Artificial Intelligence Research Papers

Ch a p t e r 12
ARtificial Intelligence
Applications in Renewable
Energy Systems and Smart
Grid – Some Novel
Applications
Bimal K. Bose
12.1 INTRODUCTION
What is artificial intelligence (AI)? AI is basically a computer emulation of human

intelligence, often called computational intelligence. A human brain with a
biological nervous system has natural intelligence and the ability to think, that is, to
learn, reason and comprehend. Can a computer really think and take intelligent
decisions? Or is it only as good as the program embedded into it? Computational
intelligence has been a debatable subject for a long time. It is true that the intelli-
gence of a computer is far inferior to the natural intelligence of a human being. But
it is also true that it can help in solving complex problems which are difficult to
solve by traditional methods. The AI technology evolution has been tremendous in
the last several decades, and it has been extensively applied in industrial process
control, geology, medicine, information management, military systems, space
exploration, finance, marketing, and so on.
AI can be basically classified into four different areas. These are: Expert
Systems (ES), fuzzy logic (FL), artificial neural network (ANN) or neural net-
work (NNW), and genetic algorithms (GA) or evolutionary computation (EC).
In recent years, AI technology has been almost synonymous with neural network
applications. In fact, the origins of AI theory go back to the days of George
Power Electronics in Renewable Energy Systems and Smart Grid: Technology and Applications,
First Edition. Edited by Bimal K. Bose.
© 2019 by The Institute of Electrical and Electronics Engineers, Inc.
Published 2019 by John Wiley & Sons, Inc.
625
626 Chapter 12 ARtificial Intelligence Applications
Boole, an English mathematician, when he theorized that human thinking is

basically based on “yes–no” principles, or “1–0” logic, and published his his-
toric paper “The Investigations on the Laws of Thoughts” in 1854. As a result of
this paper, Boolean algebra was gradually born and ushered in the modern‐day
digital computers. This also gave birth to the classical AI technique called Expert
Systems in the 1970s which is based on IF … THEN … rules or rule‐based sys-
tems, which were extensively used in the 1980s. Lotfi Zadeh, another mathema-
tician and computer scientist, invented FL in 1965, arguing that human thinking
is often fuzzy, vague, or imprecise in nature. A similar rule base is also valid in
FL, but the rules use multivalued logic, which will be discussed later. The GA
theory was proposed by John Holland in 1974. In this evolutionary method of
computation, a probabilistic method of solving optimization problems is used.
Following the principle of biological genetics, or Darwin’s theory of evolution,
biological terms such as population, offspring, cromosomes, reproduction, gen-
eration, crossover (or mating), mutation, fitness function, and convergence are
commonly used in GA theory. An initial population (a series of binary strings) is
assumed and then optimization is reached over several generations. While ES,
FL and GA follow the behavioral nature of human thinking, the NNW is the
most generic form of AI and tends to emulate the human nervous system for
problem‐solving. The NNW is the most powerful technique among all the AI
methods. Zadeh classified ES as “hard” or precise computing (because it is
based on Boolean variables), whereas FL, GA and NNW are classed as “soft” or
approximate computing.
If the control, estimation, or diagnosis of a system is generally based on AI,
it is generally called an intelligent system. For example, the traditional control of a
linear system is based on a PI (proportional‐integral) or PID (proportional‐integral‐
derivative) controller, the parameters of which are determined from mathematical
models of the system. The intelligent control may not need any mathematical
model. Such control is often defined as learning, self‐organizing, or self‐adaptive
control. In fact, many systems may not have s mathematical model, or the model
may be ill‐defined with parameter variation problems. For example, the propor-
tional gain (Kp) and integral gain (Ki) of a nonlinear parameter‐varying system can
be tuned online for a desired response from the transient behavior of the plant. An
adaptive PI‐tuned control system can be designed by ES, FL, NNW or GA. Of
course, for computer simulation study, a mathematical plant model is essential. The
experimental performance of a plant can be captured and its intelligent control can
be designed on this basis.
Suffice to say that the integration of modern power electronics, com-
puters, communications, information and cyber‐technologies with the smart
grid has already created a challenging frontier for smart grid design and imple-
mentation. This challenge will be more formidable with the emerging AI tech-
niques. Smart grid is truly an interdisciplinary technology with multiple
dimensions of complexity. This chapter will first comprehensively review the
12.2 EXPERT SYSTEMS 627
basic principles of AI (i.e., ES, FL and NNW). This background knowledge

[1–10] is essential to understanding the applications of AI in renewable energy
systems (RESs) and smart grid discussed in this chapter. The NNW will be
discussed in more detail because it appears that the next generation of AI will
be heavily based on NNW.
More specifically, the following novel applications will be covered in this
chapter:
•• ES‐based control concept of a smart grid system with real‐time simulator.
•• Permanent magnet synchronous generator (PMSG)‐based wind generation
system with FL‐based control.
•• Adaptive neuro fuzzy inference system (ANFIS)‐based health monitoring of
a wind generation system.
•• Fault identification of a smart grid subsystem by neural network.
These application principles can be extended to cover many other applications.
12.2 EXPERT SYSTEMS
In this section, the basic principles of ES and its applications will be discussed with
relevance to renewable energy systems and smart grid.
12.2.1 Expert System Principles

An ES is basically an advanced and intelligent computer program that embeds the
expertise of a human being in a certain domain. The human expert acquires
the domain knowledge by education and training over a prolonged period of time.
The objective of an ES is to replace the human expert to solve the particular problem
by computer‐embedded expertise. Figure 12.1 shows the block diagram of ES with
the different elements as follows [11–15]:
•• Knowledge base
•• Inference engine
•• User interface
•• Explanation subsystem for user education.
The core of an ES is the knowledge base, which is acquired from the domain expert
by a knowledge engineer and translated into a software program. It is assumed that
the domain expert (say, a power electronics or power systems engineer) does not
have the requisite software expertise to structure or represent their knowledge into
efficient ES software. The core of the knowledge base is the expert knowledge,
which consists of a matrix of IF… THEN rules that is supported by the database as
Explanation
subsystem
(user education)
Knowledge base
User User Inference Expert

Database
dialog interface engine knowledge
Work Knowledge
space acquisition
Domain
expert
Figure 12.1 Expert system block diagram showing the different elements.
Knowledge base
Z
X X1 X2 X3 X4
Y
Y1 A B C D
Inference
engine Y2 E F G H Data base
Y3 I J K L
Y4 M N O P
Matrix of rules
Figure 12.2 Knowledge base rule‐matrix showing interface with data base and inference
engine.
shown in Figure 12.2. In the 16‐rule knowledge base, shown in the figure, the rules
can be read as
Rule 1: IF X = X1 AND Y = Y1, THEN Z = A (12.1)

.
.
.
Rule 16: IF X = X4 AND Y= Y4, THEN Z = P (12.16)
where X, Y and Z are the rule parameters, and X1, Y1, A, and so on, are the respective
parameter values. The database that supports the knowledge base embeds the
parameter values which can be in the form of data (logical or numeric), facts or state-
ments. A rule has a conditional (premise or antecedent) part in the IF statement and
the action (consequent or conclusion) part in the THEN statement. The logical
connective can be AND, OR, and NOT operations for drawing the conclusion. A rule
is fired IF the conditional part is true, and then the action part guided by the THEN
statement is executed. A complex rule can have a large number of parameters and
corresponding values. A practical rule for a power grid control can be given as
IF photovoltaic station A output 0 AND wind station C output 800 kW,

THEN trip circuit breaker K and assign battery storage station D outpput 100 kW
(12.17)
The knowledge base rules can be designed to be complex to handle probability with
certainty factors and probability‐based models, such as a Bayesian approach. It can
also incorporate fuzzy inference systems (FIS) which will be discussed later. The
knowledge base can be adaptive in nature, depending on system changes or tech-
nology advancement. Knowledge is sometimes defined as “shallow” or “deep.”
Shallow knowledge is obtained directly from the domain expert and deep knowledge
is derived from the system model and simulation response [15], which is based on
the designer’s or researcher’s knowledge.
The inference engine is basically the controller or executive software of the
ES. It tests the rules of the knowledge base in sequence and tries to draw inferences
or conclusions. The validation of the rules can be either by “forward chaining” or
“backward chaining.” In the forward chaining principle, the conditional part is
tested first, and if it is true, then the rule is fired. In backward chaining, the infer-
ence engine assumes that the action part is true and then tests backwards the
conditional part for validity before firing the rule. The ES software uses symbolic
language like LISP or PROLOG for problem‐solving. The program can also be
developed in C language for time‐critical real‐time operation.
The inference engine communicates with the user in a very user‐friendly
dialog in natural language (English, for example). Through the inference engine,
the ES requests parameter values for the rules for problem‐solving from the user. If
the parameter values satisfy the conditional part of a rule, then it will be fired.
One very important function of the ES is user education through the explana-
tion subsystem, as shown in Figure 12.1. This subsystem responds during problem‐
solving by the HELP, WHY and HOW commands of the user. The user is usually a
semi‐skilled person and does not understand the details for solving a complex
problem by ES. With the HELP command, the ES explains to the user the technical
features of the problem with explanatory texts and figures.The WHY command
explains to the user why the ES is asking for relevant information from the user, and
the HOW command explains how the ES arrived at the particular solution. Often,
user education may be the sole purpose for designing an ES.
The knowledge base of a large system can be organized in the form of a tree
using a large number of hierarchical frame‐based structures, as shown in Figure 12.3.
Sub frame Rules

1C
Sub frame
1
Sub frame
Parameters
1D
Root
Sub frame
frame
2
Sub frame
3
Figure 12.3 Hierarchical frame‐based structure of knowledge.
Each frame has a characteristic set of rules and parameters. The frame‐based
knowledge organization changes a large system into a simple and logical modular
structure. The core knowledge of the system may be embedded in the root frame.
The first‐layer subframes (child frames) and second‐layer subframes (grandchild
frames) can be considered as subdomains of the expert knowledge, which can be
obtained by the direction of root frame. For a small system, just the root frame
alone may be sufficient. An ES can be organized to track its own operation to
improve efficiency and speed of its own knowledge base with learning. The
knowledge about its own knowledge‐base operation is defined as meta‐knowledge.
The meta‐rules will dictate the most efficient order of rule search to improve the
efficiency of knowledge‐base operation.
The ES software is normally designed with the help of a SHELL‐based soft-
ware platform,which is basically a user‐friendly development system. A number of
commercial SHELLs [11] are available for ES program development. The program
developer is required to have familiarity with the SHELL, although the production
rules are written in English‐like Abbreviated Rule Language (ARL) for rapid
development of the knowledge‐base rules. For example, the rule in (12.17) can be
represented in ARL as
IF :: PVA NO AND WINDC 800

(12.18)
THEN :: CBK 0 AND BESSD 100
where PVA, WINDC, CBK and BESSD are the parameter names with the given
assigned values. The program software is in LISP or C, but the user dialog and
explanation subsystem should be in natural English language. The external
Database Graphical
files files
Simulation
files
User ES
interface shell Interface
routines
User
Mathematical functions
program
Figure 12.4 Expert system development shell interfaced with external elements.
interface of the SHELL with USER INTERFACE, DATABASE FILES, GRAPHICS

FILES, SIMULATION FILES, INTERFACE ROUTINES, USER FUNCTIONS
and MATHEMATICAL PROGRAM for program development is shown in
Figure 12.4. A limited amount of numerical data, logical, and arithmetical capabil-
ities can be directly embedded in the ES program, but for a large amount of data,
such as product catalog information, database files should be constructed. For
complex calculations, such as solving differential equations, the program should
take help from external programs. In the same way, an external simulation program
can generate the result for the ES program. The SHELL can also integrate pictures
(such as power grid or power electronic systems) in the knowledge base. The
developed ES program can be operated from the SHELL directly, or exported to
another computer with the user interface and inference engine. When the client
program is resident in the SHELL, the program developer can easily alter or
update it, but when it is in the client environment outside the SHELL, no program
modification is possible.
12.2.2 Expert System‐Based Control of Smart Grid

