ANFIS - Identification With Special Emphasis On Training Data Minimization

Adaptive Network based Fuzzy Inference
System (ANFIS) as a Tool for System

Identification with Special Emphasis on
Training Data Minimization
MRINAL BURAGOHAIN
A
Thesis Submitted
in Partial Fulfilment of the Requirements
for the Degree of
DOCTOR OF PHILOSOPHY
by
MRINAL BURAGOHAIN
Department of Electronics and Communication Engineering

Indian Institute of Technology Guwahati
Guwahati - 781 039, INDIA
July, 2008
Adaptive Network Based Fuzzy Inference
A
Thesis Submitted
in Partial Fulfilment of the Requirements
for the Degree of
DOCTOR OF PHILOSOPHY
by
MRINAL BURAGOHAIN
Department of Electronics and Communication Engineering

Guwahati - 781 039, INDIA
July, 2008
Certificate
This is to certify that the thesis entitled “Adaptive Network based Fuzzy Inference System (AN-
FIS) as a Tool for System Identification with Special Emphasis on Training Data Minimization”,
submitted by Mrinal Buragohain, a research scholar in the Department of Electronics and Communi-
cation Engineering, Indian Institute of Technology Guwahati, for the award of the degree of Doctor
of Philosophy, is a record of an original research work carried out by him under my supervision and
guidance. The thesis has fulfilled all requirements as per the regulations of the Institute and in my
opinion has reached the standard needed for submission. The results embodied in this thesis have not
been submitted to any other University or Institute for the award of any degree or diploma.
Dated: Dr. Chitralekha Mahanta

Guwahati Associate Professor
Dept. of Electronics and Communication Engg.
Guwahati - 781 039
India
T HIS WORK IS DEDICATED TO
M Y PARENTS
Acknowledgements
I feel it is a great privilege to express my deepest and most sincere gratitude to my supervisor, Dr.
Chitralekha Mahanta for her suggestions, constant encouragement and support during the course of the
thesis work. I am also grateful to the other members of my doctoral committee, namely Prof. P.K. Bora,
Prof. S. Majhi and Dr. J. S. Sahambi for their valuable comments on my work. I take this opportunity
to thank Prof. A. K. Gogoi and Prof. S. Dandapat, the former heads of the department and the present
HOD, ECE for their kindness in allowing me to use various computing facilities of the department. I
would also like to thank all the other faculty members of the department for their encouragement and
help.
I am also grateful to all the members of the Research and technical staff of the department, namely
Jharana Rani Rabha, L. N. Sharma, Pranab Kr. Goswami, Sanjib Das, Sidananda Sonowal, Usha Bhar-
ali and Utpal Kumar Sharma without whose help I could not have completed this assignment. I thank
all my fellow research scholars and M. Tech students for their cooperation. My thanks also go out to
all my friends, namely Ahmad Ali, Amrita Ganguly, Atanu Kr Dutta, D. Senthil, Diganta Kr Gogoi,
Manas Kamal Bhuyan, P. Vinod, Prabin Kr Padhy, S. K. Mahesh, Senthil Raja, Utkal Mehta and others
who have rendered their help at different times which made my stay in this institute a memorable period
of my life.
I also acknowledge the endurance and perseverance of my wife Meghali for the last few years without
whose sacrifice, love, encouragement and assistance, I would not have been able to complete my re-
search.
Finally, I also express my heartfelt gratitude to all my family members for their unbounded love, en-
couragement and assistance throughout. I am deeply indebted to them for their countless sacrifices and
support without which my research work would not have finished.
(Mrinal Buragohain)
Abstract
Nearly two decades back nonlinear system identification consisted of several ad-hoc approaches which
were restricted to a very limited class of systems. However, with the advent of the various soft com-
puting methodologies like neural networks, the fuzzy logic and the genetic algorithm combined with
modern structure optimization techniques, a wider class of systems can be handled at present. Complex
systems may be of diverse characteristics and nature. These systems may be linear or nonlinear, con-
tinuous or discrete, time varying or time invariant, static or dynamic, short term or long term, central
or distributed, predictable or unpredictable, ill or well defined. System outputs may be measurable or
unmeasurable. Models of real systems are of fundamental importance in virtually all disciplines and
hence there is a strong demand for advanced modeling, identification and controlling schemes. This is
because models help in system analysis which in turn help to get a better understanding of the system
for predicting or simulating a system’s behavior. Also, system models facilitate application and valida-
tion of advanced techniques for controller design. Development of new processes and analysis of the
existing ones along with their optimization, supervision, fault detection, and component diagnosis are
all based on the models of the systems. As most of the real world systems are nonlinear in nature, an
endeavor is made for modeling a nonlinear system in the present work. A linear system is considered
to be a special case of the nonlinear system. The challenges involved in modeling, identification and
control of a nonlinear system are too many and attempt has been made to tackle them by applying
various soft computing methodologies. In most of the conventional soft computing methods the system
modelling results are dependent on the number of training data used. It has been found that the model-
ing results improve as the number of training data increases. But in many complex systems the number
of available training data are less and the generation of new data is also not cost effective. In such a
scenario the system has to be modelled with the available data. The proposed modeling scheme has
been devised keeping such a possibility in mind. The results obtained by applying this proposed model
are compared with the results obtained by using various statistical and genetic algorithm based fuzzy
models and finally the relative merits and demerits involved with the respective models are discussed.
The work embodied in the present thesis is concerned with optimal design of the conventionally ex-
isting soft computing based system models. The statistics based Full factorial design (FFD) and the
V-fold cross validation technique are applied to augment a conventional neuro-fuzzy technique and the
following observations are noted :
• The results obtained by applying the proposed technique are comparable and in some cases su-
perior to those obtained by using the conventional neuro-fuzzy model.
• Comparable or superior results are obtained with this proposed model even though the number
of data pairs used for system modeling here are less as compared to that used in the conventional
methods.
• It resulted in reduction of the number of computations involved. As the experiments were per-
formed by using reduced number of specifically chosen data, the number of computations re-
quired to be performed also came down.
Contents
List of Figures iii
List of Tables v
Nomenclature vi
Mathematical Notations viii
1 Introduction 1
1.1 Aim of the research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Contribution of the thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Organization of the thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2 Soft Computing based Techniques for System Identification 6

2.1 Neural Network based Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 Fuzzy Logic based Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2.1 Analysis with Fuzzy Inference System . . . . . . . . . . . . . . . . . . . . . . 9
2.2.2 Types of Fuzzy System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3 Genetic Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.4 Neuro Fuzzy Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.4.1 Adaptive Network based Fuzzy Inference System . . . . . . . . . . . . . . . 18
2.5 Forecasting Methodologies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.6 Research gap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3 Research Objectives and Methodology 24

3.1 Objectives of the present work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.2 Methodology of the present work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.2.1 ANFIS based methodology for modeling . . . . . . . . . . . . . . . . . . . . 25
3.2.2 Full factorial design based methodology for optimal data selection . . . . . . 26
3.2.3 V-fold based methodology for optimal data selection . . . . . . . . . . . . . . 26
3.2.4 Full factorial design combined with V-fold based methodology for optimal data
selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.2.5 Model building with statistical approaches for performance study . . . . . . . 27
3.2.6 Genetic algorithm based fuzzy methodology . . . . . . . . . . . . . . . . . . 27
3.2.7 Controller with the FFD-V-fold based methodology . . . . . . . . . . . . . . 27
i
CONTENTS
4 Adaptive Network based Fuzzy Inference System 28

4.1 ANFIS structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4.2 Learning Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
4.3 Derivation of the Initial Fuzzy Model . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.3.1 Extracting the initial fuzzy model . . . . . . . . . . . . . . . . . . . . . . . . 31
4.3.2 Selection of input variables and final fuzzy model . . . . . . . . . . . . . . . 35
4.3.3 Optimization of the initial fuzzy model . . . . . . . . . . . . . . . . . . . . . 37
4.4 Experimental Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.5 Experimental Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4.5.1 ANFIS Model for Gas Furnace . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.5.2 ANFIS Model for Thermal Power Plant . . . . . . . . . . . . . . . . . . . . . 41
4.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
5 Optimal Data based ANFIS Modeling 54

5.1 Full Factorial Design Based ANFIS Modeling . . . . . . . . . . . . . . . . . . . . . . 54
5.2 V-Fold Technique Based ANFIS Modelling . . . . . . . . . . . . . . . . . . . . . . . 58
5.3 FFD-V-fold Based ANFIS Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
5.4 Experimental Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
5.4.1 Modeling results with Full factorial design based ANFIS model . . . . . . . . 60
5.4.2 V-fold technique based ANFIS model . . . . . . . . . . . . . . . . . . . . . . 67
5.4.3 Modeling results with FFD-V-fold technique based ANFIS model . . . . . . . 68
5.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
6 Statistical Models and Genetic Algorithm Based Fuzzy Model 76

6.1 Statistical Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
6.1.1 Auto Regressive with Exogenous Input (ARX) Model . . . . . . . . . . . . . 77
6.1.2 Auto Regressive Moving Average with Exogenous Input (ARMAX) Model . . 77
6.1.3 Auto Regressive Integrated Moving Average with Exogenous Input (ARIMAX)
Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
6.2 Genetic Algorithm Based Fuzzy Model . . . . . . . . . . . . . . . . . . . . . . . . . 78
6.3 Experimental Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
6.3.1 Statistical models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
6.3.2 GA based fuzzy model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
6.4 Controller Using FFD-V-fold based ANFIS Model . . . . . . . . . . . . . . . . . . . 88
6.4.1 Controller Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
6.4.2 Experimental Results and Discussion . . . . . . . . . . . . . . . . . . . . . . 92
6.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
7 Conclusions and Future Work 98

7.1 Summary of the thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
7.2 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
7.3 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
7.4 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
A Sample Power Plant Data 103
Bibliography 110
ii
List of Figures
2.1 Crossover operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

2.2 Mutation operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3 GA algorithm flowchart . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
4.1 Type-3 ANFIS Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

4.2 Grid partitioned fuzzy subspaces for a type-3, 2-input ANFIS . . . . . . . . . . . . . . 35
4.3 Variable Selection Process for a four input initial model . . . . . . . . . . . . . . . . . 36
4.4 Rules for the grid partition based ANFIS model when trained with entire data set for
the gas furnace example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
4.5 Rules for the subtractive clustering based ANFIS model when trained with entire data
set for the gas furnace example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
4.6 Actual and ANFIS model predicted output with Box and Jenkins gas furnace data . . . 41
4.7 ANFIS model prediction error for Box and Jenkins gas furnace data . . . . . . . . . . 43
4.8 RMSE of the ANFIS model during training for Box and Jenkins gas furnace data . . . 44
4.9 Output surface of the data for Box and Jenkins gas furnace example . . . . . . . . . . 44
4.10 Rules for the ANFIS model when trained with half of the data set for the gas furnace
example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.11 Actual and ANFIS model predicted output with Box and Jenkins gas furnace data . . . 46
4.12 Rules for the ANFIS model when trained with half of the data set-I for the thermal
power plant example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.13 Actual and ANFIS model predicted output with the data set-I for thermal power plant . 47
4.14 Rules for the ANFIS model when trained with half of the data set-II for the thermal
4.15 Actual and ANFIS model predicted output with the data set-II for thermal power plant 48
4.16 Rules for the ANFIS model when trained with half of the data set-III for the thermal
4.17 Actual and ANFIS model predicted output with the data set-III for thermal power plant 49
4.18 Rules for the ANFIS model when trained with half of the data set-IV for the thermal
4.19 Actual and ANFIS model predicted output with the data set-IV for thermal power plant 50
4.20 Rules for the ANFIS model when trained with half of the data set-V for the thermal
4.21 Actual and ANFIS model predicted output with the data set-V for thermal power plant 52
5.1 Selection of training data in a 23 factorial design . . . . . . . . . . . . . . . . . . . . 56

5.2 Flow Chart for the FFD-V-fold Technique . . . . . . . . . . . . . . . . . . . . . . . . 60
iii
LIST OF FIGURES
5.3 Actual and FFD-V-fold based ANFIS model predicted output with the Box and Jenkins
gas furnace data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.4 Actual and FFD-V-fold based ANFIS model predicted output with data set-I of thermal
power plant . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.5 Actual and FFD-V-fold based ANFIS model predicted output with data set-II of thermal
power plant . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.6 Actual and FFD-V-fold based ANFIS model predicted output with data set-III of ther-
mal power plant . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.7 Actual and FFD-V-fold based ANFIS model predicted output with data set-IV of ther-
mal power plant . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
5.8 Actual and FFD-V-fold based ANFIS model predicted output with data set-V of thermal
power plant . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
6.1 Flow Chart for the GA based Fuzzy model . . . . . . . . . . . . . . . . . . . . . . . . 79

6.2 Actual and ARIMAX model predicted output with Box and Jenkins gas furnace data . 81
6.3 Actual and ARIMAX model predicted output with data set-I of thermal power plant . . 81
6.4 Actual and ARIMAX model predicted output with data set-II of thermal power plant . 82
6.5 Actual and ARIMAX model predicted output with data set-III of thermal power plant . 83
6.6 Actual and ARIMAX model predicted output with data set-IV of thermal power plant . 83
6.7 Actual and ARIMAX model predicted output with data set-V of thermal power plant . 84
6.8 Actual and GA based fuzzy model predicted output with Box and Jenkins gas furnace
data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
6.9 Actual and GA based fuzzy model predicted output with data set-I of thermal power plant 85
6.10 Actual and GA based fuzzy model predicted output with data set-II of thermal power
plant . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
6.11 Actual and GA based fuzzy model predicted output with data set-III of thermal power
plant . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
6.12 Actual and GA based fuzzy model predicted output with data set-IV of thermal power
plant . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
6.13 Actual and GA based fuzzy model predicted output with data set-V of thermal power
plant . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
6.14 Training phase of the inverse controller . . . . . . . . . . . . . . . . . . . . . . . . . 90
6.15 Application phase of the inverse controller . . . . . . . . . . . . . . . . . . . . . . . . 90
6.16 Inverse controller with fuzzy integrator block . . . . . . . . . . . . . . . . . . . . . . 91
6.17 Desired output and the actual output of the plant with the controller . . . . . . . . . . 93
6.18 Prediction error of the controller . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
6.19 Output of the controller . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
6.20 Desired and actual output of the plant with controller in presence of disturbance . . . . 95
6.21 Desired and actual output of the plant with controller having plant parameter variations 95
6.22 Prediction error of the controller with plant parameter variations . . . . . . . . . . . . 96
6.23 Output of the controller with plant parameter variations . . . . . . . . . . . . . . . . . 96
6.24 Desired and actual output of the controller based plant with plant parameter variations
along with disturbance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
iv
List of Tables
4.1 Comparison of various models derived for the Box and Jenkins gas furnace data. The
first 7 rows are excerpted from a table in [76] . . . . . . . . . . . . . . . . . . . . . . 42
4.2 Modeling results for the Box and Jenkins gas furnace data and the thermal power plant
data using ANFIS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
data using FFD based ANFIS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
data using V-fold based ANFIS model . . . . . . . . . . . . . . . . . . . . . . . . . . 68
data using FFD-V-fold based ANFIS . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.4 Comparison of modeling results for the Box and Jenkins gas furnace data and the ther-
mal power plant data using various ANFIS models . . . . . . . . . . . . . . . . . . . 71
6.1 Results obtained for Statistical models using the Box and Jenkins gas furnace and the
thermal power plant data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
6.2 Results obtained for the GA based fuzzy model using the Box and Jenkins gas furnace
and the thermal power plant data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
6.3 Comparison of the FFD-V-fold based ANFIS, GA based fuzzy and the statistical ARI-
MAX model obtained for the Box and Jenkins gas furnace and the thermal power plant
data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
A.1 Thermal power plant data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
v
Nomenclature
AI Artificial Intelligence
ANFIS Adaptive network based fuzzy inference system
ANN Artificial neural network
AR Auto regressive
ARIMAX Auto regressive integrated moving average with exogenous input
ARMAX Auto regressive moving average with exogenous input
ARX Auto regressive with exogenous input
BD Bad data
BP Back propagation
DAS Data acquisition system
DDEC Data driven echo canceller
EOF Effect of factor
FFD Full factorial design
FIS Fuzzy inference system
FL Fuzzy logic
FLC Fuzzy logic controller
GA Genetic algorithm
vi
LP Linear Programming
LSE Least square estimation
MA Moving average
MIMO Multi input multi output
MLP Multi layer perceptron
MSE Mean square error
NC Neuro computing
RBFN Radial basis function network
RHONN Recurrent high order neural network
RMEEI Recursive measurement error estimation identification
RMSE Root mean square error
SISO Single input single output
SMT Statistical machine translation
SOFC Self organizing fuzzy controller

Mathematical Notations
µ Membership value
ra Radius of the cluster
rb Radius defining the neighborhood of the cluster
α Constant whose value depends on ra
β Constant whose value depends on rb
viii
Chapter 1
Introduction
Most of the present day systems are large and may be considered to be complex in nature. Electrical
power, chemical, water treatment and similar large-scale industrial plants are all complex in nature.
Complex systems may be linear or nonlinear, continuous or discrete, time varying or time invariant,
static or dynamic, short term or long term, central or distributed, predictable or unpredictable, ill or
well defined. Also, system outputs may be measurable or unmeasurable. They may consist of many
interconnected systems, sub-processes or components. The processes involved in the complex systems
may possess widely varying properties. In large scale systems, every part performs a desired function
and the overall system works satisfactorily only if all the different parts work in tandem for what they
are designed for. Modeling of complex systems is of fundamental importance in almost all fields. This
is because models facilitate better understanding of the system and so help in system analysis. So pre-
diction and simulation of the system’s behavior are then possible. System model also helps to design
new processes and analyze the existing ones. The design, optimization and supervision of controllers,
fault detection and faulty component diagnosis are all based on the system model. This is because for
the improvement of the system’s performance, it is required to model the system correctly so that the
model parameters can be tuned to get the required system response. It is because of this fact that in the
last few decades, modeling of large scale, complex systems has been a special topic of interest among
the researchers of various disciplines worldwide [1]. Most of the real world systems are ill defined
in nature and hence difficult to model. Generally the performance of the system is dependent on the
accuracy of the model. Therefore it is of utmost importance to build a model which correctly reflects
the behavior of the system under consideration. The functioning of complex large-scale systems also
involves numerous tradeoff problems like cost and accuracy [2]. Hence, there is a strong demand for
developing advanced methods of system modeling and identification techniques. The conventional
1
2
methods that have been used for system modeling rely heavily on the mathematical tools which require
precise knowledge about the involved physical processes. In systems where the mathematical model
is not available, it is not possible to use the conventional methods for its analysis. In such cases, soft
computing based modeling [3] approaches provide a viable alternative for identification of the system
from the available data. The concept of soft computing [4] began to materialize near about the time
when Lotfi Zadeh was working on soft analysis of data and fuzzy logic. This gave birth to the intelligent
systems. Nearly four decades later, the intelligent system became a reality. However, initially the tech-
nology needed for building systems that possess Artificial intelligence (AI) was not available. Instead
only predicate logic and symbol manipulation techniques formed the core of the traditional AI. These
techniques could not be used for building machines which could be called intelligent from the point
of view of real world application. But today the requisite hardware, software and sensor technology
are available for building intelligent systems. In addition to these, computational tools are available
now which are far more effective for conception and design of intelligent systems. These tools are
derived from a collection of methodologies called soft computing. Unlike hard computing the essence
of soft computing is aimed at accommodating the prevalent imprecision of the real world. Therefore
soft computing helps in exploiting the tolerance for imprecision, uncertainty and partial truth so that
tractability, robustness, low solution cost and better rapport with reality can be achieved. Hence the
human mind can be considered to be a role model for soft computing. Rather than a single technique,
soft computing may be considered to be comprising of different methodologies with Neuro-computing
(NC), the Fuzzy logic (FL) and the Genetic algorithm (GA) as the principal partners. Therefore in soft
computing based system identification, instead of a single standard method, a collection of techniques
has been put forward as possible solutions to the identification problem. They can be broadly grouped
as neural network based algorithm, fuzzy logic based algorithm and the genetic algorithm. The neural
network has the inherent advantage of being able to adapt itself and also in its learning capabilities.
Similarly the salient feature that is associated with the fuzzy logic is the distinct ability to take into
account the prevailing uncertainty and imprecision of real systems with the help of the fuzzy if-then
rules. In order to exploit the advantage of the self adaptability and learning capability of the neural
network and the capability of the fuzzy system to take into account of the prevailing uncertainty and
imprecision of real systems with the help of the fuzzy if-then rules, an integrated forecasting approach
comprising of both the fuzzy logic and the neural network has been considered. This hybrid system is
called the Adaptive network based fuzzy inference system (ANFIS). Here the fuzzy system with its ex-
pert knowledge stands as a front end preprocessor for the neural network input and output layers. Based
1.1. AIM OF THE RESEARCH 3
on the historical data, the neural network learning algorithms are used to determine the parameters of
the expert knowledge based fuzzy system. The use of this hybrid system ANFIS helps to complement
the weakness of the respective systems.
1.1 Aim of the research

The main research objective that has been kept in mind while initiating and furthering the present work
is the modeling of a real world system with the help of the soft computing technique. In the quest
for developing a model for a system based on its available input output data, it has been observed that
in the conventional modeling approach the results depend on the mathematical model of the system
and its accuracy. In cases where the mathematical model is not available the system analysis becomes
very difficult. It is in this context that the soft computing approach can provide a viable alternative.
The prime inherent advantage associated with the soft computing techniques of not requiring a math-
ematical model has been a motivating factor for consideration in our present work. Motivated by this
advantageous feature of soft computing based system identification, the present work focuses on build-
ing a model for an ill defined real world system based on its available record of input-output data using
ANFIS. An endeavor is made to extract previously unknown information from the available time-series
data so that an accurate model can be built. Once built, the model can be used to predict or forecast fu-
ture values. The present research concentrates on the difficult situation when the available input-output
data for a system are very less and generation of real time data is also prohibitively expensive. In such
cases, it is really a challenging task to build a faithful model for the system using soft computing based
data driven identification technique like ANFIS. This research attempts to find a solution to this prob-
lem of modeling an ill defined real world system faithfully in the situation when available input-output
data is scanty.
1.2 Contribution of the thesis

The major contributions of this thesis can be outlined as follows:
• Building an ANFIS model for real world systems like gas furnace and thermal power plant
• Proposal of a novel technique for training the neural network in the ANFIS model by optimally
1.3. ORGANIZATION OF THE THESIS 4
selecting the available data using the Full factorial design (FFD) and the V-fold cross validation
method
• Proposal of a hybrid method for optimal data selection incorporating both the above techniques
of full factorial design and V-fold cross validation
• Proposal of a GA based fuzzy model
• Design of an ANFIS based controller to study the performance of the proposed ANFIS model
1.3 Organization of the thesis

The thesis records a detailed account of the use of the soft computing techniques for system identifica-
tion. The organization of the thesis is given below.
Chapter 2: A brief overview of the system identification techniques based on soft computing method-
ology is presented in this chapter. Available literature about model based forecasting techniques
is also reviewed in this chapter.
Chapter 3: In this chapter the objectives and methodology used in the research are discussed.
Chapter 4: In this chapter the ANFIS architecture is discussed in details. This chapter is devoted to
the modeling of systems based on ANFIS.
Chapter 5: In this chapter the use of two new techniques proposed in the present research for optimal
data selection in training the ANFIS model is explained. These two methods are the Full factorial
design (FFD) and the V-fold technique by using which the number of data required for training
the ANFIS was drastically reduced. Also another hybrid model of ANFIS combining these two
techniques has been proposed.
Chapter 6: In this chapter the proposed GA based fuzzy model for system identification is discussed.
This chapter also gives a detailed account of the statistical models namely ARX, ARMAX and
ARIMAX which are applied to the system under consideration. The results obtained from the
proposed soft computing based model are compared with those obtained by using these conven-
tional statistical models as well as GA based models.
1.3. ORGANIZATION OF THE THESIS 5
Chapter 7: The last chapter presents a brief discussion about the research work and also draws con-
clusions by analyzing the results obtained in this thesis. Also a few areas for further investigation
are suggested in this chapter.
Chapter 2
Soft Computing based Techniques for System

Identification
Contrary to the conventional hard computing techniques, the prime inherent advantage associated with
the soft computing techniques is the non requirement of a mathematical model and hence are becoming
increasingly popular as system identification methodology. Three powerful soft computing techniques
which are very popular are the Neural network, the Fuzzy logic and the Genetic algorithm (GA). A brief
overview of available methods using these techniques for identification and control of linear as well as
nonlinear dynamical systems is presented in this chapter. These soft computing based approaches are
reviewed thoroughly in this chapter since this research focuses on utilizing soft computing as a tool
for system identification for the purpose of modeling an ill defined real world system so as to forecast
future values.
2.1 Neural Network based Algorithms

In addition to being the source of natural intelligence, the human brain can process incomplete infor-
mation obtained by perception at a very rapid rate. Inspired by this biological property of the nervous
systems and the brain, researchers attempted to model the human brain resulting in the evolution of the
neural network. Here the brain has been modeled as a continuous time nonlinear dynamic system with
a connection architecture. In this architecture the neurons or the processing units which are intercon-
nected by weights are expected to mimic the human brain. This gives the neural network the capability
for learning and adaptation by adjusting the interconnection between the layers. The most important
6
2.1. NEURAL NETWORK BASED ALGORITHMS 7
characteristics of the neural network are:
• Presence of a large number of simple units
• Presence of a large number of highly parallel units
• Presence of strongly connected units
• Robustness against the failure of single units
• Learning from data
The network of any system will be considered to be an artificial neural network if the same basis
function is used throughout the network. Here the nodes of the system are called the neurons. The
layers of a neural network architecture may be subdivided into three principal groups:
• Depending upon the number of inputs all the input neurons together constitute the input layer
• Similarly all the output neurons together constitute the output layer
• All the neurons in any intermediate layer form the hidden layer. There may be more than one
hidden layer in a neural network
Some of the popular neural network architectures are the Multi-layer perceptron (MLP) network,
Radial basis function network (RBFN) and the neuro-fuzzy network. The pivotal contribution of neural
networks is a methodology for identification, learning and adaptation.
The early works on neural networks include those of McCulloch and Pitts [5], Hebb [6], Rosenblatt [7],
Widrow and Hoff [8], Minsky and Papert [9], Hopfield [10], Parker [11], Rumelhart and McClel-
land [12], Carpenter and Grossberg [13] and Kohonen [14]. Narendra and Parthasarathy [15] demon-
strated the use of neural network for the identification and control of nonlinear dynamical systems.
Polycarpo and Ioannou [16] proposed the general formulation for modelling, identification and control
of a nonlinear dynamical system. Another development was the design and analysis based on Lya-
punov stability theory [16, 17]. Pham and Liu [18] proposed the use of recurrent neural networks for
the identification of linear and nonlinear dynamic systems. The interest in the use of neural networks
for modeling and identification of static and dynamical complex system on the basis of the input-output
data pairs was a new development. Kosmatopoulos and Christodoulou [19] proposed an algorithm for
identification of nonlinear systems using Recurrent high order neural network (RHONN) based on the
2.2. FUZZY LOGIC BASED ALGORITHMS 8
extension of Hopfield [20] and Cohen-Grossberg [21] works. Neural networks with Radial basis func-
tion (RBF) are used due to their excellent classification property. Sanner and Slotine [17] presented an
approach with a Gaussian radial-basis function adaptive dynamical system with unknown nonlineari-
ties. Hong and Xinkuo [22] proposed a neural network approach by combining the equivalence between
RBF and the Fuzzy inference system (FIS) for identification of a nonlinear system. Ahmad et al. [23]
presented a nonlinear Multi-input-multi-output (MIMO) system identification scheme which is based
on the radial basis function network. Similarly Selmic and Lewis [24] presented a multi model identifi-
cation scheme by using nonlinear system identification technique with the RBF based neural network.
Azam and Valandham [25] presented a RBF based neural network which uses the log-sigmoid as the
basis function for identification purpose. This function eliminates the risk of mathematical instabil-
ities which are found while using Gaussian radial basis function based networks. Robert et al. [26]
proposed a class of additive dynamic connectionist model for the identification of unknown dynamic
systems with the help of two online parameter adaptation algorithms. One of these algorithms is based
on gradient descent [27–29] technique and sensitivity analysis while the other is based on the varia-
tional calculus. To deal with the problem of time variation of disturbance and system parameters, Song
and Soh [30] proposed an adaptive and robust identification algorithm. This method can overcome the
disturbance problem by the selection of a robust adaptive dead zone scheme. The main drawback of the
feed forward Multilayered perceptron (MLP) with Back-propagation (BP) algorithm is the requirement
of intensive computation and the slow rate of error convergence. To remove this drawback, Patra and
Chen [31, 32] proposed the Chebyschev functional link artificial neural network (C-FLANN) compris-
ing of the Chebyschev polynomials. This method was found to have superior performance as compared
to the MLP requiring less computation for the task of nonlinear system identification. Ren et al. [33]
proposed an algorithm for the online identification and control of a class of continuous time higher
order nonlinear system using dynamic neural network.
2.2 Fuzzy Logic based Algorithms

