07 chapter 1

Chapter I
With growing industrialization the transmission and distribution of electric power has gained
tremendous importance. The trend is towards an integrated power network that will be immune to
most of the man made and natural disturbances. In other words the stability and speed of response
are the two conflicting criteria being imposed on such systems. To improve the system stability
(constant voltage, constant frequency) a fast power flow control scheme is highly essential. In the
recent years the HVDC transmission technology has attracted a lot of attention in is this regard. It
has the capability of fast power flow control along with other advantages like;
a) Handling bulk power
b) Power flow under asynchronous conditions
c) Power flow under faulted conditions
d) Can supply Reactive Power with forced commutation
e) Can improve the stability being a part of the AC Network
But the control and co-ordination of HVDC systems is still a challenge for the Design
engineers. The point-to-point two terminal HVDC links are gaining popularity day-by-day. The
trend is towards building two terminal and parallel AC-DC-links to enhance the power
transmission capability. Still the research has to go a long way on the control aspects of DC
links to exploit all its advantages. Many researchers have presented their contributions to the
DC link control on various aspects. The following reviews the various control strategies for
HVDC links to improve AC-DC system performance during normal and abnormal operating
1.2 HVDC Control
As mentioned one of the major advantages of using an HVDC link is the rapid
controllability of transmitted power by the control of firing angles of the converters. The
system control in an HVDC link tends to be quite complex with a hierarchy of controller
structures (Fig. 1.1). With the advent of very fast micro-processors the complexity of such
schemes are being minimised. Elaborate literature available in DC link control [1.1] have
established the following control schemes for some obvious reasons. However, the standard is
to subject the rectifier end with constant current and the inverter with constant extinction angle
control. Although the standard controllers of the HVDC systems are well documented, a brief
review is presented here for the sake of completeness.
Fig. 1.1 Hierarchial Control Structure of a DC link
1.2.1 Rectifier Control
The pole controIler(Fig.l.2) at the rectifier has two control modes i.e. minimum firing
angle control and constant current control.
Minimum Firing Angle Control
The minimum firing angle is selected on the basis of minimum reactive power demand
and sufficient positive voltage across non-conducting valves to ensure successful commutation.
Typical value of 'a' ranges between 10° to 18° [1.2]. This control mode operates for low
voltage levels on the rectifier terminal.
Fig. 1.2 Block diagram of Converter Control
Current Control
The rectifier generally operates in this control mode during steady-state. The current is
kept constant by making appropriate changes in the firing angle for changes in system voltage.
1.2.2 Inverter Control
The inverter may operate in one of the following control modes
1. Extinction Angie Control
2. Current Control
Extinction Angle Control (Figl.2)
The inverter normally operates with this control in steady state. The advantage of using
this control is primarily to avoid commutation failure and to minimise reactive power
requirement. Typical value of extinction angle (yref) used is approximately 18°.
Current Control
During disturbances such as DC line faults, starting of the system and power reversal,
both rectifier and inverter operate in the current control mode. It has been demonstrated [1.1]
that this control works efficiently for power reversal. The probability of commutation failure is
minimised in this control mode because the extinction angle is always greater than the
minimum limit(around 5°). Normally, the inverter is provided with current margin of 10%-20%
to allow the current controller to operate at both the rectifier and the inverter ends
independently [1.1],
1.2.3 Auxiliary Controls
Most of the existing HVDC system have been designed for the well known economic
and practical reasons. When integrated into a larger network, the fast power flow control
capability of the DC links also offer an attractive method for enhancing overall system
performance in eliminating AC system dynamic oscillations and improving transient stability
performance [1.1].
The Power in a DC link can be modulated by the quantities derived from the AC system
(such as frequency) in order to improve the security of the overall system. This is achieved by
an additional power (or current) order derived from an emergency power controller or auxiliary
Power/Frequency Control
When the DC link is used as a tie between two power systems, the frequency bias can
be used to adjust the power flow over the tie to assist the system in difficulty. It is to be noted
that an HVDC link has no inherent sensitivity to system frequency unless it is deliberately
introduced in the control system. Without it, a constant powerflow can overstep a receiving
system which has become separated from part on all of its load, even if governors cut-off all
primary energy flow to the local generation. Similarly, a sending system can be brought
eventually to a halt, if more power is required from it than the collective generation can
produce. Thus, although it would first appear that there is little benefit in the frequency control
for DC link between two systems so large that the link cannot be expected to influence the
frequency of either; it is prudent to incorporate an element of frequency control which in the
event of a partial break-up of a system, will prevent over-speed or under-speed [1.1].