A smart or intelligent grid is basically a vision of an advanced electric power grid
of tomorrow using state‐of‐the‐art technologies in power systems, power elec-
tronics, control systems, computers, communications, information, AI, cyber, and
so on, that will improve system availability, reliability, power quality, energy
efficiency, and security, with optimum resource utilization and economical
electricity to the consumers [16–25]. Suffice to say that control and protection of
smart grid are extremely complex, particularly if the grid is large. Currently, with
the present status of technology, we do not have a clear understanding of the control
strategy for a future smart grid.
Regional controls
Regional Regional Regional Regional

controller controller controller input
signals
Local control
override
ES based Real time

Operator System
master simulation
interface signals
controller (opal-RT)
Local control
override
Regional
Regional Regional Regional
input
controller controller controller
signals
Regional controls
Figure 12.5 Expert system based simplified control block diagram of smart grid with real
time simulator.
It is convenient to centralize the control system of the whole grid with the
integration of the latest technologies, as mentioned above. As the respective tech-
nology advances, incremental update can be made on the system. Figure 12.5 shows
a simplified conceptual control block diagram of the smart grid [1, 2, 26, 27] based
on ES principles. The grid incorporates a large segment of renewable energy sources
along with bulk power generation by fossil and nuclear power plants. There are
bulk energy storage plants (mainly to support the RES), HVDC systems, FACTS
(flexible AC transmission systems), VAR compensating STATCOMs (static com-
pensators), and MMMCs (Marquardt modular multilevel converters) distributed
strategically throughout the grid.
The whole grid is segmented into a number of regions which are under the
control of centralized master controller. The master controller (shown at the center
of the figure) has an ES oriented knowledge base, where the rules are formulated on
the basis of extensive a priori offline analysis, design and simulation of the whole
system. This is a massive task that requires a large number of iterations. The ES rules
usually have a large number of parameters with different parameter values. The
knowledge base can be updated continuously as more knowledge is gained about the
system. A simple typical rule, as mentioned before, may have the structure as:
IF photovoltaic station A output 0 AND wind station C output 800 kW,

THEN trip circuit breaker K and assign battery storage station D outpput 100 kW
(12.17)
The knowledge base can also incorporate rules based on a fuzzy expert system (or
fuzzy inference system) which will be discussed in the next section.
The master controller has an operator interface on the left and a real‐time
simulator (RTS) interface on the right, and controls all the regional controllers, as
shown on the figure. The RTS can be designed, for example, on the basis of OPAL‐
RT technology [28] which will be described later. All the regional controllers have
local override in case of a problem with the master controller. The RTS is based on
a cluster of supercomputers which embeds the dynamic model of the whole grid
including the fast response power electronic systems. It is a virtual grid which runs
independently from the physical system.
The RTS helps the master controller to generate the command and control
signals for the regional controllers based on a real‐time solution of the virtual (or
simulated) grid. The RTS receives the whole system signals, whereas the regional
controllers receive only the regional signals. A large number of grid signals can be
obtained through system‐wide PMUs (phasor measurement units). The signals may
be sensor‐based or sensorless. Redundant signals may be needed for unreliable
signals.
The operator interface in the central control station is assisted with large
display boards (not shown) that monitor all the key signals (such as system fre-
quency, bus voltages, active and reactive power flows, and status of the system
operation) for information and assistance of the operators. The virtual grid opera-
tion in RTS is similar to a full‐order real‐time observer of a high‐performance
machine drive system [9]. Note that AI‐based (ES, FL and NNW) controls can also
be used extensively within the elements of the smart grid system.
The RTS can also be used for initial planning, analysis and design of the grid,
and to formulate the optimum default control strategy, typically on a 24‐hour period
basis with the predicted loading, generation and energy storage capability curves,
including the planned outages of the equipment. This offline default control strategy
can remain valid if the grid topology with the parameters remain unchanged and the
predicted load and generation curves remain the same. In practical grid operation,
all of these will change to some extent. For example, unpredictable weather condi-
tions in some regions of the grid can cause equipment outages and substantially
lower wind and PV generation. There may also be equipment outages due to unpre-
dicted faults. For these reasons, control laws and parameters require continuous
iteration based on real‐time simulation for optimum grid performance under all
conditions. The RTS can also be used to study future expansion of the grid.
The broad functions of the control system can be briefly summarized as
follows:
•• Initialize the system control gracefully to the default state, assuming normal
state of the grid, with the predicted load and available generation curves based
on offline system analysis and simulation on 24‐hour period basis, as men-
tioned above. For a certain pattern of system loading, the generators are
assigned along with the routing of P and Q so that generation, transmission
and distribution are most economical without any overloading, and the system
losses are minimized. The system frequency and bus voltages are controlled
appropriately.The cost and loss models of the elements of the system should
be available for real‐time computation.
•• After initialization of the controllers, iterate the system control typically on
an hourly basis based on real‐time simulation, considering the deviation from
the predicted generation and load curves, forced equipment outages, and
other system topology and parameter variations, so that the system operation
remains healthy with minimum cost and optimum efficiency. Of course, for
any major equipment commissioning or outage, faster control iteration is
desired.
•• From the actual usage of the equipment, depreciation and loss computations,
calculate the dynamic tariff rate and promote demand‐side energy management
with the help of smart meters and smart appliances.
•• For transient system loading/unloading and/or generator power shedding, sta-
bilize the system frequency and bus voltages with the help of fast response
PWM converters working with energy storage.
•• Perform hardware‐in‐loop (HIL) automated testing of deteriorating equip-
ment and reconfigure them for fault‐tolerant operation (self‐healing), or iso-
late them for repair. In addition, perform HIL testing of new equipment before
commissioning the system.
•• Extensive system monitoring, diagnostics, fault‐tolerant control and protec-
tion strategies are embedded in the system. However, for any hard fault, the
equipment should be protected from damage by local protection.
The reliable operation of hardware and software of the controllers and the RTS are
extremely important. For this reason, some redundancy in hardware and software
should be built in. However, the system tends to be prohibitively expensive if very
high reliability is demanded.
OPAL‐RT Based Real‐Time Simulation Technology

OPAL‐RT Company commercializes supercomputer‐based hardware and software
platforms for real‐time simulation and HIL testing of equipment for electrical, elec-
tromechanical and power electronic systems for a complex power grid. The
ePOWER grid products, as shown in Figure 12.6, permit simulation of everything
from fast electromagnetic phenomena to the transient stability of large power sys-
tems, optimization of security in power systems, efficiency and performance evalu-
ation of microgrids, renewable energy sources and large interconnected smart
power grids. The selection of a particular technology depends on the size and com-
plexity of the power system, and the desired speed of computation. The ePHASOR-
sim can be used in a large system with a maximum number of single‐phase nodes,
and slower transient phenomena. Then, HYPERSIM, eMEGAsim and eFPGAsim
can be used as the speed of simulation increases with the corresponding reduction
Choose the technology that is right for you by looking at the frequency
of the transient phenomena simulated as well as the numbers of nodes.
Number of
single-phase nodes
30,000
20,000
10,000
2,000
1,000
500
100
10
1 Hz 100 Hz 20 KHz 100 KHz 1 MHz 10 MHz 100 MHz

(1 s) (10 ms) (50 μs) (10 μs) (1 μs) (100 ns) (10 ns)
Frequency and period of transient phenomena simulated
Figure 12.6 OPAL-RT based real-time simulation products family for power grids (with
permission from OPAL‐RT). (See electronic version for color representation of this figure.)
in number of nodes, as shown in the figure. Of course, the use of multiple units of
the same type can increase the size of the system with the same speed range. Or
else, different types of the technology can be strategically mixed in a system. Let
us consider only the eMEGAsim‐based simulator, whicht has a wide frequency
range and can include fast power electronics‐based converters such as HVDC
converters, FACTS, STATCOMS, MMCs, along with electrical and electrome-
chanical components.
The basic features of eMEGAsim technology are summarized in Table 12.1.
The hardware of eMEGASIM is based on an Intel high‐speed parallel processing
system along with high speed XLINK FPGAs to provide small time‐steps in the
range of 100 ns to 10 µs. More units can be used as a cluster for large grid simu-
lation. The real‐time software platform (RT‐LAB) uses a mathematical model of
the grid consisting of electrical, electromagnetic, electromechanical and power
electronic systems. The real‐time simulation of eMEGAsim is based on MATLAB/
Simulink and SimPowerSystem (MathWorks), where for an increase in computa-
tion speed, segmented parallel computation is used and the results are then
combined at the output. As usual, the libraries of models for machines, motor
drives, transmission/distribution lines, transformers, power electronic converters
(two‐level and multilevel), high‐frequency switches, and so on, are embedded in
the system. The supercomputer system permits hardware‐in‐loop (HIL) testing of
electrical, electronic and electromechanical systems. Besides normal operation,
various types of faults and disturbances in the system can be studied. The simu-
lator can also be used as a great tool for education and training of power system
engineers.
TABLE 12.1 Features of eMEGAsim supercomputer‐based real‐time smart grid digital

simulator [28].
• Core hardware can scale up to 96 Intel 3.3 GHz processors with up to 48‐processor parallel
INTEL computing system with time‐step of 10 µs, and XILINK 6 FPGA chips with time‐step
as low as 100 ns and timing accuracy of 10 ns.
• Extra INTEL and XILINK boards can be added.
• No. of signal I/O channels: 128 analog and 256 digital.
• For a large power system, a cluster of eMEGAsims can be used with interconnections.
• Can be hybridized with ePHASORsim, HYPERSIM and eFPGAsim technologies.
• Real‐time integrated software platform (RT‐LAB) uses mathematical models of dynamic
system.
• Uses MATLAB/Simulink and SimPowerSystems real‐time simulation with graphical user
interface. To enhance computation speed, segmented parallel simulation is used and then
combined.
• Softwares include model block libraries of machines, drives, lines, transformers, load, power
systems, protection equipment, relays, power electronic components, protection in equipment
and control systems, and so on.
• Multi‐rate sampling and hundreds of three‐phase buses can be used.
• Provides interface with MATLAB, Plecs, SimScape, EMTP‐RV, etc.
• Can emulate ES, FL and ANN‐based software in real time.
• Permits real‐time simulation and control of HVDC, FACTS, STATCOMS, storage, generation,
transmission and distribution systems.
• Permits HIL testing of electrical, power electronic and electromechanical systems.
12.3 FUZZY LOGIC
Since the invention of FL by Lotfi Zadeh in 1965, its applications have expanded
widely in both engineering and non‐engineering areas. The first application to con-
trol a dynamic process was reported by Mamdani in 1974, and by Mamdani and
Assilian in 1975. These were extremely significant contributions because they
stirred widespread interest by later workers in the field for many other applications.
Mamdani and Assilian were concerned with the control of a small laboratory steam
engine. The control problem was to regulate the engine speed and boiler steam
pressure by means of the heat applied to the boiler and the throttle setting of the
engine. The process was difficult because it was nonlinear, noisy, and strongly cou-
pled, and no mathematical model was available. The fuzzy control designed purely
from the operator’s experience by a set of IF… THEN… rules was found to per-
form well and was better than manual control. In 1976, Kickert and Lemke exam-
ined the fuzzy control performance of an experimental warm water plant, where the
problem was to regulate the temperature of water leaving a tank at a constant flow
rate by altering the flow of hot water in a heat exchanger contained in the tank. The
success of Mamdani and Assilian’s work led King and Mamdani (1977) to attempt
to control the temperature in a pilot‐scale batch chemical reactor by a fuzzy
12.3 FUZZY LOGIC 637
algorithm. These results indicated that fuzzy control was very useful for complex
processes and gave superior performance over the conventional PID control, partic-
ularly if a good mathematical model of the plant is not available.
In this section, fuzzy inference system principles and their applications will
be discussed for modern wind generation systems.
12.3.1 Fuzzy Inference System Principles

A fuzzy inference system (FIS) (or fuzzy system (FS)) [8, 9, 29, 30] basically emu-
lates fuzzy human thinking with the help of multivalued fuzzy variables, as men-
tioned before. This concept can be extended to many applications. Each fuzzy
variable is defined by a number of fuzzy sets which are linguistic in nature, and
each fuzzy set is defined by a multivalued membership function (MF).
Similar to ES, a FIS can be described by a set of IF…. THEN… rules, a
simplified form of which is shown in Figure 12.7. Here, X, Y and Z are the fuzzy
variables, where X and Y are the inputs and Z is the output. The fuzzy variables can
be unipolar or bipolar, and the range of values of each is defined as a “universe of
discourse.” For convenience, the fuzzy variables are normally defined on a per unit
(pu) basis. In Figure 12.7, there are three simple rules which can be defined as:
Rule 1: IF X is negative small NS AND Y is zero ZE

THEN Z is positive small PS (12.19)
Rule 2 : IF X is zero ZE AND Y is zero ZE
THEN Z is zero ZE (12.20)
Rule 3 : IF X is zero ZE AND Y is positive small PS
THEN Z is negative small NS (12.21)
Because of analogy, a fuzzy system is often defined as a fuzzy expert system.
X
NS ZE
Y
ZE PS ZE
PS NS
Figure 12.7 Simple rule matrix for fuzzy inference system.