The pivotal contribution of fuzzy logic is a methodology for computing with words which can deal
with imprecision and granularity. The human brain can interpret and process imprecise and incomplete
sensor information which are received from the perceptive organs. Analogously the fuzzy set theory
can also provide a systematic approach to deal with such information linguistically. It can also perform
numerical computation by using membership function for the stipulated linguistic labels. The Fuzzy
inference system (FIS) is based on the concepts of fuzzy set theory, fuzzy if-then rules and fuzzy
reasoning. The framing of the fuzzy if-then rules forms the key component in FIS. FIS is a very
popular technique and has been widely applied in different fields like data classification, automatic
control, expert system, decision making, robotics, time series analysis, pattern classification, system
identification etc. The basic structure of a fuzzy inference system consists of three principal components
viz a rule base comprising of the selected fuzzy rules, a database defining the membership functions of
the fuzzy rules, and a reasoning mechanism which performs a fuzzy reasoning inference with respect
to the rules so as to derive a reasonable output or conclusion.
2.2.1 Analysis with Fuzzy Inference System
For the analysis of a fuzzy system whose inputs and outputs are described by linguistic variables, the
following steps have to be carried out:
• Fuzzification :-The linguistic variables of the fuzzy rules are expressed in the form of fuzzy sets
where these variables are defined in terms of degree of their associated membership functions.
This method of calculating the degree of belongingness of the crisp input in the fuzzy set is
called the fuzzification. The membership functions may be triangular, trapezoidal, gaussian or
bell shaped. As the information about the degree of the membership is used for further process-
ing, considerable amount of information may be lost during the course of fuzzification. This is
because the procedure can be seen as a nonlinear transformation of the inputs. For example in
the case of triangular or trapezoidal membership functions information is lost in the regions of
membership functions where the slope is zero, as at these points the membership functions are
not differentiable. Therefore fuzzy systems having triangular or trapezoidal membership function
can encounter problems of learning from data. Smoother membership functions like gaussian or
bell function may be used to overcome this difficulty.
• Aggregation :-After the degree of each linguistic statement is evaluated, they are combined by
logical operators such as AND and OR. The conjunction of these linguistic statements is carried
out by logical t-norm and the t-conorm operator to a large number of linguistic statements. Max
and Min operators are used for classification task. For the purpose of approximation and identifi-
cation the product and algebraic product operators are better suited due to their smoothness and
differentiability. Similarly the bounded sum and difference operators offer several advantages to
some neuro-fuzzy learning schemes.
• Activation :-Here the degree of rule fulfilment is used to calculate the output activations of the
rules.
• Accumulation :-In this step the output activations of all the rules are joined together to give rise
to the fuzzy output of the system.
• Defuzzification :-If a crisp value of the system is required, the final fuzzy output has to be de-
fuzzified. This can be done by different methods like center of gravity, bisector of area, mean
of maximum (mom), smallest (absolute) of maximum (som) and largest (absolute) of maximum
(lom).
2.2.2 Types of Fuzzy System
A fuzzy system may be of three principal types, namely:
• Mamdani fuzzy system :-This type of system is also known as the linguistic fuzzy system.
• Singleton Fuzzy system :-The complexity of defuzzification of a linguistic fuzzy system can be
simplified by restricting the output to a singleton membership function. Since no integration has
to be carried out numerically, this results in reducing the computational demand for the evaluation
and learning of the fuzzy system. Therefore a singleton fuzzy system is most widely applied in
industry.
• Takagi-Sugeno Fuzzy system :- This system may be considered to be an extension of the singleton
fuzzy system. Here the function f is not a fuzzy set. But the premise of a Takagi-Sugeno
fuzzy system [34,35] is linguistically interpretable. For a dynamic process modelling the Takagi-
Sugeno models possess an excellent interpretation. A singleton fuzzy system can be recovered
from a Takagi-Sugeno fuzzy system if the function f is chosen to be a constant. As the constant
can be seen as a zeroth order Taylor series expansion of the function f , it is also called the
zeroth order Takagi-Sugeno fuzzy system. However, in most of the applications, the first order
Takagi-Sugeno fuzzy system is more common.
Since the introduction of the fuzzy logic concept by Zadeh [36, 37], research was continuing for ap-
plication of fuzzy system theory for system identification [38–40]. This is because in many complex
and ill-defined systems where precise mathematical models are difficult to build, their fuzzy models
can be obtained easily which reflect the uncertainty of the system in a proper way. There are numerous
applications of Fuzzy logic controller (FLC) in industrial processes [41–43]. The interpolation of the
fuzzy rule base provided by human experts governs the performance of a FLC. Self-organizing fuzzy
controller (SOFC) was proposed by Mamdani and Procyk [44] where the fuzzy rule base is modified
by evaluating the system performance that is expressed linguistically. Park and Lee [45] proposed an
SOFC where the rules are generated by input-output data. Here the rules get updated by self learning
procedure. Graham and Newell [46] presented another approach where the fuzzy model of the process
is identified online and the control input is calculated based on the identified model of the process and a
heuristically determined performance measure. Here the system is considered to be linear. It has been
observed that a lot of researchers are paying attention to the fuzzy relational model. This is because
the fuzzy relational equations reflect the fuzzy nature of the system effectively. These relations are
universal approximators which can perform nonlinear mappings between input-output variables which
are treated as fuzzy sets. There are two aspects of fuzzy relational equation method. One aspect is
based on the resolution of composite fuzzy relational equation [47, 48] and the other is the linguistic
model identification as proposed by Tong [49]. But as Tong’s model could not be extended to a higher
dimensional system, Li and Liu [50] proposed an adaptive mechanism based on the decision table form
of simple linguistic models. Pedrycz [48] proposed another fuzzy compositional rule based system
identification algorithm which uses the concept of referential fuzzy set and Zadeh’s conditional pos-
sibility distribution. Babu and Sachidanand [51] presented another method for the identification of a
nonlinear system based on the fuzzy relational model. Here the input is considered to be linear and the
output is considered to be nonlinear in nature. Lee et. al. [52] proposed a combined approach to fuzzy
model identification which combines the linguistic approach and the numerical resolution of fuzzy re-
lational equation. Moore and Harris [53] proposed an indirect adaptive fuzzy control method which
uses a first order fuzzy model for the identification of the plant. Xu [54] showed the application of the
fuzzy system for the identification of a nonlinear system. Up till now, there are three different kinds
of fuzzy models which have been in use from the fuzzy control viewpoint. In the first class the fuzzy
basis function approximation [55] has been used, which may be considered as a mapping between the
input and the output space. This model suffers from the disadvantage that some important dynami-
cal behavior of the system can not be represented. In the second type of fuzzy model as described
by Takana and Sugeno [56], global function approximation can be achieved from a set of local linear
equations. But this model suffers from the disadvantage that it cannot be used for controller design.
The third model is called the fuzzy dynamic model [57] which requires accurate determination of the
upper bound of the local model. Huaguang and Yongbing [58] proposed a fuzzy hyperbolic model for
2.3. GENETIC ALGORITHM 12
a class of complex systems which is difficult to model. Sugeno and Yasukawa [59] presented a general
approach for qualitative modeling based on fuzzy logic. Cipriano and Montoya [60] proposed a fuzzy
model for the identification of nonlinear systems. Similarly Gaweda and Zurada [61] presented a fuzzy
identification system with relational input partition. Saez and Cipriano [62] proposed the application
of fuzzy models for the control of a combined cycle power plant boiler. Elshafei and Karray [63] pro-
posed a fuzzy model for the identification of a class of black-box type nonlinear systems. Similarly
Saez and Cipriano [64] proposed a fuzzy model for the representation of the nonlinearity of a process.
Again Chen and Linkens [65] proposed a method for the fuzzy modeling from numerical data. Flores
et al. [66] applied fuzzy logic for controlling a solar power plant.
The fuzzy if-then rules contain the structured knowledge representation of the fuzzy inference sys-
tem. But this does not provide the adaptive capability to the fuzzy inference system for dealing with
the changing external environment which is found in a neural network.
2.3 Genetic Algorithm

Genetic Algorithm was envisaged by John Holland [67] at the university of Michigan, in the 1970s.
It is based on the Darwinian evolution theory of survival of the fittest which states that the fitter and
stronger individuals in a population have a higher chance of creating offsprings for the next generation
by random mutation and natural selection. It can be implemented as an optimization search procedure
which uses the principles of genetics and natural selection by modelling possible solutions to a search
problem as strings of zeroes and ones. Each point in a parameter or solution space can be encoded
by Genetic algorithm (GA) in a binary bit string, which is called the chromosome. For example, if a
particular point (8, 5, 7) in a three dimensional space, is to be transformed, it can be represented by a
concatenated binary string of 1000, 0101, 0111. Here each coordinate is encoded by a gene which is
represented by four binary bits. Encoding plays a key role in determining the GA’s performance as
it helps to translate problem specific knowledge directly to the GA framework. After the creation of
a generation comprising of these points, each of these points is also assigned a fitness value which is
given by the evaluated value of the objective function at that point. Generally positive values of the
objective function are preferred. So if the fitness value is not positive some kind of monotonical scaling
and/or translation can be done [68]. The problem can also be sometimes tackled by calculating the
fitness value in terms of the ranking of the members in a population. A collection of these points is kept
as a gene pool or population which can be repeatedly evolved towards a better fitness value in the next
successive generations, by upgrading the entire population. After the completion of the encoding and
the fitness evaluation steps, the GA constructs a new population in each successive generation by using
three basic genetic operators. They are:
• Selection:- The selection operation determines the parents which will be participating in mating
to produce offsprings for the next generation. This is analogous to the survival of the fittest in the
natural selection. Starting from a possible solution strings, pairs of individuals from the current
population are allowed to mate to produce offsprings for the next generation. This selection
procedure is based on the strategy of survival of the fittest. Generally the selection of members
for mating are based on their selection probability which is proportional to their fitness value.
Usually the roulette wheel selection strategy is used where depending upon the string’s relative
fitness size a string is assigned a slot in the simulated wheel. The selection strategy results in the
replacement of members with below-average fitness value with members having above-average
fitness value, to take part in the mating process.
• Crossover:-The crossover operation is carried out for the generation of new chromosomes which
will be able to retain the good features from the previous generation. Selected pairs of parents
having a probability equal to a given crossover rate are considered for this operation. In this
operation genetic materials (i.e bit-values) between two parent strings are swapped so that gen-
erated offsprings represented by highly fit strings can have a greater probability to be selected
in the subsequent generation. A single point crossover is the most basic operator. In a single
point crossover along a randomly selected bit position the genetic materials of the two parent
chromosome strings cross over. Similarly for a two-point crossover, the genetic materials be-
tween two randomly selected crossover points in two chromosome strings are interchanged to
generate two children. The effect of crossover is analogous to mating in the evolutionary process
in which the parents pass segments of their own chromosomes to their children. These children
can outperform their parents if they inherit good genes from them. This operation is shown in
Fig. 2.1.
• Mutation:- A certain pool of population may not be able to solve a particular problem if it does
not contain all the encoded information. To rectify this problem the mutation operator is used so
that spontaneously generated new chromosomes can be added to the existing pool. This operation
alters a few more selected bit values in randomly selected strings, with a probability equal to a
Crossover
point
100 11110 100 10010
101 10010 101 11110
(Single−point crossover)
1 0011 110 1 0110 110
1 0110 010 1 0011 010
(Two−point crossover)
Figure 2.1: Crossover operation
very low mutation rate, after the crossover operation. The mutation rate is kept low so that good
chromosomes obtained from crossover operation are not lost. The mutation operator can prevent
any single bit from converging to a value throughout the entire population. In addition to this the
population is prevented from converging or getting stagnant in any local optima. This operation
enhances the ability of the GA to find a solution which is near optimal by searching the entire
solution space for the best solution. This operation is shown in Fig. 2.2.
Mutated bit
10011110 10011 0 10
Figure 2.2: Mutation operation
The pivotal contribution of genetic algorithm is a methodology for systematized random search and
optimization. Researchers attempt to simulate complex biological evolutionary processes to discover
how evolution can propel living beings towards a higher level of intelligence which has resulted in the
concept of the genetic algorithm. The flow chart for this algorithm which is based on the evolutionary
principle of natural selection has been shown in Fig. 2.3.
Kumon et al. [69] proposed the use of genetic algorithm for nonlinear system identification. Simi-
larly Akramizadeh et al. [70] proposed the use of genetic algorithm for the identification of a nonlinear
2.4. NEURO FUZZY APPROACH 15
Begin
Generate initial population
Calculate fitness
Is Yes
Stopping criteria Stop
met?
No
Pass the best individual

to the next generation
Genetic operators
Crossover
Mutation
Figure 2.3: GA algorithm flowchart
Hammerstein model. Juang [71] presented an algorithm which was based on recurrent neural network
with genetic algorithm for nonlinear system identification. Kim and Lee [72] proposed the application
of multi-objective fuzzy optimization for the parameter identification of a nonlinear system. Vazquez
et. al. [73] used genetic programming for the structure identification of nonlinear dynamic systems.
2.4 Neuro Fuzzy Approach

Fuzzy logic, neural network and genetic algorithm are complementary rather than competitive for sys-
tem identification. Therefore it is advantageous to use these techniques in combination amongst them-
selves rather than exclusively. This gives rise to what is called the hybrid intelligent systems. One of
the popular combinations that has been used extensively is the neuro fuzzy hybrid system. The essen-
tial part of neuro-fuzzy modelling comes from a common framework called adaptive network which
unifies the neural network and the fuzzy model. In this resultant hybrid intelligent system, the neural
network has the ability to recognize patterns and adapt themselves to cope with changing environment.
On the other hand the fuzzy inference system incorporates human knowledge and performs inferencing
and decision making. The modeling by neuro-fuzzy method is concerned with model extraction from
numerical data which represents the dynamic behavior of the system. System modelling based on this
methodology can serve two purposes:
• The behavior of the system under consideration can be predicted from the derived model
• The derived model can be used for the design of a controller
The important steps of the neuro-fuzzy modelling approach are:
• Fuzzification of the input physical variables
• Computation of the degree of satisfaction for the available linguistic terms
• Conjunction of the premise and the fuzzy inferred parameters
• Defuzzification of the output
All the above mentioned steps are realized in sequentially arranged layers of the neural network which
has an architecture to adjust the weights in the form of the parameters of the extracted rules. Using
the neuro-fuzzy modelling as a backbone, the characteristics of soft computing may be classified as
follows :-
• Human expertise:- For solving practical problems, soft computing can be used to utilize the
human expertise in the form of conventional knowledge representation as well as fuzzy if-then
rules.
• Biologically inspired computing models:- Inspired by the biological neurons, the artificial neural
network can be used in soft computing methodology to deal with many practical problems like
pattern recognition, nonlinear regression, perception and classification.
• Innovative approach:- Soft computing methodology applies innovative optimization techniques

imbibed from various sources. They are genetic algorithm as motivated from the evolution and se-
lection process, simulated annealing which is inspired from thermodynamics, the random search
method and the simplex downhill method.
• Numerical computation:- Unlike symbolic artificial intelligence, soft computing mainly relies on
numerical computation.
• New application domains:- Because of being software based, the soft computing method is in-
creasingly finding applications in new domains like adaptive signal processing, adaptive control,
nonlinear system identification etc.
• Model free learning:- In the absence of system models, the fuzzy inference system and the neural
network have the ability to construct models from the available sampled data.
• Intensive computation:- Neuro-fuzzy and soft computing methods rely heavily on the high speed
number crunching computation to find rules or regularity in data sets, rather than too much back-
ground knowledge on the problem being solved.
• Fault tolerance:- The deletion of a neuron in a neural network or a rule in a fuzzy inference
system does not stop these models to perform because of their parallel and redundant architecture.
However, the performance degrades gradually.
• Goal driven characteristics:- Soft computing and neuro-fuzzy systems are goal oriented. How-
ever, a little deviation from their assigned goal will not deter them from achieving their goal
finally because of their domain specific knowledge.
• Real world applications:- Because of the complexity and uncertainty involved with the real world
problems, the conventional approaches of problem solving require a mathematical model or a
detailed description of the problem which is being solved. But with the integrated approach
of soft computing, it usually utilizes the specific techniques within the subtasks to constitute a
satisfactory solution to the real world problems.
Jang and Sun [74] discussed the problems of neuro-fuzzy modeling and also the direction for its future
use. Lin and Cunningham III [75] presented a method where the network was viewed as a fuzzy model
which gave insight into the real system and also provided a method to simplify the neural network.
Jang [76] proposed a simple method for the selection of inputs for the neuro-fuzzy model in identifying
a nonlinear system. Similarly Chiu [77] proposed another method for the selection of inputs of the
neuro-fuzzy model built for nonlinear system identification. Denai et al. [78,79] showed the application
of the neuro-fuzzy method for the modeling of nonlinear systems. Ishibuchi et al. [80] proposed a
method where the neural network is trained by utilizing the numerical data and also human expert
knowledge that is represented by the fuzzy if and then rules. Juang and Lin [81] proposed a method
for the identification of a dynamic system with the help of a Takagi-Sugeno-Kang (TSK) type fuzzy
rule based model which also possesses the learning ability of the neural network. Sanchez et al. [82]
proposed a method for the identification of a nonlinear system using a fast and stable neuro-fuzzy
method having error minimization. Li et al. [83] presented a neuro-fuzzy method for the identification
of a nonlinear system where in the first step the structure identification task is accomplished and in
the next step the parameter identification is carried out. Kawaji and Chen [84] also proposed a soft
computing based approach for the identification of a nonlinear system. Wu et al. [85] proposed an
approach of generating the fuzzy rules using a generalized dynamic fuzzy neural network which is
built on the ellipsoidal basis function. Gao and Joo [86] proposed a robust adaptive fuzzy neural model
for the identification of a certain class of multi input-multi output (MIMO) systems. This model has a
fast online learning capability where the fuzzy rules are generated or deleted automatically. Panella and
Galeo [87] proposed a clustering approach which is applied to a joint input-output space for the neuro-
fuzzy modelling of nonlinear systems. Thangavel et al. [88] proposed the use of an intelligent controller
for reactive power control. Abraham and Nath [89] used the neuro-fuzzy approach for modeling the
electricity demand in Victoria. Kamia et al. [90] discussed the use of soft computing technique for the
modelling of large scale plants.
2.4.1 Adaptive Network based Fuzzy Inference System
A neuro-fuzzy technique called Adaptive network based fuzzy inference system (ANFIS) [68, 74,
91] has been used as a prime tool in the present work. Adaptive network based fuzzy inference system
(ANFIS) is a neuro fuzzy technique where the fusion is made between the neural network and the fuzzy
inference system. In ANFIS the parameters can be estimated in such a way that both the Sugeno and
Tsukamoto fuzzy models [92] are represented by the ANFIS architecture. Again with minor constraints
the ANFIS model resembles the Radial basis function network (RBFN) functionally [93]. This ANFIS
methodology comprises of a hybrid system of fuzzy logic and neural network technique. The fuzzy
logic takes into account the imprecision and uncertainty of the system that is being modeled while the
neural network gives it a sense of adaptability. Using this hybrid method, at first an initial fuzzy model
along with its input variables are derived with the help of the rules extracted from the input output data
of the system that is being modeled. Next the neural network is used to fine tune the rules of the initial
fuzzy model to produce the final ANFIS model of the system. In this proposed work ANFIS is used as
the backbone for the identification of real world systems. Detailed description of ANFIS is presented
in Chapter 4.
2.5. FORECASTING METHODOLOGIES 19
After building a model for the system based on ANFIS, the model can be used for forecasting future
values using a suitable forecasting technique. A brief description of various forecasting methodologies
available is presented below.
2.5 Forecasting Methodologies

Some of the data based time series forecasting methodologies that were carried out in the recent past are
described in this section. Hard computing based methods like linear and nonlinear auto-regression have
been traditionally used for system forecasting. But due to the highly nonlinear relationship between
the determining factors and the demand forecast, it made the understanding of the system difficult.
This made the auto regression approach extremely tedious and doubts also rose on the accuracy of the
results.
The self adaptive data driven neural networks [94] scored a major point in the sense that it can approx-
imate any arbitrary continuous function to any given accuracy [95] even with very little knowledge on
the structural relationship between the demand forecast and the determining factors.
Fuzzy logic can be used to approximate any continuous function to a required degree of accuracy [96].
Both the neural network and the fuzzy systems have been found to have outperformed the auto-
regression approach with higher forecasting accuracy [96]. It is because of this fact that these ap-
proaches have been well accepted in many practical utilities, like load forecasting [97, 98].
Studies also show that attempts were made to combine multiple neural networks with recursive least
squares (RLS) algorithm to utilize the advantages of both the methods [97,99]. In some utilities the dif-
ferent combination of fuzzy logic, neural networks and hard computing based techniques like statistical
methods, fourier transform [100], bayesian curve fitting [101], Box-Jenkins Autoregressive integrated
moving average with exogenous input(ARIMAX) method [102], extrapolation technique [103], wavelet
transform [104], B-spline [105], state space [106] approach were also tried to checked if any implicit
information that may be embedded in the available data can be extracted for use in the forecasting
model. In certain cases it has been found that these approaches make the forecasting accuracy improve
substantially. Applications were also found where the past knowledge of experienced forecasters were
utilized to improve the forecasting accuracy [107].
In many cases it has been found that the presence of spurious data has a considerable effect on the
forecasting accuracy [108]. Significant research work continued in the direction of identification, de-
tection and estimation of bad data [109–111]. In order to overcome the drawbacks encountered by
conventional learning methods like statistics based ones, the use of the evolutionary methods were also
explored [105].
Traditionally, hard computing based methods like linear and nonlinear auto-regression methods have
been applied to build forecasting models [112–114]. However, as the relationships between these de-
mand forecasts and the determining factors are highly nonlinear, they are not properly understood. This
made the auto regression approach extremely tedious and doubts also surfaced on the accuracy of the
results.
On the other hand, even with a very little knowledge on the structural relationship between the demand
forecast and the determining factors, the self adaptive data driven neural networks can approximate any
arbitrary continuous function to any given accuracy [94]. Zhang et al. [115] studied the application of
neural networks for forecasting in the electricity load consumption study.
Similarly the fuzzy logic can also be used to approximate any continuous function to a required de-
gree of accuracy [95]. Performancewise both the neural network and the fuzzy systems can be put at
the same level [96, 116, 117]. Both these methods have been found to have outperformed the auto-
regression approach with higher forecasting accuracy [96] and hence have been well accepted in many
practices by utilities for load forecasting [97, 98].
In addition to the soft computing methods, different hard computing methods were also found to have
been applied for generating the different time series forecasting models.
Verleysen et al. [118] proposed the method of fractal projection forecasting for engineering applica-
tions. Amjady [107] used an approach in which he utilized the past knowledge of experienced fore-
casters to improve the forecasting accuracy, where the hard computing based Box-Jenkins Autoregres-
sive integrated moving average with exogenous input(ARIMAX) method outperformed the stand alone
neural network approach. Adya et al. [103] proposed the rule-based forecasting approach which is an
expert system that uses features of time series alongwith weight extrapolation techniques. Gao and
Tsoukalas [104] proposed the wavelet based forecasting method and used it for load identification and
forecasting. Wang et al. [119] proposed a nonparametric smoothing technique to build a kernel pro-
jection forecasting model for a given small seasonal time series data. Meade [101] used the logistic
Bayesian curve fitting model to study and investigate the effect of the assumed error structure on the
forecasting accuracy. Saito and Abe [106] used the Kalman filter to derive a state space model for
traffic forecasting. Wang et al. [119] proposed a nonparametric smoothing technique to build a kernel
projection forecasting model for a given small seasonal time series data. Meade [101] used the logistic
Bayesian curve fitting model to study and investigate the effect of the assumed error structure on the
forecasting accuracy.
Work was also carried out combining different existing methodologies for generating the forecasting
models. In many utilities the forecasting approaches are based on multiple neural networks being
combined with Recursive least squares (RLS) algorithm [97, 99]. This approach helps in extracting
the various load patterns which are implicitly embedded in the training data. Khotanzad et al. [97]
explored such an approach. This multiple forecaster approach is also supported by many hard comput-
ing based forecasting applications and theories. This is because this approach makes the forecasting
accuracy improve substantially [120]. Kim et al. [121] proposed the use of linear combination of a
fuzzy system with a neural network for improving the forecasting accuracy. Abraham and Nath [122]
discussed the use of Evolving fuzzy neural network (EFuNN) and the feed forward Artificial neural
network (ANN) as well as a conventional statistical approach for modeling electricity demand. Mora-
bito and Versaci [123] proposed the use of fuzzy neural networks for identification and forecasting the
experimental urban air pollution data. Makiko and Yoshitsugu [100] proposed a new demand forecast-
ing method using the neural network and Fourier transform.
In time series forecasting, spurious or Bad data (BD) affect the forecasting performance drastically. So,
identification of bad data is a challenging task here. Milli et al. [108] gave a comparative assessment
of the evaluating techniques for identifying Bad data (BD), i.e. data gathered from grossly erroneous
measurements. These techniques are first classified, then explored and compared. Abur [110] proposed
another algorithm for detecting the bad data using the measurement residuals of the measurements
rejected by the Linear programming (LP) estimator. Then the bad measurements are identified and
eliminated by estimating the measurement errors of the zero residual measurements. The residuals ob-
tained from this second estimation step are made use of for this purpose. Long and Ling [111] proposed
a new method of estimating the impulse response of a complex system based on its complex input and
only the real part of its output data values. Zhang and Lo [109] proposed a recursive measurement error
estimation and identification algorithm for identifying multiple interacting bad data in the power set in
power system static state estimation. As an extension for further analysis of bad data, Zhang et al. [124]
proposed an efficient bad data identification method for a state estimator which was implemented in
real-time for a power system control centre in Northeast China. The proposed algorithm, called the Re-
cursive measurement error estimation identification (RMEEI) is powerful and efficient having a high
computational speed. Kandemir and Ramanujam [125] presented an abstraction, called data relation
vectors, to improve the data access characteristics and memory layouts in regular computations, by
defining a relation between the data elements accessed by close-by iterations and using this relation to
2.6. RESEARCH GAP 22
guide a number of optimizations for array-based computations. Ahmed and Cruz [126] proposed the
modified version of this algorithm and proved that it yields a least squares solution, which is compa-
rable to the Recursive least squares (RLS) method and initialized it for a Data-driven echo canceller
(DDEC). Erdogmus and Principe [127] investigated error-entropy-minimization in adaptive systems
training, by minimization of a Csiszar distance measure between the densities of desired and system
outputs. Didenko and Movchan [128] investigated the evaluation of the uncertainty of A/D conver-
sion with reasonable accuracy by considering the minimum number of parameters of the metrological
model of the Data acquisition system (DAS). Joachims [129] presented an approach to automatically
optimize the retrieval quality of search engines using clickthrough data. Babcock et al. [130] discussed
one adaptive strategy for processing bursty streams of data which fluctuate over time without sacrific-
ing the system performance as in many applications, systems are required to produce rapid or realtime
query responses. Weekes and Forgel [131] proposed using of evolutionary training of Artificial neural
networks (ANNs) to generate predictive models of quantitative structure, based on the data of activity
relationships between a set of molecular descriptors and activity. In order to overcome the drawbacks
encountered by conventional learning methods, Coelho and Krohling [105] proposed the use of an in-
telligence methodology called swarm optimization methodology to provide a stochastic global search
of B-spline networks for nonlinear system identification. Meekhof and Heckendorn [132] explored the
possibility of using string space transformations to reduce the perplexity of the modeling problem and
thereby improve model performance of a Markov-based classifier on the problem of classifying Eng-
lish and Spanish character strings, where training set size is arbitrarily limited. Pekar and Stecha [133]
proposed a real time system parameter estimation from the set of input-output data by minimization
of quadratic norm errors of system equations. Lii et al. [134] proposed translation model training in
Statistical machine translation (SMT) using off line data optimization and online model optimization.
2.6 Research gap

