters are updated by
       O3,i = wi fi = wi ri                      i=1,2,3            (9)   Gradient Descent (GD) and node outputs go forward until
Where ri is the parameter set of this node. Parameters in this            layer 3 and the consequent parameters are identified by the
layer are referred to as consequent parameters.                           Least Squares (LS) method. This hybrid learning is organized
                                                                          as follows:
Layer 4                                                                        a) Linear and nonlinear parameters are distinguished
The single node in this layer is a fixed node which computes                   b) Each iteration (epoch) of GD update the nonlinear
the overall output as the summation of all incoming signals:                       parameters
                                                                               c) LS method follows to identify the linear parameters.
                           wi f i
Overall output = O4i  i                         i=1,2,3           (10)                III. SIMULATION ENVIROMENT
                               wi
                              i                                              The simulation environment based on MATLAB software
From the ANFIS architecture shown in Figure (1), it is                    package [12] is selected. It is used as the main engineering
observed that when the values of the premise parameters are               tool for performing modeling and simulation of power
fixed, the overall output can be expressed as a linear                    systems and relays, as well as for interfacing the user and
combination of the consequent parameters. In symbols, the                 appropriate simulation programs. ATP [13] is used for
final output in Layer 4 can be rewritten as:                              detailed modeling of power network and simulation of
                                                                          interesting events. It possesses excellent power networks
                                                                          modeling capabilities, exceptional libraries of elements and
                                                                          provides fast and accurate simulation results. Scenario setting
                                                                          and neural network relaying algorithm will be implemented in
                                                                          MATLAB and interfaced with the power network model

Reference Number: W09-0013                                                                                                                     75
The Online Journal on Electronics and Electrical Engineering (OJEEE)                                              Vol. (1) – No. (2)

implemented in ATP. MATLAB has been chosen due to                     ii) All type of faults (i.e. single phase to ground, phase to
availability of the powerful set of programming tools, signal             phase, double phase to ground or three-phase fault)
processing, numerical functions, and convenient user-friendly        iii) Inception fault time (Tf) 2 msec
interface. In this specially developed simulation environment,       iv) Fault resistances (Rf) 0, 25, 50 and 100 ohms.
the evaluation procedures can be easily performed.
                                                                    There are 444 training data. The input data to the ANFIS of
  So the power system model was simulated and the different         the locating unit are the impedances of the three phases
fault situations were performed by using ATP. Then the              (magnitude and phase i.e. 6 inputs) after dividing them by
voltage and current measurements have been sent to                  their non fault values. They are taken from the fundamental
MATLAB to demonstrate the ANFIS protective relay.                   values of the voltage and current measurements after
                                                                    evaluating Fourier transform every 20 msec. The output data
             IV. THE PROTECTION SCHEME                              from the ANFIS are the normalized fault distance value.
     A single line diagram for the protected transmission line      c) The ANFIS Locator:
  (T.L) is illustrated in Figure (3). It consists of two circuits   The ANFIS locator consists of six neurons in the input layer
  of 80 km length, 66 kV voltage level and 2 GVA short              (i.e. N=6), four triangular membership functions for each
  circuit level.                                                    input (i.e. F=4), and constant membership function for the
                                                                    d) Testing Data:
                                                                    The testing data are chosen at different fault conditions which
                                                                    are carried out at different fault distances, different fault
   Figure (3): Single line diagram for the Transmission line        resistances, different fault inception times and different fault
                                                                    types which are not chosen for the training data. Besides that
The overall protection scheme can be demonstrated as in             a white noise in introduced in the testing data to model the
Figure (4). Where:                                                  errors in the voltage and current measurements. Some of the
                                                                    simulation results are shown in Table 1.
     Vabc (VFabc) and Iabc (IFabc) are the instantaneous
                                                                    Table 1 can be explained as follows; the first three columns
         values of the three phase's voltage and current
         respectively (at fault condition).
                                                                             Fault inception time (Tf);
     V*abc (VF*abc) and I*abc (IF*abc) are the
                                                                             Fault resistance (Rf); and
         fundemantal compontents (peak values and the
                                                                             Fault type respectively.
         phases) of the three phases voltage and current
         respectively after Fourier transformation (at fault        Then the next six columns are impedances (magnitude and
         condition).                                                phase) of the three phases and these six values are used as
     Z*abc (ZF*abc) are the fundemantal compontents                input to the ANFIS detector. Then target fault distance (Df
         (magnitudes and the phases) of the three phases            p.u), finally the output of the ANFIS locator is shown in the
         impedances (at fault condition).                           next column which is the estimated per unit fault distance and
     IoF is the zero sequence current at fault condition.          the final column is the percentage error between the accurate
     CU is the control unit that receives the outcomes of          value and the estimated one where:
         the two units and only activates the fault classifier                          Dactual  Destimated
                                                                            % Error                         * 100%           (18)
         block diagram when a fault is detected.                                                Dtotal
                                                                      The testing data is chosen taking into consideration the
                                                                      faults on transmission lines are quite random in nature with
                                                                      respect to the time of occurrence, location, type and fault
                                                                      resistance. So, the testing data are taken randomly with
                                                                      random fault distances, fault resistances, fault inception
                                                                      times and fault types in each training vector.
          Figure (4): The proposed protection Layout
                                                                                          V. CONCLUISON
a) Fault Locating Unit:                                                A new digital distance relaying technique based on ANFIS
The fault Location unit is built at different situations of all     technique has been developed. ANFIS as control technique
fault types (i.e. single line to ground, double lines, double       was used to implement this relay. The relay has been tested
lines to ground and three lines fault). After that, it is tested    for different fault resistances, fault locations, fault types and
using different situations of the faulted power system.             different system conditions. In all these test cases, the
b) Training data                                                    maximum error was found to be less than 8%. The proposed
The training data used to train the ANFIS of the fault location     relaying technique has the ability to provide accurate and
unit are taken at:                                                  vigorous estimation for the fault distance in transmission
   i) Fault distance (Df) 5%, 10%, 15%, 20%, 30%, 40%,              lines.
       50%, 60%, 70% and 80%