Stabilization ofAC Ties
When a DC link is connected to a system with weak AC ties to neighbouring systems,
DC link power can be varied quickly and automatically, to balance the load flows and maintain
stability if one of the AC ties trips.
A DC tie used in parallel with an AC tie can be employed to damp the low frequency
inter-area oscillations in the AC tie. The control signal used can be - the rate of change of AC
tie line power (or current) or the phase angle difference across the AC tie. The DC tie can also
provide frequency control for one end if the AC tie gets disconnected and can thus permit
resynchronisation of the AC tie.
Sub-Synchronous Damping Control
A radial HVDC link connected to a thermal generating station can contribute to the
negative damping of the torsional oscillations at subsynchronous frequency due to the
interaction with the current controller. This problem usually surfaces when there are no parallel
AC links. The power modulation controller mentioned earlier which is designed to damp low
frequency rotor oscillations can aggravate the problem. However, a suitably designed
subsynchronous damping controller (SSDC) with control signal derived from the rotor velocity
can help to damp torsional frequency oscillations.
Emergency Control
An HVDC link employing the simplest of control strategies, buffers one system from
the disturbances on the other. Power flow can continue unchanged in the worst-possible
condition, but the option is available to vary power flow to assist the system in trouble to the
extent to which the healthy system can be allowed without putting itself in difficulty, subject
only to the rating of the link. In general, with suitable control, a disturbance originating in
either system can be shared in a predetermined manner and the oscillations occurring in both
the systems can be damped simultaneously. Substantial damping can be achieved with a very
small amount of DC power modulation.
1.3 The Problem
As described, an HVDC system irrespective of point to point or multiterminal, is
equipped with one or more of the complex control schemes. To overcome AC-system
problems, a DC link is required to have fast power flow control, while on the other hand sharp
rise and fall times may cause extensive stress on the converter valves and other associated
equipments. Therefore, for enhancing the performance of the DC link controls it is often
required to adapt alternative methods to conventional fixed gain PI/PID controllers to have an
optimal compromise between the above requirements. It is a well known fact that, due to the
high degree of non-linearity and uncertainty it is difficult to obtain any accurate mathematical
description of the HVDC systems for designing any fixed gain controllers. Therefore, to obtain
optimum system performance it is essential to go for some advanced control strategies, where
the wide variations in the operating points and uncertain system configurations can be taken
care of. Keeping the above factors in view extensive research has already been done in the area
of DC link control.
A number of HVDC control design.techniques have already been reported, which make
use of optimal control theory. Alexandridis and Galanos [1.3] have proposed a Kaman filter
based approach for an optimal current regulator for rectifier side. This approach leads to
decentralized optimal control systems based on the decomposition of the initial system into
interconnected and independent subsystems, and the design of a Kalman filter capable of
reconstructing the states of each subsystem and the unknown inputs at the interconnection
points. In essence the method tries to find out the control using some fast hand information
regarding the plant model and disturbances. The performance of the plant when subjected to
such control schemes is better as long as the plant model does not change appreciably.
However, there is no guarantee as to the effectiveness of such schemes under more uncertain
and large signal conditions.
Fig. 1.3 The Kalman Filter based optimal current regulator
When integrated into a larger AC network, the fast power flow control capability of the
DC link also offers an attractive method for enhancing overall system performance, for
eliminating AC system dynamic oscillations or for improving transient stability performance.
Fig. 1.4 Controlling the behaviour of generators by HVDC link
To and David [1.5] have described a systematic investigation of the full potential of using
HVDC link to ensure the stability of operation and enhance the performance of the
interconnected AC systems [Fig. 1.4],
Yet in another interesting application of adaptive control Alvis et al,[1.6] have described
a microprocessor based self tuning control for the rectifier current regulator of an HVDC link.
The main contribution of this paper is to choose the proportional, integral and derivative gains
for the rectifier current regulator so as to maintain the damping factor for the dominant mode of
the ac/dc system. The identification algorithm adopted was based on a recursive least-square
method with factorisation of the covariance matrix [Figl.5]. As a whole the control strategy
comprises three distinct stages; the modeling of the ac/dc system, the choice of a recursive
identification algorithm and a control design block.