In all the data based time series forecasting methods that have been discussed in the previous section, the
availability of a sufficient number of data is a must for achieving a good forecasting model. However,
in many cases it has been observed that the number of available data is not sufficient for modeling
purpose. In such a scenario one alternative solution lies in the generation of more data so that modeling
of the system under consideration can be carried out with good accuracy. But generation of more data
is not always possible as it may be a very costly affair, particularly in a production environment, like
2.6. RESEARCH GAP 23
a thermal power plant. This is because in order to get the training data, the power plant has to be kept
running for long period of time. In addition, the thermal power plants have the inherent problem of
taking a long time to start before it can be fully loaded to get the requisite data. So the training of the
ANFIS has to be carried out with the available limited number of input output data with good modeling
accuracy. This necessitates the development of a proper procedure to choose the data critically and
optimally for training the neural network. With these objectives in mind it has been proposed to utilize
the salient feature of self-adaptability in approximating an arbitrary function with a good accuracy of
the neural network and the inherent property of the fuzzy logic to take into account the imprecision
and uncertainty of a system which is considered for modeling, with the help of an ANFIS model. A
novel method for optimal choice of the available data for training the ANFIS network is proposed
in the present work. The major objective of the present work is to formulate an efficient method to
choose the available data optimally from the existing data set to build an ANFIS model for the system
under consideration and use the model for faithful prediction of future values. The research focuses
on modeling of real world systems where input output data required for modeling are either scanty or
difficult to generate. The research work then aims to design a controller based on ANFIS so that the
controller performs satisfactorily even with plant parameter variations and disturbance.
The methodology followed for achieving the above goal is described in the next chapter.
Chapter 3
Research Objectives and Methodology
In the quest for developing a model for a system based on its available input output data, it has been
observed that in the conventional modeling approach the results depend on the mathematical model of
the system and its accuracy. In cases where the mathematical model is not available the system analysis
becomes very difficult. It is in this context that the soft computing approach can provide a viable
alternative for system modeling in cases where the conventional hard computing techniques cannot be
used. In the present work it is proposed to use these soft computing techniques for time series modeling
of a real world system with the help of its input-output data.
3.1 Objectives of the present work

In data driven modeling techniques like ANFIS, the number of time series data that are available
for system modeling has to be in abundance, as this number has a direct bearing on the modeling
result. It has been found that more is the number of available input-output data the better will be the
modeling result. This is because with more number of training data the neural network of the hybrid
ANFIS system will have a better training session which will result in a good modeling validation and
prediction. But in most of the real time systems the available data is generally less, so modeling has
to be carried out with the available small pool of data only. The primary concern of the thesis is for
this type of situations where sufficiently large quantity of input-output data is not available or difficult
to generate and the model of the system has to be built for practical applications like forecasting. This
research focuses on these real life systems having scanty input-output data and attempts to identify the
model using the data selected critically such that these optimally selected data can be used to build
the system model faithfully. Hence optimal selection of the available minimal data set for training the
24
3.2. METHODOLOGY OF THE PRESENT WORK 25
ANFIS model is the key objective of this research. The key objectives of the present research can be
highlighted as follows:
• ANFIS based modeling of real world systems
• Building of ANFIS model for real world systems where generation of training data is difficult
and time consuming
• Building of ANFIS model for real world system where generation of data is expensive
• Building of ANFIS model for real world system where available input output data are scanty
• Designing of an ANFIS based controller which performs satisfactorily even in the presence of
plant parameter variations and disturbances
The building of an ANFIS model faithfully with available scarce input output data is studied in this
research work and a novel solution is offered so that this research gap can be bridged.
3.2 Methodology of the present work

ANFIS is the backbone for the present work where a small set of time series data critically selected
from a real time system are used for the modeling of the system. The modeling methodology used in
the present work is briefly mentioned in the following subsections.
3.2.1 ANFIS based methodology for modeling
The Adaptive network based fuzzy inference system (ANFIS) [68,74,91] is a hybrid system comprising
of the neural network and the fuzzy logic. It is a data driven procedure which can be used to provide
the solution of function approximation problems in a neural network platform. Here at first a fuzzy
inference system comprising of an initial fuzzy model is formed, based on the fuzzy rules extracted
from the input output data set. In the next step the neural network is used to fine tune the rules of the
initial fuzzy model that was built. Using ANFIS methodology the network is trained. The number of
training data used in the ANFIS is drastically reduced by applying an optimal data selection criterion.
Two novel techniques are attempted for selecting the optimal input-output data pairs. One of these
methods is known as the Full factorial design (FFD) [135, 136] and is based on statistical design
of experiments. The other technique is a cross validation technique known as the V-fold technique
[137–139]. These two techniques are described briefly in the next three subsections. This work is
an endeavour to augment the ANFIS as a modeling technique by incorporating the above two novel
techniques for accomplishing the training of the network faithfully when the available data is scarce.
3.2.2 Full factorial design based methodology for optimal data selection
From the statistical point of view, in experiments where a large number of tests is involved, the order
of selection of the test specimens has to be randomized, so that each specimen has an equal chance
of being selected for the test. The training of the neural network can also be replicated to a statistical
experiment involving a large number of tests for optimization. In such an experimental setup the iden-
tification of the important variables that affects the experimental results forms an important aspect. The
experimental variables controlled by the operators on which the system’s response depends are called
the factors. The effect of the response of one factor may or may not depend on the levels of the other
factors. The number of factors and their levels required can be fixed depending upon the complexity
of the experiment. The statistical technique based Full factorial design (FFD) [135, 136] methodology
identifies the important factors and levels of the experiments conducted to model a real time system.
This FFD methodology is used for selecting the critical data set for training the ANFIS model of the
real system.
3.2.3 V-fold based methodology for optimal data selection
If a neural network is built using a specific learning data set, it has to be tested with a data set which
is independent of the data set that was used to train the network. It has often been found that obtain-
ing an independent test data set is very difficult. The neural network can also be trained by another
accurate method called the V-fold technique [137–139]. This method possesses the distinct advantage
of not requiring a separate and independent data set for testing the accuracy of the network. In this
methodology the available data set is subdivided into some subgroups. All but one of the subgroups is
used for training the neural network. Next the remaining subgroup is used for testing. This procedure
is continued by testing the network with a new subgroup every time. This approach of the V-fold cross
validation technique is used in our research work to select the optimal data set for training the ANFIS
network in the model of the real time system considered in our study.
3.2.4 Full factorial design combined with V-fold based methodology for optimal
data selection
Both the Full factorial design (FFD) and the V-fold technique are combined to select the critical data
set for training the ANFIS model.
3.2.5 Model building with statistical approaches for performance study
The conventional statistical methods of modeling are also studied in the present research and applied
to build the system models to study their performance against the ANFIS model. The three statisti-
cal models namely Autoregressive with exogenous input (ARX) [140], Autoregressive moving average
with exogenous input (ARMAX) [140] and Autoregressive integrated moving average with exogenous
input (ARIMAX) [140] are used for modeling the real world system and the results obtained are com-
pared with those obtained by using the proposed models.
3.2.6 Genetic algorithm based fuzzy methodology
A Genetic algorithm(GA) [67,141,142] based fuzzy model is build for studying its performance against
the ANFIS model. In this methodology the genetic algorithm is used to update the consequent parame-
ters of the fuzzy model of the system under consideration.
3.2.7 Controller with the FFD-V-fold based methodology
This research also studies the utility of the proposed ANFIS model for designing a controller. A con-
troller based on the proposed FFD-V-fold based ANFIS method is built and its performance is tested
against disturbance and plant parameter variation. Out of the different ANFIS based models which are
proposed, the model showing the best result is used for the design of a controller.
In the next chapter, the ANFIS based methodology for modeling is discussed in details.
Chapter 4

System
The adaptive network based fuzzy inference system (ANFIS) [68, 74, 91] is a data driven procedure
representing a neural network approach for the solution of function approximation problems. Data
driven procedures for the synthesis of ANFIS networks are typically based on clustering a training set of
numerical samples of the unknown function to be approximated. Since introduction, ANFIS networks
have been successfully applied to classification tasks, rule-based process control, pattern recognition
and similar problems. Here a fuzzy inference system comprises of the fuzzy model [34, 35] proposed
by Takagi, Sugeno and Kang to formalize a systematic approach to generate fuzzy rules from an input
output data set.
4.1 ANFIS structure

For simplicity, it is assumed that the fuzzy inference system under consideration has two inputs and one
output. The rule base contains the fuzzy if-then rules of Takagi and Sugeno’s type [143] as follows:
If x is A and y is B then z is f(x,y)
where A and B are the fuzzy sets in the antecedents and z = f (x, y) is a crisp function in the
consequent. Usually f (x, y) is a polynomial for the input variables x and y. But it can also be any other
function that can approximately describe the output of the system within the fuzzy region as specified
by the antecedent. When f(x,y) is a constant, a zero order Sugeno fuzzy model is formed which may be
considered to be a special case of Mamdani fuzzy inference system [144] where each rule consequent
is specified by a fuzzy singleton. If f(x,y) is taken to be a first order polynomial a first order Sugeno
28
4.1. ANFIS STRUCTURE 29
fuzzy model is formed. For a first order two rule Sugeno fuzzy inference system, the two rules may be
stated as:
Rule 1: If x is A1 and y is B1 then f1 = p1 x + q1 y + r1
Rule 2: If x is A2 and y is B2 then f2 = p2 x + q2 y + r2
Here type-3 fuzzy inference system proposed by Takagi and Sugeno [143] is used. In this inference
system the output of each rule is a linear combination of the input variables added by a constant term.
The final output is the weighted average of each rule’s output. The corresponding equivalent ANFIS
structure is shown in Fig. 4.1.
A x y
1
x
w w
1 1
A
2 π N w f
1 1
Σ f
B1
w f
y π N 2 2
w
2 w
2
B2
x y
Layer 1 Layer 2 Layer 3 Layer 4 Layer 5
Figure 4.1: Type-3 ANFIS Structure
The individual layers of this ANFIS structure are described below :
Layer 1: Every node i in this layer is adaptive with a node function
Oi1 = µAi (x) (4.1.1)
where, x is the input to node i, Ai is the linguistic variable associated with this node function and
µAi is the membership function of Ai . Usually µAi (x) is chosen as
1
µAi (x) = (4.1.2)
1+ [( x−c
ai
i 2 bi
)]
4.2. LEARNING ALGORITHM 30
Or
x − ci 2
µAi (x) = exp {−( )} (4.1.3)
ai
where x is the input and {ai , bi , ci } is the premise parameter set.
Layer 2: Each node in this layer is a fixed node which calculates the firing strength wi of a rule.
The output of each node is the product of all the incoming signals to it and is given by,
Oi2 = wi = µAi (x) × µBi (y), i = 1, 2 (4.1.4)
Layer 3: Every node in this layer is a fixed node. Each ith node calculates the ratio of the ith
rule’s firing strength to the sum of firing strengths of all the rules. The output from the ith node is the
normalized firing strength given by,
wi
Oi3 = wi = , i = 1, 2 (4.1.5)
w1 + w2
Layer 4: Every node in this layer is an adaptive node with a node function given by
Oi4 = wi fi = wi (pi x + qi y + ri ), i = 1, 2 (4.1.6)
where wi is the output of Layer 3 and {pi , qi , ri } is the consequent parameter set.
Layer 5: This layer comprises of only one fixed node that calculates the overall output as the
summation of all incoming signals, i.e.
X P
wi fi
Oi5 = overall output = wi fi = Pi (4.1.7)
i i wi
4.2 Learning Algorithm

In the ANFIS structure, it is observed that given the values of premise parameters, the final output can
be expressed as a linear combination of the consequent parameters. The output f in Fig. 4.1 can be
4.3. DERIVATION OF THE INITIAL FUZZY MODEL 31
written as
w1 w2
f = f1 + f2
w1 + w2 w1 + w2
= w1 f1 + w2 f2
= (w1 x)p1 + (w1 y)q1 + (w1 )r1 + (w2 x)p2 + (w2 y)q2 + (w2 )r2 . (4.2.1)
where f is linear in the consequent parameters (p1 , q1 , r1 , p2 , q2 , r2 ).
In the forward pass of the learning algorithm, consequent parameters are identified by the least squares
estimate. In the backward pass, the error signals, which are the derivatives of the squared error with re-
spect to each node output, propagate backward from the output layer to the input layer. In this backward
pass, the premise parameters are updated by the gradient descent algorithm [27–29].
4.3 Derivation of the Initial Fuzzy Model

As described earlier, in ANFIS based system modeling for a set of rules with fixed premise para-
meters, identification of an optimal fuzzy model with respect to the training data reduces to a linear
least-squares estimation problem. A fast and robust method for identification of fuzzy models from
input-output data was proposed by S.L.Chiu [77]. This method selects the important input variables
when building a fuzzy model from data by combining cluster estimation method with a least squares
estimation algorithm. The method follows in two steps : i) First step involves extraction of an initial
fuzzy model from input output data by using a cluster estimation method incorporating all possible in-
put variables. ii) In the next step the important input variables are identified by testing the significance
of each variable in the initial fuzzy model.
4.3.1 Extracting the initial fuzzy model
In order to start the modeling process, an initial fuzzy model has to be derived. This model is required
to find the number of inputs, number of linguistic variables and hence the number of rules in the final
fuzzy model. The initial model is also required to select the input variables for the final model and also
the model selection criteria, before the final optimal model can be derived. This initial fuzzy model can
be selected based on the fuzzy rules framed by either using the subtractive clustering technique [145]
or the grid partitioning method [74] [68, 91].
Subtractive Clustering Technique
As a first step towards extracting the initial fuzzy model by subtractive clustering, this technique is
applied to the input output data pairs, which are obtained from the system which is to be modeled. The
cluster estimation technique helps in locating the cluster centers of the input output data pairs. This in
turn helps in the determination of the rules which are scattered in input output space, as each cluster
center is an indication of the presence of a rule. In addition to this it also helps to determine the values
of the premise parameters. This is important because an initial value, which is very close to the final
value, will eventually result in the quick convergence of the model towards its final value during the
training session with neural network. In this clustering technique the potentials of all the input output
data points are calculated as functions of their Euclidian distances from all the other data points. The
points having a potential above a certain preset value are considered as cluster centers. After the cluster
centers are ascertained the initial fuzzy model can be subsequently extracted as the centers will also
give an indication of the number of linguistic variables. The cluster estimation method for determining
the number of rules and initial rule parameters [77] is briefly described below.
Let us consider a collection of n data points {x1 , x2 , ..., xn } in an M dimensional space. The data
points are assumed to be normalized in each dimension so that they are bounded by a unit hypercube.
Each data point is considered to be a potential cluster center. Pi is a measure of the potential of data
point xi to serve as a cluster center and is defined as
X
n
2
Pi = e−αkxi −xj k (4.3.1)
j=1
where
4
α= . (4.3.2)
ra2
k.k denotes the Euclidean distance and ra is a positive constant. Thus measure of the potential for a
data point is a function of its distances to all other data points. Here ra is the radius effectively defining
a neighborhood. Data points outside this radius have little influence on the potential. After the potential
of every data point has been computed, the data point with the highest potential is selected as the first
cluster center. Let x∗1 be the location of the first cluster center and P1∗ be its potential value. Then the
potential of each data point xi is revised by the formula
∗ 2
Pi = Pi − P1∗ e−βkxi −x1 k (4.3.3)
where
4
β= (4.3.4)
rb2
and rb is a positive constant. Thus an amount of potential is subtracted from each data point as a function
of its distance from the cluster center. The constant rb is effectively the radius defining the neighborhood
which will have measurable reduction in potential. Typically chosen value of rb = 1.25ra [77].
When the potentials of all data points have been revised according to Eq. 4.3.3, the data point with the
highest remaining potential is selected as the second cluster center. The potential of each data point is
further reduced according to their distance to the second cluster center. In general, after the k th cluster
center has been obtained, the potential of each data point is revised by the formula
∗ 2
Pi = Pi − Pk∗ e−βkxi −xk k (4.3.5)
where x∗k is the location of the k th cluster center and Pk∗ is the potential value.
The process of acquiring new cluster center and revising potentials repeats until the stopping criterion
Pk∗ < 0.15P1∗ [145] is satisfied. Each cluster center as derived above is in essence a data point that
describes a characteristic input-output behaviour of the system we wish to model. Hence each cluster
center can be used as the basis of a rule that describes the system behaviour.
It is considered that {x∗1 , x∗2 , ..., x∗c } is a set of c cluster centers in an M dimensional space. It is again
considered that the first N dimensions correspond to input variables and the last M − N dimensions
correspond to output variables. Each vector x∗i is decomposed into two component vectors yi∗ and zi∗
where yi∗ is the location of the cluster center in input space and zi∗ is the location of the cluster center
in output space. Therefore x∗i may be represented as
x∗i = [yi∗ ; zi∗ ]
Each cluster center x∗i is considered as a fuzzy rule, “if input is near yi∗ then output is near zi∗ ”, to
describe the system behaviour. Given an input vector y, the degree to which rule i is fulfilled is defined
as
∗ 2
µi = e−αky−yi k (4.3.6)
where α is a constant defined by Eq. 4.3.2. Output vector z is computed as
Pc
µi zi∗
z = Pi=1
c (4.3.7)
i=1 µi
This computational scheme can be viewed in terms of an inference system employing fuzzy if-then
rules. Each rule has the following form :
IF Y1 is Ai1 and Y2 is Ai2 and ... THEN Z1 is Bi1 and Z2 is Bi2 ...
where Yj is the jth input variable and Zj is the jth output variable; Aij is an exponential member-
ship function in the ith rule with the jth input and Bij is a singleton in the ith rule associated with the
jth output. For the ith rule that is represented by cluster center x∗i , Aij and Bij are given by :
∗
Yj −yij
−0.5( )2
Aij (Yj ) = e σij
(4.3.8)
Bij = zij∗ (4.3.9)
1
where yij∗ is the jth element of yi∗ and zij∗ is the jth element of zi∗ and σij2 = 2α
[77].
Grid Partitioning Technique
The second method which can be used for framing the rules of the initial fuzzy model is by grid parti-
tioning [74] [68, 91]. This method is used when the number of inputs and their membership functions
are less. Here the input space are partitioned into a number of fuzzy regions to form the antecedents of
the fuzzy rules. The Grid partitioned fuzzy space for a two input model, with each input having three
membership functions each is shown in Fig. 4.2. The two dimensions represent the abscissa and the
ordinate of the input space. The rules obtained from either of the two methods are then optimized by
using ANFIS methodology developed by Jang [91]. This method involves optimization of the premise
membership functions by gradient descent algorithm combined with optimization of the consequent
equations by linear least squares estimation.
Y Y
3 6 9
B3
B2 2 5 8
B1 1 7
4
X
1
A1 A2 A3
1
Figure 4.2: Grid partitioned fuzzy subspaces for a type-3, 2-input ANFIS
4.3.2 Selection of input variables and final fuzzy model
Any good criteria which is used for model selection based on some inputs cannot guarantee that the
model will be producing the optimal result unless the results from all possible combinations of the input
variables in the model are compared. But this is a tedious task as even for a moderate number of input
variables N , 2N possible number of combinations of the variables arises. From the modeling viewpoint,
incorporation of only the important variables results in a practical model which is simpler, more reliable
and useful for application. This is because now only a fewer variables will have to be measured. For
achieving this, the importance of each input variable can be ascertained from the initial fuzzy model.
In the proposed ANFIS model this has been accomplished by choosing the model showing the least
modeling error from the models obtained using the following two methods:
• First method:-In this method the fuzzy rule framework provides an easy mechanism to test the
importance of each input variable without having to generate new models. The basic idea is to
remove all antecedent clauses associated with a particular input variable from the rules and then
evaluate the performance of the model by applying the checking error criterion [77]. If this de-
creases the modeling error the process is repeated by eliminating another input variable. If the
modeling error increases, the eliminated variable is retained and another variable is eliminated in
its place. This process is continued until the modeling error can be decreased no further by elim-
inating any more extra input variable. The criterion that is used for selection of the final fuzzy
model is the Root mean square error (RMSE). The ANFIS structure of the system which is being
modeled is considered as the final model for which the RMSE is the minimum. For example,
suppose the initial model has four inputs, with rules of the form :
IF Y1 is Ai1 and Y2 is Ai2 and Y3 is Ai3 and Y4 is Ai4 THEN Z1 is Bi1
The importance of the Y3 variable in the model can be tested by temporarily removing the an-
tecedent clauses that involve Y3 , thus truncating the rules to the form :
IF Y1 is Ai1 and Y2 is Ai2 and Y4 is Ai4 THEN Z1 is Bi1
If the resultant model performance does not degrade with respect to the performance measure
which is the RMSE of the output corresponding to an independent set of checking data, then Y3
can be eliminated from the possible important variables. The variable selection process for a four
input initial model is shown in Fig. 4.3.
Model with input variables
1 2 3 4
Remove 3
2 3 4 1 3 4 1 2 4 1 2 3
Remove 1
2 4 1 4 1 2
Remove 2
2 4
Remove 4
Model with no input variables
Figure 4.3: Variable Selection Process for a four input initial model
4.4. EXPERIMENTAL DATA 37
• Second method:- For simplicity, the second model assumes that ANFIS consists of only two
inputs. For choosing these inputs a set of 10 dynamic modeling inputs [77] is selected which
comprises of 6 from the historical inputs of the system and 4 from the historical outputs of the
system to be modeled. Out of these 10 dynamic modeling candidates, for each of the 6 inputs
there can be 4 different combinations with the 4 different outputs. So there will be a total of
6 × 4 = 24 different combinations of input candidate pairs. Each of these different combinations
will lead to the generation of 24 different initial fuzzy models as follows:
{x(t − 1), y(t − 1)}, {x(t − 1), y(t − 2)}, {x(t − 1), y(t − 3)}, {x(t − 1), y(t − 4)},
{x(t − 6), y(t − 1)}, {x(t − 6), y(t − 2)}, {x(t − 6), y(t − 3)}, {x(t − 6), y(t − 4)};
From the above 24 models, the model with the least RMSE is chosen as the initial fuzzy model.
In this method the number of membership functions per input can be determined by applying
either the subtractive clustering or the grid partition based technique.
4.3.3 Optimization of the initial fuzzy model
The ANFIS structure of the system which is being modeled is considered as the final model for which
the RMSE is the minimum. The consequent parameters of the initial fuzzy model are updated by us-
ing the Least squares estimation (LSE) algorithm. Similarly, the rules which are obtained from the
clustering or the grid partition based method are updated by neural network which uses back propaga-
tion learning method with gradient descent algorithm. This updation leads to the optimization of the
premise parameters of the fuzzy membership functions to give the final fuzzy model.
4.4 Experimental Data

The data that has been used for the experimental purpose are taken from two different sets.
In the first set the data are taken from the bench mark problem of Box and Jenkins [102]. It comprises
4.5. EXPERIMENTAL RESULTS AND DISCUSSION 38
of the input-output data that are collected from a gas furnace. In the dataset the input is the gas feed
rate to the furnace in cubic feet per minute (ft3 /min)and the output is in the form of percentage of
carbon-dioxide (CO2 ) concentration in the outlet gas.
The second data set comprises of five different data sets that are obtained by physically visiting a ther-
mal power plant under North East Electric Power Corporation (NEEPCO) Limited situated at Kathal-
guri, Assam, India. This real time data have been collected from the past records maintained by the
power plant authorities in hard form. In the first four data-sets the input is the gas flow rate in cubic
kilometer per hour (km3 /hr) and the output is in the form of generated power in Gigawatt (GW). These
are the numerical records of the daily data that the power plant authority records after every half hour.
These were subdivided into four different sets with the half hourly data covering a few months at dif-
ferent periods of the year.
The power plant authority also keep a record of data on a monthly basis by aggregating the half hourly
data over the whole month. The fifth subset of the thermal power plant data has been collected from
the records of this monthly data. In this subset the input is the gas feed to the plant in Million metric
standard cubic metre (Mmscum)(106 m3 ) and the output is in the form of generated energy in Million
unit (MU), where 1 unit = 1 kWh (kilo-watt-hour).
These numerical data in the same form as collected from the plant’s record are analyzed for determin-
ing the different levels and folds for the Full factorial design (FFD) and the V-fold based experiments
respectively. The different levels of the Full factorial design experiments are determined by noting
the numerical values around the highest, lowest, mid-values and cluster centers of the collected data.
Next, these data are used for training and validation of the proposed model. A sample data set that was
collected from the thermal power plant is presented in the Appendix A.
4.5 Experimental Results and Discussion