The shape of the MF may be triangular, trapezoidal, Gaussian, or other types, and
may be symmetrical or asymmetrical about the vertical axis. The computations in
FIS consist of the following five steps to determine the inference or conclusion:
1. Fuzzification of the input crisp variables (X, Y, etc.);
2. Application of fuzzy operator (AND, OR, NOT) in the IF (or antecedent) part
of the rule;
3. Implication from the antecedent to the consequent (THEN) part of the rule;
4. Aggregation of the consequents of the rules;
5. Defuzzification to convert the fuzzy output to crisp value.
There are a number of fuzzy implication methods in the literature [9]. The
frequently used types are:
•• Mamdani method;
•• Lusing‐Larson method; and
•• Sugeno (or Takagi‐Segeno‐Kang) method.
The Sugeno method is again subdivided into zero‐order and first‐order
methods. The Mamdani method is most commonly used in the literature.
Mamdani‐Type Fuzzy Inference System

Figure 12.8 illustrates the Mamdani‐type FIS using symmetrical triangular MFs
which corresponds to the rules in Figure 12.7. All the three rules are shown in the
consecutive horizontal rows with the input values of X = −3 and Y = 1.5. Since all
the rules use the AND operator, the DOFs (degrees of fulfilment) from the ante-
cedent parts of the rules are calculated as DOF1 = 0.6, DOF2 = 0.4 and DOF3 = 0.4,
respectively. The corresponding fuzzy output MFs (shaded) are PS’, ZE’ and NS’,
respectively from the consequent part of the rules. The total fuzzy output, as shown
in the figure, is the aggregation of the consequents of the rules. The defuzzification
calculates the center of the area (COA) which is generally given by the expression
Z. out Z
Z0 (12.22)
Z dZ
out
where Z0 is the defuzzified or crisp output. The overlap areas in the output are not
considered. Note that a practical FIS can have a large number of input and output
variables and each variable may have a large number of MFs.
Fuzzy Speed Control of Induction Motor Drive

Let us now illustrate application of a Mamdani type FIS for speed control of a
vector drive with induction motor, as shown in Figure 12.9.
*
In the figure, the variable X e r r is the speed loop error, Y = CE =
dE/dt, the rate of change of the speed loop error, and Z dU iqs* is the incremental
μ(X) μ(Y) μ(Z)

1 NS 1 ZE PS
0.8 1
0.6 PSʹ
Rule-1 DOF1 = 0.6
–6 0 6 –6 0 6 –6 0 6
μ(X) μ(Y) μ(Z)
ZE ZE
1 1 1 ZE
0.6
Rule-2 0.4 DOF2 = 0.4 ZEʹ
–6 0 6 –6 0 6 –6 0 6
μ(X) μ(Y) μ(Z)
ZE PS 1.0 NS
1 1 1
Rule-3 0.4 DOF3 = 0.4 NSʹ
–6 0 x 6 –6 0 Y 6 –6 0 Z 6
x = –3 Y = 1.5
μ(Z)
μout (Z)
1
–6 0 6
Z0
Figure 12.8 Mamdani implication method of fuzzy inference system showing output

defuzzification.
ℐ
d CE
dt VC
Fuzzy DU U
controller ∫ and Induction
ω*r (Δi*qs) (i*qs) inverter motor
+ E
–
ωr
Figure 12.9 Fuzzy speed control of a vector‐controlled induction motor drive.
active component of current signal. The iqs* is integrated to generate the active
current signal iqs* for the vector drive. The equations for the control can be written as
*
e r r (12.23)
de e ce
(12.24)
dt t ts
dU K1e K 2ce (12.25)
U K1 edt K 2 e (12.26)

where variation of K1 (integral gain) and K2 (proportional gain) in the fuzzy
controller makes the vector control adaptive for parameter (such as inertia J and
load torque disturbance TL) variation, giving robustness of the drive system.The
fuzzy P‐I control principle, as indicated by Eq. (12.26) and shown in Figure 12.9,
can be applied in Figure 12.15 to give robust speed control of the wind generation
system, which will be described later. A FIS can be designed with the help of a
Fuzzy Logic Toolbox [31] and simulated in the MATLAB/Simulink environment
with the dynamic system model.
The MFs of the fuzzy speed controller and the corresponding rule matrix for
a practical system are given in Figures 12.10 and 12.11, respectively, where all the
fuzzy variables are defined on a pu basis. The input variables, e and ce, have seven
MFs, whereas the output variable (du) has nine MFs. There are altogether 7 × 7 = 49
rules, as shown in Figure 12.11. This indicates (by square box) the validation of the
four rules at the typical operating point e = 0.08 and ce = 0.13, and the corresponding
fuzzy outputs are indicated by the shaded areas in Figure 12.10). The defuzzifica-
tion gives the output du0 = 0.427.
μ(e)
(a)
NB NM NS Z PS PM PB
0.6
0.4
–1 –0.8 –0.6 –0.4 –0.2 0 0.2 0.4 0.6 0.8 1 e(pu)
e = 0.08
(b)
μ(ce)
NB NM NS Z PS PM PB
0.7
0.3
–1 –0.8 –0.6 –0.4 –0.2 0 0.2 0.4 0.6 0.8 1 ce(pu)

ce = 0.13
(c) μ(du)
NB NM NS NVS PVS PS PM PB
Z
0.6
0.3 0.3 0.4 0.7
–1 –0.8 –0.6 –0.4 –0.2 0 0.2 0.4 0.6 0.8 1 du(pu)

duo = 0.427
Figure 12.10 Membership functions for fuzzy speed controller: (a) error (e); (b) change
in error (ce); and (c) change in output (du).
du(pu)
e(pu)
NB NM NS Z PS PM PB
ce(pu)
NB NVB NVB NVB NB NM NS Z
NM NVB NVB NB NM NS Z PS
NS NVB NB NM NS Z PS PM
Z NB NM NS Z PS PM PB
PS NM NS Z PS PM PB PVB
PM NS Z PS PM PB PVB PVB
PB Z PS PM PB PVB PVB PVB
Figure 12.11 Rule matrix for fuzzy speed control of induction motor.
Input Space Output Space
Black
Box
Excitation Response
Figure 12.12 Input–output nonlinear mapping property of fuzzy and neural systems.
A FIS can be interpreted as a general input–output nonlinear mapping

problem, as indicated in Figure 12.12. This is a very important concept for both
fuzzy systems and neural networks, which will be discussed later. The input
information is defined in the input space, it is processed in the “Black Box,” and
the solution appears in the output space. This mapping can be interpreted as a
pattern matching or pattern recognition problem. For a human being, for example,
when we see an apple, we recognize that it as an apple by its characteristic color
and shape. This is pattern recognition by the associative memory property of the
human brain, which is represented by the Black Box in the figure. The mapping
can be static or dynamic with linear or nonlinear relationships, as determined by
the Black Box’s characteristics. In general, the Black Box can represent not only
a fuzzy system, but also ES, NNW, or a mathematical function, where in general
an excitation creates the response. In Figure 12.9, the Black Box is the fuzzy
controller that gives a robust control algorithm with inertia variation and load
torque disturbance of the drive system. In solving static nonlinearity in engi-
neering problems, the black box is often implemented by a mutidimensional
DSP‐based look‐up table.
Sugeno Method of Fuzzy Inference System

Figure 12.13 illustrates the Sugeno zero‐order method of implication in FIS, where
the input MFs are the same as those in the Mamdani method in Figure 12.8, but the
output MFs are given by singleton MFs at constant values of Z, that is, Z1 = K1,
Z2 = K2 and Z3 = K3 in the universe of discourse, as shown in the figure. In the conse-
quent part of each rule, the respective singleton is truncated with the respective
DOF1, DOF2 and DOF3 values of each rule, and are then aggregated to constitute
the total fuzzy output shown in the lower right of the figure. The rule base and the
MFs are considered the same as that of the Mamdani method for simplicity. A typ-
ical rule in the figure can be read as:
Rule 1 : IF X is NS AND y is ZE
THEN Z K1 (12.27)
μ(X) μ(Y) μ(Z)

NS ZE
1 0.8 1 0.6 1
DOF1 = 0.6
–6 0 6 –6 0 6 –6 0 6
Z1
μ(X) μ(Y) μ(Z)
1 ZE 1 ZE ZE
1
0.6
0.4 DOF2 = 0.4
–6 0 6 –6 0 6 –6 0 6
Z2
μ(X) μ(Y) μ(Z)
ZE PS 1.0 1
1 1
0.4 DOF3 = 0.4
–6 0 X 6 –6 0 Y 6 –6 Z3 0 Z 6
X = –3 Y = 1.5 μ(Z)
1 DOF1 = 0.6
DOF2 = 0.4
DOF3 = 0.4
–6 Z3 Z2 Z1 6
Z0
Figure 12.13 Sugeno (zero‐order) implication method of fuzzy inference system.

X Y
μ(X) μ(Y)
NS ZE
1 1
0.8
Rule-1
0.6
Z1 = A01+ A11X +A21 Y
DOF1 = 0.6
Z1·DOF1
–6 0 6 –6 0 6
μ(X) μ(Y)
ZE ZE X Y Z2·DOF2
1 1
0.6
Rule-2
0.4 DOF2= 0.4

Z2 = A02 + A12X +A22Y Defuzzification Z0
–6 0 6 –6 0 6
μ(X) μ(Y) X Y
ZE PS 1.0
1 1 Z3·DOF3
Rule-3
DOF3= 0.4
0.4 Z3 = A03+ A13X +A23Y
–6 0 x 6 –6 0 y 6
X = –3 Y = 1.5
Figure 12.14 Sugeno (first‐order) implication method of fuzzy inference system.
The fuzzy output can be converted to crisp output by the defuzzification formula as
K1. dof1 K 2 . dof2 K 3 . dof3

Z0 (12.28)
dof1 dof2 dof3
A more general first‐order Sugeno method of implication is illustrated in

Figure 12.14. In this case, the the Ks of the zero‐order method move in the universe
of discourse linearly as a function of input signals X and Y with the constants A01,
A02, etc., so that the respective rule output is:
Z1 A01 A11 X A21y (12.29)
Z2 A02 A12 X A22y (12.30)
Z3 A03 A13 X A23y (12.31)
The constants in the above equations can be iterated to optimize the FIS performance.
The output defuzzification can be given by the formula:
Z1. dof1 Z 2 . dof2 Z 3 . dof3

Z0 (12.32)
dof1 dof2 dof3
where Z1, Z2 and Z3 are the functions of the input signals as shown in the figure. It
can be mentioned here that a FIS can be implemented by a NNW which will be
discussed later, and the Sugeno methods are convenient for such implementation.
12.3.2 Fuzzy Logic Control of a Modern Wind