Plant Parameters
Fig. 1.5 Self Tuning Regulator Scheme
The main drawbacks of these type of controllers are
the identification procedure assumes a fixed plant model
this cannot accommodate large variations in the operating conditions
the noise rejection property is limited
However, for some of the application particularly when the AC systems at both the
ends are strong this type control is good enough. But it is difficult to use such schemes when
either end of the dc link is a weak system. In this case the weaker AC system demands
additional attention pertaining to the voltage instability and reactive power requirement at the
converter bus.
In its new role as a major element embedded in the AC network, an HVDC scheme
must actively participate in the instantaneous scheduling of power according to the network
conditions, which is a feature inherent to the AC systems. To and David[1.7] have presented a
bang-bang-optimal control scheme to take care of the first swing instability and the subsequent
instability caused by poor damping. The recursive least square algorithm with a variable
forgetting factor has been used for on-line identification of the AC/DC system parameters. A
bang-bang optimal controller is used for resetting the optimal control except for a bang-bang
action during the first swing for certain categories of faults, has been introduced. The decision
when to implement the bang action and the timing of its duration are an important consideration
for the controller design.
From the above literature, it is apparent that, the HVDC transmission link is a non­
linear system that poses a difficult control problem because of the following reasons :
Presence of non-linear system components such as power transformers, converters
and surge arrestors.
Variable nature and topology of the power system
Insufficient knowledge and data about the system
Presence of AC/DC filters which can often form resonant circuits with the power
Generation of harmonics by the converters, which can interact with the controllers
Control strategy employed that has many operational modes due to system
protection requirements
Microprocessor-based adaptive controllers are increasingly being used within the
industry to adapt controller parameters as functions of operating angle, system topology, and
prior knowledge of system dynamics. However, there is scope for further improvement since
very little on-line adaptation has been performed.
In the brief literature review presented, it is found that most of the researchers have tried
a number of schemes based on various optimal and adaptive control methods. And in all of
them the model of the plant has been presumed to be fixed or only the parameters are identified
(the order remaining same). But as mentioned, the HVDC system along with the associated AC
systems is highly non-linear and uncertain.
Therefore, the design of controllers can not be carried out with any of the standard
mathematical procedures. Rather a more intelligent human like control method which uses the
approximate I/O inference methods should be developed and tested for its effectiveness.
1.4 Necessity of Intelligent Control
Recently a lot of research in the application of knowledge based and Intelligent
techniques to the control problems hints their effectiveness in HVDC links. Changes in
environments and performance criteria, unmeasurable disturbances and component failures are
some of the characteristics which necessitate intelligent control. Adaptive controller is the only
choice when parameters uncertainty exists. However, adaptive controllers do not have long
term memory and hence do not remember the optimal control parameters corresponding to
various configurations of the plant. In such cases pattern recognizers have been used to classify
the plant and choose the corresponding control parameters. Learning controllers can be used to
choose between different adaptive algorithms, reference models or performance criteria. As
more intelligent control systems are designed, it becomes more necessary to combine
adaptation, learning and pattern recognition in novel ways to make decisions at various levels.
On one hand one may use deep knowledge perhaps in the form of quantitative models along
with qualitative reasoning for decision making and on the other hand, a simple rule-base
derived from the experience of the process operator can also be used. The only requirement is
that the decision making is automated so that the KBS system does not give advice on what
action is best suited but actually goes ahead and implements the decision without checking with
an operator. Control systems based on knowledge based systems, neural networks, genetic
algorithms are all viable alternatives to those derived from conventional control theory.
The Fuzzy controller is one such simple rule-based control system. The knowledge used
for this are derived not only from an expert operator, but also from the designer of the systems.
One of the main advantages of using a Fuzzy approach that the Fuzzy logic provides the best
technique for knowledge representation that could possibly be devised for encoding knowledge
about continuous variables.
On the other hand Artificial Neural Networks with its massive parallelism and ability to
learn any kind of nonlinearity are used to address some of the very practical control problems.