The ANFIS based model identification method is applied to the well known example of Box and Jenk-
ins gas furnace data [102]. Next the modeling of a thermal power plant based on real data is considered.
This real time data set has been obtained from a thermal power plant under North East Electric Power
Corporation (NEEPCO) Limited situated at Kathalguri, Assam, India. The modeling error is calculated
as the difference between the model output and the real system output and either the Mean square error
(MSE) or the Root mean square error (RMSE) has been used as the comparative index. The details of
these experiments and the results obtained are discussed in the following subsections.
4.5.1 ANFIS Model for Gas Furnace
In this experiment 296 input-output data pairs from Box and Jenkins [102] gas furnace data are consid-
ered for modeling with ANFIS. In the dataset the input x(t) is the gas feed rate in cubic feet per minute
(ft3 /min)and the output y(t) is in the form of percentage of carbon-dioxide (CO2 ) concentration in the
outlet gas.
In order to extract the dynamic process model for the prediction of y(t), 10 candidates are considered
as input variables following standard method [77]. These 10 input variables are :
{x(t − 1), x(t − 2), x(t − 3), x(t − 4), x(t − 5), x(t − 6), y(t − 1), y(t − 2), y(t − 3), y(t − 4)}
After converting the data so that each training data considers
{x(t − 1), x(t − 2), x(t − 3), x(t − 4), x(t − 5), x(t − 6), y(t), y(t − 1), y(t − 2), y(t − 3), y(t − 4)},
the number of effective data points reduces to 290. Using a cluster radius of ra = 0.5 [77], the modeling
was started with these ten prospective input candidates in the initial fuzzy model. Consequently in the
final Takagi-Sugeno type fuzzy model which is derived by using Chiu’s [77] input variable selection
method, the number of input variables reduced to 2 and the number of rules narrowed down respec-
tively to 3 and 4 for the subtractive clustering and the grid based method. The optimally selected input
variables are x(t − 3) and y(t − 1). The final optimized ANFIS model of the gas furnace process is
obtained after the updation of the consequent parameters by LSE algorithm and updation of the premise
parameters by the back propagation gradient descent algorithm. In order to provide the same basis for
comparison with other published results, the entire dataset of 290 input output pairs is used for training.
Fig. 4.4 and Fig. 4.5 show the rules and parameters of the grid partition based and subtractive clustering
based ANFIS models for the gas furnace data. In Fig. 4.6 the actual output and the output predicted by
the ANFIS model are plotted vs. sample number. In Fig. 4.7 the prediction error of the ANFIS model
for the training data is plotted vs. sample number. Fig. 4.8 shows the RMSE of the ANFIS model
during training vs epochs. The surface graph corresponding to the output is shown in Fig. 4.9.
If x(t−3) is and y(t−1) is then y(t) is
1
y(t)=−0.3601x(t−3)
+1.142y(t−1)
Rule 1 −8.268
0 2.834
−2.716 45.6 60.5
1 y(t)=−1.354x(t−3)
+0.3158y(t−1)
Rule 2 +39.43
0 2.834
−2.716 45.6 60.5
1
y(t)=−1.925x(t−3)
+0.5233y(t−1)
Rule 3 +27.34
0 2.834
−2.716 45.6 60.5
1
y(t)= 1.469x(t−3)
Rule 4 +1.597y(t−1)
−35.48
0 2.834 60.5
−2.716 45.6
Gbell MF parameters:− [1.577, 3.02, −1.516]; [3.68, 2.07, 1.932]; [7.338, 4.732, 46.69]; [8.223, 0.9815, 60.01;]
Figure 4.4: Rules for the grid partition based ANFIS model when trained with entire data set for the
gas furnace example
1
y(t)=−2.275x(t−3)
−5.151y(t−1)
Rule 1 +305.6
0 2.834
−2.716 45.6 60.5
1 y(t)= 2.995x(t−3)
−0.2808y(t−1)
Rule 2 +104.8
0 2.834
−2.716 45.6 60.5
1
y(t)=−0.04683x(t−3)
+0.5213y(t−1)
Rule 3 +20.95
0 2.834
−2.716 45.6 60.5
Gauss MF parameters:− [1.518, −0.1355]; [1.595, −0.3007]; [2.131, 0.1226,]; [3.255, 54.24]; [3.057, 56.2]; [2.601, 49.82];
Figure 4.5: Rules for the subtractive clustering based ANFIS model when trained with entire data set
for the gas furnace example
62
actual output
predicted output
60
58
% of CO2 in the outlet gas

56
54
52
50
48
46
44
0 50 100 150 200 250 300
sample number
Figure 4.6: Actual and ANFIS model predicted output with Box and Jenkins gas furnace data
In Table 4.1 the 2-input based ANFIS model is compared with other models available in literature.
The first 7 rows are excerpted from a table in [77]. The comparison shows that the modeling result of
the gas furnace data has improved with the proposed two inputs subtractive clustering and grid based
ANFIS models as compared with the other models both in terms of the number of rules required and
the RMSE. The proposed grid based model is also showing better performance as compared to Chiu’s
two input model [77] but with the requirement of one more rule.
Another experiment was performed using the first half of the Box and Jenkins gas furnace data for
training the neural network and the second half of the data set for validation of the model. Fig. 4.10
shows the rules and parameters of this model. The modeling result is shown in Fig. 4.11 where the
actual output and the output predicted by ANFIS model are plotted vs. sample number.
4.5.2 ANFIS Model for Thermal Power Plant
In this experiment five different data sets, comprising of input-output data pairs collected from the ther-
mal power plant are considered for modeling with ANFIS. In the first four data-sets the input x(t) is the
gas flow rate in cubic kilometer per hour (km3 /hr) and the output y(t) is in the form of generated power
in Gigawatt (GW). These data sets are based on the daily data collected from the thermal power plant
at an interval of every half hour. These half hourly data have been classified into four different data sets
depending on their collection time which is spread over different periods of the year. In the fifth data
Table 4.1: Comparison of various models derived for the Box and Jenkins gas furnace data. The first 7
rows are excerpted from a table in [76]
Number Model Error
Sl. No. Model Input variables
of Rules (mean square)
1 Tong’s Model [1980] x(t − 4), y(t − 1) 19 0.4690
2 Pedrycz’s Model [1984] x(t − 4), y(t − 1) 81 0.3200

Xu’s model
3 x(t − 4), y(t − 1) 25 0.3280
[Xu and Yong 1987]
x(t − 1), x(t − 2),
Takagi-Sugeno model
4 x(t − 3), y(t − 1), 2 0.0680
[Sugeno and Tanaka 1991]
y(t − 2), y(t − 3)
Sugeno’s position-
5 gradient model x(t − 3), x(t − 4) 6 0.1900
[Sugeno and Yasukawa 1993]
Takagi-Sugeno model x(t − 3)
6 [3-input] y(t − 1) 3 0.0720
[Chiu 1996] y(t − 3)
Takagi-Sugeno model
7 [2-input] x(t − 3), y(t − 1) 3 0.1460
[Chiu 1996]
0.1322
3 (subtractive
clustering based)
Proposed model
8 x(t − 3), y(t − 1)
[2-input]
0.1277
4 (grid partition
based)
2.5
1.5
prediction error
0.5
−0.5
−1
−1.5
0 50 100 150 200 250 300
sample number
Figure 4.7: ANFIS model prediction error for Box and Jenkins gas furnace data
set the input x(t) is the gas feed to the plant in Million metric standard cubic metre (Mmscum)(106 m3 )
and the output y(t) is in the form of generated energy in Million unit (MU), where 1 unit = 1 kWh
(kilo-watt-hour). This fifth data set has been collected from the thermal power plant on the basis of the
monthly data for a certain period of time.
The dynamic process model for the above thermal power plant is extracted as in the previous case for
the gas furnace data. The 10 candidates which are considered for the input variables are the following:
{x(t − 1), x(t − 2), x(t − 3), x(t − 4), x(t − 5), x(t − 6), y(t − 1), y(t − 2), y(t − 3), y(t − 4)}
In all the five data sets, modeling of the thermal power plant was considered with these 10 prospective
input candidates in the initial fuzzy model. Both the grid partition method and the subtractive clustering
method were applied with a cluster radius of ra = 0.5. The number of input variables finally reduced to
2 and the number of rules narrowed down to 4. After the updation of the consequent parameters by LSE
algorithm and updation of the premise parameters by the back propagation gradient descent algorithm,
the final optimized ANFIS model of the thermal power plant is obtained. The rules of the different
models of ANFIS and their experimental results are shown in Figs. 4.12 – 4.21 where the predicted
output of the ANFIS models are plotted vs. sample number for the five data sets of the thermal power
plant. In addition, the fuzzy rules and the parameters of the models for the thermal power plant data
sets are also given.
0.363
0.362
0.361
RMSE
0.36
0.359
0.358
0.357
0 20 40 60 80 100 120 140 160 180 200
epochs
Figure 4.8: RMSE of the ANFIS model during training for Box and Jenkins gas furnace data
65
% of CO2 in the output gas
60
55
50
45
60
58
56 2.5
2
54 1.5
1
52 0.5
0
50 −0.5
−1
48 −1.5
−2
46 −2.5
% of CO2 in the output gas gas feed rate
Figure 4.9: Output surface of the data for Box and Jenkins gas furnace example
1
y(t)=−0.3797x(t−3)
+1.009y(t−1)
Rule 1 +0.759
0 2.834
−2.716 45.6 60.5
1 y(t)=−1.222x(t−3)
+0.7175y(t−1)
Rule 2 +14.45
0 2.834
−2.716 45.6 60.5
1
y(t)=−1.424x(t−3)
+0.6379y(t−1)
Rule 3 +21.57
0 2.834
−2.716 45.6 60.5
1
y(t)= 0.7164x(t−3)
Rule 4 +2.165y(t−1)
−67.22
0 2.834 60.5
−2.716 45.6
Gbell MF parameters:− [2.705, 1.845, −2.911]; [2.269, 1.994, 3.31]; [7.78, 2.992, 46.39]; [7.28, 1.301, 60.35];
Figure 4.10: Rules for the ANFIS model when trained with half of the data set for the gas furnace
example
Case 1: Modeling Results for the first set of the thermal power plant data with ANFIS
In this experiment the data set comprises of 1405 input-output pairs which finally reduced to 1399 ef-
fective data pairs as a result of the dynamic process model under consideration [77]. Here the optimally
selected input variables are x(t − 6) and y(t − 1). The number of rules required finally narrowed down
to 2 and 4 respectively for the clustering and the grid based method. Fig. 4.12 shows the rules and
parameters for this model. In this example the first 699 data pairs are used for training purpose with the
remaining half being used for validation. The actual output and the ANFIS model’s predicted output
for this data set is plotted against sample number as shown in Fig. 4.13.
Case 2: Modeling Results for the second set of the thermal power plant data with ANFIS
In this experiment the data set comprises of 1357 input-output pairs which finally reduced to 1351
effective data pairs [77]. The optimally selected input variables are x(t − 3) and y(t − 1). The number
of rules required finally narrowed down to 2 and 4 respectively for the clustering and the grid partition
based method. Fig. 4.14 shows the rules and parameters for this model. In this example the first 675
data pairs are used for training purpose while the other half is used for testing purpose. The modeling
result for this data-set is shown in Fig. 4.15.
62
actual output
predicted output
60
58

56
54
52
50
48
46
44
0 50 100 150 200 250 300
sample number
Figure 4.11: Actual and ANFIS model predicted output with Box and Jenkins gas furnace data
Case 3: Modeling Results for the third set of the thermal power plant data with ANFIS
In this experiment the data set comprises of 1432 input-output pairs which finally reduced to 1426
effective data pairs [77]. Here the optimally selected input variables are x(t − 1) and y(t − 1). The
number of rules required finally narrowed down to 3 and 4 respectively for the clustering and the grid
based method. Fig. 4.16 shows the rules and parameters for this model. In this example the first half of
the 713 data pairs are used for training and the remaining 713 data pairs are used for testing purpose.
The modeling result with this data-set is shown in Fig. 4.17.
Case 4: Modeling Results for the fourth set of the thermal power plant data with ANFIS
In this experiment the data set comprises of 994 input-output pairs which finally reduced to 988 effec-
tive data pairs as a result of the dynamic process model [77]. Here the optimally selected input variables
are x(t − 1) and y(t − 1). The number of rules required finally narrowed down to 2 and 4 respectively
for the clustering and the grid based method. Fig. 4.18 shows the rules and parameters for this model.
In this example the first 494 data pairs are used for training the model and the next 494 data pairs are
used for testing purpose. The modeling result for this data-set is shown in Fig. 4.19.
1
y(t)= 0.001565x(t−6)
+0.6354y(t−1)
Rule 1 −0.005981
0 58.6
23.34 0.015 0.219
1 y(t)=−0.0006048x(t−6)
+0.9293y(t−1)
Rule 2 +0.04373
0 58.6
23.34 0.015 0.219
Gauss MF parameters:− [6.252, 51.12]; [6.215, 50.21]; [0.01955, 0.2571]; [0.01185, 0.1784];
Figure 4.12: Rules for the ANFIS model when trained with half of the data set-I for the thermal power
plant example
0.25
actual output
predicted output
0.2
power output (GW)
0.15
0.1
0.05
0
0 200 400 600 800 1000 1200 1400
sample number
Figure 4.13: Actual and ANFIS model predicted output with the data set-I for thermal power plant
1
y(t)= −0.0008299x(t−3)
+1.263y(t−1)
Rule 1 −0.008002
0 57.95
33.11 0.096 0.216
1 y(t)= 0.001169x(t−3)
+0.3943y(t−1)
Rule 2 +0.05423
0 57.95
33.11 0.096 0.216
Figure 4.14: Rules for the ANFIS model when trained with half of the data set-II for the thermal power
plant example
0.22
0.2
0.18
power output (GW)
0.16
0.14
0.12
0.1
actual output
predicted output
0.08
0 200 400 600 800 1000 1200 1400
sample number
Figure 4.15: Actual and ANFIS model predicted output with the data set-II for thermal power plant
1
y(t)= 0.0008325x(t−1)
+1.227y(t−1)
Rule 1 −0.1408
0 63.52
5.02 0.03 0.323
1 y(t)= 0.0008978x(t−1)
+0.1598y(t−1)
Rule 2 +0.1508
0 63.52
5.02 0.03 0.323
1
y(t)= 0.00135x(t−1)
+0.7551y(t−1)
Rule 3 +0.02299
0 63.52
5.02 0.03 0.323
Gauss MF parameters:− [10.34, 47.04]; [10.34, 53.44]; [10.34, 41.09]; [0.0518, 0.181]; [0.0518, 0.214]; [0.0518, 0.157];
Figure 4.16: Rules for the ANFIS model when trained with half of the data set-III for the thermal power
plant example
0.35
actual output
predicted output
0.3
0.25
power output (GW)
0.2
0.15
0.1
0.05
0
0 500 1000 1500
sample number
Figure 4.17: Actual and ANFIS model predicted output with the data set-III for thermal power plant
1
y(t)= 0.001457x(t−1)
+4.0574y(t−1)
Rule 1 −0.5178
0 62.92
34.59 0.117 0.256
1 y(t)= 0.0004708x(t−1)
+0.8967y(t−1)
Rule 2 −0.003516
0 62.92
34.59 0.117 0.256
Gauss MF parameters:− [5.008, 47.93]; [5.009, 54.66]; [0.00276, 0.1448]; [0.01059 0.2078];
Figure 4.18: Rules for the ANFIS model when trained with half of the data set-IV for the thermal power
plant example
0.26
actual output
predicted output
0.24
0.22
power output (GW)
0.2
0.18
0.16
0.14
0.12
0.1
0 100 200 300 400 500 600 700 800 900 1000
sample number
Figure 4.19: Actual and ANFIS model predicted output with the data set-IV for thermal power plant
4.6. CONCLUSIONS 51
1
y(t)= 0.05748x(t−1)
+0.5502y(t−1)
Rule 1 +2.432
0 1.744
0.1338 0.5415 6.088
1 y(t)= 3.54x(t−1)
−0.08726y(t−1)
Rule 2 −0.26
0 1.744
0.1338 0.5415 6.088
Figure 4.20: Rules for the ANFIS model when trained with half of the data set-V for the thermal power
plant example
Case 5: Modeling Results for the fifth set of the thermal power plant data with ANFIS
In this experiment the data set comprises of 645 input-output pairs which finally reduced to 639 effective
data pairs as a result of the dynamic process model [77]. Here the optimally selected input variables
are x(t − 1) and y(t − 1). The number of rules required finally narrowed down to 2 and 4 respectively
for the clustering and the grid based method. Fig. 4.20 shows the rules and parameters for this model.
In this example the first 320 data pairs are used for training the model and the next 319 data pairs are
used for validation purpose. The validation result for this data-set are shown in Fig. 4.21.
For the above experiments, initially, the first half of the data set and then the entire data set is used
for training the model and the remaining half of the data set is used for validation of the model. The
modeling results with our 2-input based ANFIS model using the Box and Jenkins gas furnace and the
thermal power plant data are tabulated in Table 4.2. It is observed that when the number of training
data is more the RMSE of the test data reduces.
4.6 Conclusions
In this chapter, the neuro-fuzzy modeling of systems using ANFIS has been demonstrated using the
input-output data pairs collected from the Box and Jenkins gas furnace example and a thermal power
plant. The modeling results obtained by using the gas furnace data are compared with some of the
existing results in Table 4.1. In order to keep the same platform for comparing the results with some of
4.6. CONCLUSIONS 52
7
actual output
predicted output
energy output (MU) 5
0
0 100 200 300 400 500 600 700
sample number
Figure 4.21: Actual and ANFIS model predicted output with the data set-V for thermal power plant
Table 4.2: Modeling results for the Box and Jenkins gas furnace data and the thermal power plant data
using ANFIS
RMSE (testing)
Max No of
(Grid (Subtractive
Sl. No. Data No of Input variables training
partition clustering
rules data
based) based)
Box and Jenkins 145 0.5382 0.5724
1 4 x(t − 3), y(t − 1)
gas furnace 290 0.3574 0.3636
Thermal power plant 699 0.0114 0.0111
2 4 x(t − 6), y(t − 1)
(Data set-I) 1399 0.0108 0.0110
3 4 x(t − 3), y(t − 1)
(Data set-II) 1351 0.0085 0.0089
4 4 x(t − 1), y(t − 1)
(Data set-III) 1426 0.0072 0.0071
5 4 x(t − 1), y(t − 1)
(Data set-IV) 988 0.0033 0.0034
6 4 x(t − 1), y(t − 1)
(Data set-V) 639 0.4725 0.5176
4.6. CONCLUSIONS 53
the existing modeling results obtained from the first 7 rows as excerpted from a table in [77], the entire
gas furnace data set is used for training the network. From the modeling results shown in Table 4.1,
it can be observed that the mean square error (MSE) of the model obtained by using the subtractive
clustering based ANFIS is less than the two input Takagi-Sugeno model proposed by Chiu [77]. The
MSE for this model obtained by using the grid partitioning based ANFIS model is also less than that
obtained by using Chiu’s [77] model at the expense of only one more rule. As the addition of one
more rule has a very negligible effect on the computational cost, the model may be considered to be
comparable to the most accurate model. The modeling results of the Box and Jenkins gas furnace and
the thermal power plant are shown in Figs. 4.10 - 4.21. These results show that the 2-input ANFIS
model has good prediction capability. Table 4.2 represents the modeling results of the Box and Jenkins
gas furnace and the thermal power plant data using at first the initial half and then the entire data set
for training. From the results it can be concluded that the RMSE of the ANFIS based model reduces as
the number of training data increases. In the ANFIS model, back propagation learning method is used
for training the neural network for optimization of the fuzzy rules. A sufficient number of data should
be used to guarantee good training. There is still no formula to estimate the number of data required to
train a neural network. This number can vary greatly depending on the complexity of the problem and
the quality of the data, but many neural networks have been trained successfully with smaller number
of data. However, it is not always possible to find the number of training data in abundance. Here the
optimal selection of the data set is a major issue which is still a challenge in the field of neural networks.
In the present work, an effort is made in that direction keeping in mind the need for the choice of an
optimal training data set for the purpose of modeling. In the previous section, the modeling results
of the conventional ANFIS model are shown where the training data set comprises of either the entire
portion or half of the available data set. The proposed ANFIS model with the optimal choice of the
modeling data for training the ANFIS model is the topic of discussion in the next chapter.
Chapter 5
Optimal Data based ANFIS Modeling
In Chapter 4 it has been observed that the performance of the ANFIS based model improves as the
number of training data increases. But in many large scale systems the number of available training
data is less and the generation of new data is a costly affair. In such a scenario the system has to be
modeled with the available limited data only. This chapter proposes three different types of ANFIS
based system modeling schemes where the number of data pairs employed for training is minimized by
application of the Full factorial design (FFD) technique, the V-fold technique and the combined FFD-
V-fold technique in conjunction with ANFIS. Optimal choice of dataset for training is the key step here,
subsequent to which the modeling procedure is the the same as that of the ANFIS. These techniques
help in selection of the data pairs for training the ANFIS network optimally. The above techniques for
optimization of the training data set are described in the subsequent sections.
5.1 Full Factorial Design Based ANFIS Modeling

The statisticians and engineers can make a combined effort to reap the maximum benefit from statistical
analysis. For achieving this objective the experiments can be planned in advance to ensure that the
proper choice of experimental data can be made in a way that will provide the most unbiased and
precise results commensurate with the desired expenditure of time and money. But in many complex
systems the number of available training data for modeling is less. In that case, the available data have
to be optimally used for training. In this section a system modeling scheme is presented where out
of the available data set, a small number of data is critically chosen based on a statistically designed
experiment. This statistical design method is called the Full factorial design (FFD) [135, 136]. The
full factorial design method is used to select data optimally from the available data set for training the
54
5.1. FULL FACTORIAL DESIGN BASED ANFIS MODELING 55
ANFIS.
There are manifold advantages of statistically designed experiments. Some of the important advantages
are as follows:
• As compared to unplanned experimentation more information per experiment can be extracted
• The collection and analysis of information can be done in a more organized manner
• The conclusion from statistically designed experiments is very often evident without extensive
statistical analysis
• Credibility is awarded to the conclusions of an experimental analysis when the variability and the
sources of experimental error are based on statistical analysis
• These experiments can discover the interaction between experimental variables
For an experimental program involving a large number of tests, the order of selecting the specimens for
testing has to be randomized so that each specimen has an equal chance for being selected for testing.
The next step is to reduce the large number of possible variables so as to restrict the variables to a few
most important ones. The training of the neural network can be visualized as an experiment involving
a large number of tests for optimization. An important part of planning an experimental programme
is to identify the important variables that affect the response and deciding how to exploit them in the
experiment. The experimental variables that are controlled by the investigator are called the factors.
The important factors that affect the response have to be identified to use them in the experiment.
These factors may be independent i.e the level of one factor may be independent of the levels of the
other factors. But the effect on the response of one variable may also depend upon the levels of the
other variables due to the interaction of two or more factors.
Factorial designs are experiments in which all levels of each factor in an experiment are combined with
all levels of every other factor. In a factorial experiment several factors can be controlled to investigate
their effect at each of two or more levels. The experimental design consists of making an observation
at each of all possible combinations which can be formed for the different levels of the factors. Every
different combination is called a treatment combination.
The simplest and most common type of factorial design is one that uses 2 levels, n factors, i.e. 2n
factorial design. If we consider a 2 level, n factor system, training data have to be selected so that they
lie at each of the corners of a n -dimensional space. So in a 23 factorial design the training data should
x
B
3 4
7 8
1 2
x
A
5 6
x
C
Figure 5.1: Selection of training data in a 23 factorial design
be selected such that they lie at each corner of a 3-dimensional hypercube as illustrated in Fig. 5.1. A
2n factorial design is the simplest type of factorial design as it uses two levels and thereby reduces the
number of experimental conditions. But the disadvantage that is associated with a two level factorial
design is its inability to distinguish between linear and higher order effects. So the number of levels in
a factorial design experiment has to be judiciously chosen.
In addition to the number of levels, the Effect of a factor (EOF) has a significant bearing on the
modelling outcome. Effect of a factor may be defined as a change in response produced by a change in
the level of a factor and is given by,
P P
responses at high xi − responses at low xi
EOF (xi ) = (5.1.1)
Half the number of runs in experiment
where xi is a factor. The values of the factors corresponding to the level are called its responses. If
m is the number of levels and n is the number of factors, then there will be mn number of runs in the
experiment.
In the present experiments only two factors have been considered where the input and the output are
the two factors. The different levels that have been considered in the experiments are viz 2, 3 and 4.
When 3 level experiment is considered then there are 32 = 9 number of runs of the experiment. Once
the runs for the factors are decided, the corresponding responses are divided into two equal groups in
descending order of their values. The differences between the sum of the responses of each individual
group determines the numerator of the EOF. This value, when divided by half the number of runs fixes
the value of the EOF.
So a parameter having a higher EOF should have a greater representation on the training data set.
This is done by increasing the number of data related to that parameter in the training data set. As the
modeling of all the ANFIS based models has been carried out with two inputs, the FFD experiments
are performed with two factors only with different experimental levels. The FFD based ANFIS model
chosen for modeling the complex system is the one with the level having the least RMSE for its model-
ing result. The different experimental levels that have been used in the modeling schemes for building
the test ANFIS network are as follows:
• 2-level full factorial design
At first only two levels viz the minimum and the maximum of the input-output data pairs are
considered. Therefore with full factorial design at first only 22 = 4 data pairs are selected. These
4 data pairs are taken around the minimum and the maximum input-output pairs with equal
representation (2 pairs around minimum and 2 pairs around maximum).
• 3-level(1) full factorial design
Now one more level in the form of the mid-value of the data range is added to the former two
levels. So now the 3 levels are the minimum, the maximum and the middle values of the data
range. It leads to factorial design of 32 = 9 data pairs for the first run.
• 3-level(2) full factorial design
For another 3-level combination the cluster center is added instead of the mid-value of the data
range in the previous experiment. It again leads to factorial design of 32 = 9 data pairs initially.
Another experiment is performed by taking all the 4 different levels of the dataset viz the max-
imum and the minimum values, the middle value of the data range and the cluster center. This
5.2. V-FOLD TECHNIQUE BASED ANFIS MODELLING 58
leads to factorial design of 42 = 16 data pairs for the first run.
5.2 V-Fold Technique Based ANFIS Modelling

This section presents another ANFIS based complex system modelling method where the number of
data pairs employed for training the ANFIS network can be chosen by application of a technique called
the V-fold technique [137–139]. This method can be used when the number of available training data
are less.
V-fold cross validation technique is a highly accurate method for training a neural network and it has
the advantage of not requiring a separate, independent dataset for assessing the accuracy of the neural
network. If a network is built using a specific learning dataset, it is necessary to have test data samples
independent of the learning dataset that was used to train the network. However, it is difficult or
expensive to obtain independent test data frequently and moreover it is undesirable to hold back data
from the learning dataset to use for a separate test because that weakens the learning dataset. V-fold
cross validation is a technique for performing independent tests without requiring separate test datasets
and without reducing the data used to build the network.
The general idea of this method is to divide the overall sample into a number of folds, say V. The same
type of analysis is then successively applied to the observations belonging to all the V-1 folds (training
samples), and the results of the analysis are applied to the testing sample, which is the Vth sample (the
sample or fold that was not used). This is repeated until V random samples are drawn from the data
for the analysis. The results for the V replications are aggregated (averaged) to yield a single measure
of the stability of the respective model, i.e. the validity of the model for predicting new observations.
Thus, this technique allows the analyst to evaluate the overall accuracy of the respective prediction
model or method in repeatedly drawn random samples. Here the training and validation process using
the V-fold technique for building an ANFIS model process are accomplished as follows :
The learning dataset is partitioned into a number of groups called “folds”. The partitioning is done using
stratification methods so that the distribution of categories of the target variable are approximately the
same in the partitioned groups. In “V-fold cross classification” V is the number of groups that the
dataset is partitioned into. Research has established that 10 partitions are optimum and using more than
10 partitions does not yield any significant improvement in results [137–139]. As such 10 partitions are
formed in the experiment. Out of these 10 partitions, 9 are grouped into a new pseudo-learning dataset.
An ANFIS model is built using this pseudo-learning dataset. The quality of the network built with this
5.3. FFD-V-FOLD BASED ANFIS MODELING 59
new pseudo-learning data set will in general, be a little inferior to the model obtained by fitting the full
learning dataset, because only 90% of the data is used to build it. Since 10% (1 out of 10 partitions) of
the data is held back from being used by the network, it can be used as an independent test sample for
the ANFIS. The 10% of the data that was held back when the test ANFIS was built is now run through
the test network and the classification error for that data is computed.
A different set of 9 partitions is now collected into a new pseudo-learning dataset. The partition that
was held back previously is selected this time so that it is different from the partition held back for the
first test ANFIS. A second ANFIS is built and its classification error is computed using the data that
was held back when it was built. This process is repeated 10 times, building 10 separate networks. In
each case, 90% of the data is used to build the network and 10% is held back for independent testing.
A different 10% is held back for each test network. The V-fold based ANFIS model showing the least
modeling error is considered as the final model for the system.
5.3 FFD-V-fold Based ANFIS Modeling

In this proposed modeling scheme the full factorial design and the V-fold methods have been combined
together to generate the optimal data set for training the ANFIS model. This model is basically a
V-fold based ANFIS model where the full factorial design method is used to determine the data for
the different folds. The use of the full factorial design method also ensures that similar kind of data
are selected to the same folds of the V-fold based ANFIS model. This combined method presents a
systematic way of choosing and grouping the data for forming the different folds for the V-fold based
ANFIS model. The flow chart for this method is drawn in Fig. 5.2.