Generation System
Let us now discuss fuzzy logic‐based control of a modern wind generation system.
It is well known that electrical energy generated from wind is renewable, safe,
economical, environmentally clean (or “green”), and the cost is generally cheaper
than fossil‐fuel or nuclear‐based energy generation. Recently, technological
advances in power electronics, variable speed drives and wind turbines have made
it even more economical. The wind (as well as photovoltaic) energy resources are
abundant, and tapping only a small part of it can essentially meet all the energy
demands of the world. However, since renewable energy is generally sporadic and
regional, adequate storage and transmission capabilities at economical prices are
essential for their utilization. A modern wind generation system [9] uses variable‐
speed horizontal‐axis wind turbines that are usually coupled to a generator shaft
with or without a speed‐up gear. The generator can be either an induction or a
synchronous type machine. Many configurations of wind generation systems are
possible. Figure 12.15 shows a modern variable‐speed wind generation system [32,
33] that uses a high‐efficiency interior (or buried) permanent magnet (IPM) (with
NdFeB magnet) type sinusoidal synchronous generator, where the control system is
based on fuzzy logic. Basically, the system can be looked upon as a four‐quadrant
AC drive, where the machine operates continuously in regenerative braking mode.
Ipm
sync. PWM PWM
generator rectifier inverter
Vd Bus
P0
Vw Q0
Ids, Iqs
Wind
turbine Ia, Ib
P.S. Pd
θe Svm,
Cos(θe + δ) Svm,
Sin(θe + δ) sync.
sync. Sinθe
IM current
IT current Cosθe
control Q0
control
Te ψs ωr I M* IT* Ids* Iqs*
P-I P-I P-I P-I
Te
ψs* ψs* + Q0* + – Q0 P0
ψs Te*
– + Pd + –
Te +P0*
Fuzzy PI
MPPT-2 Vd P-I
FLC-3 –
FLC-2 +
Vd* –1
P0
MPPT-1 FLC-1
ωr*
+
ωr
–
Figure 12.15 Permanent magnet synchronous generator (PMSG)‐based wind generation

system showing fuzzy logic‐based MPPT control.
Description of Power Circuit

The horizontal axis wind turbine at variable speed with varying wind velocity (Vw)
couples to the shaft of the generator directly without any speed‐up gear. Such a direct‐
drive system eliminates the cost, weight, power loss, noise and loss of reliability
due to the gear, and lowers the cut‐in speed of the generator. Of course, this causes
some cost and weight penalty to the machine. With interior mounting of magnets, the
machine is robust at high speed. The machine is salient pole with the direct axis air
gap larger than that of the quadrature axis (Ldm < Lqm), and the armature reaction effect
becomes dominant due to a smaller effective air gap. The variable‐frequency, vari-
able‐voltage generator output is rectified to regulated DC voltage (Vd) by a three‐
phase PWM IGBT converter. The machine current waves are sinusoidal. The active
component of the current corresponds to machine output power, but the reactive (or
magnetizing) component of current controls the stator flux (Ψs) for light‐load
efficiency optimization. The DC‐link voltage is converted to constant‐frequency,
constant‐voltage output by the line‐side three‐phase sinusoidal PWM IGBT converter
that couples to the utility bus. For higher power generation, multilevel converters
(such as Marquardt MMCs) can be used. A transformer can be used to boost the
output voltage with isolation, if necessary. Both the active (P0) and reactive power(Q0)
(leading or lagging) can be controlled by the line‐side c onverter. Such a four‐quadrant
dual converter system can be operated in generating mode as well as in motoring
mode, where the machine can be controlled as a variable speed motor. Since the avail-
able wind energy is sporadic in nature due to variation in wind velocity, a battery
storage system in the DC‐link can provide compensation for the sporadic power.
Wind Turbine Characteristics

Figure 12.16 shows the family of torque–speed curves of the wind turbine at differ-
ent wind velocities with a fixed blade pitch angle. The turbine speed (ωr) is con-
trolled independently by the drive system to tune with the wind velocity (Vw) in order
to maximize the turbine aerodynamic efficiency. Assume that initially the wind
speed is constant at Vw1 and the generator speed is ωr1 so that the operating point is
at A. If ωr is decreased, the turbine shaft torque increases and reaches a maximum
value at point B, and then decreases with further decrease of ωr. Since turbine power
PT = Te.ωr, the maximum power point is at C corresponding to ωr2, as shown. The
locus of maximum power delivery ( pT K r3 ) for different values of Vw is shown in
the figure. This means that as the wind speed varies, the turbine speed is to be varied
to track with it to capture the maximum power. The power output of the turbine can
be reduced at a fixed Vw by reducing the blade pitch angle. The turbine can be
stopped completely by control (feathering) of the pitch angle. The torque–speed
curves can easily be converted to power‐speed curves (shown in Figure 12.20).
Description of Control
The optimum control of the wind generation system is quite complex and is shown
by the simplified block diagram of Figure 12.15. Both the machine‐side and line‐
side converters use space vector modulation (SVM), and the converters always
er
ow
tp
an
st
on
Turbine torque (Te)
c
of
Locus of maximum
s
ne
power delivery
Li
C
B
A
Turbine torque/speed
VW1
curves for increasing
wind speed
ωr2 ωr1
Turbine rotational speed (ωr)
ρπRw2
TM = CP(λ) 0.5· η Vw2
GEAR
Figure 12.16 Torque‐speed curves of wind turbine at different wind velocity showing

fuzzy control to maximize output power.
operate in undermodulation mode. The generator speed (ωr) is controlled by close

loop which generates the torque command te* for the machine, which is also con-
trolled by close loop, as shown.The close loop torque control as well as the fuzzy
controller FLC‐3 gives robustness to the drive system with pulsating torque and
wind gusts. The machine uses direct vector control with stator flux orientation and
synchronous current control, as explained by the phasor diagram in Figure 12.17,
where the torque component of current IT is aligned in phase with the stator voltage
Vs and the stator‐fed magnetizing current IM is aligned in the direction of stator flux
Ψs, as shown. The command IM* is generated from the stator flux Ψs control loop,
where the flux command s* is generated from torque command through the flux
program, as indicated, for light‐load efficiency improvement. The flux command
curve is prior‐computed by the machine analysis.
Following the phasor diagram in Figure 12.17, the following equations can be
derived:
1
ids ia sin e ia 2ib cos e (12.33)
3
1
iqs ia cos e ia 2ib sin e (12.34)
3
vq b qs b lqsiqs (12.35)
qe
deʹ
Vf
B
ψqs
ψs
qeʹ Is δ
Vs ψa
iqs
ϕ C δ
IT Magnetic flux
IM ψf
–ids de
D θe ψds A
ds
ψs
ψa
Figure 12.17 Phasor diagram of IPM machine with vector control.
vd b f ldsids (12.36)
vq
sin (12.37)
ˆ
b s
vd
cos (12.38)
ˆ
b s
ˆs vd 2 vq 2 (12.39)
b
cos e cos e cos sin e sin (12.40)
sin e sin e cos cos e sin (12.41)

3 p 1
te vd iqs vq ids (12.42)
2 2 b
iT iqs cos ids sin (12.43)
iM iqs sin ids cos (12.44)
where ia, ib = machine phase currents,

θe= field flux angle with ds axis
ids, iqs = de‐qe axes currents
Ψf = field flux
Ψds,Ψqs = de‐qe axes stator fluxes
ωb = base speed
δ = machine torque angle

cos( e ), sin( e ) = unit vectors for vector control
P = number of machine poles
Lds, Lqs = machine stator inductances.
There are three fuzzy controllers (FLC‐1, FLC‐2 and FLC‐3) in the control
system which will be described later. The line‐side converter in Figure 12.15 is also
vector‐controlled with synchronous current control in order to control the line
active (IP) and reactive (IQ) currents to control the active (P0) and reactive powers
(Q0), respectively. The P0 that corresponds to the generated power is controlled to
regulate the DC‐link voltage Vd by close loop, as shown in the figure. The tight con-
trol of Vd requires feed‐forward injection of DC‐link power Pd in the P0 control
loop, as indicated. The following equations can be derived from Figure 12.18,
where IP and v̂s are oriented to the de axis, and IQ (leading) is oriented to the qe axis:
vqss
sin (12.45)
vˆ
e
s
vdss
cos (12.46)
vˆs
e

iP ids iqss sin e idss cos e (12.47)
iQ iqs iqss cos e idss sin e (12.48)
p0 3vs i P (12.49)
Q0 3vs i Q (12.50)
The wind generation system in Figure 12.15 uses three fuzzy controllers
(FLC‐1, FLC‐2 and FLC‐3), as shown, with Mamdani‐type FIS in all of them. The
qe
iqs = IQ
θe
iqss vqss
qs
θe ids = IP
idss
vdss Vˆ s = Vds
de
ds ωe
Figure 12.18 Phasor diagram for line‐side converter with vector control.

Rule
matrix
ωr
∆P0 +
∆ω*r ω*r
+
Membership
functions
Z−1
L∆ω*r
Figure 12.19 Fuzzy MPPT control principle (FLC‐1).
F
-1
Po jump by FLC FLC-2
Line power (Po)
Vw change E
I
D G Vw1
H FLC-1
C
FLC-2 Vw2
-1 B Vw3
FLC Wind
A
Vw4 velocity
ωr1 ωr2 ωr3 ωr4

Generator speed (ωr)
Figure 12.20 Output power maximization of wind generation system with fuzzy MPPT‐1
(or FLC‐1) and MPPT‐2 (or FLC‐2) controls.
FLC‐1 is an online‐search based MPPT (maximum power point tracking) (MPPT‐1)

control, where optimum turbine speed (ωr) is searched for maximum line power
(P0) at constant Vw so that the turbine operates at maximum aerodynamic efficiency.
This was explained in Figure 12.16. A simplified fuzzy MPPT control principle is
explained in Figure 12.19, where the inputs are incremental power ΔP0 and last
increment/decrement of Δωr (L Δωr), and the output is the command speed
increment Δωr*. This is added with the previous ωr signal to obtain the command
speed r* . The FLC‐2 operates on the same (MPPT‐2) principle to set the optimum
light‐load flux command s* as a function of te* for best light‐load generator
efficiency. The operations of FLC‐1 and FLC‐2 are explained on the power–speed
curves of the generator shown in Figure 12.20. Assume that initially the wind
velocity is Vw4 and generator speed is ωr1, which gives the operating point A on the
power curve. The FLC‐1 control brings the operating point to B. Then, FLC‐2
control will further increase the output power to point C at the same Vw4 and ωr2. If
Vw now increases to Vw2, the operating point jumps to D. Further operation of the
fuzzy controllers will bring the operating point to F. The effect on decrease of Vw
is also shown in the figure. The FLC‐3 is basically robust fuzzy PI control of the
generator speed using the principle explained in Figure 12.9.
12.4 NEURAL NETWORKS
The principles and applications of neural networks will be discussed in this