A neuro-controller (neural networks based control system) in general, performs a specific form
of adaptive control, with the controller taking the form of a multilayer network and the
adaptable parameters being defined as the adjustable weights. In general, neural networks
represent parallel and distributed processing structures, which make them prime candidates for
use in multivariable control systems. The neural network approach defines the problem of
control as the mapping of measured signals and control actions.
Thus, there are primarily two concepts prevailing over the domain of Intelligent Control
and they are;
1. Fuzzy Logic based Control
2. Neural Network based Control
In the first case, the controller is represented as a set of rules which accepts the input in the
form of linguistic variables and gives the output in the form of linguistic variables. The
advantages of such a controller are,
The approximate knowledge about the plant is required (not an accurate model like
other optimal and adaptive control strategies)
Knowledge representation and inference is relatively simple and easy
Real time implementation is easier.
In the second case, the controller is represented as a mapping between the inputs and
outputs. Depending on a specific plant the map in the form of a network can be trained to
implement any kind of control strategy. The network can' also be trained on-line to adapt to the
changing conditions of the plant.
The advantages of this controller are,
Parallel architecture makes faster implementation
Any kind of non-linear mapping is possible
Training possible for various operating conditions, therefore, it can be adaptable to
The simple Fuzz1/ controller though represents a good non-linear controller cannot adapt
its structure if the situation demands. Sometimes the Fuzzy controllers with fixed structures,
fail to stabilise the plant under wide variations in the operating conditions. These type of
controllers also lack the parallelism of Neural Controllers.
On the other hand the Neural networks are very much adaptive to situations by adjusting
their weights accordingly. The parallel architecture
enables faster implementation of the
control algorithm. However, in the presence of noise and other uncertainties the performance
may deteriorate. Sometimes, in certain Neural Controller structures the model of the plant is
required. But in case of plants whose model becomes highly uncertain it is difficult to use
neural networks with fixed structure. In both the type of controllers the greatest advantage is the
non-linearity which, if properly expanded (for a simple case) serves as a very good non-linear
self tuning PI controller.
To extract advantages of both the types of controllers it is wiser to use the combination
of both, which leads to Fuzzy-Neural controllers. The new hybrid structure can be named as an
Adaptive Fuzzy Controller.
There are many structures of the aforesaid controllers used in practice. Some of them
will be discussed in the following paragraphs. However, most of the above find wide use in
small plants. For large-non-linear dynamic systems they are yet to be tested. In case of HVDC
links as discussed, the situation demands such type controllers. The literature available in the
Application of Intelligent Controllers to HVDC links is limited. V.K.Sood et al.[1.8] attempted
to replace the rectifier side current regulator by a Neural Network based Controller [Fig. 1.6].
The paper discusses the feasibility and effectiveness of the controller and gives a comparative
picture with the conventional controller.
Here the rectifier side current controller has been completely replaced by a pair of
Neural networks. One of them represents the fixed structure controller with a fixed set of
weights. It is trained offline to adjust the weights for optimum system performance. The other
network is adaptable to the dynamic conditions of the plant and the v/eights automatically
adjust themselves for various operating conditions, on-line. The results obtained for large and
small signal disturbances are encouraging.
1.4.1 Intelligent Control Structures
The failure of conventional fixed gain controllers in large non-linear dynamic systems
has urged researchers to find alternate methodologies. Adaptive control is one such method
where the changing conditions of the plants are taken into account to adjust the effective PID
(proportional, derivative and integral) gains of the controllers. However, due to nonavailability
of sufficient knowledge about the system the controllers do not work efficiently. Even some­
times the adaptive control degrades the performance if the plant model is incorrect. These are
the reasons for finding more intelligent human like controllers for the plants described above.
HVDC link is one such plant which needs efficient, fault tolerant, robust control strategies for
reliable operation under normal and abnormal conditions. As discussed Neural networks and
Fuzzy logic are the modem tools being used for control purposes. Several structures of the
above controllers have been proposed in the literature . Some of them are in fact in use for
controlling small systems. Extensive research can only tell the feasibility of such controllers in
large systems.
1.4.2 Neural Controllers
Models of dynamic systems and their inverses have immediate usage for control. In the
literature on neural network architectures for control a large number of structures have been
proposed and used. Some of the well-established and well analysed structures are presented
Supervised Control [1.9] (Fig. 1.7) ■
In some situations it may be desirable to design an automatic controller which mimics
the action of the human operator. This has been called as supervised control. A neural network
provides one possibility for this. Training the network is similar in principle to learning a
system model. In this case however, the network inputs correspond to the sensory input
information received by the human. The network target outputs used for training correspond to
the human control input to the system.