In the experiment the same data sets which were used in the previous chapter have been used for
modeling. These data are the Box and Jenkins gas furnace data and the NEEPCO thermal power plant
data. From these available data sets the data for training the subtractive clustering and the grid partition
based ANFIS models are chosen optimally by applying the Full factorial design and the V-fold based
techniques. While using the subtractive clustering method a cluster radius of 0.5 was selected to extract
the initial fuzzy model. The results and observations have been presented in the subsequent subsections.
Start
Framing of fuzzy rules

by
Subtractive clustering/Grid partitioning
Application of FFD technique

for
forming the data folds
Using the data folds

for
the V−fold technique
Application of V−fold technique

for
training the ANFIS model
Stop
Figure 5.2: Flow Chart for the FFD-V-fold Technique
5.4.1 Modeling results with Full factorial design based ANFIS model
The optimally selected data obtained by applying the full factorial design technique have been used in
these experiments to train the ANFIS model. While training the ANFIS model, the effect of selecting
an input variable xi corresponding to different levels in the range of operation of the input variable
is investigated. Considering the input-output pairs (xi , yi ) and 2, 3 and 4 levels of operation, the ob-
tained factorially designed data are22 , 32 and 42 respectively for use in training. These experiments are
described below:
Case 1:- Modeling with Box and Jenkins gas furnace data
The Box and Jenkins gas furnace data [102] are used in this experiment. Optimal choice of dataset for
training is the key step here subsequent to which the modelling procedure is the same as that of ANFIS.
In this case, x(t − 3) and y(t − 1) were found to be the optimal input variables by using the RMSE
criterion. So these two inputs are used in the ANFIS model. The method for choosing the number of
modeling data for the different levels of the FFD based models is explained below.
Considering only the minimum and the maximum values of the input-output data pairs, at first
only 22 = 4 data pairs are selected. These 4 data pairs are taken around the minimum and the
maximum valued input-output pairs with equal representation (2 pairs around the minimum and
2 pairs around the maximum). Now EOF for both xi and yi (from Eq. 5.1.1) are calculated and
found to be 1 for both the cases. So 4 (1 each around the minimum and the maximum values of
both xi and yi ) additional data pairs are now needed and these are taken around the minimum and
the maximum valued input-output data pairs. Again EOF is calculated and is found to be 1. So 4
additional data pairs are now selected around the minimum and the maximum input-output pairs
with equal representation (2 pairs around the minimum and 2 pairs around the maximum) for
making a balanced representation of data around the minimum and the maximum values. EOF is
again calculated and is found to be 1. Therefore 4 more input-output data pairs are taken around
the minimum and the maximum of the input-output pairs. Finally 4 + 4 + 4 + 4 = 16 data pairs
are selected for training.
• 3(1)-level full factorial design
Now one more level in the form of the mid-value of the data range is added to the former two
extreme levels. So now the 3 levels are the minimum, the maximum and the middle value of
the data range. It leads to factorial design of 32 = 9 data pairs. Proceeding in a similar manner
as in the 2-level case, EOF values of 2, 1 and 1 are obtained for yi , xi and both of xi and yi
respectively, in 3 successive evaluations. So 6 (2 each around yi of the minimum, the maximum
and the centroid), 3 (1 each around xi of the minimum, the maximum and the centroid) and 6 (1
each around xi and yi of the minimum, the maximum, and the centroid) additional data pairs are
required. Therefore for this experiment 9 + 6 + 3 + 6 = 24 data pairs are used for training.
For another 3-level combination the cluster centers are added instead of the mid-value of the data
range which was the case in the previous experiment. After three successive evaluations of the
32 = 9 data pairs, EOF values of 1, 1 and 1 are obtained for xi . Therefore in this factorially
designed experiment additional data pairs of 4, 4 and 4 have to be considered, thereby requiring
a total of 9 + 4 + 4 + 4 = 21 data pairs for training.
Another experiment is performed by taking all the 4 different levels of the data set viz the mini-
mum, the maximum, the middle value of the data range the cluster center. After three successive
evaluations on the 42 = 16 data pairs, EOF values of 1, 1 and 1 are obtained for yi , xi and (xi , yi )
of the different levels. Therefore with 1 data pair each around the minimum, the maximum, the
centroid and each of the two cluster centers for yi , xi and (xi , yi ), a value of 5, 5 and 10 additional
data pairs are taken making a total number of 16 + 5 + 5 + 10 = 36 training data pairs.
Case 2:- Modeling with thermal power plant data
(i) Data set-I

The first set of the NEEPCO’s thermal power plant data is used in this experiment. In this case, x(t − 6)
and y(t − 1) are used in the ANFIS model as they were found to be the optimal input variables by using
the RMSE criterion. The procedure for choosing the optimal number of training data is given below.
Considering only the minimum and the maximum valued input-output data pairs, only 22 = 4
data pairs are selected initially. After three successive evaluations, EOF values of 1, 1 and 1
were obtained for the two successive (xi , yi ) values and lastly for yi . Therefore with 4, 4 and 2
additional data pairs, a total of 4 + 4 + 4 + 2 = 14 input-output pairs are used for training.
Now with the addition of one more level in the form of the mid-value of the data range to the
former two levels, a factorial design of 32 = 9 data pairs, comprising of the minimum, the maxi-
mum and the middle value of the data range is obtained. Again after two successive evaluations
for yi , the values of EOF obtained are 2 and 3 respectively. So with the inclusion of 6 and 9
additional data pairs, a total of 9 + 6 + 9 = 24 input-output pairs are used in this experiment for
training.
For another 3-level combination the cluster center is added instead of the mid-value of the data
range. After three successive evaluations of the 32 = 9 data pairs, EOF values of 1,1 and 2 are
obtained for yi . Therefore in this factorially designed experiment, 9+4+4+8 = 25 input-output

data pairs are used for training.
Another experiment is performed by taking all the 4 different levels of the dataset, viz the max-
imum and the minimum values, the middle value of the data range and the cluster center. After
two successive evaluations on the 42 = 16 data pairs, EOF values of 1 and 2 were obtained for
(xi , yi ). So with 10 and 20 additional data pairs, 16 + 10 + 20 = 46 input-output data pairs are
used for training.
(ii) Data set-II

The second set of the thermal power plant data was used in this experiment. In this case, x(t − 3) and
y(t−1) were found to be the optimal input variables. So these two inputs are used in the ANFIS model.
The number of data pairs required for the different levels of the experiment are explained below.
Here at first 22 = 4 data pairs are selected considering the minimum and the maximum values of
the input-output data pairs. With three successive evaluations of EOF, the values obtained are 1, 1
and 1 respectively for (xi , yi ). So the total number of data used for training are 4+4+4+4 = 16.
Starting with 32 = 9 data pairs initially for the three levels of the maximum, the minimum and
the mid-values of the data pairs, the values of EOF obtained are 1, 1 and 1 after three successive
evaluations for (xi , yi ). So a total of 9 + 6 + 6 + 6 = 27 data pairs are used for training in the
ANFIS model.
With 32 = 9 data pairs used initially for the minimum, the maximum, and the cluster center
levels, after three successive evaluations the values of EOF found are 1, 1 and 1 for xi , (xi , yi )
and yi . This results in a total number of 9 + 4 + 8 + 4 = 25 training data pairs.

Starting initially with 42 = 16 data pairs and evaluating the EOF after two successive runs, the
values of 2 and 1 are obtained for yi and (xi , yi ) respectively. Therefore the total number of data
pairs used for training is 16 + 10 + 10 = 36.
(iii) Data set-III

In the third set of the thermal power plant data, x(t − 1) and y(t − 1) were found to be the optimal
inputs. The number of training data chosen for the different models are shown below.

With 22 = 4 data pairs chosen initially and EOF values of 1, 1 and 1 for (xi , yi ) respectively, the
training for this model was carried out with 4 + 4 + 4 + 4 = 16 data pairs.

Starting with 32 = 9 data pairs and EOF values of 1, 1 and 1 for (xi , yi ), a total of 9+6+6+6 = 27
training data are used for training the ANFIS model.

In this experiment, with 32 = 9 data taken initially, the values of EOF are obtained as 1, 1 and 1
respectively for (xi , yi ). So the total number of data pairs used is 9 + 8 + 8 + 8 = 33.

With 42 = 16 data pairs selected initially, the evaluated values of EOF are 1, 1 and 1 for input-
output pair (xi , yi ). So the total number of data pairs used in training is 16 + 10 + 10 + 10 = 46.
(iv) Data Set-IV

In the fourth set of the thermal power plant data, x(t − 1) and y(t − 1) were found to be optimal inputs.
The number of modeling data chosen for the different models are as follows.

With 22 = 4 data pairs selected initially and computed EOF values of 1, 1 and 2 for two suc-
cessive evaluations of (xi , yi ) and yi , the total number of data pairs used in training the ANFIS
model is 4 + 4 + 4 + 4 = 16.

Starting with 32 = 9 pairs of initial data and corresponding EOF values of 1, 1 and 1 for (xi , yi ),
the total number of data pairs needed for training is 9 + 6 + 6 + 6 = 27.

For this experiment, EOF values of 1, 1 and 1 were obtained for 32 = 9 initial data for yi and
then for two successive input-output pairs (xi , yi ). Hence the number of data pairs required for
training is 9 + 4 + 8 + 8 = 29.

For this case, the values of EOF for xi , (xi , yi ) and yi are 1, 1 and 1 for 42 = 16 input-output data
pairs. Therefore the ANFIS model requires 16 + 5 + 10 + 5 = 36 data pairs.
(V) Data set-V

In the fifth set of the thermal power plant data, x(t − 1) and y(t − 1) were found to be optimal and
hence are used in our model.

Using 22 = 4 number of data initially, the EOF values obtained are 1, 1 and 1 for (xi , yi ). This
resulted in the requirement of 4 + 4 + 4 + 4 = 16 data pairs for training.

The calculated values of EOF obtained for 32 = 9 data pairs for two successive yi and (xi , yi )
are 2, 2 and 1 respectively . So 9 + 6 + 6 + 6 = 27 data pairs are needed for training.

With 32 = 9 initial data pairs, the values of EOF obtained are 1, 1 and 1 for (xi , yi ). For these
values of EOF, the total number of data pairs required for training is 9 + 8 + 8 + 8 = 33.

Here 42 = 16 input-output data pairs are chosen initially. Corresponding values of EOF are 1, 1
and 1 for (xi , yi ). Therefore the number of data pairs needed for training is 16+10+10+10 = 46.
The modeling results so obtained with the FFD based ANFIS model for the Box and Jenkins gas furnace
and the thermal power plant data are tabulated in Table 5.1. Here the ANFIS model used is of 2-input
4-rule type with the FFD based optimally selected data for training. The second half of the data set is
used for validation of the model.
using FFD based ANFIS
RMSE (testing)
No of
(Grid (Subtractive
Sl. No. Data Input variables No of Model training
rules FFD data
based) based)
Box and 2-level 16 0.6473 0.6437
Jenkins 3(1)-level 24 0.7254 0.6875
1 x(t − 3), y(t − 1) 4
gas 3(2)-level 21 0.5607 0.6229
furnace 4-level 36 0.5589 0.5972
Thermal 2-level 14 0.0264 0.0275
power plant 3(1)-level 24 0.0152 0.0205
2 x(t − 6), y(t − 1) 4
(Data 3(2)-level 25 0.0216 0.0117
set-I) 4-level 46 0.0151 0.0125
Thermal 2-level 16 0.0175 0.0200
3 x(t − 3), y(t − 1) 4
(Data 3(2)-level 25 0.0140 0.0119
set-II) 4-level 36 0.0097 0.0090
Thermal 2-level 16 0.0107 0.0175
4 x(t − 1), y(t − 1) 4
(Data 3(2)-level 33 0.0090 0.0083
set-III) 4-level 46 0.0099 0.0095
Thermal 2-level 16 0.0038 0.0036
5 x(t − 1), y(t − 1) 4
(Data 3(2)-level 29 0.0036 0.0035
set-IV) 4-level 36 0.0061 0.0053
Thermal 2-level 16 0.5890 0.9613
6 x(t − 1), y(t − 1) 4
(Data 3(2)-level 33 1.0259 0.5347
set-V) 4-level 46 0.8392 0.9149
5.4.2 V-fold technique based ANFIS model
In the following experiment the V-fold method is used to choose the data set for training the ANFIS
model. After the selection of the data for the different folds, which is around one-eighth of the training
data used in the conventional model, the results so obtained are presented below.
Case 1:- Modeling with Box and Jenkins gas furnace data
The data set for the Box and Jenkins gas furnace example is again used for modeling the V-fold based
ANFIS. The optimal inputs so chosen are x(t − 3) and y(t − 1) along with a cluster center of 0.5. The
different folds for applying the V-fold technique were constituted with 18 data pairs which is around
one-eighth of the 145 input-output pairs used for training the conventional ANFIS model.
Case 2:- Modeling with thermal power plant data
(i) Data set-I

With x(t − 6) and y(t − 1) as the optimal inputs and a cluster center of 0.5 for the subtractive clustering
method, the modeling is carried out with 81 data pairs, which is around one-ninth of the 699 data pairs
required in the conventional ANFIS model.
(ii) Data set-II
Here modeling is carried out with 72 data pairs which is around one-ninth of the 675 input-output data
pairs . The optimal inputs are x(t − 3) and y(t − 1).
(iii) Data set-III
The optimal inputs for this model are x(t − 1) and y(t − 1). For this data set the V-fold based ANFIS
model is built with about one-tenth of the 713 input-output data pairs i.e 72 data pairs.
(iv) Data set-IV
For this data set, x(t − 1) and y(t − 1) are found to be the optimal inputs. The V-fold based ANFIS
model is built with 72 data pairs which is around one-seventh of 494 data pairs.
(v) Data set-V
Here the optimal inputs are x(t − 1) and y(t − 1). For the purpose of modeling, around one-ninth of
320 i.e 36 input-output data pairs are chosen.
The modeling results obtained by using the V-fold technique based ANFIS model are tabulated in
Table 5.2.
using V-fold based ANFIS model
No of RMSE (testing)
Number
Sl. No. Data training Input variables (Grid partition (Subtractive
of rules
data based) clustering based)
Box and
1 Jenkins 18 x(t − 3), y(t − 1) 4 0.5943 0.5556
gas furnace
Thermal
2 power plant 81 x(t − 6), y(t − 1) 4 0.0110 0.0115
(Data set-I)
Thermal
3 power plant 72 x(t − 3), y(t − 1) 4 0.0092 0.0104
(Data set-II)
Thermal
4 power plant 72 x(t − 1), y(t − 1) 4 0.0078 0.0082
(Data set-III)
Thermal
5 power plant 72 x(t − 1), y(t − 1) 4 0.0034 0.0036
(Data set-IV)
Thermal
6 power plant 36 x(t − 1), y(t − 1) 4 0.6508 0.6097
(Data set-V)
5.4.3 Modeling results with FFD-V-fold technique based ANFIS model
The V-fold technique is now combined with the FFD based ANFIS. The following experiments are
carried out with this combined model.
Case 1:- Modeling results for Box and Jenkins gas furnace data
The optimally selected data set obtained from the Box and Jenkins gas furnace data set is used for
modeling at various levels like 2-level, 3(1)-level, 3(2)-level and 4-level. The modeling is carried out
by using both the grid partition based method and subtractive clustering technique by choosing a cluster
radius of 0.5. The best modeling result is observed in the case of the 2-level FFD-V-fold based ANFIS
model both for the grid and the subtractive clustering based models by using x(t − 3) and y(t − 1) as
the two optimal inputs.
Case 2:- Modeling results for the thermal power plant data
(i) Data set-I

Here the modeling is carried out by using x(t − 6) and y(t − 1) as the two optimal inputs. For this data
set the best result was obtained for the 2-level FFD-V-fold based ANFIS model.
(ii) Data set-II
The two optimal inputs used here are x(t − 3) and y(t − 1). The 3(1)-level FFD-V-fold based ANFIS
model produced the best result.
(iii) Data set-III
The optimal inputs used here for modeling are x(t − 1) and y(t − 1). The FFD-V-fold based ANFIS
model with the 3(2) level showed the least MSE.
(iv) Data set-IV
Here x(t − 1) and y(t − 1) are the optimal inputs used for modeling and the least MSE was obtained
for the 2-level FFD-V-fold based ANFIS model.
(vi) Data set-V
Here the modeling is carried out by using the optimal inputs x(t − 1) and y(t − 1). The 3(1)-level
FFD-V-fold based ANFIS model shows the least MSE.
The results obtained by using the combined FFD-V-fold technique based ANFIS model are tabulated in
Table 5.3. In Table 5.4, the results obtained by using the different models viz conventional ANFIS, FFD
based ANFIS, V-fold based ANFIS and FFD-V-fold based ANFIS are compared with respect to their
RMSE. From Table 5.4, it is observed that the best modeling results are obtained with the FFD-V-fold
based ANFIS model as the RMSE (testing) is the least in this case for all the different types of data.
The modeling results for the different data sets as obtained by using different modeling methodologies
are shown in Figs. 5.3 - 5.8. In Fig. 5.3 the actual output is plotted vs the predicted output for the
FFD-V-fold based ANFIS model using the Box and Jenkins gas furnace data. Fig. 5.4 shows the actual
and the predicted output for the FFD-V-fold based ANFIS model using the data set-I for the thermal
power plant. Fig. 5.5 shows the actual and the predicted output for the FFD-V-fold based ANFIS model
using the data set-II for the thermal power plant. Fig. 5.6 shows the actual and the predicted output for
the FFD-V-fold based ANFIS model using the data set-III for the thermal power plant. Fig. 5.7 shows
the actual and the predicted output for the FFD-V-fold based ANFIS model using the data set-IV for
the thermal power plant. Fig. 5.8 shows the actual and the predicted output for the FFD-V-fold based
ANFIS model using the data set-V of the thermal power plant.
using FFD-V-fold based ANFIS
RMSE (testing)
No of Model No of
(Grid (Subtractive
Sl. No. Data Input variables rules FFD-V-fold training
data
based) based)
Box and 2-level 18 0.5378 0.5332
Jenkins 3(1)-level 18 0.5705 0.6017
1 x(t − 3), y(t − 1) 4
gas 3(2)-level 18 0.6153 0.5373
furnace 4-level 18 0.5627 0.5684
Thermal 2-level 81 0.0110 0.0108
2 x(t − 6), y(t − 1) 4
(Data 3(2)-level 81 0.0109 0.0112
set-I) 4-level 81 0.0112 0.0114
Thermal 2-level 72 0.0092 0.0092
3 x(t − 3), y(t − 1) 4
(Data 3(2)-level 72 0.0091 0.0093
set-II) 4-level 72 0.0088 0.0091
Thermal 2-level 72 0.0084 0.0079
4 x(t − 1), y(t − 1) 4
(Data 3(2)-level 72 0.0076 0.0075
set-III) 4-level 72 0.0083 0.0082
Thermal 2-level 72 0.0033 0.0034
5 x(t − 1), y(t − 1) 4
(Data 3(2)-level 72 0.0035 0.0035
set-IV) 4-level 72 0.0035 0.0036
Thermal 2-level 36 0.5740 0.5293
6 x(t − 1), y(t − 1) 4
(Data 3(2)-level 36 0.6651 0.5900
set-V) 4-level 36 0.5174 0.5312
Table 5.4: Comparison of modeling results for the Box and Jenkins gas furnace data and the thermal
power plant data using various ANFIS models
RMSE (testing)
No of
(Grid (Subtractive
Sl. No. Data Input variables No of Model training
rules data
based) based)
Box and ANFIS 145 0.5382 0.5724
Jenkins FFD 18 0.5589 0.5972
1 x(t − 3), y(t − 1) 4
gas V-Fold 18 0.5943 0.5556
furnace FFD-V-fold 18 0.5378 0.5332
Thermal ANFIS 699 0.0114 0.0111
power plant FFD 46, 25 0.0151 0.0117
2 x(t − 6), y(t − 1) 4
(Data V-Fold 81 0.0110 0.0115
set-I) FFD-V-fold 81 0.0109 0.0108
Thermal ANFIS 675 0.0118 0.0111
power plant FFD 36 0.0097 0.0090
3 x(t − 3), y(t − 1) 4
(Data V-Fold 72 0.0092 0.0104
set-II) FFD-V-fold 72 0.0086 0.0089
Thermal ANFIS 713 0.0081 0.0076
4 x(t − 1), y(t − 1) 4
(Data V-Fold 72 0.0078 0.0082
set-III) FFD-V-fold 72 0.0076 0.0075
Thermal ANFIS 494 0.0037 0.0035
5 x(t − 1), y(t − 1) 4
(Data V-Fold 72 0.0034 0.0036
set-IV) FFD-V-fold 72 0.0033 0.0034
Thermal ANFIS 319 0.5591 0.5361
power plant FFD 16, 27 0.5890 0.5347
6 x(t − 1), y(t − 1) 4
(Data V-Fold 36 0.6508 0.6097
set-V) FFD-V-fold 36 0.5174 0.5118
62
actual output
predicted output
60
58

56
54
52
50
48
46
44
0 50 100 150 200 250 300
sample number
Figure 5.3: Actual and FFD-V-fold based ANFIS model predicted output with the Box and Jenkins gas
furnace data
0.25
actual output
predicted output
0.2
power output (GW)
0.15
0.1
0.05
0
0 200 400 600 800 1000 1200 1400
sample number
Figure 5.4: Actual and FFD-V-fold based ANFIS model predicted output with data set-I of thermal
power plant
0.22
0.2
power output (GW) 0.18
0.16
0.14
0.12
0.1
actual output
predicted output
0.08
0 200 400 600 800 1000 1200 1400
sample number
Figure 5.5: Actual and FFD-V-fold based ANFIS model predicted output with data set-II of thermal
power plant
0.35
actual output
predicted output
0.3
0.25
power output (GW)
0.2
0.15
0.1
0.05
0
0 500 1000 1500
sample number
Figure 5.6: Actual and FFD-V-fold based ANFIS model predicted output with data set-III of thermal
power plant
0.26
actual output
predicted output
0.24
0.22
power output (GW)
0.2
0.18
0.16
0.14
0.12
0.1
0 100 200 300 400 500 600 700 800 900 1000
sample number
Figure 5.7: Actual and FFD-V-fold based ANFIS model predicted output with data set-IV of thermal
power plant
7
actual output
predicted output
6
4
energy output (MU)
−1
−2
0 100 200 300 400 500 600 700
sample number
Figure 5.8: Actual and FFD-V-fold based ANFIS model predicted output with data set-V of thermal
power plant
5.5. CONCLUSION 75
5.5 Conclusion
From Table 5.4 it is observed that the FFD-V-fold based ANFIS model shows the best performance
even though only around one-eighth of the dataset used in the conventional ANFIS model has been
selected for training. The FFD-V-fold based ANFIS model was built on the basis of optimally chosen
data for modeling. This shows that the ANFIS model based on the training data set selected by using
the FFD-V-fold technique shows the best prediction capability. To further test the performance of the
proposed FFD-V-fold based ANFIS model, the modeling results of this chapter are compared with
standard statistical models as well as another soft computing based model in the next chapter. The two
types of models that are chosen for the comparative analysis are the statistical model and the GA based
fuzzy model.
Chapter 6
Statistical Models and Genetic Algorithm

Based Fuzzy Model
From the preceding chapters it is observed that by using an optimally selected training data set, the
performance of the conventional ANFIS model can be enhanced. The proposed FFD-V-fold technique
augments the conventional ANFIS model and it even outperforms the conventional ANFIS model with
lesser RMSE but utilizing far fewer training data. In order to determine the efficacy of this proposed
FFD-V-fold based ANFIS model, it is compared with conventional statistical models like ARX, AR-
MAX and ARIMAX in this chapter. The proposed model is also compared with the genetic algorithm
(GA) based fuzzy model which is a new technique in the emerging area of artificial intelligence. Lastly
a controller is designed using the proposed FFD-V-fold based ANFIS model and its performance is
studied.
6.1 Statistical Models

Engineers and scientists have been using probability and statistics as a working tool in many areas of
engineering practices. Most often, during engineering design, many problems crop up due to poorly
defined situation or having to use data with low precision. This problem can be solved by using sta-
tistical models [140, 146] by proper application of statistical analysis. The statistical models help in
making important decisions in engineering. Here the observed samples are used to estimate the statis-
tical population whose properties provide the basis for decision making. The results from the analysis
of experimental data can be unequivocally described by appropriate statistical parameters. Many of the
techniques used for data analysis are based on univariate and multivariate statistics.
76
6.1. STATISTICAL MODELS 77
In this chapter some of the existing statistical models are fitted into the systems under study and they
are compared with the models proposed in the previous chapters. Some of the statistical models which
are used for comparison with the soft computing based models proposed in the previous chapter are
described in the successive subsections. The statistical models that have been used for modeling are
the Auto regressive with exogenous input (ARX) model, Auto regressive moving average with exoge-
nous input (ARMAX) model and Auto regressive integrated moving average with exogenous input
(ARIMAX) model.
6.1.1 Auto Regressive with Exogenous Input (ARX) Model
The Auto regressive (AR) [140] is a very common model with the help of which one can shape the
frequency characteristics of the model with a few linear parameters. AR is a very powerful tool for
analysis of weakly damped oscillatory systems which may be hidden under a high noise level. The
Auto regressive with exogenous input (ARX) [140] model is the most widely applied linear dynamic
model because of its easily computable parameters. The model uses linear least square technique as the
prediction error is linear in the parameters.
6.1.2 Auto Regressive Moving Average with Exogenous Input (ARMAX) Model
The Moving average (MA) [140] model has lesser significance in practical engineering applications
than the AR model because it cannot be used to model oscillations with a few parameters. In addition
to this, the MA model is nonlinear in its parameters when the prediction error approach is considered. In
the Auto regressive moving average (ARMA) [140] model, the MA and the AR models are combined
together so that the flexibility of the AR model can be enhanced. After the ARX model, the Auto
regressive moving average with exogenous input (ARMAX) [140] model is the next most popular
model as it possesses an extended noise model and thereby becomes more flexible.
6.1.3 Auto Regressive Integrated Moving Average with Exogenous Input (ARI-
MAX) Model
In many time series data set, the homogeneity property is reflected even though the series behaves as
though they have no fixed mean. Apart from local level or a combination of local level and trend, one
part of such series is found to be like any other part. Such homogeneous non stationary behavior of
the data can be described by a model with the assumption that some suitable difference (say dth ) of the
6.2. GENETIC ALGORITHM BASED FUZZY MODEL 78
process is stationary. The autoregressive moving average with exogenous input model for which the dth
difference is stationary, is called the Autoregressive integrated moving average with exogenous input
(ARIMAX) [102] model.
6.2 Genetic Algorithm Based Fuzzy Model

Applications requiring the optimization of a multi-dimensional function can be successfully solved by
using evolutionary algorithms. These algorithms are based on the evolution of a population towards
the solution of the application problem. The population of possible solutions which evolves in one
generation is progressively taken to the next successive generations until it converges at a satisfactory
solution of the problem. The genetic algorithm (GA) [141, 142] is one such evolutionary algorithm.
This algorithm was envisaged by Holland [67] in the 1970s. It is based on the Darwinian theory of
survival of the fittest, which states that the fitter and stronger individuals in a population have a higher
chance of creating offsprings for the next generation. It can be implemented as an optimization search
procedure which uses the principles of genetics and natural selection by modeling possible solutions to
a search problem as strings of zeroes and ones. This algorithm comprises of basic genetic operators.
They are selection, crossover and mutation. The flow chart for this method is presented in Fig. 6.1.
A genetic algorithm based fuzzy model is studied in this chapter. Here the genetic algorithm is
fused with fuzzy logic to model a system. In this GA based model, the process of deriving the initial
fuzzy model is the same as that of the ANFIS model. After the initial fuzzy model is formed, the
genetic algorithm is used to update the consequent parameters of the generated fuzzy rules, so as to
produce a final GA based fuzzy model of the system.