section.
12.4.1 Neural Network Principles

Among all the AI techniques, neural networks [9, 34–36] are most important, and
therefore their principles, characteristics and applications will be discussed in some
detail. In fact, modern AI technology is generally synonymous with neural network
applications. It is often said that the invention of neural networks is as significant as
the invention of the transistor. There are predictions that AI technology will signif-
icantly impact our future civilization. The application potential of neural networks
appears to be tremendous. We have hardly scratched the surface because we do not
understand it well, and suitable microchips are not yet available for economical
applications.
What is a neural network? A neural network (also called neurocomputer or
connectionist system) is the most generic form of AI for emulation of human
thinking, compared to rule‐based ES or FL, which emulate the behavioral nature of
human thinking. A neural network tends to mimic the biological neural network of
a human brain. The structure of our nervous system is not well understood, but it is
said to contain around 100 billion nerve cells or biological neurons. Although
biological neuron structure is known, the way they are interconnected is not well
known. Our memory and intelligence, and the corresponding thinking process, are
generated by the action of a biological neural network. An artificial NNW tends to
emulate a biological nervous system in an approximate way by using an electronic
circuit or computer program. It helps us to solve pattern recognition and image
processing type problems, which are difficult to solve by conventional methods.
The core of neurocomputation is nonlinear multidimensional input–output
mapping (see Figure 12.12) or pattern recognition by auto‐associative memory
property of human brain. It is the property by which we recognize a person when
we see their face, or recognition of an alphabet character by a child. Another prop-
erty is pattern classification by which we can differentiate a Chinese face from an
African face by looking at the skin color and facial patterns. This pattern recogni-
tion or classification property can be obtained by learning or supervised training
through input–output example data‐sets, which will be discussed later.
Before we proceed to neural network principles and applications, it would be
interesting to touch upon the historical evolution of this important technology. In
12.4 NEURAL NETWORKS 651
fact, neural network evolution pre‐dates modern digital computers and other AI
techniques. The history essentially started with the pioneering work of McCulloch
and Pitts in 1943, who were trying to model the nervous system of the human brain
by binary valued artificial neurons that were capable of performing simple threshold
logic computations. McCulloch was a psychiatrist and neuroanatomist, whereas
Pitts was a mathematician. The next major development in neural networks came in
1949 when Hebb proposed his synaptic learning mechanism of biological neurons
in his book The Organization of Behavior. The book had immense influence among
psychologists, but was practically ignored by the engineering community. In the
1950s, the dominant personality in neural network research was psychologist Frank
Rosenblatt of Cornell Aeronautical Laboratory, who invented Perceptron (see
Figure 12.23) and its learning theory to model a biological sensory organ, such as
an eye. In 1960, Widrow and Hoff proposed Adaline and Madaline. Unfortunately,
the lack of expected performance of these networks, coupled with the glamour of
the von Neumann digital computer in the late 1960s and 1970s, practically camou-
flaged the neural network evolution. Many researchers, except those in psychology
and neurosciences, practically deserted the field almost for a decade. Of course,
during the 1970s and early 1980s, attention was diverted for extensive research of
expert system techniques and their applications, as mentioned before. The modern
era of neural networks with rejuvenated research was started by Hopfield in 1982,
a professor at CalTech, who presented his research at the National Academy of
Science. Since the beginning of the 1990s, neural networks have captivated the
interest of a large segment of the scientific community. It is amazing to see how the
results of research mainly from psychologists, psychiatrists, neuroscientists and
mathematicians can be of such importance to the engineering community.
Biological and Artificial Neurons

A NNW consists of a cluster of artificial neurons interconnected together. The
structure of an artificial neuron is inspired by the concept of a biological neuron, as
shown in Figure 12.21a. A neuron is the basic processing element (PE) in the ner-
vous system of the brain which receives and combines signals from other similar
neurons through thousands of input paths called dendrites. Each input signal
(electrical in nature) flowing through a dendrite passes through a synapse or syn-
aptic junction. The junction gap is filled with neurotransmitter fluid which either
accelerates or retards the flow of the signal. The signals are then summed up in
nucleus and nonlinearly modified at the output before flowing to other neurons
through axon, as shown in the figure. The adjustment of the impedance of the syn-
aptic gap by the neurotransmitter fluid contributes to the “memory” or “intelli-
gence” of the brain. According to the theory of the neuron, we believe that our brain
has distributed associative memory or intelligence characteristics which are con-
tributed by the synaptic junctions of the cells. This means that we do not have a
computer‐like centralized memory.
The model of an artificial neuron that closely matches the biological neuron
is shown in Figure 12.21b. Basically, it is an op‐amp summing structure. The
(a)
Dendrites
Nucleus
Axon
Synapse
(b)
Bias source
X1 Synaptic +1.0
weights
Wb
X2 W1
b
W2 F(S)
N S F(S) Neuron
Inputs X3 W3 S = ΣXKWK + b
output Y
K= 1
WK Activation
Summing function
XK
WN node (sigmoidal)
XN
Figure 12.21 (a) Biological neuron structure; (b) artificial neuron model.
artificial neuron is also known as a processing element (PE), neurode, node, or cell.
Each input signal (continuous variable or discrete pulses) flows through a gain or
weight (called synaptic weight or connection strength) which can be positive (exci-
tory) or negative (inhibitory), integer or noninteger. The summing node accumu-
lates all the input‐weighted signals, adds to the weighted bias signal b, and then
passes to the output F(S) through the nonlinear (or linear) activation or transfer
function (TF), as shown in the figure.
Some common type activation functions are shown in Figure 12.22. These
are defined, respectively, as linear bipolar, threshold, signum, sigmoidal (or log‐
sigmoid), and hyperbolic tan (or tan‐sigmoid). Another type is Gaussian, which is
not included here. The magnitude of these functions varies between 0 and 1, or −1
to +1, as indicated. The linear function can be unipolar or bipolar. The sigmoidal
and hyperbolic tan functions are commonly used because they are differentiable
(a) (b) (c)

F(S) F(S) F(S)
+1 +1 +1
Bias
0 S 0 S 0 S
–1 –1
Linear Threshold Signum
(d) (e)
F(S) F(S) α =1
+1 +1
1 α=1
F(S)= eαS – e–αS α =2
1 +e–αS 0.5 F(S) = α = 0.5
eαS + e–αS
0 S 0 S
–1
Sigmoidal Hyperbolic tan

Figure 12.22 Activation (or transfer) functions of artificial neuron.
and slope or gain can be adjusted by varying the coefficient α. All the functions are
squashing in nature. The nonlinear activation functions of neurons permit nonlinear
input–output mapping of NNW.
Neural Network Structures

The objective of NNW applications is to solve scientific, engineering, and many
other real‐life problems as human beings solve them using the a biological neural
network. As mentioned before, scientists understand well the structure of individual
neurons, but their interconnection is not understood well. Therefore, many models
have been proposed in the literature. A model is useful only if it can solve problem(s)
and it can be trained properly. Neural networks are generally classified as feed‐
forward and feedback (or recurrent) types, some of which are listed below [9]:
Feed‐forward types:
•• Perceptron
•• Adaline and Madaline
•• Backpropagation (BP) Network
•• Radial basis function network (RBFN)
•• General regression network (GRN)
•• Modular neural network (MNW)
•• Learning vector quantization network (LVQN)

•• Probabilistic neural network (PNN)
•• Fuzzy neural network (FNN)
Recurrent types:
•• Real‐time recurrent NNW
•• Kohonen’s Self‐Organizing Feature Map (SOFM)
•• Adaptive resonance theory (ART) network
•• Elman network
•• Recirculation network
•• Brain‐state‐in‐box (BSB)
•• Hopfield network
•• Boltzmann machine
•• Bi‐directional associative memory (BAM) network
In the feed‐forward class (see Figure 12.23), the signals flow only in the forward
direction, whereas in a recurrent neural network (RNN), the signals can flow in for-
ward as well as in backwards or lateral directions. A network can be defined as
static or dynamic, depending on whether it emulates a static or dynamic system. For
static input–output mapping, feed‐forward topologies are important, whereas for
dynamic temporal mapping, the RNNs are important. A description of all the topol-
ogies and their applications is beyond the scope of this chapter. Currently, around
90% of applications use the feed‐forward BP‐type networks (see Figure 12.24), and
therefore this topology will be emphasized.
Perceptron Network
A perceptron network [8] is a simple, single‐layer, parallel feed‐forward cluster of
neurons which is used for classification of input signal patterns that are linearly
separable. The general structure of a perceptron network (multiple perceptrons)
with a number of neurons S that use unipolar threshold (hard‐limit) TFs is shown in
Figure 12.23a, while Figure 12.23b shows the performance for the upper percep-
tron only. The general parallel input signal vector p has R elements that are multi-
plied by weight matrix W (W1,1, W2,1, etc.) given by (RXS) before connecting to the
summing junction. The resulting Wp vector with the added bias b constitutes the
vector given by n = Wp + b. Each bias signal is generated from +1 source through a
weight (b1, b2, etc.), as shown. The output vector a is then given by a = hardlim
(Wp + b). The lower figure shows the output with p1 and p2 inputs only, when
W1,1 = −1 and W1,2 = +1. The shaded area classifies p1 and p2 loci for n1 > 0 and the
unshaded area for n1 < 0. The classification boundary is given by line L = Wp + b = 0.
The bias signal shifts the line L horizontally, whereas the input weights change its
slope. For more than two input signals, the classification boundary is given by a
hyperplane. The boundary hyperplane of each perceptron can be designed by its
(a)
W1,1 n1 a1
∑
p1
b1
p2 1
n2 a2
∑
p3
• • b2 • •
• • • •
• • 1 • •
pR
nS aS
∑
WS,R
bS
1
(b)
p2 L
W Wp + b > 0
+1 a=1
–b/w1,1 Wp + b = 0
a=0
+1 p1
–1
Wp + b < 0
–b/w1,2 a=0
–1
Where... w1,1 = –1 and b = +1
w1,2 = +1
Figure 12.23 (a) Perceptron network model structure; (b) explanation of the pattern
classification property for upper neuron only.
weights and bias. For the desired classification boundaries, the perceptron weights
can be designed by a supervised computer training algorithm (perceptron learning
rule). The training of NNW will be discussed later. A simple application of a per-
ceptron is to identify the response charactetistics of a drive system when the time‐
delayed samples of the response curve (see Figure 12.27) are placed at the inputs of
a perceptron. Another simple application is solving a Boolean function [8].
The Adaline or Madaline (multiple Adalines) is another feed‐forward NNW
that has the same structure as the perceptron except its TF is bipolar linear. The
linear network can give only linear input–output mapping. In addition to its use for
pattern classification, it can be used for linearizing nonlinear functions (linear
function approximation) or pattern association.
Backpropagation Network
Figure 12.24 shows the topology of the most commonly used feed‐forward multi-
layer backpropagation (backprop or BP) network. The name “backpropagation”
Bias source
+1.0
Hidden
layer (b)
Link
4 V4
Input W14 W49 Output
layer (a) V5 layer (c)
X1 1 5 9 Y1
V6 Output
Input X2 2 6
Descaling or
V7 normalization
X3 3 7 10 Y2
Node
V8 Weight
Scaling or W38 Actual
normalization 8 W8 10 output
Y1 Y2 pattern
Weight E
adjustment
algorithm
Error Desired
Z1 Z2 output
pattern
Figure 12.24 Feed‐foward neural network structure showing backpropagation training.
comes from the method of supervised training for the NNW shown by the lower
two blocks in the figure, which will be discussed later. The network is commonly
called multilayer perceptron (MLP), although the activation function can be differ-
ent from the threshold function. The circles represent the neurons which have an
associated TF (not shown in the figure). The dots indicate the synaptic weights of
the neurons which are often omitted for simplicity. The figure shown has three
layers of neurons: an input layer, a hidden layer and the output layer. With five neu-
rons in the hidden layer as shown, it is defined as a 3‐5‐2 network. The input layer
of neurons essentially distributes the signals to the hidden layer. Therefore, this
topology is often called a two‐layer network. If the signals are bipolar, the hidden
layer neurons have hyperbolic‐tan TF, and the output layer has bipolar linear TF.
Otherwise, for unipolar input signals, these TFs can be sigmoidal and unipolar,
respectively. Sometimes, the output layer uses nonlinear TF also. The NNW nor-
mally processes the signals in per unit (pu) manner. Therefore, the input signals
scaling (or normalization) and output signals descaling (or denormalization) are
used. A fixed bias signal normally couples to all the hidden layer and output layer
neurons through weights, but the output layer coupling is not shown in the figure.
The NNW input and output signals can be continuous, logical unipolar or logical
bipolar. Although a majority of applications use three layers of neurons, sometimes

more than three layers are used.
The artificial NNW, as shown in Figure 12.24, has analogies with a biological
neural network. Similar to a biological network, where intelligence or memory is
contributed to in a distributed manner by the synaptic junctions of neurons (see
Figure 12.21a), the NNW synaptic weights contribute to distributed intelligence.
This intelligence permits the basic input–output nonlinear mapping or pattern rec-
ognition property of a NNW. This is the associative memory property, which is
similar to that of a biological nervous system.
A NNW can be implemented by serial or parallel‐connected DSPs, FPGAs
(field‐programmable gate arrays), or application‐specific ICs (ASIC chips).
Massive parallel computation contributes not only high‐speed execution but also
fault tolerance and noise immunity of signals.
One unique characteristic of NNW is that it generally requires supervised
backpropagation training with example input–output data‐sets instead of traditional
programming for solving a problem in a computer. This is analogous to supervised
alphabet training (or learning) of a child with the help of a tutor. It may be inter-
esting to note that backpropagation theory was originally proposed by Paul Werbos
in his doctoral thesis at Harvard University in 1974, but it went unnoticed by the
scientific community for almost a decade. In 1986, the modern backpropagation
supervised training algorithm was reported by Rumelhart, Hinton and Williams,
and has been used extensively for feed‐forward multilayer‐perceptron (MLP)‐type
networks. This type of network is the most commonly used in recent years, as men-
tioned before.
Backpropagation Training Method