Reference Number: W09-0013                                                                                                        76
The Online Journal on Electronics and Electrical Engineering (OJEEE)                                         Vol. (1) – No. (2)

                       Table 1: Testing data of the Fault Location Unit and their percentage errors.

                               fault     Za                Zb                   Zc                Df
              Tf       Rf      type      p.u    Za ph      p.u      Zb ph       p.u     Zc ph     p.u    % Error
             0.013     91       DL      0.98    142.3     2.59      147.1      1.38     -250     0.76      1.2
             0.007     54       TL      0.47      21      0.67          25.3   0.61     21.3     0.56      5.4
             0.002     78      DLG      0.09     75.4     1.02      138.7      0.14      5.9     0.79      3.6
             0.012     22       TL      0.17     15.4     0.21          -344   0.19      20       0.4      2.5
             0.006     86       DL        1      -220     4.95          94.9   1.03     92.8     0.51      0.5
             0.014     18      DLG      0.01     -8.7     0.01      128.4       0.8     -141     0.05      2.8
             0.003     61       DL      1.03    134.3     1.15          4.2    0.53     69.8     0.18      2.1
             0.002      0      DLG      1.06    144.4     0.09          5.3    0.06     56.2     0.56      1.9
             0.012     100     DLG      0.58     103      0.03           -1    0.02     -272.    0.15      1.1
             0.002     37       TL       0.3     16.5     0.39          18.4   0.36     19.9      0.5       2
             0.003     90       DL      0.55     70.4       1       141.5      2.51     -8.7     0.15      3.6
             0.016     27      SLG      0.53      37      0.98      143.7      0.98     -209     0.39      3.7
             0.009     60       DL        1      -10.4    0.64          75     0.99     -206     0.35      5.4
             0.001     66      SLG      1.01     53.1     0.98      145.2      0.97     -209     0.15      0.6
             0.006     58       DL      0.59     66.5       1       141.2      2.67     -26.3    0.53      1.1
             0.004     66      DLG      0.03     -6.7     0.02          82.9   0.92     -169     0.17      0.9

                                                                 [10]    Transactions on Power Delivery, Vol. 13, No. 1, pp.
                     VI. REFERENCE                                       102-108, 1998.
                                                                 [11]    Huisheng Wang and W. W. Keerthipala, "Fuzzy-Neuro
[1]   Sunil S. Rao, "Switchgear and Protection", 10th                    Approach to Fault Classification for Transmission
      Edition, KHANNA Publishers, Delhi, 1994.                           Line Protection" IEEE Transactions on Power
[2]   W. Mark Carpenter "IEEE Guide for Protective Relay                 Delivery, Vol. 13, No. 4, pp.1093-1104, 1998.
      Applications to Transmission Lines", IEEE Std              [12]    M. Jayabharata Reddy and D.K. Mohanata,
      C37.113, 1999.                                                     "Performance Evaluation of Adaptive Network Based
[3]   Jacek M. Zurada, "Introduction to Artificial Neural                Fuzzy Inference System Approach for Location of
      Systems",1st Edition, PWS Publishing Company,                      Faults on Transmission Lines Using Monte Carlo
      Boston, 1995.                                                      Simulation" This paper has been accepted for
[4]   Kevin M. Passino and Stephen Yurkovich, "Fuzzy                     publication in a future issue of IEEE journal, but has
      Control", 1st Edition, Addison Wesley Longman, Inc.,               not been fully edited, 2007.
      California, 1998.                                          [13]    MATLAB R2008a, 2008 The Math Works,
[5]   Dalia Farouk Mohamed, "A New Design of an                          Inc. (MATLAB and Simulink are registered
      Intelligent Digital Distance Protective Relay" PhD                 trademarks of The Math Works, Inc.)
      Dissertation Submitted to the Office of Graduate           [14]    ATP Draw version 3.5
      Studies of Cairo University, 2007.
[6]   Slavko Vasilic, "Fuzzy Neural Network Pattern
      Recognition Algorithms For Classification Of The
      events In Power System Network", Ph. D. Dissertation
      Submitted to the Office of Graduate Studies of Texas
      A&M University 2004.
[7]   Abeer Galal Saad, "Digital Relaying of High Voltage
      Transmission Lines by Artificial Neural Networks",
      Master Dissertation Submitted to the Office of
      Graduate Studies of Cairo University 2004.
[8]   P. K. Dash, A. K. Pradhan, and G. Panda "A Novel
      Fuzzy Neural Network Based Distance Relaying
      Scheme", IEEE Transactions on Power Delivery, Vol.
      15, No. 3, pp.902-907, 2000
[9]   D. V. Coury and D. C. Jorge, "Artificial Neural
      Network Approach to Distance Protection" IEEE

Reference Number: W09-0013                                                                                                  77