The statistical and the GA based fuzzy model for the Box and Jenkins gas furnace and the thermal
power plant data are presented in the following two subsections. In each of these cases, the first half of
the data set is used for training and the next half is used for validation purpose. For the Box and Jenkins
gas furnace data, the input x(t) is the gas feed rate in cubic feet per minute (f t3 /min) and the output
y(t) is in the form of percentage of carbon-dioxide (CO2 ) concentration in the outlet gas. Similarly for
the first four sets of thermal power plant data, the input x(t) is the gas flow rate in cubic kilometer per
hour (km3 /hr) and the output y(t) is in the form of generated power in Gigawatt (GW). In the fifth set
Start
Formation of
an
initial fuzzy model
Determination
of
Premise parameters by fuzzy logic
Determination
of
Consequent parameters by GA
Stop
Figure 6.1: Flow Chart for the GA based Fuzzy model
of the thermal power plant data, the input x(t) is the gas feed in Million metric standard cubic metre
(Mmscum) and the output y(t) is in the form of generated energy in Million unit (MU), where 1 unit =
1kWh (kilo-watt-hour).
6.3.1 Statistical models
For the model identification purpose, three different statistical models have been used. These are the
ARX, ARMAX and the ARIMAX model. The modeling results obtained from these three statistical
models are arranged in a tabular form in Table 6.1.
From Table 6.1 it is observed that the performance of the ARIMAX model is the best having the least
RMSE as compared to the ARX and the ARMAX model. Hence, the ARIMAX model is selected to
compare with the proposed model. The ARIMAX model is built for both the Box and Jenkins gas
furnace and the thermal power plant data using the optimally selected inputs.
Table 6.1: Results obtained for Statistical models using the Box and Jenkins gas furnace and the thermal
power plant data
No of RMSE
Sl. No. Data training Model with
data testing data
Box and ARX 1.5796
1 Jenkins 145 ARMAX 1.5353
gas furnace ARIMAX 1.2305
Thermal ARX 0.0676
2 power plant 699 ARMAX 0.0533
(Data set-I) ARIMAX 0.0110
Thermal ARX 0.0980
(Data set-II) ARIMAX 0.0163
Thermal ARX 0.1302
(Data set-III) ARIMAX 0.0113
Thermal ARX 0.0947
(Data set-IV) ARIMAX 0.0096
Thermal ARX 0.5871
(Data set-V) ARIMAX 0.5451
Case 1:- Box and Jenkins gas furnace data
Here, the 145 Box and Jenkins gas furnace data pairs are used for building the ARIMAX model with
x(t − 3) and y(t − 1) as the two optimal inputs. Fig. 6.2 shows the actual output and the ARIMAX
model predicted output which are plotted versus sample number.
Case 2:- Thermal power plant data
(i) Data set-I

The first set of the thermal power plant data comprises of 699 data pairs. The optimal inputs used are
x(t − 6) and y(t − 1). Fig. 6.3 shows the ARIMAX model’s predicted output along with actual output
versus sample number.
(ii) Data set-II
For this data set the ARIMAX model is built with x(t − 3) and y(t − 1) as the optimal inputs and by
using 675 input output data pairs. The actual and the ARIMAX model predicted output for this data set
are shown in Fig. 6.4.
(iii) Data set-III
62
actual output
predicted output
60
58
% 0f CO2 in the outlet gas
56
54
52
50
48
46
44
0 50 100 150 200 250 300
sample number
Figure 6.2: Actual and ARIMAX model predicted output with Box and Jenkins gas furnace data
0.25
actual output
predicted output
0.2
power output (GW)
0.15
0.1
0.05
0
0 200 400 600 800 1000 1200 1400
sample number
Figure 6.3: Actual and ARIMAX model predicted output with data set-I of thermal power plant
0.24
actual output
predicted output
0.22
0.2
power output (GW)

0.18
0.16
0.14
0.12
0.1
0.08
0 200 400 600 800 1000 1200 1400
sample number
Figure 6.4: Actual and ARIMAX model predicted output with data set-II of thermal power plant
In this data set x(t − 1) and y(t − 1) are used as the optimal inputs and the ARIMAX model is built with
713 input-output data pairs. The actual and the ARIMAX model’s predicted output plotted vs sample
number are shown in Fig. 6.5.
(iv) Data set-IV
With x(t−1) and y(t−1) as the optimal inputs, the ARIMAX model is built with 494 input-output data
pairs. Fig. 6.6 shows the actual and the ARIMAX model predicted output plotted vs sample number.
(v) Data set-V
The ARIMAX model for this data set is built with x(t − 1) and y(t − 1) as the optimal inputs and using
320 data pairs. The output predicted by the ARIMAX model along with the actual output plotted vs
sample number is shown in Fig. 6.7.
6.3.2 GA based fuzzy model
In GA based fuzzy model, the initial fuzzy model is built in the same line as that of the ANFIS model.
After the initial fuzzy model is developed based on both the subtractive clustering and the grid based
partitioning, the updation of the consequent parameters to fine tune the fuzzy model is carried out by
using genetic algorithm. The genetic algorithm searches the entire solution space for the best fit of the
consequent parameters for the fuzzy model. The GA based fuzzy model is built for both the Box and
Jenkins gas furnace and the thermal power plant data.
0.35
actual output
predicted output
0.3
0.25
power output (GW)
0.2
0.15
0.1
0.05
0
0 500 1000 1500
sample number
Figure 6.5: Actual and ARIMAX model predicted output with data set-III of thermal power plant
0.26
actual output
predicted output
0.24
0.22
power output (GW)
0.2
0.18
0.16
0.14
0.12
0.1
0 100 200 300 400 500 600 700 800 900 1000
sample number
Figure 6.6: Actual and ARIMAX model predicted output with data set-IV of thermal power plant
7
actual output
predicted output
energy output (MU)

4
0
0 100 200 300 400 500 600 700
sample number
Figure 6.7: Actual and ARIMAX model predicted output with data set-V of thermal power plant
The different parameters that have been used in all GA based fuzzy models are as follows:-
The number of generations are restricted to a maximum of 1000, with a crossover fraction of 0.8, a
migration fraction of 0.2, migration interval of 20, stall generation limit of 100, a stall time limit of 50
seconds, migration interval of 20 and a penalty factor of 100. The fitness function in this methodology
scales the raw scores based on the rank of each individual. The rank of an individual is its position
in the sorted scores. The rank of the fittest individual is assigned a value of 1, the next most fittest is
assigned a value of 2, and so on. Rank fitness scaling removes the effect of the spread of the raw scores.
Case 1:- Box and Jenkins gas furnace data
This GA based fuzzy model is built using the 145 gas furnace data pairs, where x(t − 3) and y(t − 1)
are the two optimal inputs. The actual output and the output predicted by the GA based fuzzy model
are shown in Fig. 6.8.
Case 2:- Thermal power plant data
(i) Data set-I

The first data set of the thermal power plant comprising of 699 data pairs is used here. The two optimal
inputs are x(t − 6) and y(t − 1). The actual and the GA based fuzzy model’s predicted output plotted
vs sample number are shown in Fig. 6.9.
62
actual output
predicted output
60
58
56
54
52
50
48
46
44
0 50 100 150 200 250 300
sample number
Figure 6.8: Actual and GA based fuzzy model predicted output with Box and Jenkins gas furnace data
0.7
actual output
predicted output
0.6
0.5
power output (GW)
0.4
0.3
0.2
0.1
0
0 200 400 600 800 1000 1200 1400
sample number
Figure 6.9: Actual and GA based fuzzy model predicted output with data set-I of thermal power plant
0.7
0.6
0.5
0.4
power output (GW)

0.3
0.2
0.1
−0.1 actual output

predicted output
−0.2
0 200 400 600 800 1000 1200 1400
sample number
Figure 6.10: Actual and GA based fuzzy model predicted output with data set-II of thermal power plant
(ii) Data set-II

The second data set comprises of 675 input output data pairs. Here x(t − 3) and y(t − 1) are the optimal
inputs. Fig. 6.10 shows the actual output and the GA based fuzzy model’s predicted output plotted vs
sample number.
(iii) Data set-III
In this experiment 713 input-output data pairs are used. The optimal inputs are x(t − 1) and y(t − 1).
The output predicted by the GA based fuzzy model and the actual output are plotted vs sample number
in Fig. 6.11.
(iv) Data set-IV
Here 494 input-output data pairs are used with x(t − 1) and y(t − 1) as the optimal inputs. The actual
and the GA based fuzzy model predicted output are shown in Fig. 6.12.
(v) Data set-V
In this experiment 320 data pairs are used with x(t − 1) and y(t − 1) as the optimal inputs.The actual
and the GA based fuzzy model’s predicted output vs sample number is shown in Fig. 6.13.
The results obtained with the GA based fuzzy model for the Box and Jenkins gas furnace and the
thermal power plant data are tabulated in Table 6.2. The table shows that for the Box and Jenkins gas
furnace data set, the grid partition based method produces slightly better result whereas for the thermal
0.5
actual output
predicted output
0.45
0.4
0.35
power output (GW)
0.3
0.25
0.2
0.15
0.1
0.05
0
0 500 1000 1500
sample number
Figure 6.11: Actual and GA based fuzzy model predicted output with data set-III of thermal power
plant
0.3
actual output
predicted output
0.28
0.26
0.24
power output (GW)
0.22
0.2
0.18
0.16
0.14
0.12
0.1
0 100 200 300 400 500 600 700 800 900 1000
sample number
Figure 6.12: Actual and GA based fuzzy model predicted output with data set-IV of thermal power
plant
6.4. CONTROLLER USING FFD-V-FOLD BASED ANFIS MODEL 88
7
actual output
predicted output
energy output (MU)

4
0
0 100 200 300 400 500 600 700
sample number
Figure 6.13: Actual and GA based fuzzy model predicted output with data set-V of thermal power plant
power plant data the subtractive clustering based model shows better performance.
The proposed FFD-V-fold based ANFIS model is now compared with the GA based fuzzy model
as well as the statistical model. The performance of the proposed model is studied along with the GA
based fuzzy model and the ARIMAX model. The ARIMAX model is chosen among the three different
statistical models as it turns out to be the best among the three. The results obtained for these three
models are now arranged in a tabular form in Table 6.3. From the table it can be observed that the
proposed FFD-V-fold based ANFIS model outperforms the GA based fuzzy model as well as the ARI-
MAX model having the least RMSE even though the training data used by this proposed model is only
about one-eighth of that used by the other two models.
6.4 Controller Using FFD-V-fold based ANFIS Model

In this section a controller based on the proposed FFD-V-fold based ANFIS model is designed. Here
the ANFIS based inverse model controller is used with an open loop strategy in which the controller
is the inverse of the ANFIS based plant [79]. The inverse model of the plant [79] is obtained from the
input-output data that has been used to obtain the ANFIS based model for the system. This controller
uses the inverse learning or general learning in two different stages.
Table 6.2: Results obtained for the GA based fuzzy model using the Box and Jenkins gas furnace and
the thermal power plant data
Sl. No. Data training (Grid partition (Subtractive
Box and
1 Jenkins 145 0.5830 0.5998
gas furnace
Thermal
2 power plant 699 0.0430 0.0360
(Data set-I)
Thermal
3 power plant 675 0.0964 0.0644
(Data set-II)
Thermal
4 power plant 713 0.1509 0.0173
(Data set-III)
Thermal
5 power plant 494 0.3879 0.0174
(Data set-IV)
Thermal
6 power plant 320 0.5562 0.5399
(Data set-V)
Table 6.3: Comparison of the FFD-V-fold based ANFIS, GA based fuzzy and the statistical ARIMAX
model obtained for the Box and Jenkins gas furnace and the thermal power plant data
Sl. No. Data Model training (Grid partition (Subtractive
Box and FFD-V-fold 18 0.5378 0.5332
1 Jenkins ARIMAX 145 1.2305 1.2305
gas furnace GA based fuzzy 145 0.5830 0.5998
Thermal FFD-V-fold 81 0.0109 0.0108
2 power plant ARIMAX 699 0.0110 0.0110
(Data set-I) GA based fuzzy 699 0.0430 0.0360
3 power plant ARIMA 675 0.0163 0.0163
(Data set-II) GA based fuzzy 675 0.0964 0.0644
(Data set-III) GA based fuzzy 713 0.1509 0.0173
(Data set-IV) GA based fuzzy 494 0.3879 0.0174
(Data set-V) GA based fuzzy 320 0.5562 0.5399
• In the learning phase the inverse model of the ANFIS based plant is obtained with the available
input-output data. This is called the training phase of the inverse controller. Fig. 6.14 illustrates
this training phase.
• In the next phase the inverse model of the ANFIS based plant is used to initiate the control action.
This is called the application phase of the ANFIS based controller and it is illustrated in Fig. 6.15.
u(k) PLANT
y(k)
ANFIS MODEL
+ eu
−1
INVERSE Z
ANFIS MODEL
Figure 6.14: Training phase of the inverse controller
u(k) PLANT
r(k) CONTROLLER
ANFIS MODEL y(k)

ANFIS MODEL
−1
Z
Figure 6.15: Application phase of the inverse controller
6.4.1 Controller Design
The controller designed using the general or the off-line training method does not take into consid-
eration the aspect of minimizing the output error to ensure reference tracking. This results in a large
Controller
r(k) Inverse u1 (k)
ANFIS model
u(k)
+ ANFIS based plant y(k)
+
Fuzzy
Integrator
u (k)
2
−1
Z
Figure 6.16: Inverse controller with fuzzy integrator block
steady state error. To reduce the sustained steady state error of the plant a fuzzy integrator block is in-
corporated along with the ANFIS based controller to form the total controller block. The integral action
increases the control signal if there is a small positive error. Similarly for a negative error the integral
action decreases. The prediction error is used as an input to the fuzzy integrator which is illustrated in
Fig. 6.16. This is done to generate the integrating constant Ki so that the final steady state error can
be brought nearly to zero. The fuzzy rules for the integrator block are framed heuristically to tune the
value of Ki , as there is no systematic method for framing the fuzzy rules for building the controller. In
the present work these rules have have been framed based on the results of the simulations that were
conducted with different values of the controller inputs.
The rules for the integrator block are of the following form:
Control rule b:
IF e1 (k) is M1b and e2 (k) is M2b and . . . eN (k) is MNb THEN
P
u2 (k) = Ki kj=1 ej Ts ,
b = 1, 2, ...q and n = 1, 2, ..., N
where
Mnb is the fuzzy set (n = 1, 2, ...N ), Ts is the sampling time and error ej = r(k) − y(k),
r(k) being the reference input and y(k) is the system’s output.
For the above rule, the fuzzy integral controller system may be inferred as follows:
Given a combination of (Ki , ej , Ts ), the final output of fuzzy integrator system may given as fol-
lows:
X
q
X
k
u2 (k) = µb (e(k))Ki ej Ts , (6.4.1)
b=1 j=1
b = 1, 2, ..., q (6.4.2)
where
X
q
µb (e(k)) = wb (e(k))/ wb (e(k))
b=1
and
Y
N
wb (e(k)) = Mnb (e(k)).
n=1
Mnb (e(k)) is the degree of the membership of e(k) in Mnb . Here wb (k) ≥ 0, for b = 1, 2, ..., q and
Pq Pq
b=1 wb (k) > 0 for all k. Therefore, µb (e(k)) ≥ 0 for b = 1, 2, ..., q and b=1 µb (e(k)) = 1. The
output values of the available input-output data set range from 0.0150 − 0.219. So the fuzzy rules for
the ANFIS based controller is framed to suit this range. The reference input which is a step signal is
also chosen to suit the available data set.
6.4.2 Experimental Results and Discussion
The performance of the controller is illustrated in Figs. 6.17 - 6.24. In Fig. 6.17 the system output is
shown against the reference step input and the overshoot here is found to be 1.6%. Fig. 6.18 shows
the prediction error of the controller. This error is the difference between the desired output and the
actual output of the plant. In Fig. 6.19 the control signal i.e the output of the controller is shown. In
6.5. CONCLUSION 93
0.22
Desired
Actual
0.2
0.18
System output
0.16
0.14
0.12
0.1
0 500 1000 1500
Samples
Figure 6.17: Desired output and the actual output of the plant with the controller
order to check the performance of the controller in presence of disturbance, two separate noises (step
inputs) were injected at around the 400th and 1000th sample. The result in Fig. 6.20 demonstrates the
fast recovery of the controlled system to its final steady state after being disturbed with noise.
In order to check the performance of the controller under variable plant dynamics, a random varia-
tion of the plant parameters by 10% is allowed. The results under the changing plant dynamics are
illustrated in Figs. 6.21 - 6.24. The controller’s performance under variable plant dynamics along with
disturbance is plotted in Fig. 6.24.
6.5 Conclusion
It is observed that the FFD-Vfold based ANFIS controller can track the reference step input quite
faithfully even under plant parameter variation alongwith disturbance.
6.5. CONCLUSION 94
0.1
0.08
0.06
Prediction Error
0.04
0.02
−0.02
0 500 1000 1500
Samples
Figure 6.18: Prediction error of the controller
52
50
48
Controller Output
46
44
42
40
38
36
0 500 1000 1500
Samples
Figure 6.19: Output of the controller

6.5. CONCLUSION 95
0.22
Desired
Actual
0.2
0.18
System output
0.16
0.14
0.12
0.1
0 500 1000 1500
Samples
Figure 6.20: Desired and actual output of the plant with controller in presence of disturbance
0.22
Desired
Actual
0.2
0.18
System output
0.16
0.14
0.12
0.1
0 500 1000 1500
Samples
Figure 6.21: Desired and actual output of the plant with controller having plant parameter variations
6.5. CONCLUSION 96
0.1
0.08
0.06
Prediction Error
0.04
0.02
−0.02
0 500 1000 1500
Samples
Figure 6.22: Prediction error of the controller with plant parameter variations
52
50
48
Controller Output
46
44
42
40
38
36
0 500 1000 1500
Samples
Figure 6.23: Output of the controller with plant parameter variations

6.5. CONCLUSION 97
0.22
Desired
Actual
0.2
0.18
System output
0.16
0.14
0.12
0.1
0 500 1000 1500
Samples
Figure 6.24: Desired and actual output of the controller based plant with plant parameter variations
along with disturbance
Chapter 7
Conclusions and Future Work
In this concluding chapter the salient contributions of the thesis are summarized. Also a few aspects
which may be explored for further research are outlined here.
7.1 Summary of the thesis

The major contribution in this thesis is towards newer development in the ANFIS methodology. The
main emphasis is to build an ANFIS model based on an optimally selected data set. The advantage of
the proposed method has been demonstrated in terms of improved modeling results as compared to the
conventional ANFIS model as well as conventional statistical models. The proposed method is used to
design a controller and study its performance.
In Chapter 1 the aim of the present research work is discussed. Chapter 2 presents a brief overview of
different system identification techniques based on soft computing method. Here some important fore-
casting techniques are also discussed briefly. Chapter 3 outlines the research objectives and method-
ology followed in the present work. In Chapter 4 the ANFIS model which forms the backbone of the
proposed method for system identification is explained in details. Chapter 5 describes the two differ-
ent techniques namely the Full factorial design and the V-fold method which are used to augment the
ANFIS model for optimal selection of the training data. In Chapter 6 the results obtained by using
the proposed model are compared with the GA based fuzzy model and statistical models. Chapter 7
concludes the thesis with a discussion on the important findings of the research work.
The major contributions of the thesis are summarized below:
• The statistical Full factorial design technique is used to choose an optimal data set for training
the neural network in the conventional ANFIS model. The available data is selected optimally
98
7.2. DISCUSSION 99
for training the subtractive clustering and grid partition technique based ANFIS models.
• The V-fold technique which is used for cross validation purpose is applied for selecting the opti-
mal data for training. By using this technique, it has been possible to build an ANFIS model using
a training data set which is around one-eighth of the size of the data set used in the conventional
ANFIS model.
• The full factorial design technique and the V-fold method are combined for optimal data selection
in training the ANFIS model. The obtained results show that performance of the combined FFD-
V-fold based ANFIS model is better than the individual FFD based and V-fold based ANFIS
models.
• A controller is designed using the proposed FFD-V-fold based ANFIS model which performs
faithfully even with plant parameter variations and disturbance.
7.2 Discussion
The prime advantage of the soft computing tools which have been used in the present work is that the
requirement of a mathematical model is not a prerequisite. The model of the system under consider-
ation can be built around the available input-output data of the system. The inspiring features of the
neural network in its ability to learn and adapt and the capability of a fuzzy system to take into account
the imprecision and prevailing uncertainty are vouched to be the prime tools in the present research
work. To exploit the advantages that are associated with the respective methods, the Adaptive network
based fuzzy inference system (ANFIS) model comprising of both the neural network and the fuzzy
logic has been used in the present work. ANFIS model requires a sufficient pool of input-output data
for training. As the absence of sufficient number of input-output data in many a real world systems
poses a major hurdle in modeling, an endeavor is made to use the available data, to its best possible
ability such that a small critical data set can be selected optimally and used for training the ANFIS
model. For this purpose, two novel techniques namely Full factorial design and V-fold cross validation
technique are proposed for optimal data selection to train the ANFIS.
The experimental results obtained by applying the proposed method are encouraging enough. The pro-
posed model shows reliable prediction performance so far as the actual values of the available data
are concerned. The performance of the proposed model was also found to be at par with conventional
statistical models used in forecasting. The comparable results were achieved even though a very small
7.3. CONCLUSION 100
critically selected training data set was used in training the ANFIS. The strength of the input output
data used is only around one-eighth of the total number of the available data set.
The results obtained from the proposed model can be very encouraging to industries associated with
the real time systems, like a thermal power plant that has been considered in the present work. These
results are expected to give a boost to the economic front as precious fossil fuel whose reserves are
depleting day by day can be saved as the power plant need not be run for long hours to generate the
modeling data for future load prediction. Besides this the time required for starting the thermal power
plant before it can be fully loaded can be saved as it takes a pretty long time to start real world systems
like the thermal power plant. Therefore in terms of human resources management also the results can
prove to be a boon as a great many number of man hours can be saved, as the plant does not have to be
run for long periods of time just to generate the modeling data.
So far as the implementation of the present findings are concerned, it is not expected to be too dif-
ficult a proposition, as the soft computing model can be connected in parallel to the hard computing
and manual control system that is already there in place. Even in the remote possible case of the pro-
posed model failing, the hard computing based tools or the manual controller will always be there to
act as a backup protection system.
The results with the proposed model are, however bounded with certain limitations. This is because
the modeling results are fairly accurate with respect to the choice of the critical training data set. In the
FFD based ANFIS model, the levels selected in the available training data set are crucial. The proper
selection of the FFD levels are important as the ANFIS model is decided based upon the optimal choice
of these levels. Rigorous experiments need to be performed to determine the best FFD levels to fit into
the optimal ANFIS model. Hence it requires time and the model cannot be predetermined. Another
limitation of the model is the possible presence of outliers and noise in the data set that is used for the
system modeling. The cross checking for a possible model which is a hybrid between an evolutionary
methodology and the proposed model also could have produced a more better result.
7.3 Conclusion
System identification is the primary step in modeling of a system as its model should correctly repre-
sent the system for further analysis and control. A practical important utility of modeling is prediction
7.4. FUTURE WORK 101
of future values. This aspect of forecasting is even more significant in real world systems like thermal
power plants. The research work carried out in this thesis studied modeling of a real world system
like gas furnace and thermal power plant based on actual input output data for which a soft computing
methodology ANFIS is used. This methodology has been considered to make the best use of the inher-
ent salient features of the neural network of self adaptability and that of the fuzzy logic in its ability
to take into account the imprecision and uncertainty of systems which are considered for modeling. It
has been a traditionally accepted fact that for successful training of ANFIS, the number of data used
should be sufficiently large. However, it may be difficult to get a large number of data in those systems
like thermal power plant where generation of input output data is quite expensive as precious fuel is
involved. Hence, this research work attempted to build a model for such a system using the existing
small pool of input output data. For this purpose, the emphasis was on optimal selection and use of
the available resource of data. It has been observed from experimental results that it is also possible
to build a successful model for those types of systems where the available data set is scanty. This
thesis proposed two novel techniques for optimal selection of data and applied the same in the model-
ing methodology. Performance of the model built using these two proposed techniques was tested for
prediction and the results were compared with traditional statistical models commonly used in litera-
ture. The results obtained are promising and encouraging enough deserving further investigation by the
concerned industry. The experimental results obtained verify the competence of the proposed model-
ing technique. Hence conclusions can be drawn that the proposed ANFIS based system identification
technique augmented by optimal data selection strategy for training is comparable with other existing
conventional techniques of modeling.
There is a scope for further exploring the possibility of bettering the present modeling results by
taking into account the possible presence of outliers in the model’s training data. The scope of strength-
ening the proposed ANFIS model by creating its hybrid with another evolutionary technique is another
direction which might be worth looking into.
7.4 Future Work

Following are the potential directions which can be explored for future research:
• Developing a proper methodology for framing and determining the number of fuzzy rules for the
proposed ANFIS model
7.4. FUTURE WORK 102
• Preprocessing the data before modelling is carried out so as to remove the outliers
• Exploring the possibility of developing the ANFIS model strengthened by other evolutionary
algorithms
• Developing a formal methodology to decide an optimal data size for the V-folds so that modeling
can be carried out with the least possible minimum number of data
Appendix A
Sample Power Plant Data
One sample data set that was collected from the thermal power plant under North East Electric Power
Corporation (NEEPCO) Limited, located in Kathalguri, Assam, India has been presented here. This set
of the input output data consists of an input to the thermal power plant in the form of the gas feed to the
plant in Million metric standard cubic metre (Mmscum)(106 m3 ). Similarly the output from the thermal
power plant is in the form of generated energy in Million unit (MU), where 1 unit = 1 kWh (kilo-watt-
hour). The thermal power plant data set comprising of the inputs and the outputs are collected from the
hard copy of the records maintained by the power plant authorities. This data set has been shown in
tabular form in Table A.1 which spreads over the next six successive pages.
103
104
Table A.1: Thermal power plant data