The backpropagation training method of NNW is shown by a flowchart in
Figure 12.25. In the beginning, input–output example data patterns are obtained by
simulation with the mathematical model of the plant, if such a model is available.
An initial NNW configuration is created by the desired input and output layer neu-
rons dictated by the number of signals, a hidden layer with a few neurons, and
appropriate TFs. Small random weights are selected to prevent saturation of sig-
nals. With one input signal pattern, the output is calculated and compared with the
desired output pattern. The weights are then altered in the backward direction by a
backpropagation algorithm (gradient descent technique) until the error between the
calculated pattern and the desired pattern is very small and acceptable, which is
determined by the equation:
p Q
2
SSE dj p y j p (12.51)
p 1j 1
where SSE is the total sum of squared error, P is the number of training patterns, Q
is the output layer neurons, djp is the desired output of the jth neuron of the output
layer, and yjp is the corresponding actual output.
Get input-output example

data patterns from experimental
or simulation results
Select ANN topology with no. of

layers, nodes, and activation function
Initialize with
random weights
Select an
Change number of input-output pattern
neurons in hidden
layer or no, of layers
Calculate output
and compute error
Is error N Change weights

acceptable? by training algorithm
Y
Train network with
other patterns
N Is error
acceptable?
Y
Test network
performance
Network is
ready for use
Figure 12.25 Flowchart explaining the backpropagation training algorithm of a feed‐
forward neural network.
A number of BP algorithms have been proposed, but the L‐M (Levenberg‐

Marquardt) algorithm is frequently used because of its rapid convergence. Once the
NNW is trained properly, it should be adequately tested with intermediate data to
verify the quality of the training before acceptance. Normally, a computer program
(such as MATLAB‐based Neural Network Toolbox [38]) performs the training to
determine the number of hidden layer neurons and the network weights.
Real‐time Recurrent Network

In the previous sections, we have discussed the feed‐forward neural networks that
give only static nonlinear input–output mapping. In many applications, a neural
network is required to emulate a dynamic system with temporal behavior, such as
identification of a machine model. Figure 12.26a shows a recurrent or feedback

network with time delays Z–1 that can emulate a dynamic system. The output of the
network not only depends on the present input, but also prior inputs, thus giving
temporal behavior of the network. The network can emulate nonlinear differential
equations that are characteristic of a nonlinear dynamic system (such as a dynamic
model of an induction motor).
The model of the network can be given as:
Q1 w11 . . wn m 1 U1
. . . . . .
(12.52)
. . . . . .
Qn w1n . . wn m n
Un m

yj k 1 f Q j k (12.53)

where Q = neuron output before the TF
F(Q) = neuron activation function
N = number of output neurons
M = external input connections
M + N = total network inputs
X(k) = external input vector
Y(k + 1) = vector neuron output at one step later at discrete time (k + 1)
W’s = weights of neurons that define the parameters of the system
The network can be trained by a dynamic backpropagation (real‐time
temporal supervised learning) algorithm, where the desired time domain output
from the reference dynamic system can be used to force the network to track by
tuning the weights dynamically sample by sample, as indicated in Fig.12.26b. The
rea‐time simulation of an IPM synchronous generator in Figure 12.15, for example,
can be made by the real‐time recurrent network type of Figure 12.26a.
Time‐delayed Network
In addition to the RNN structure of Figure 12.26a, a time‐delayed neural network
(TDNNW), as shown in Figure 12.27, is also popular for identification and control
of a dynamic system. In this case, a single input X(k) is fed to a multiple‐input static
network through a tapped delay line. The delay line generates a sequence of signals
with unit time delay (Z−1). The signals are multiplied by the respective weights
within the ANN and then generate the output Y(k) through the TF as shown in the
following equation:
n
y k f wnk X k n (12.54)
n 0
(a) Delayed feedback
Z–1
•
•
•
Z–1
1 W11
Y1(k) • Q1
• W1N 1 F(Q1) Y1(k+1)
YN(k) •
N • TF
• Output
• Signals
X1(k) 1
Intput • QN
signals • Y(N+M)1 N F(QN) YN(k+1)
•
XN(k) M TF
Y(N+M)N
(b)
X(k) y(k)
PLANT
+
y′(k)
ANN
–
e(k)
Figure 12.26 (a) Real‐time recurrent network structure; (b) block diagram for training.
where wnk w0 k ,w1k wnk is the network weight vector and F is the
activation function of the neuron. The network can be trained by BP algorithm
as discussed before.
Inverse Model Identification of Plant by Neural Network

Previously, we discussed identification of the dynamic model of a plant with the
help of a recurrent neural network. It is also possible to identify the inverse plant
model by training of a RNN, as shown in Figure 12.28a. In this case, the plant
response Y(k) is impressed at the input of the NNW and its calculated output is com-
pared with the plant input, which is the data of the target. The resulting error ξ(k) is
X(k)
Z–1
X(k–1)
Z–1 STATIC
ANN Y Y(k)
X(k–2)
Z–1 F(·)
X(k–3)
Z–1
X(k–N)
N
Y(k) = F Σ WnkX(k–n)
n=0
Figure 12.27 Time‐delayed network with tapped delay line.
(a)
F(·)
X(k) Plant y(k)
+
ξ(k)
– F–1(·)
Inverse
model
(b)
F–1(·)
ANN Main signal
inverse
model
u’(k)
F(·)
+
e + u(k)
r(k) G Plant y(k)
+
–
Supplemental
signal
Figure 12.28 (a) Training of inverse dynamic model of a plant; (b) inverse dynamic
model‐based adaptive control of a plant.
used to train the network so that the error becomes reduced to the acceptable
minimum value. After satisfactory training and testing, the NNW represents the
inverse dynamic model of the plant. This model can then be placed in series as the
actual controller of the plant shown in Figure 12.28b. The inverse model ideally can-
cels the forward model, i.e., F−1(.) F(.) = 1. Therefore, the output signals follow the
input signals and no feedback control is necessary. However, in practical cases, the
output will deviate from the input because the inverse model may not be perfect, and
there may be parameter variation effect. Therefore, a feedback loop is added where
the supplemental error signal e is generated as shown. The signal e can also be used
for online tuning of the NNW inverse model.
12.4.2 Neural Network Applications

Before we proceed to NNW applications in power systems, let us discuss its appli-
cation in a simple non‐engineering optical character recognition (OCR) of an
alphabet [35] to clarify the ideas. Figure 12.29 illustrates mapping of the letter “A”
by the 5‐bit binary code (10101) using a BP NNW. However, the general problem
is to represent all the 26 alphabet characters by a 5‐bit binary code. In the figure, A
is represented by a 5 × 7 matrix of inputs consisting of logical 0 s and 1 s, where the
shaded squares correspond to 1 and unshaded squares represent 0. The input vector
of 35 logical signals is connected to the input of a 35‐20‐5 BP network without any
bias signals. The NNW uses sigmoidal TFs in all the neurons and the output is
clamped to logical 0 or 1. The input–output mapping is peformed by supervised
learning of all the weights. The NNW has altogether 800 weights, which give it 800
degrees of freedom for pattern mapping.
The letter B is to be coded by 10001. After training for A, if B is given as the
input, the NNW output will be totally distorted. Another round of training tunes the
Input Output
1 0 Hidden layer
1
0 Binary
35 lines 1 code
10101
0
1
7×5 35-20-5 5 ×1
Matrix Feed forward neural Matrix
network
Figure 12.29 Neural network‐based pattern recognition principle by alphabet character
identification.
weights so that B gives the desired code 10001. However, if A is impressed again,
the output will be distorted. The NNW is trained back and forth with all the 32 input
vectors and the corresponding binary codes so that each vector generates the
corresponding code after successful training. Evidently, the nonlinearity of the
NNW with so many degrees of freedom and logical clamping at the output permits
such recognition of patterns. This example gives a clear idea of pattern recognition
properties of NNWs. Note that inverse mapping of the codes to generate the
corresponding input character by a NNW is also possible, as shown in Figure 12.30a.
(a) 1 0
5-20-35
5×1 Feedforward 7×5
Matrix neural Matrix
network
(b) Distorted Clean

A A
Distorted 35-20-35
7×5 Neural 7×5
Matrix network Matrix
Figure 12.30 (a) Inverse mapping of letter “A”; (b) autoassociative mapping of “A”.
It is also possible to cascade these NNWs to recover the original character (i.e.
35‐20‐35 NNW topology) (Figure 12.30b). The advantage of these heteroassocia-
tive or autoassociative NNWs is that if the input pattern is distorted, the output
pattern will be clean. This is the noise filtering property of NNWs.
Three‐Phase Sine Wave Generation

Let us now consider a simple example application of NNW for generation of three‐
phase sine waves in some detail. Figure 12.31a shows the network that generates
the sine functions from the input signal X in radians that varies from 0 to 2π. The
output peak amplitude is set to 1 for normalized data processing. The network uses
a three‐layer BP NNW with the topology of 1‐5‐3, using tan‐sigmoid TF in the
hidden layer and linear bipolar TF in the output layer. For generality, a bias of +1 is
supplied to the hidden and output layers, as shown. The training data of X (with a
step size of 0.1 radian) and the corresponding output target data were generated by
MATLAB sine functions. The input–output data file in MATLAB was imported to
a GUI (graphical user interface) for training. The network was trained by MATLAB‐
based Neural Network Toolbox [38] using an L‐M algorithm. After training the
network, it was converted to MATLAB‐Simulink program for testing. A sawtooth
wave with X was impressed at the input and the corresponding three‐phase waves
were generated at the output. The training error converged to the desired goal of
0.001 after 178 epochs (number of forward and reverse pass cycles). The final
matrices of weights and biases are given below:
w111 0.9746 b111 4.9313

w121 1.0423 b121 6.5538
w1 w131 0.7937 b1 b131 2.5381
w141 0.9571 b141 1.1780
w151 0.9504 b151 0.1443
w112 w212 w312 w412 w512

w2 w12 2 w22 2 w32 2 w42 2 w52 2
w132 w232 w332 w432 w532
2.7096 4.1398 9.5330 1.4385 5.3972
8.2493 2.1779 4.4349 7.7484 8.3220
5.5379 6.3177 5.0981 9.1869 2.9248
b112 3.4111
2
b b12 2 10.8206
b132 14.2317
(a)
+1
BIAS
B1 B2
1
W2
W1
2 1 Y1 = AsinX
X = (0–2pi)
INPUT 2 Y2 = Asin(X–2pi/3)
SIGNAL 3
3 Y3 = Asin(X+2pi/3)
4
OUTPUT
SIGNALS
5
(b)
IW{1,1} LW{2,1}
b{1} b{2}
1 5 3
(c)
Figure 12.31 Neural network training example for three‐phase sine wave generator.
(a) NNW topology; (b) symbolic form in the toolbox; (c) output waves after training.
Figure 12.31b shows the NNW in symbolic form and Figure 12.31c shows the
output waves. Note that the same functions could also be generated by DSP‐based
look‐up tables, but the NNW has the advantage of precise interpolation.
Neural Network Applications in Renewable Energy Systems and Smart Grid