Sample Input Output Sample Input Output Sample Input Output
No (Mmscum) (MU) No (Mmscum) (MU) No (Mmscum) (MU)
1 1.324 4.9452 41 0.574 1.3966 81 1.32 5.6243
2 1.322 4.6551 42 0.618 1.3997 82 1.4534 5.6538
3 1.313 4.8245 43 0.673 1.5822 83 1.4387 5.6143
4 1.191 4.5939 44 0.73 1.7254 84 1.4749 5.6229
5 1.188 4.2019 45 0.598 1.4773 85 1.4641 5.6235
6 1.241 4.5838 46 0.455 1.023 86 1.3119 5.3734
7 1.28 4.8133 47 0.497 1.1173 87 1.3136 5.325
8 1.25 4.973 48 0.529 1.2861 88 1.319 5.6387
9 1.153 3.9096 49 0.447 0.986 89 1.3993 5.2355
10 1.232 3.9021 50 0.374 0.9672 90 1.4414 5.3554
11 1.129 3.9072 51 0.437 0.8112 91 1.4045 5.5192
12 1.159 3.9627 52 0.412 0.8423 92 1.3778 5.4554
13 1.169 3.8762 53 0.421 0.838 93 1.4796 5.6182
14 1.157 3.9246 54 0.325 0.7497 94 1.4598 5.5868
15 1.083 3.9547 55 0.283 0.7372 95 1.4679 5.4311
16 1.025 3.0296 56 0.372 0.7779 96 1.4582 5.6565
17 0.922 2.4153 57 0.379 0.8284 97 1.4408 5.4803
18 0.819 2.7417 58 0.382 0.7776 98 1.3682 5.5866
19 0.77 2.2473 59 0.378 0.8381 99 1.3574 5.4574
20 0.678 1.9082 60 0.257 0.6241 100 1.3831 5.4638
21 0.804 2.4771 61 0.262 0.6413 101 1.3372 5.4873
22 0.724 2.4917 62 1.402 5.5288 102 1.3671 5.2863
23 0.674 2.4805 63 1.4491 5.3996 103 1.4481 5.4758
24 0.598 2.0008 64 1.4677 5.5825 104 1.4554 5.3959
25 0.439 1.1264 65 1.4577 5.582 105 1.3766 5.5891
26 0.556 1.9277 66 1.4161 5.468 106 1.3897 5.5449
27 0.485 1.4958 67 1.3967 5.6255 107 1.433 5.5731
28 0.493 1.4993 68 1.4723 4.9201 108 1.3962 5.5036
29 0.515 1.461 69 1.4039 5.2882 109 1.2972 4.436
30 0.627 2.2222 70 1.4499 5.5317 110 1.3581 5.466
31 0.66 2.3257 71 1.3176 5.5103 111 1.4581 5.6001
32 0.521 1.9292 72 1.4471 5.6107 112 1.3911 5.4636
33 0.663 2.1237 73 1.4267 5.591 113 1.3568 5.2911
34 0.461 1.4513 74 1.4003 5.5735 114 1.4164 5.3258
35 0.401 0.7558 75 1.4198 5.3997 115 1.4649 5.5799
36 0.422 0.9432 76 1.2299 5.6397 116 1.4609 5.6018
37 0.404 0.7583 77 1.2666 5.6239 117 1.391 5.6248
38 0.54 1.2023 78 1.2331 5.4488 118 1.3226 4.9621
39 0.494 1.181 79 1.0844 4.5987 119 1.4091 5.7229
40 0.559 1.3888 80 1.3462 5.65 120 1.4565 5.6937
105

No Mmscum MU No Mmscum MU No Mmscum MU
121 1.4046 5.6481 161 1.4514 5.6148 201 1.0818 3.7636
122 1.4227 5.7434 162 1.4216 5.6523 202 0.9669 2.885
123 1.4581 5.7434 163 1.4487 5.5285 203 0.794 2.3875
124 1.4078 5.7161 164 1.3801 5.5787 204 0.638 2.0173
125 1.4615 5.643 165 1.3978 5.6097 205 0.7455 2.3181
126 1.5922 5.6752 166 1.371 5.5655 206 0.6064 1.624
127 1.4414 5.6831 167 1.4916 5.5533 207 0.5539 1.6297
128 1.5922 5.7081 168 1.4093 5.4241 208 0.5714 1.3992
129 1.4079 5.5961 169 1.4812 5.3844 209 0.8361 1.9668
130 1.4641 5.6839 170 1.4714 5.3092 210 0.9027 3.2116
131 1.4744 5.5765 171 1.4681 5.6352 211 0.9078 2.4865
132 1.4162 5.6774 172 1.4598 5.6986 212 0.8212 2.5966
133 1.3626 5.0974 173 1.491 5.835 213 0.9907 3.4888
134 1.3944 5.1586 174 1.4121 5.47 214 1.0112 3.0417
135 1.4749 5.4782 175 1.3523 5.5049 215 1.1095 3.2544
136 1.3827 4.7539 176 1.4041 5.1959 216 0.8682 2.7095
137 1.4534 5.5895 177 1.4264 5.4507 217 0.801 2.4917
138 1.4126 5.4194 178 1.5057 5.7048 218 0.787 2.2152
139 1.4518 5.5629 179 1.4833 6.0147 219 0.7407 2.3094
140 1.4009 5.3345 180 1.576 6.0291 220 0.7879 2.0827
141 1.4354 5.1913 181 1.4829 5.8037 221 0.7294 1.9713
142 1.4416 5.5908 182 1.4271 5.4411 222 0.9181 2.4715
143 1.4146 5.6031 183 1.3938 4.941 223 0.9833 2.9014
144 1.4078 5.7035 184 1.3375 4.4442 224 1.027 3.15
145 1.3782 5.5866 185 1.3653 4.8132 225 0.8044 2.922
146 1.4371 5.5992 186 1.3583 4.9125 226 0.7579 2.0179
147 1.4322 4.9094 187 1.5534 4.8933 227 0.7683 1.7303
148 1.4076 5.4534 188 1.4916 5.0371 228 0.7597 1.9127
149 1.3406 5.5334 189 1.1866 3.8338 229 0.7235 1.7416
150 1.3542 5.4973 190 1.3944 5.0443 230 0.8203 2.1081
151 1.4172 5.6085 191 1.4521 5.1634 231 0.7821 2.3469
152 1.3743 5.4574 192 1.2189 4.7144 232 0.7907 2.103
153 1.4078 5.6656 193 1.1288 3.3662 233 0.6632 1.553
154 1.4045 5.603 194 1.2875 3.6034 234 0.7855 1.7264
155 1.3844 5.3959 195 1.4923 5.1222 235 0.925 2.2885
156 1.3961 5.6544 196 1.3559 4.9382 236 1.0508 2.808
157 1.4146 5.685 197 1.1572 4.0445 237 1.2034 3.5663
158 1.4075 5.7383 198 1.2138 3.9089 238 1.2067 4.0265
159 1.3173 4.9094 199 1.2549 3.8814 239 1.0577 2.6875
160 1.4598 5.641 200 1.3406 4.8714 240 0.8854 2.2007
106

241 1.0559 2.8302 281 1.3064 4.5745 321 1.5001 5.3927
242 0.9588 2.6181 282 1.3609 5.0043 322 1.4315 5.6101
243 0.8492 2.2116 283 1.4632 5.4324 323 1.5087 5.5803
244 1.4039 3.7935 284 1.4497 5.4443 324 1.5447 5.3973
245 1.3961 4.664 285 1.4179 5.1507 325 1.6115 5.6272
246 1.335 4.7778 286 1.4229 5.3323 326 1.5841 5.6337
247 1.3264 4.5189 287 1.3609 5.1421 327 1.5978 5.7309
248 1.1766 4.2765 288 1.4607 4.7898 328 1.5532 5.59
249 1.2671 4.0429 289 1.3852 5.1154 329 1.5687 5.6709
250 1.1706 4.2399 290 1.2498 4.9163 330 1.5978 5.4148
251 1.1662 3.7326 291 0.9913 5.0368 331 1.6269 5.5838
252 1.1438 3.5132 292 1.324 3.2511 332 1.5532 5.7224
253 1.1524 3.5352 293 1.4281 5.4863 333 1.492 5.625
254 1.1126 2.985 294 1.4347 5.2931 334 1.5346 5.7142
255 1.3166 4.4993 295 1.4495 5.3268 335 1.5322 5.6245
256 1.2018 4.547 296 1.4675 5.1976 336 1.4889 5.5269
257 1.1092 3.513 297 1.4658 5.3189 337 1.519 5.6984
258 1.0252 3.0348 298 1.4624 5.3774 338 1.4577 5.609
259 1.2858 4.2664 299 1.4864 5.4225 339 1.5104 5.6215
260 1.2138 3.6258 300 1.4607 5.4206 340 1.3603 5.6538
261 1.2084 3.6389 301 1.4281 5.3925 341 1.4884 5.6918
262 1.2926 4.229 302 0.1338 5.0432 342 1.5427 5.6148
263 1.371 4.0629 303 1.4401 5.0623 343 1.5233 5.512
264 1.3475 4.6036 304 1.5134 5.433 344 1.5389 5.5724
265 1.3458 4.8307 305 1.3869 5.3003 345 1.4951 5.5488
266 1.13 3.6346 306 1.5961 5.4927 346 1.5261 5.5808
267 1.3928 4.9537 307 1.4665 5.6163 347 1.6153 5.5724
268 1.1761 4.6311 308 1.4534 5.7034 348 1.5516 5.6513
269 1.0525 2.8417 309 1.4845 5.5608 349 1.6195 5.5318
270 1.2419 3.9535 310 1.5369 5.7098 350 1.5313 5.6522
271 1.3408 4.6652 311 1.5789 6.0245 351 1.5009 5.606
272 1.4916 5.0152 312 1.563 5.8668 352 1.476 5.6187
273 1.4833 5.1074 313 1.46 5.2192 353 1.5786 5.6733
274 1.381 4.7367 314 1.4142 5.3613 354 1.4831 5.5214
275 1.3207 4.7058 315 1.3733 5.4151 355 1.5199 5.5845
276 1.3425 4.5717 316 1.5344 5.739 356 1.5922 5.4624
277 1.49 4.9783 317 1.5447 5.7368 357 1.5887 5.6533
278 1.4665 5.3421 318 1.6269 5.9413 358 1.4968 5.6832
279 1.4146 5.2614 319 1.5978 5.8773 359 1.6589 5.5447
280 1.4028 5.1238 320 1.5838 5.8336 360 1.6763 5.4693
107

361 1.5254 5.9156 401 1.5102 5.6807 441 1.5214 5.2991
362 1.5685 5.8613 402 1.5461 5.6802 442 1.516 5.2847
363 1.593 5.6776 403 1.5919 5.6983 443 1.561 5.3937
364 1.57 5.6831 404 1.4594 5.4908 444 1.4002 4.3372
365 1.6136 5.5962 405 1.4795 5.5897 445 1.4143 5.1852
366 1.6965 5.7247 406 1.381 5.6056 446 1.5418 4.9834
367 1.7126 5.7391 407 1.4862 5.6184 447 1.4326 5.0081
368 1.6633 5.761 408 1.509 5.6667 448 1.4615 4.8435
369 1.6589 5.8093 409 1.5133 5.7054 449 1.5218 5.0014
370 1.5945 5.8111 410 1.5168 5.7218 450 1.4862 5.2092
371 1.5198 5.8198 411 1.4689 5.7706 451 1.5232 5.2754
372 1.6374 5.8631 412 1.1808 3.9978 452 1.5003 5.3092
373 1.5735 5.7124 413 0.916 4.0574 453 1.437 5.198
374 1.5655 5.31 414 1.4765 5.3341 454 1.5179 5.5755
375 1.4845 5.4675 415 1.4656 5.3112 455 1.2531 4.8364
376 1.4816 5.6235 416 1.4932 5.5365 456 1.418 5.2137
377 1.5458 5.6525 417 1.4135 5.5452 457 1.426 5.4378
378 1.5226 5.7043 418 1.4613 5.5509 458 1.4 5.2222
379 1.5473 5.6844 419 1.46 5.5727 459 1.3889 4.9787
380 1.4173 5.6001 420 1.4659 5.5711 460 1.4028 4.6367
381 1.5267 5.6298 421 1.4925 5.5565 461 1.3546 4.6151
382 1.5938 5.6656 422 1.4943 5.5805 462 1.4103 4.1426
383 1.7443 5.8414 423 1.4781 5.4996 463 1.4563 4.7997
384 1.5018 5.7291 424 1.513 5.6461 464 1.4252 4.737
385 1.5968 5.7892 425 1.7171 5.6773 465 1.4413 4.7756
386 1.5377 5.4932 426 1.5543 5.7158 466 1.4349 4.9599
387 1.4734 5.7086 427 1.5696 5.7249 467 1.1952 4.1722
388 1.5027 5.6201 428 1.5681 5.6764 468 1.2605 3.94
389 1.5688 5.4349 429 1.5774 5.6057 469 1.4092 4.8165
390 1.5029 5.3902 430 1.5659 5.6933 470 1.4017 4.4719
391 1.5298 5.7404 431 1.5867 5.6658 471 1.3182 4.0426
392 1.4183 5.668 432 1.5178 5.612 472 0.9063 2.8749
393 1.4655 5.5867 433 1.565 5.4375 473 1.1171 3.4271
394 1.4751 5.5357 434 1.5505 5.3128 474 1.1321 2.2888
395 1.4581 5.5635 435 1.5014 5.1009 475 1.0871 3.2424
396 1.4222 5.4617 436 1.542 5.3041 476 1.1246 2.8123
397 1.4619 5.5473 437 1.659 5.6067 477 1.3792 4.1522
398 1.5727 5.5732 438 1.5997 5.6964 478 1.4178 4.3501
399 1.5699 5.619 439 1.5175 5.5347 479 1.3942 4.246
400 1.4833 5.6923 440 1.4492 5.2899 480 1.3717 4.2139
108

481 1.4081 4.2123 521 1.2348 4.424 561 1.2947 4.8076
482 1.3921 4.245 522 1.284 4.7181 562 1.1984 4.2725
483 1.5194 4.4006 523 1.2145 4.3307 563 1.2808 4.7571
484 1.5964 4.7967 524 1.3225 4.899 564 1.2851 4.621
485 1.6157 5.1056 525 1.3717 5.2216 565 1.2829 4.6642
486 1.6264 5.0966 526 1.1717 4.8175 566 1.1513 3.7889
487 1.636 5.2014 527 1.0005 2.6721 567 1.3332 4.718
488 1.2776 4.5022 528 1.1503 4.0102 568 1.314 4.7808
489 1.3075 4.3023 529 1.0465 3.0405 569 1.3215 4.8193
490 1.5868 5.4289 530 1.1428 4.3035 570 1.3471 5.0174
491 1.5376 5.5534 531 1.1866 3.9606 571 1.2305 4.2833
492 1.4691 5.2084 532 1.1588 3.6403 572 1.3225 4.7482
493 1.636 5.8577 533 1.1995 3.9429 573 1.2348 4.3296
494 1.2904 4.5422 534 1.3364 4.8332 574 1.3215 4.8044
495 1.6382 4.6743 535 1.3621 5.0755 575 1.2177 4.1353
496 1.6478 5.4469 536 1.284 4.9141 576 1.3075 4.8193
497 1.498 4.9146 537 1.1781 3.8019 577 1.3086 4.8132
498 1.6489 5.9131 538 1.2016 3.9578 578 1.3182 4.9313
499 1.6521 5.9033 539 1.208 4.1534 579 1.2872 4.7966
500 1.6724 5.7017 540 1.2284 4.2113 580 1.3097 4.7426
501 1.6628 5.6355 541 1.284 4.5927 581 1.3182 4.8939
502 1.6296 5.4745 542 1.3215 4.7669 582 1.3161 4.8427
503 1.6553 5.54 543 1.3429 4.9341 583 1.2284 4.368
504 1.6488 5.6014 544 1.3204 4.8847 584 1.2626 4.4404
505 1.6628 5.5977 545 1.3536 5.1006 585 1.2808 4.5692
506 1.6703 6.0881 546 1.3022 4.921 586 1.3033 4.5241
507 1.6125 5.8677 547 1.3493 4.868 587 1.3022 4.5065
508 1.4413 4.6036 548 1.3696 5.0596 588 1.2337 4.3886
509 1.4413 5.3241 549 1.3482 5.0562 589 1.3033 4.6095
510 1.376 4.4069 550 1.3493 5.0899 590 1.3193 4.8156
511 1.452 4.6676 551 1.2915 4.8116 591 1.2551 4.4135
512 1.3354 5.8898 552 1.2647 4.6472 592 1.2819 4.1596
513 1.3664 3.526 553 1.2808 4.7844 593 1.3707 4.8925
514 1.5494 5.7493 554 0.7008 3.4926 594 1.2391 4.4775
515 1.4028 4.8649 555 0.5339 0.5415 595 1.2273 4.2588
516 1.3418 4.608 556 1.2861 4.7351 596 1.2883 3.8693
517 1.3546 4.5833 557 1.2669 4.3244 597 1.36 4.9252
518 1.2787 4.257 558 1.2883 4.534 598 1.3311 4.608
519 1.3503 4.9746 559 1.2487 4.417 599 1.314 4.7617
520 1.3193 4.9995 560 1.3065 4.8166 600 1.3161 4.7271
109

601 1.3097 4.5577 617 1.3022 4.8304 633 1.2979 4.8111
602 1.3621 5.0799 618 1.3643 5.1449 634 1.1909 4.4569
603 1.3108 4.7921 619 1.3568 5.0312 635 1.1952 4.291
604 1.2851 4.5264 620 1.3974 5.2211 636 1.1609 4.2125
605 1.2337 4.3612 621 1.4049 5.2859 637 1.1214 4.2567
606 1.3536 4.9453 622 1.3075 4.777 638 1.1224 4.2445
607 1.3728 5.0658 623 1.2594 4.5762 639 1.1695 4.2372
608 1.3652 4.2713 624 1.2829 4.413 640 1.3461 4.9516
609 1.3803 4.602 625 1.4017 5.3109 641 1.3461 5.1956
610 1.3964 5.4194 626 1.2637 4.4816 642 1.391 5.3655
611 1.2198 4.2685 627 1.1224 4.3469 643 1.4049 5.4376
612 1.33 4.697 628 1.2155 4.2982 644 1.4606 5.5094
613 1.3525 5.0446 629 1.193 4.2245 645 1.4017 5.4505
614 1.2005 4.2995 630 1.0903 4.0368 646 1.3386 5.4017
615 1.3311 4.8712 631 1.0411 3.8126
616 1.3225 4.8863 632 1.1984 4.0609
Bibliography
[1] M. Jamshidi, Large-Scale Systems : Modelling, Control and Fuzzy Logic. Prentice Hall PTR,
1997.
[2] A. Kamiya, S. J. Ovaska, and R. Roy, “Fusion of soft computing and hard computing for large-
scale plants : a general model,” J. Applied Soft Computing, vol. 5, pp. 265–279, 2005.
[3] V. Cherkassky, Fuzzy Inference Systems : A Critical Review, Computational Intelligence: Soft
Computing and Fuzzy-Neuro Integration with Applications, O. Kayak and L. A. Zadeh, Eds.
Springer, Berlin, 1998.
[4] Y. Dote and S. J. Ovaska, “Industrial applications of soft computing : a review,” in Proc. of the
IEEE, 2001, pp. 1243–1265.
[5] W. S. McCulloch and W. Pitts, “A logical calculus of the ideas imminent in nervous activity,”
Bull. of Math. Biophys., vol. 5, pp. 115–133, 1943.
[6] D. O. Hebb, The Organization of Behaviours. John Wiley & Sons, New York, 1949.
[7] F. Rosenblatt, Principles of Neuro-Dynamics: Perceptrons and the theory of Brain Mechanisms.
New York: Spartan, 1962.
[8] B. Widrow and M. E. Hoff, “Adaptive switching circuits,” IRE WESTCON Convention Record,
vol. 4, pp. 96–104, 1960.
[9] M. L. Minsky and S. Papert, Perceptrons. Cambridge, MA:M.I.T. Press, 1969.
[10] J. J. Hopfield, “Neural networks and physical systems with emergent collective computational
abilities,” in Proc. Nat. Acad. Sci., USA, vol. 79, pp. 2554–2558, 1982.
[11] D. B. Parker, “Learning logic,” Massachusetts Institute of Technology, Centre for Computational
Research in Economics and Management Sci., Cambridge, MA., Tech. Rep. TR-47, 1985.
110
BIBLIOGRAPHY 111
[12] D. E. Rumelhart and J. L. McClelland, Parallel Distributed Processing: Exploration in the Mi-
crostructure of Cognition. Foundations MIT Press, Cambridge MA, 1986, vol. 1.
[13] G. A. Carpenter and S. Grossberg, “Art2 self organisation of stable category recognition codes
for analog input patterns,” Applied Optics, vol. 26, pp. 4919–4930, 1987.
[14] T. Kohonen, Self Organisation and Associative Memory, 3rd ed. Springer Verlag, Heidelberg,
1989.
[15] K. S. Narendra and K. Parthasarathy, “Identification and control of dynamic system using neural
networks,” IEEE Trans. on Neural Netw., vol. 1, no. 1, pp. 4–27, March 1990.
[16] M. M. Polycarpou and P. A. Ioannou, “Identification and control of nonlinear systems using
neural network models: Design and stability analysis,” University of Southern Cal. Los Angeles,
Tech. Rep. 91-09-01, September 1991.
[17] R. M. Sanner and J. J. E. Slotine, “Gaussian networks for direct adaptive control,” in Proc.
American Control Conf., ACC91, 1991, pp. 2153–2159.
[18] D. T. Pham and X. Liu, “Identification of linear and nonlinear dynamic systems using recurrent
neural networks,” Artificial Intell. in Engineering, vol. 8, pp. 67–75, 1993.
[19] E. B. Kosmatopoulos, P. Ioannou, and M. A. Christodoulou, “Identification of nonlinear sys-

tems using dynammic neural network structures,” in Proc. 31st. Conf. on Decision and Control,
Tucson, Arizona, December 1992, pp. 20–25.
[20] J. J. Hopfield, “Neurons with graded response have collective computational properties like those
of two state neurons,” in Proc. Nat. Acad. Sci., vol. 81, 1984, pp. 3088–3092.
[21] M. A. Cohen and S. Grosberg, “Absolute stability of global pattern formation and parallel mem-
ory storage by competitive neural networks,” IEEE Trans. Syst., Man, Cybern., vol. SMC 13, pp.
815–826, 1983.
[22] B. Hong, X. Yun, and C. Xinkuo, “Generalized fuzzy RBF networks and nonlinear system iden-
tifications,” in Proc. of the 4th World Congress on Intelligent Control and Automation, Shanghai,
China, June 10–14 2002, pp. 2508–2512.
BIBLIOGRAPHY 112
[23] S. M. Ahmad, M. H. Shaheed, A. J. Chipperfield, and M. O. Tokhi, “Nonlinear modelling of a

twin rotor mimo system using radial basis function netw.” in Proc. of the IEEE Nat. Aerospace
and Electronics Conf. (NAECON), 2000, pp. 313–320.
[24] R. R. Selmic and F. L. Lewis, “Multi model neural networks identification and failure detec-
tion of nonlinear systems,” in Proc. of the 40th IEEE Conf. on Decision and Control, Orlando,
Florida USA, Dec 2001, pp. 3128–3133.
[25] F. Azam and H. F. Vanlandingham, “An alternate radial basis function neural network model,” in
Proc. of the IEEE Int. Conf. on Syst., Man and Cybern., 2000, pp. 2679–2684.
[26] R. Grino, G. Cembrano, and C. Torras, “Nonlinear system identification using additive dynamic
neural networks- two on-line approaches,” IEEE Trans. Circuits and Syst.-Part 1: Fundamental
Theory and Appl., vol. 47, no. 2, pp. 150–165, February 2000.
[27] S. Haykin, Neural Networks - A Comprehensive Foundation, 4th ed. Pearson Education (Sin-
gapore) Pvt. Ltd., Indian Branch, 2003.
[28] J. M. Zurada, Introduction to Artificial Neural Systems. Jaico Publishing House 121, Mumbai,
1999.
[29] M. T. Hagan, H. B. Demuth, and M. H. Beale, Neural Network Design, 2nd ed. PWS Publishing,
Boston, MA, USA, 1996.
[30] Q. Song, L. Yin, and Y. C. Soh, “Robust adaptive identification of nonlinear system using neural
network,” in Proc. IEEE Signal Processing Society Workshop, 2000, pp. 95–104.
[31] J. C. Patra, A. C. Kot, and Y. Q. Chen, “Chebyschev functional link artificial neural networks for
nonlinear dynamic system identification,” in Proc. IEEE Conf. on Syst., Man and Cybern., Oct
2000, pp. 2655–2660.
[32] J. C. Patra and A. C. Kot, “Nonlinear dynamic system identification using Chebyshev functional
link artificial neural networks,” IEEE Trans. Syst., Man and Cybern., vol. 32, no. 4, pp. 505–511,
Aug. 2002.
[33] X. M. Ren, A. B. Rad, P. T. Chan, and W. L. Lo, “Identification and control of continuous-time
nonlinear systems via dynamic neural networks,” IEEE Trans. Industrial Electron., vol. 50, no. 3,
pp. 478–486, June 2003.
BIBLIOGRAPHY 113
[34] M. Sugeno and G. T. Kang, “Structure identification of fuzzy model,” Fuzzy Sets and Syst.,
vol. 28, pp. 15–33, 1988.
[35] T. Takagi and M. Sugeno, “Fuzzy identification of systems and its applications to modeling and
control,” IEEE Trans. Syst., Man and Cybern., vol. 15, pp. 116–132, 1985.
[36] L. A. Zadeh, “Fuzzy sets,” Inform. and Control, vol. 8, pp. 338–353, 1965.
[37] ——, “Outline of a new approach to the analysis of complex systems and decision process,”
IEEE Trans. Syst., Man and Cybern., vol. 3, pp. 28–44, 1973.
[38] R. M. Tong, “Analysis and control of fuzzy systems using finite discrete relations,” Int. J. Con-
trol, vol. 27, pp. 431–440, 1978.
[39] E. Czogala and W. Pedrycz, “On identification in fuzzy systems and its applications in control
problems,” Fuzzy Sets Syst., vol. 6, pp. 73–83, 1981.
[40] W. Pedrycz, “An identification algorithm in fuzzy relational systems,” Fuzzy Sets and Syst.,
vol. 13, pp. 153–167, 1984.
[41] E. H. Mamdani, “Applications of fuzzy algorithms for control of simple dynamic plants,” Proc.
IEE., vol. 121, pp. 1585–1588, 1974.
[42] T. T. Lie, G. B. Shrestha, and A. Ghosh, “Design and application of a fuzzy logic control scheme
for transient stability enhancement in power systems,” Electric Power Syst. Research, vol. 33,
pp. 17–23, 1995.
[43] P. S. Babu, A. Ghosh, and Sachchidanand, “An optimal fuzzy PI controller for a permanent
magnet dc motor,” Proc. Int. Conf. Automation IC-AUTO’95, vol. 121, pp. 713–716, 1995.
[44] T. J. Procyk, “A linguistic self organizing process controller,” Automatica, vol. 15, pp. 15–30,
1979.
[45] Y. M. Park, U. C. Moon, and K. Y. Lee, “A self organising fuzzy logic controller for dynamic
systems using a fuzzy auto regressive moving average model,” IEEE Trans. Fuzzy Syst., vol. 3,
pp. 75–82, 1995.
[46] B. P. Graham and R. P. Newell, “Fuzzy identification and control of liquid level rig,” Fuzzy Sets
and Syst., vol. 26, pp. 253–273, 1983.
BIBLIOGRAPHY 114
[47] E. Sanchez, “Resolution of composite relation equations,” Inf. and Control, vol. 30, pp. 38–48,
1976.
[48] W. Pedrycz, “Numerical and applicational aspects of fuzzy relational equations,” Fuzzy Sets and
Syst., vol. 11, pp. 1–18, 1983.
[49] R. M. Tong, “Synthesis of fuzzy models for industrial processes,” Int. J. Gen. Syst., vol. 4, pp.
143–162, 1978.
[50] B. S. Li and Z. J. Liu, “Identification of fuzzy models using fuzzy theory,” Inf. and Control,
vol. 9, no. 3, 1980.
[51] P. S. Babu, A. Ghosh, and Sachchidanand, “Fuzzy identification and control of a class of nonlin-
ear systems,” in Proc. European Control Conf. (ECC), no. 8217, 1997.
[52] Y. C. Lee, C. Hwang, and Y. P. Shih, “A combined approach to fuzzy model identification,” J.
Fuzzy Sets and Syst., vol. 24, no. 5, pp. 736–744, 1994.
[53] C. G. Moore and C. J. Harris, “Indirect adaptive fuzzy control,” Int. J. Control, vol. 56, pp.
441–468, 1992.
[54] C. W. Xu, “Fuzzy system identification,” in Proc. IEE, vol. 136, July 1989, pp. 146–150.
[55] L. X. Wang and J. M. Mendel, “Fuzzy basis functions, universal approximation and orthogonal
least squares learning,” IEEE Trans. Neural Netw., vol. 3, no. 5, pp. 807–814, 1992.
[56] K. Takana and M. Sugeno, “Stability analysis and design of fuzzy control systems,” Fuzzy Sets
and Syst., vol. 45, no. 2, pp. 135–156, 1992.
[57] S. G. Cao, N. W. Rees, and G. Feng, “Analysis and design for a class of complex control systems-
part II: Fuzzy controller design,” Automatica, vol. 33, pp. 1029–1039, 1997.
[58] Z. Huaguang and Q. Yongbing, “Modelling, identification and control of a class of nonlinear
systems,” IEEE Trans. Fuzzy Syst., vol. 9, no. 2, pp. 349–354, 2001.
[59] M. Sugeno and Y. Yasukawa, “A fuzzy-logic-based approach to qualitative modeling,” IEEE

Trans. Fuzzy Syst., vol. 1, no. 1, pp. 7–31, February 1993.
BIBLIOGRAPHY 115
[60] A. Cipriano, M. Ramos, and F. Montoya, “A new method for fuzzy model identification,” in
Proc. IEEE 21st Int. Conf. Indus. Electron., Control and Instrumentation (IECON), vol. 2, 1995,
pp. 1514–1519.
[61] A. E. Gaweda and Z. M. Zurada, “Data-driven design of fuzzy system with relational input
partition,” in Proc. 10th IEEE Int. Conf. Fuzzy Syst., vol. 2, no. 3, 2001.
[62] D. Saez and A. Cipriano, “Fuzzy models based economic predictive control for a combined
cycle power plant boiler,” in Proc. IEEE Int. Symp. on Intell. Control/Intell. Syst. and Semiotics,
Cambridge, MA, September 1999, pp. 417–422.
[63] A. L. Elshafei and F. Karray, “Fuzzy based sliding manifolds for identification of a class of
nonlinear systems,” in Proc. 9th IEEE Int. Conf. Fuzzy Syst., vol. 2, 2000, pp. 841–846.
[64] D. Saez and A. Cipriano, “Design of a supervisory predictive controller based on fuzzy models,”
in Proc. IEEE Int. Conf. Fuzzy Syst., 2001, pp. 1004–1007.
[65] M. Y. Chen and D. A. Linkens, “A systematic method for fuzzy modeling from numerical data,”
in Proc. IEEE Int. Conf. Syst., Man and Cybern., 2001, pp. 28–33.
[66] A. Flores, D. Saez, J. Araya, M. Berenguel, and A. Cipriano, “Fuzzy predictive control of a solar
power plant,” IEEE Trans. Fuzzy Syst., vol. 13, no. 1, pp. 58–68, February 2005.
[67] J. H. Holland, Adaptation in Natural and Artificial Systems. MIT Press, Cambridge MA, 1975.
[68] J. S. R. Jang, C. T. Sun, and E. Mizutani, Neuro-Fuzzy and Soft Computing: A Computational
Approach to Learning and Machine Intelligence. Prentice Hall Inc., 1997.
[69] T. Kumon, M. Iwasaki, T. Suzuki, T. Hashiyama, N. Matsui, and S. Okuma, “Nonlinear system
identification using genetic algorithm,” in Proc. 6th IEEE Conf. Indus. Electron. Society, vol. 4,
2000, pp. 2485–2491.
[70] A. Akramizadeh, A. A. Farjami, and H. Khaloozadeh, “Nonlinear hammerstein model identifi-

cation using genetic algorithm,” in Proc. IEEE Int. Conf. Artificial Intell. Syst. (ICAIS’02), 2002,
pp. 351–356.
[71] J. G. Juang, “Application of genetic algorithm and recurrent network to nonlinear system identi-
fication,” in Proc. IEEE Conf. Control Appl., vol. 1, 2003, pp. 129–134.
BIBLIOGRAPHY 116
[72] D. H. Kim and H. Lee, “Intelligent control of nonlinear power plant using immune algorithm
based multiobjective optimization,” in Proc. IEEE Int. Conf. Networking, Sensing & Control,
March 2004, pp. 1388–1393.
[73] K. R. Vazquez, C. M. Fonseca, and P. J. Fleming, “Identifying the structure of nonlinear dynamic
systems using multiobjective genetic programming,” IEEE Trans. Syst., Man and Cybern.-Part
A: Syst. and Humans, vol. 34, no. 4, pp. 531–545, July 2004.
[74] J. S. R. Jang and C. T. Sun, “Neuro-fuzzy modeling and control,” Proc. IEEE, vol. 83, no. 3, pp.
378–406, March 1995.
[75] Y. Lin and G. A. Cunningham-III, “A new approach to fuzzy-neural system modelling,” IEEE
Trans. on Fuzzy Syst., vol. 3, no. 2, pp. 190–198, May 1995.
[76] J. S. R. Jang, “Input selection for ANFIS learning,” in Proc. IEEE Int. Conf. Fuzzy Syst., vol. 2,
1996, pp. 1493–1499.
[77] S. Chiu, “Selecting input variables for fuzzy models,” J. Intell. and Fuzzy Syst., vol. 4, no. 4, pp.
243–256, 1996.
[78] M. A. Denai, F. Palis, and A. Zeghbib, “ANFIS based modelling and control of nonlinear sys-
tems: A tutorial,” in Proc. of the IEEE Int. Conf. on Syst., Man and Cybern., vol. 4, 2004, pp.
3433–3438.
[79] ——, “Modelling and control of nonlinear systems using soft computing techniques,” Applied
Soft Computing, vol. 7, pp. 728–738, 2007.
[80] H. Ishibuchi, R. Fujioka, and H. Tanaka, “Neural networks that learn from fuzzy if-then rules,”
IEEE Trans. on Fuzzy Syst., vol. 1, no. 2, pp. 85–97, May 1993.
[81] C. F. Juang and C. T. Lin, “An on-line self constructing neural fuzzy inference network and its
application,” IEEE Trans. on Fuzzy Syst., vol. 6, no. 1, pp. 12–32, Feb 1998.
[82] E. G. Sanchez, J. M. C. Izquierdo, M. J. A. Bravo, Y. A. Dimitriadis, and J. L. Coronado,

“Adaptive IMC using fuzzy neural networks for the control on nonlinear systems,” in Research
Paper. ESPRIT Project No. 22416 ”MONNET”, Leon, Spain.
BIBLIOGRAPHY 117
[83] Y. Li, X. Zhao, and L. Jiao, “A nonlinear system identification approach based on neurofuzzy
networks,” in Proc. of the ICSP., vol. 3, 2000, pp. 1594–1597.
[84] S. Kawaji and Y. Chen, “Soft computing approach to nonlinear system identification,” in Proc.
IEEE Conf. on Indus. Electron. Society, vol. 3, 2000, pp. 1803–1808.
[85] S. Wu, M. Joo, and Y. Gao, “A fast approach for automatic generation of fuzzy rules by gener-
alized dynamic fuzzy neural networks,” IEEE Trans. on Fuzzy Syst., vol. 9, no. 4, pp. 578–594,
Aug 2001.
[86] Y. Gao and M. Joo, “Nonlinear identification and control using a generalized fuzzy neural net-
work,” in Proc. 41st IEEE Conf. on Decision and Control, Las Vegas, Nevada USA, vol. 2, Dec
2002, pp. 1363–1368.
[87] M. Panella and A. S. Gallo, “An input-output clustering approach to the synthesis of ANFIS
networks,” IEEE Trans. on Fuzzy Syst., vol. 13, no. 1, pp. 69–81, Feb 2005.
[88] S. Thangavel, V. Palanisamy, K. Duraiswamy, and S. C. Pandian, “Fuzzy identification and

modeling of an intelligent controller for adaptive control of reactive power in a utility system
using anfis,” The J. CPRI., vol. 3, no. 1, pp. 29–35, September 2006.
[89] A. Abraham and B. Nath, “A neuro-fuzzy approach for modelling electricity demand in victoria,”
J. Applied Soft Computing, vol. 1, pp. 127–138, 2001.
[90] A. Kamia, S. J. Ovaska, R. Roy, and S. Kobayashi, “Fusion of soft computing and hard com-
puting for large scale plants: A general model,” J. Applied Soft Computing, vol. 5, pp. 265–279,
2005.
[91] J. S. R. Jang, “ANFIS: Adaptive-network-based fuzzy inference systems,” IEEE Trans. on Syst.,
Man and Cybern., vol. 23, no. 3, pp. 665–684, May/June 1993.
[92] Y. Tsukamoto, M. M. Gupta, R. K. Ragade, and R. R. Yager, “An approach to fuzzy reasoning
method,” in Advances in Fuzzy Set Theory and Application, M. M. Gupta, R. K. Ragade, and
R. R. Yager, Eds., North-Holland, Amsterdam, 1979, pp. 137–149.
[93] J. S. R. Jang and C. T. Sun, “Functional equivalence between radial basis function networks
and fuzzy inference systems,” IEEE Trans. on Neural Netw., vol. 4, no. 1, pp. 156–159, January
1993.
BIBLIOGRAPHY 118
[94] A. Pinkus, “Approximation theory of the MLP in neural networks,” ACTA Numerica, vol. 8, pp.
143–196, 1999.
[95] H. Ying, Y. Ding, and S. L. S. Shao, “Comparison of necessary conditions for typical Takagi-
Sugeno and Mamdani fuzzy systems as universal approximators,” IEEE Trans. Syst. Man Cy-
bern. Part A, vol. 29, pp. 508–514, 1999.
[96] K. Liu, S. Subbarayan, R. R. Shoults, M. T. Manry, C. Kwan, F. I. Lewis, and J. Naccarino,

“Comparison of very short term load forecasting techniques,” IEEE Trans. Power Syst., vol. 11,
pp. 877–882, 1996.
[97] A. Khotanzad, R. A. Rohani, and D. Maratukulam, “ANNSTLF-Artificial neural network short

term load forecaster-generation,” IEEE Trans. Power Syst., vol. 13, pp. 1413–1422, 1998.
[98] H. S. Hipert, C. E. Pedreira, and R. C. Souza, “Neural networks for short term load forecasting:
a review and evaluation,” IEEE Trans. Power Syst., vol. 16, pp. 41–55, 2001.
[99] K. S. Swarup and B. Satish, “Integrated ANN approach to forecast load,” IEEE, Comput. Appl.
Power, vol. 15, pp. 46–51, 2002.
[100] S. Makiko and K. Yoshitsugu, “Demand forecasting by the neural network with fourier trans-
form,” in Proc. IEEE Int. Joint Conf. on Neural Netw., vol. 4, 2004, pp. 2759–2763.
[101] N. Meade, “Forecasting with growth curves: The effect of error structure,” J. Forecast., vol. 7,
no. 4, pp. 235–244, 2006.
[102] G. E. P. Box, G. M. Jenkins, and G. C. Reinsell, Time Series Analysis : Forecasting and Control,
3rd ed. Prentice-Hall Englewood, 1994.
[103] M. Adya, F. Callopy, J. S. Armstrong, and M. Kennedy, “Automatic identification of time series
features for rule based forecasting,” Int. J. of Forecast., vol. 17, no. 10, pp. 143–157, 2001.
[104] R. Gao and L. Tsoukalas, “Neural-wavelet methodology for load forecasting,” J. Intell. and
Robotic Syst., vol. 31, no. 1, 2001.
[105] L. D. S. Coelho and R. A. Krohling, “Nonlinear system identification based on B-spline neural
network and modified particle swarm optimization,” in Int. Joint Conf. on Neural Netw., July
2006, pp. 3748–3753.
BIBLIOGRAPHY 119
[106] H. Saito and T. Abe, “A traffic forecasting method taking into account of outliers,” Electron. and
Communications in Japan (part I:Communications), vol. 69, no. 10, pp. 85–92, March 2007.
[107] N. Amjady, “Short term hourly load forecasting using time series modeling with peak load esti-
mation capability,” IEEE Trans. Power Syst., vol. 16, pp. 789–805, 2001.
[108] L. Mili, T. V. Cutsem, and M. R. Pavella, “Bad data identification methods in power system state
estimation: A comparative study,” IEEE Trans. on Power Apparatus and Syst., vol. PAS-104,
no. 11, pp. 3037–3049, November 1985.
[109] B. M. Zhang and K. L. Lo, “A recursive measurement error estimation identification method for
bad data analysis in power system state estimation,” IEEE Trans. on Power Syst., vol. 6, no. 1,
pp. 191–198, February 1991.
[110] A. Abur, “A bad data identification method for linear programming state estimation,” IEEE
Trans. on Power Syst., vol. 5, no. 3, pp. 894–901, August 1990.
[111] G. Long and F. Ling, “A new complex system identification method and its application to echo
canceller fast initialization,” in IEEE Int. Conf. on Acoustics, Speech and Signal Processing,
ICASSP-90, vol. 3, April 1990, pp. 1671–1674.
[112] A. Pankratz, Forecasting with Univariate Box-Jenkins Models: Concepts and cases. Wiley,
1983.
[113] A. D. Papalexopoulos and T. C. Hesterberg, “A regression based approach to short term system
load forecasting,” IEEE Trans. Power Syst., vol. 5, pp. 1535–1547, 1990.
[114] J. G. D. Gooijer and K. Kumar, “Some recent developments in nonlinear time series modeling,
testing and forecasting,” Int. J. Forecast., vol. 8, pp. 135–156, 1992.
[115] G. Zhang, B. E. Patuwo, and M. Y. Hu, “Forecasting with artificial neural networks: the state of
the art,” Int. J. Forecast., vol. 14, pp. 35–62, 1998.
[116] P. A. Mastorocostas, J. B. Theocharis, and A. G. Bakirtzis, “Fuzzy modeling for short term load
forecasting using the orthogonal least squares method,” IEEE Trans. Power Syst., vol. 14, pp.
29–36, 1999.
BIBLIOGRAPHY 120
[117] M. R. Khan, A. Abraham, and C. Ondrusek, “Short term load forecasting models in Czec. Repub-
lic using soft computing techniques,” Int. J. Knowledge-Based Intell. Engineering Syst., vol. 7,
no. 4, pp. 172–179, 2003.
[118] M. Verleysen, E. D. Bodt, and A. Lendasse, Engineering applications of Bio-inspired Artificial

Neural Networks. Springer Berlin / Heidelberg, 1999, vol. 1607.
[119] Z. J. Wang, Y. Zhao, C. J. Wu, and Y. T. Li, “Application of kernel smoothing to time series
data,” ACTA Mathematicae Applicatae sinica, vol. 22, no. 2, pp. 219–226, April 2006.
[120] R. T. Clemen, “Combining forecasts: a review and annotated bibliography,” Int. J. Forecast.,
vol. 5, pp. 559–583, 1989.
[121] K. H. Kim, H. S. Youn, and Y. C. Kang, “Short term load forecasting for special days in anom-
alous load condition using neural networks and fuzzy inference method,” IEEE Trans. Power
Syst., vol. 15, pp. 559–565, 2000.
[122] A. Abraham and B. Nath, “A neural-fuzzy approach for modelling electricity demand in Victo-
ria,” Applied Soft Computing, vol. 1, pp. 127–138, 2001.
[123] F. C. Morabito and M. Versaci, “Fuzzy neural identification and forecasting techniques to process
experimental urban air pollution data,” Neural Netw., vol. 16, no. 3–4, pp. 493–506, 2003.
[124] B. M. Zhang, S. Y. Wang, and N. D. Xiang, “A linear recursive bad data identification method
with real time application to power system state estimation,” Trans. on Power Syst., vol. 7, no. 3,
pp. 1378–1385, August 1992.
[125] M. Kandemir and J. Ramanujam, “Data relation vectors: A new abstraction for data optimiza-
tions,” IEEE Trans. on Computers, vol. 50, no. 8, pp. 798–810, August 2001.
[126] S. A. Ahmed and J. R. Cruz, “Complex system identification methods for fast echo canceller ini-
tialization,” in IEEE Int. Conf. on Acoustics, Speech and Signal Processing, ICASSP-92, March
1992, pp. 525–528.
[127] D. Erdogmus and J. C. Principe, “An error entropy minimization for supervised training of non-
linear systems,” IEEE Trans. on Signal Processing, vol. 50, no. 7, pp. 1780–1786, July 2002.
BIBLIOGRAPHY 121
[128] V. I. Dedenko and A. L. Movchan, “Minimization of number of metrological parameters for

data acquisition systems,” IEEE Trans. on Instrumentation and Measurement, vol. 51, no. 1, pp.
88–91, February 2002.
[129] T. Joachims, “Optimizing search engines using clickthrough data,” in Proc. of the ACM,
SIGKDD 02, 2002.
[130] B. Babcock, S. Babu, M. Datar, and R. Motwani, “Chain: Operator scheduling for memory
minimization in data stream systems,” in Proc. of ACM, SIGMOD 2003, June 2003.
[131] D. Weekes and G. B. Fogel, “Evolutionary optimization, backpropagation and data preparation
issues in QSAR modeling of HIV inhibition by HEPT derivatives,” Bio. Sys., vol. 72, pp. 149–
158, 2003.
[132] T. Meekhof and R. B. Heckendorn, “Using evolutionary optimization to improve markov based
classifcation with limited training data,” in Proc. of ACM GECCO’05, 2005, pp. 2211–2212.
[133] J. Pekar and J. Stecha, “Identification and predictive control by p-norm minimization,” Acta
Polytechnica, vol. 46, no. 1, pp. 33–39, 2006.
[134] Y. Lii, J. Huang, and Q. Liu, “Improving statistical machine translation performance by training
data selection and optimization,” in Proc. Joint Conf. on Empirical Methods in Natural Language
Processing and Comput. Natural Language Learning, June 2007, pp. 343–350.
[135] G. E. Dieter, Engineering Design: A Material and Processing Approach. McGraw-Hill Inc,
1991.
[136] A. Kohli and U. S. Dixit, “A neural-network-based methodology for the prediction of surface
roughness in a turning process,” The Int. J. Advanced Manufact. Technology, vol. 25, no. 1-2,
pp. 118–129, January 2005.
[137] M. Stone, “Cross validation choice and assessment of statistical predictions,” J. Royal Statistical
Society, vol. 36 (B), pp. 111–147, 1974.
[138] C. L. Sabharwal, “An implementation of hybrid approach to indexing image data bases,” in The
ACM Symp. on Applied Comput., Feb 28–Mar 2 1999, pp. 421–426.
BIBLIOGRAPHY 122
[139] P. Zhang, “Model selection via multifold cross validation,” Annals of Statistics, vol. 21 (1), pp.
299–313, 1993.
[140] O. Nelles, Nonlinear System Identification. Springer-Verlag, Berlin Heidelberg, Germany, 2000.
[141] D. E. Goldberg, Genetic Algorithms in Search, Optimization and Machine Learning, 3rd ed.
Addison Wesley Longman, Inc., 2000.
[142] L. C. Jain and N. M. Martin, Fusion of Neural Networks, Fuzzy Sets and Genetic Algorithm:
Industrial Applications, Int. Series on Computational Intelligence ed. CRC Press, 1999.
[143] T. Takagi and M. Sugeno, “Derivation of fuzzy control rules from human operator’s control
action,” in Proc. IFAC Symp. Fuzzy Inform., Knowledge Representation and Decision Analysis,
July 1983, pp. 55–60.
[144] E. H. Mamdani and S. Assilian, “An experiment in linguistic synthesis with a fuzzy logic con-
troller,” Int. J. Man-Machine Studies, vol. 7, no. 1, pp. 1–13, 1975.
[145] S. L. Chiu, “A cluster estimation method with extension to fuzzy model identification,” in Proc.
3rd IEEE Conf. on Fuzzy Syst., World Congress in Comput. Intell., vol. 2, June 1994, pp. 1240–
1245.
[146] M. H. Hayes, Statistical Digital Signal processing and Modelling. New York: John Wiley &
Sons Inc., 1996.
Related Publications by the Author
Journal Publication:
1. M. Buragohain and C. Mahanta, “ A Novel Approach for ANFIS Modeling based on Full Facto-
rial Design,” Applied Soft Computing, vol. 8, Issue 1, Jan 2008, pp. 609-625.
2. M. Buragohain and C. Mahanta, “V-fold Technique based ANFIS Model for Complex Large
Scale Systems,” International Journal of Artificial Intelligence, special issue on Theory and Ap-
plication of Soft Computing, vol. 1, Issue A08, Autumn 2008, pp. 34-55.
3. M. Buragohain and C. Mahanta, “Data Optimization based Modeling of Complex Large Scale
Systems: A Comparative Analysis with Soft Computing and Statistical Techniques,” Journal of
Systems Science and Engineering, Vol. 17, No. 1, June 2008.
Communicated to Journals:
1. M. Buragohain and C. Mahanta, “ANFIS based Modelling of Thermal Power Plant Using Full
Factorial Design,” to Engineering Application of Artificial Intelligence.
Conference Publications:
1. M. Buragohain and C. Mahanta, “ Data Optimization based Modeling of Complex Systems:

A Comparative Analysis with Soft Computing and Statistical Techniques,” in Proceedings of
National Systems Conference, Manipal, Dec, 2007.
2. M. Buragohain and C. Mahanta, “ Grid Partition based Neuro-Fuzzy Technique and Statistical
Modeling of Complex Systems,” in Recent Trends in IT and Soft Computing(ITSC), IMT Nagpur,
February 2007, pp. 47-58.
3. M. Buragohain and C. Mahanta, “ Complex Systems Modeling with Grid Partitioning based
Soft Computing Technique,” in Recent Trends in IT and Soft Computing(ITSC), IMT Nagpur,
February 2007, pp. 89-97.
4. M. Buragohain and C. Mahanta, “ ANFIS Modeling of Nonlinear System based on V-fold Tech-
nique,” in Proceedings of IEEE International Conference on Industrial Technology, Mumbai,
15-17 Dec. 2006, pp. 2178-2183.
5. M. Buragohain and C. Mahanta, “ ANFIS Modeling of Nonlinear Systems based on Combined
FFD-V-fold Technique,” in Proceedings of IEEE International Conference on Industrial Tech-
nology, Mumbai, 15-17 Dec. 2006, pp. 2462-2467.
6. M. Buragohain and C. Mahanta, “ Full Factorial Design based ANFIS Model for Complex Sys-
tems,” in Proceedings of IEEE Annual India Conference, New Delhi, Sept. 2006.
7. M. Buragohain and C. Mahanta, “ ANFIS Modeling of Nonlinear Systems based on Subtractive

Clustering and V-fold Technique,” in Proceedings of IEEE Annual India Conference , New Delhi,
Sept 2006.
8. M. Buragohain and C. Mahanta, “ ANFIS Modeling of Nonlinear Systems based on Subtrac-

tive Clustering and Combined FFD-V-fold Technique,” in Proceedings of IEEE Annual India
Conference , New Delhi, Sept. 2006.
9. M. Buragohain and C. Mahanta, “ Modeling of Complex Systems with Soft Computing Tech-
niques,” in Proceedings of National Systems Conference , Goa, Nov 2006.
10. M. Buragohain and C. Mahanta, “ Statistical and Neuro-Fuzzy Technique based Modeling of
Complex Systems ,” in Proceedings of National Systems Conference , Goa, Nov 2006.
11. M. Buragohain and C. Mahanta, “ Modeling of Thermal Power Plants Using Full Factorial Design
based ANFIS ,” in Proceedings of IEEE International conference on Cybernetics and Intelligent
Systems, Bangkok, June 2006.
Bio-data of the Author
Mr. M. Buragohain received the Bachelor of Engineering degree in Electrical Engineering. in 1992
from Guwahati University, Guwahati. He obtained the Master of Engineering in High Voltage En-
gineering from Jadavpur University in 2000. He next joined the PhD program in Electronics and
Communication Engg. in IIT Guwahati in 2003. His research interests lie in nonlinear system identifi-
cation and soft computing. He is currently working as a Senior Lecturer in the Department of Electrical
Engineering, Jorhat Engineering College, Assam.

ANFIS - Identification With Special Emphasis On Training Data Minimization

Uploaded by

Copyright:

Available Formats

ANFIS - Identification With Special Emphasis On Training Data Minimization

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

ANFIS - Identification With Special Emphasis On Training Data Minimization

Uploaded by

Copyright:

Available Formats

Adaptive Network based Fuzzy Inference

System (ANFIS) as a Tool for System

Department of Electronics and Communication Engineering

Department of Electronics and Communication Engineering

Dated: Dr. Chitralekha Mahanta

List of Figures iii

Mathematical Notations viii

2 Soft Computing based Techniques for System Identification 6

3 Research Objectives and Methodology 24

4 Adaptive Network based Fuzzy Inference System 28

5 Optimal Data based ANFIS Modeling 54

6 Statistical Models and Genetic Algorithm Based Fuzzy Model 76

7 Conclusions and Future Work 98

A Sample Power Plant Data 103

2.1 Crossover operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

4.1 Type-3 ANFIS Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

5.1 Selection of training data in a 23 factorial design . . . . . . . . . . . . . . . . . . . . 56

6.1 Flow Chart for the GA based Fuzzy model . . . . . . . . . . . . . . . . . . . . . . . . 79

A.1 Thermal power plant data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104

ANFIS Adaptive network based fuzzy inference system

ANN Artificial neural network

ARIMAX Auto regressive integrated moving average with exogenous input

ARMAX Auto regressive moving average with exogenous input

ARX Auto regressive with exogenous input

DAS Data acquisition system

DDEC Data driven echo canceller

EOF Effect of factor

FFD Full factorial design

FIS Fuzzy inference system

FLC Fuzzy logic controller

LSE Least square estimation

MIMO Multi input multi output

MLP Multi layer perceptron

MSE Mean square error

RBFN Radial basis function network

RHONN Recurrent high order neural network

RMEEI Recursive measurement error estimation identification

RMSE Root mean square error

SISO Single input single output

SMT Statistical machine translation

SOFC Self organizing fuzzy controller

ra Radius of the cluster

rb Radius defining the neighborhood of the cluster

α Constant whose value depends on ra

β Constant whose value depends on rb

1.1 Aim of the research

1.2 Contribution of the thesis

• Proposal of a GA based fuzzy model

1.3 Organization of the thesis

Soft Computing based Techniques for System

2.1 Neural Network based Algorithms

characteristics of the neural network are:

• Presence of a large number of simple units

• Presence of a large number of highly parallel units

• Presence of strongly connected units

• Robustness against the failure of single units

• Learning from data

2.2 Fuzzy Logic based Algorithms