Among all the AI techniques, NNW appears to have the maximum potential for
application in renewable energy systems and smart grid. Currently, our knowledge‐
base in these applications is very inadequate. Exploring the full potential of NNW
applications in these areas will take a prolonged effort by many scientists. Some of
these applications can be mentioned as follows [39, 40]:
•• Consumer load forecasting on the grid
•• Forecasting of wind and PV generation curves
•• Online fault diagnostics and fault‐tolerant control of power system
•• Sensorless robust estimation of feedback signals
•• Noise and delay‐free filtering of signals
•• Neural network modeling of static and dynamic system elements and real‐
time simulation by DSPs/FPGAs/ASIC chips
•• Intelligent scheduling of generation and storage
•• High‐performance intelligent control of system elements
•• Real‐time pricing predictions of electricity with demand‐side management.
In fact, NNW‐based AI can be applied to every phase of the smart grid for
performance improvement. Let us now discuss some specific application
examples.
Health Monitoring of Wind Generation System by ANFIS

What is ANFIS? ANFIS [41], or “adaptive neuro fuzzy inference system” is
basically a fuzzy inference system implementation by a NNW. As mentioned
before, both FIS and NNW are basically multidimensional, nonlinear input–output
mappings. A NNW uses input–output example data‐tables to train a network,
typically by a supervisory backpropagation technique. On the other hand, a fuzzy
system MFs and rule tables are usually determined by trial and error that may be
time‐consuming. Therefore, a fuzzy system for the desired input–output mapping
can be trained quickly by neural network training methods.
Both Mamdani and Sugeno FIS were discussed before. The Sugeno (zero‐
order and first‐order) methods are usually popular in ANFIS, although higher‐order
systems are also possible. In this section, as shown in Figure 12.32, we will review
a simple first‐order ANFIS with only two input variables (X and Y) and one defuzzi-
fied output (F). Figure 12.32a shows the FIS with triangular MFs (A1, A2, B1 and
B2) for the antecedent part of the rules, and f1 and f2 are the output singleton MFs.
Being a first‐order system, f1 and f2 (singleton bases) vary linearly with the input
(a)
Singleton MFs
1 A1 1 B1
μA1(x) μB1(Y)
w1
f1 w1 w2
F= f + f F
A2 B2 w1+ w2 1 w1 + w2 2
1 1
Defuzzified
output
μA2(X) w2 Where
μB2(Y)
b f1 = A01+A11X+A21Y
X Y f2 f2 = A02+A12X+A22Y
a First-order
Takagi-Sugeno method
(b)
Backpropagation
f1 algorithm
μA1(x)
A1
W1 W1
X Π N Π
W1f1 e
μB1(Y) W1
A2
Input Σ
variables W2 f2 F –
B1 output +
μA2(X)
W2 W2 W2f2 Fd
Y Π N Π
AND Normalizer Desired
B2 output
μB2(Y)
MFs
Layer 1 Layer 2 Layer 3 Layer 4 Layer 5
Figure 12.32 (a) Sugeno (first‐order) fuzzy inference system with two rules; (b) ANFIS
structure for (a).
variables X and Y, as shown by the equations on the right of the figure. The two
rules which are considered in the FIS are
Rule 1 : IF X is A1 AND y is b1 , THEN Z f1 (12.55)
Rule 2 : IF X is A2 AND y is b2 , THEN Z f2 (12.56)
where f1 = A01 + A11X + A21Y
f2 = A02 + A12X + A22Y
and A01, A11, etc. are the coefficients.
w1 f1 w2 f2 Z1. dof1 Z 2 . dof2

The defuzzified output f (12.57)
w1 w2 dof1 dof2
where DOF1 = W1 and DOF2 = W2.

The corresponding ANFIS structure is shown in Figure 12.32b, where the

functions A1, A2, B1, B2, f1 and f2 are tuned by the backpropagation algorithm. The
feed‐forward NNW has five layers which can be summarized as follows:
1. Layer 1 – Generate the membership values from the two rules
A1 X , A2 X , b1 y , b2 y
2. Layer 2 – Generate the firing strength by multiplication (π) or AND

operation
w1 A1 X . b1 y
and w2 A2 X . b2 y
3. Layer 3 – Normalize the firing strengths
w1
w1
w1 w2
w2
w2
w1 w2
4. Layer 4 – Calculate the rule outputs by multiplying with the consequent
parameters
w1 f1 w1 . f1
and w2 f2 w2 . f2
5. Layer 5 – Sum all the component outputs
f w1 f1 w2 f2
Note that each triangular MF again has two unknown parameters a (peak) and
b (support), as indicated in the figure. The equation of a function in terms of
a and b parameters is:
X a
X 1 , if X a 0.5b, 0 else (12.58)
0.5b
This means that each MF is tuned by the parameters a and b.
In the training, as shown in the figure, the calculated output F of the network
is compared with the desired value Fd, and the error signal e is used to train the net-
work parameters by the backpropagation algorithm. The MATLAB‐based Fuzzy
Logic Toolbox can be used to design an ANFIS. Usually, f1 and f2 are trained first,
which require determining all the coefficients of f1 and f2. This is then followed by
the design of the a and b parameters of the triangular MFs. Other types of MFs,
such as Gaussian, can also be considered. A Gaussian‐type MF is particularly con-
venient because it is continuous, differentiable, and always has some finite value.
Instead of a first‐order Sugeno system, a simpler zero‐order can also be considered.
The principle of ANFIS in Figure 12.32 has been illustrated with only two input and
one output signals. For a practical application, the ANFIS can have a large number
of input and output signals.
Let us now consider application of the ANFIS given in Figure 12.32b for
health monitoring of a wind generation system. A practical wind generation system
(see Figure 12.15) can have a large number of system signals for monitoring and
protection. The signals may be classified as follows:
Wind signals:
Velocity
Wind direction
Turbulence
Turbine signals:
Blade speed
Shaft speed
Pitch angle
Pitch angle control signal
Bearing temperatures
Vibration of blade
Yaw angle
Shaft torque
Mechanical brake signal
Tip speed ratio
Gear box:
Oil temperature
Oil viscosity
Noise intensity
Vibration
Nacelle temperature
Generator:
Bearing temperatures
Shaft vibration
Stator winding temperatures
Rotor magnet temperatures

Shaft torque
Stator voltages
Phase sequence
Stator currents
Stator frequency
Active power
Reactive power
Converter:
Converter temperatures
Cooling fluid velocity
DC‐link voltage
DC‐link current
DC‐link power
AC line voltages
Output frequency
Phase unbalance of voltages
AC line currents
Phase unbalance of currents
Active power
Reactive power
Motoring/regeneration mode
The signals can be monitored with the help of sensors (or sensorless estimation) to
determine the general health of the wind generation system. Although the system in
Figure 12.15 is gearless, the gearbox signals have been included for generality. All
the signals vary within certain ranges. The health condition or health index of the
system can be defined as “excellent” if variation of the signals remains confined
within a highly satisfactory range. If some signals go beyond this range, but are yet
very safe, the system can be defined as “very good”. Similarly, for other ranges, the
health index can be classified as “good,” “fair,” “poor,” “unsafe,” and so on. If some
signals degrade, the diagnostic messages for the signals can be generated indepen-
dently. If any signal goes beyond the safe range into a fault condition, the system
can be shut down for protection. Fault‐tolerant control can be implemented, as
desired. Figure 12.33 shows the ANFIS implementation of the health monitoring
system. Similar health monitoring principles can be extended to PV or other sys-
tems also. Once the system design is complete, it can be implemented by DSPs or
FPGAs, as previously discussed.
Wind
signals Trip signals
Turbine
signals
Excellent
Unsafe
Gear box
signals Very
Poor
Anfis good
Fair
Generator
signals
Good Health
index
Converter
signals
Diagnostic
messages
Figure 12.33 ANFIS‐based health monitoring for wind generation system (see

Figure 12.15).
ANN‐based Fault Pattern Identification of Smart Grid Subsystem

This application of NNW is concerned with the identification of fault and power
quality problems in a smart grid subsystem (or microgrid) as indicated in
Figure 12.34. The three‐phase voltage and current wave signals of a grid give the
footprint or characteristic signatures of different types of faults in a system. As
shown in the figure, the voltage and current signals are sensed and their space vec-
tors are computed. These space vectors are then analyzed to identify different types
of faults in the system. For example, in a healthy harmonic‐free balanced system
with the rated sinusoidal voltage and current waves, the space vectors are circles of
fixed diameters. For overvoltage or overcurrent conditions, the respective circle
diameters will increase proportionately. Similarly, for a decrease in voltage or
current, the respective circle diameter will decrease proportionately. Therefore, the
quantitative deviations from the rated values can be determined by measurement of
the circle diameter. For system imbalance or harmonics in the waves, the space vec-
tors will be distorted characteristically. For every abnormal condition in the system,
the space vector signals will have some distortion, the signature of which can be
analyzed to identify the fault condition quantitatively. The space vector patterns are
displayed on optical grids with m × n matrix, and the characteristic logic signal 1 s
(shaded) and 0 s (unshaded) can be generated similar to the alphabetic character
recognition system in Figure 12.29. These matrices of the characteristic logic
signals are impressed at the input of a feed‐forward neural network, as shown
in the figure. The NNW is trained with example data patterns to identify the
corresponding fault in the system. This is essentially the image processing or
pattern identification property of NNW discussed before. Some of the identified
faults are indicated in the figure. Fault‐tolerant control based on NNW can also be
designed and implemented in the system.
SMART GRID
SUBSYSTEM
va ia
vb ib
vc ic
OPTICAL
GRID
m×n
MATRIX
m×n m×n
Bias source
+1.0
Hidden
layer (b)
Link
4 V
Input W Output
W
layer (a) V layer (c)
Thd
X 1 5 9 Y
Line short
V Output Single phasing
Input X 2 6
Descaling or 3 ph short
V normalization
Over loading
X 3 7 10 Y
Over voltage
Node
V Weight
Scaling or W
W
Actual Under voltage
normalization 8 output
Y Y pattern Over current
Weight E
adjustment
algorithm
Error Desired
Z Z output
pattern
Feedforward
backpropagation
neural network
Figure 12.34 Fault identification of smart grid subsystem by neural network mapping.
12.5 CONCLUSION
The advent of powerful AI technology has opened a new frontier for modernization
of smart power grids with renewable energy systems. Generally, this is an ambi-
tious and challenging task that will take many years for implementation with the
integrated efforts of many scientists specialized in this area. It will help in meeting
the desired SG goals of system availability, reliability, power quality, energy
efficiency, and system security with optimum resource utilization and economical
electricity, as mentioned before. The chapter includes a brief and comprehensive
review of different elements of AI, such as ES, FL and NNWs. ES has been dis-
cussed extensively in the literature and applied widely in both industrial and non‐
industrial applications, although their application in power and power electronic
systems is not that common. The ES applications have been extended in this chapter
for the control of smart grid based on a real‐time simulator. Fuzzy inference sys-
tems have been considered in various applications. A modern PMSG‐based wind
REFERENCES 673
generation system has been discussed in detail, using FL‐based optimization of

wind turbine aerodynamic efficiency for maximum energy capture, generator
efficiency improvement at light load by flux programming, and robust turbine‐
generator operation with wind gust pulsating torque control. Among all the AI
techniques, the importance of NNW is most significant, and therefore it has been
discussed in some detail. The feed‐forward backpropagation network and its appli-
cations have been emphasized because of its wide popularity. Different applications
of NNW have been discussed for smart grid and renewable energy systems. These
include adaptive neuro fuzzy inference system (ANFIS) for health monitoring
of wind generation systems, and fault identification in a smart grid subsystem by
analyzing the voltage and current wave patterns. It is interesting to note that, among
all the AI techniques, NNWs are now getting maximum emphasis for future
applications. A new kind of industrial revolution is expected in the future [42, 43]
by widespread applications of AI with the advent of powerful and economical
application‐specific NNW microchips.
ACKNOWLEDGMENT
The author is grateful to Prof. Joao Pinto of Federal University of Mato Grosso Sul,
Brazil, and Prof. Marcelo Simoes of Colorado School of Mines, USA, for their help
with the contents of this chapter when they were doctoral students in my laboratory
at the University of Tennessee.
REFERENCES
[1] B. K. Bose. “Artificial intelligence techniques in smart grid and renewable energy systems – some
example applications,” Proc. IEEE (Special Issue on Power Electronics in Smart Grid and Renewable
Energy Systems, Guest Edited by B. K. Bose), vol. 105, no. 11, pp. 2262–2273, Nov. 2017.
[2] B. K. Bose. “Power electronics, smart grid, and renewable energy systems,” Proc. IEEE (Special
Issue on Power Electronics in Smart Grid and Renewable Energy Systems, Guest Edited by B. K.
Bose), vol. 105, no. 11, pp. 1997–2285, Nov. 2017.
[3] B. K. Bose. “Expert system, fuzzy logic, and neural network applications in power electronics and
motion control,” Proc. IEEE (Special Issue on Power Electronics and Motion Control, Guest Edited
by B. K. Bose), vol. 82, no. 8, pp. 1303–1323, August 1994.
[4] M. N. Cirstea, A. Dinu, J. G. Khor and M. McCormick. Neural and Fuzzy Logic Control of Drives
and Power Systems, Burlington, MA: Newnes, Elsevier, 2002.
[5] D. A. Waterman. A Guide to Expert Systems, Reading, MA: Addison‐Wesley, 1986.
[6] G. K. Venayagamoorty. “Potentials and promises of computational intelligence for smart grid,”
IEEE PES Gen. Meet. Conf. Rec., pp. 1–6, 2009.
[7] W. T. Zhang, F. Y. Xu and L. Zhou. “Artificial neural network for load forecasting in smart grid,”
Intl. Conf. Machine Learning and Cybernetics, pp. 3200–3205, 2010.
[8] B. K. Bose. Power Electronics and Motor Drives – Advances and Trends, Burlington, MA: Academic
Press, 2006.
[9] B. K. Bose. Modern Power Electronics and AC Drives, Upper Saddle River, NJ:
Prentice‐Hall, 2001.
[10] B. K. Bose. “Artificial intelligence techniques in smart grid and renewable energy systems,”
Keynote presentation in Qatar Workshop on Smart Grid and Renewable Energy (SGRE2015),
Doha, March 22, 2015.
[11] R. G. Vedder. “PC based expert system shells: some desirable and less desirable characteristics,”
Expert Syst., vol. 6, pp. 28–42, Feb. 1989.
[12] C. Ramos and C. C. Liu. “AI in power systems and energy markets,” IEEE Intelligent Systems, vol.
26, no. 2, pp. 5–8, March/April 2011.
[13] S. D. Ramchurn, P. Vytelingum, A. Rogers and N. R. Jennings. “Putting the smarts into the smart
grid: a grand challenge for artificial intelligence,” Communications of the ACM, vol. 55, pp. 86–97,
April 2012.
[14] C. Daoshen and B. K. Bose. “Expert system based automated selection of industrial ac drives,”
IEEE/IAS Annu. Meet. Conf. Rec., pp. 387–392, 1992.
[15] S. M. Chhaya and B. K. Bose. “Expert system aided automated design, simulation and controller
tuning of ac drive system,” IEEE IECON Conf. Rec., pp. 712–718, 1995.
[16] A. Ipakchi and F. Albuyeh. “Grid of the future,” IEEE Power and Energy Magazine, vol. 7, no. 2,
pp. 52–62, 2009.
[17] H. Gharavi and R. Ghafurian (Guest Editors). “Smart Grid: The Electric Energy System of the
Future,” Proc. of the IEEE Special Issue, vol. 99, no. 6, June 2011.
[18] X. Fang, S. Misra, G. Xue and D. Yang. “Smart grid – the new and improved power grid,” IEEE
Communication Surveys & Tutorials, vol. 14, no. 4, pp. 944–980, 2011.
[19] M. Simoes, R. Roche, E. Kyriakides, A. Miraoui, B. Blunier, K. McBee, S. Suryanarayanan, P.
Nguyen and P. Ribeiro. “Smart‐grid technologies and progress in Europe and the USA,” IEEE
Energy Conversion Congress and Exposition (ECCE 2011), pp. 383–390, 2011.
[20] S. F. Bush. Smart Grids, Piscataway, NJ: IEEE Press/Wiley, 2014.
[21] H. Abu‐Rub, M. Malinowski and K. Al‐Haddad (Eds). Power Electronics for Renewable Energy
Systems, Transportation and Industrial Applications,” Piscataway, NJ: IEEE Press/John Wiley,
2014.
[22] A. Ferreira, P. Leitao and P. Vrba. “Challenges of ICT and artificial intelligence in smart grids,”
IEEE Intl. Workshop on Intelligent Energy System, pp. 6–11, 2014.
[23] L. J. Ricalde, E. Ordonez, M. Gamez and E. N. Sanchez. “Design of a smart grid management
system with renewable energy generation,” IEEE Symp. on Comp. Intelligence Appl. In Smart
Grid, pp. 1–4, 2011.
[24] A. A. Moghaddam and A. R. Seifi. “Study of forecasting renewable energies in smart grids using
linear predictive filters and neural networks,” IET Renew. Power Gener., vol. 5, pp. 470–480,
April 2011.
[25] NIST Framework and Roadmap for Smart Grid Interoperability Standards, NIST, September,
2014.
[26] B. K. Bose. “Technology advances, applications and trends of power electronics and motor drives,”
Keynote presentation in Conference on Power, Control, Communication and Computational
Technologies for Sustainable Growth (PCCTSG2015), Kurnool, India, December 12, 2015.
[27] B. K. Bose. “Energy, environment and power electronics,” IEEE Industrial Electronics Society
Distinguished Lecture presentation, Colorado School of Mines, Denver, November 23, 2015.
[28] OPAL‐RT Technologies. “eMEGAsim: Real‐Time digital simulator for power system engineers.”
Available online at Opal‐rt.com, 2014.
[29] B. K. Bose. “Artificial Intelligence Techniques in Power Electronics and Motor Drives,” ECCE2011
Tutorial, sponsored by IEEE Power Electronics and Industry Application Societies, Phoenix,
Arizona, September 17–22, 2011.
[30] T. Takagi and M. Sugeno. “Fuzzy identification of a system and its applications to modeling and
control,” IEEE Trans. Syst. Man and Cybern., vol. 15, pp. 116–132, Jan./Feb. 1985.
[31] Math Works Inc. Fuzzy Logic Toolbox User’s Guide, Jan. 1998.
[32] B. K. Bose. “A high performance inverter‐fed drive system of an interior permanent magnet
synchronous machine,” IEEE Trans. on Ind. Appl., vol. 24, pp. 989–997, Nov./Dec. 1988.
REFERENCES 675
[33] M. G. Simoes, B. K. Bose and R. J. Spiegel. “Design and performance evaluation of a fuzzy logic
based variable speed wind generation system,” IEEE Trans. Ind. Appl., vol. 33, pp. 956–985, July/
Aug. 1997.
[34] S. Haykin. Neural Networks, New York: Macmillan, 1994.
[35] L. H. Tsoukalas and R. E. Uhrig. Fuzzy and Neural Approaches in Engineering, New York: John
Wiley, 1997.
[36] B. K. Bose. “Neural network applications in power electronics and motor drives – an introduction
and perspective (Invited Paper),” IEEE Trans. Ind. Electronics (Special Section on Neural Network
Applications in Power Electronics and Motor Drives), vol. 54, no. 1, pp. 14–33, Feb. 2007.
[37] B. K. Bose. “Guest editorial,” IEEE Trans. Ind. Electronics (Special Section on Neural Network
Applications in Power Electronics and Motor Drives), vol. 54, no. 1, pp. 12–13, Feb. 2007.
[38] Math Works Inc. Neural Networks Toolbox User’s Guide, 1998.
[39] M. Z. C. Wanik, I. Erlich, A. Mohamed and H. Shareef. “Predictive var management of distributed
generators,” IPEC 2010, pp. 619–624.
[40] L. Ciabattoni, G. Ippoliti, G. Longhi and W. M. Ceralietti. “Online tuned neural networks for fuzzy
supervisory control of PV‐battery system,” IEEE PES, pp. 1–6, 2013.
[41] J. S. R. Jang. “ANFIS: Adaptive‐network‐based fuzzy inference system,” IEEE Trans. Systems,
Man. and Cybernetics, vol. 23, pp. 665–685, May/June 1993.
[42] C. Campbell/Beijing. “Baidu’s brain (Russian President Vladimir Putin recently said that whoever
masters AI will become ruler of the world),” Time Magazine, pp. 42–45, Jan. 29, 2018.
[43] Cade Metz. “Bets on AI open a new chip frontier,” New York Times, pp. B1–B3, Jan. 15, 2018.

Artificial Intelligence Research Papers

Uploaded by

Copyright:

Available Formats

Artificial Intelligence Research Papers

Uploaded by

Document Information

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

Artificial Intelligence Research Papers

Uploaded by

Copyright:

Available Formats

Ch a p t e r 12

What is artificial intelligence (AI)? AI is basically a computer emulation of human

Boole, an English mathematician, when he theorized that human thinking is

basic principles of AI (i.e., ES, FL and NNW). This background knowledge

12.2.1 Expert System Principles

User User Inference Expert

Figure 12.1 Expert system block diagram showing the different elements.

Rule 1: IF X = X1 AND Y = Y1, THEN Z = A (12.1)

IF photovoltaic station A output 0 AND wind station C output 800 kW,

Sub frame Rules

Figure 12.3 Hierarchical frame‐based structure of knowledge.

IF :: PVA NO AND WINDC 800

Figure 12.4 Expert system development shell interfaced with external elements.

interface of the SHELL with USER INTERFACE, DATABASE FILES, GRAPHICS

12.2.2 Expert System‐Based Control of Smart Grid

Regional Regional Regional Regional

ES based Real time

IF photovoltaic station A output 0 AND wind station C output 800 kW,

OPAL‐RT Based Real‐Time Simulation Technology

1 Hz 100 Hz 20 KHz 100 KHz 1 MHz 10 MHz 100 MHz

TABLE 12.1 Features of eMEGAsim supercomputer‐based real‐time smart grid digital

12.3 FUZZY LOGIC

12.3.1 Fuzzy Inference System Principles

Rule 1: IF X is negative small NS AND Y is zero ZE

Because of analogy, a fuzzy system is often defined as a fuzzy expert system.

Figure 12.7 Simple rule matrix for fuzzy inference system.

Mamdani‐Type Fuzzy Inference System

Fuzzy Speed Control of Induction Motor Drive

μ(X) μ(Y) μ(Z)

Figure 12.8 Mamdani implication method of fuzzy inference system showing output

Figure 12.9 Fuzzy speed control of a vector‐controlled induction motor drive.

dU K1e K 2ce (12.25)

–1 –0.8 –0.6 –0.4 –0.2 0 0.2 0.4 0.6 0.8 1 ce(pu)

–1 –0.8 –0.6 –0.4 –0.2 0 0.2 0.4 0.6 0.8 1 du(pu)

NB NVB NVB NVB NB NM NS Z

PB Z PS PM PB PVB PVB PVB

Figure 12.11 Rule matrix for fuzzy speed control of induction motor.

Input Space Output Space

Figure 12.12 Input–output nonlinear mapping property of fuzzy and neural systems.

A FIS can be interpreted as a general input–output nonlinear mapping

Sugeno Method of Fuzzy Inference System

μ(X) μ(Y) μ(Z)

Figure 12.13 Sugeno (zero‐order) implication method of fuzzy inference system.

0.4 DOF2= 0.4

Figure 12.14 Sugeno (first‐order) implication method of fuzzy inference system.

K1. dof1 K 2 . dof2 K 3 . dof3

A more general first‐order Sugeno method of implication is illustrated in

Z1 A01 A11 X A21y (12.29)

Z2 A02 A12 X A22y (12.30)

Z3 A03 A13 X A23y (12.31)

Z1. dof1 Z 2 . dof2 Z 3 . dof3

12.3.2 Fuzzy Logic Control of a Modern Wind

Figure 12.15 Permanent magnet synchronous generator (PMSG)‐based wind generation

Description of Power Circuit

Wind Turbine Characteristics

Figure 12.16 Torque‐speed curves of wind turbine at different wind velocity showing

operate in undermodulation mode. The generator speed (ωr) is controlled by close

Figure 12.17 Phasor diagram of IPM machine with vector control.

cos e cos e cos sin e sin (12.40)