Direct Inverse Control [1.10] (Fig.1.8)
Neural Net based
This utilises an inverse system model. The inverse model is simply cascaded with the
controlled system in order that the composed system results in an identity mapping between
desired response (i.e. the network inputs) and the controlled system output. Thus the network
£Cts directly as the controller in such a configuration. This is very common in robotics
applications. However, this model clearly relies on the fidelity of the inverse model used as the
Model Reference Control [1.11] (Fig. 1.9)
Here, the desired performance of the closed loop system is specified through a stable
reference model M, which is defined by its input-output pair {r{t),yr (t)}. The control system
attempts to make the plant output yp(t) match the reference model output asymptotically ;
finally r(/) - .^(Oll ^ £ >f°r some specified constant
e> 0.
In this structure the error defined above is used to train the network acting as the controller.
Clearly, this approach is related to the training of inverse plant model. In the case, when the
reference model is the identity mapping the two approaches coincide. In general, the training
procedure will force the controller to be a "detuned" inverse, in a sense defined by the reference
NN based Reference
Internal Model Control (IMC) [1.12] (Fig. 1.10)
In this case the role of the system forward and inverse models is emphasised. The system
forward and inverse model are used directly as elements within the feedback loop. IMC has
been thoroughly examined and shown to yield transparently to robustness and stability analysis.
In internal model control a system model is placed in parallel with the real system. The
difference between the system and model outputs is used for feedback purposes. This feedback
signal is then processed by a controller subsystem in the forward path; the properties of IMC
dictate that this part of the controller should be related to the system inverse. Given network
models for the system forward and inverse dynamics the realization of IMC using neural
networks is straight forward; the system model and the controller are realised using the neural
network model.
Predictive Control [1.13] (Fig. 1.11)
In the realm of optimal and predictive control methods, the receding horizon technique
has been introduced as a natural and computationally feasible feedback law. In this approach a
neural network provides prediction of future plant response over the specified horizon. The
predictions supplied by the network are passed to a numerical optimization routine which
attempts to minimize a specified performance criteria in the calculation of a suitable control
Fig. 1.12 Structure for Optimal Control
In this application, the state space is partitioned into regions (feature space)
corresponding to various control situations (pattern classes). The realisation of the control
surface is accomplished through a training procedure. Since the time-optimal surface in general
is non-linear, it is necessary to use an architecture capable of approximating a non-linear
surface. One possibility is to first quantize the state space into elementary hyper-cubes in which
the control action is assumed to be constant. This process can be carried out with an
LVQ(Leaming Vector Quantisation) architecture. It is then necessary to have another network
acting as a classifier. If continuous signals are required a standard back-propagation
architecture can be used.
Adaptive Linear Control [1.15] (Fig. 1.13)
The eonneetionist approach can be used in nonlinear control but also as a part of a
controller for linear plants. The Hopfield network may be utilised as a dynamic controller for
linear systems. In this case, elements of variable structure theory were utilised to construct the
controller, and as a result the proposed controller is characterised by its robustness. It is also
possible to include adaptation using some standard method.
Fig. 1.13 Structure for Feedback Networks as Controllers
The Hopfield network can also be used as a part of the adaptation mechanism. In this
case the network is induced in the adaptation path and is used to minimize simultaneously
square error rates of all the states. Here, the output of the network represents the parameters of a
linear model for the plant.
Reinforcement Learning Control [1:16]
This kind of network makes use of low quality (plant error) feedback signal. While the
performance measure for a supervised system is defined in terms of a set of target by means of
a known error, reinforcement learning addresses the problem using any measure whose value
can be supplied to the learning algorithm. Instead of trying to determine target controller
outputs from target plant responses, one tries to determine target controller output that would
lead to increase in a measure of plant performance. The "critic" block is capable of evaluating
the plant performance and generates an "evaluation signal" which is used by the reinforcement
learning algorithm. This approach is appropriate when there is a genuine lack of knowledge
required for applying more specialized learning methods.
Gain Scheduling [1.17] (Fig. 1.14)
This is an old control engineering technique which uses process variables related with
its dynamics to compensate the effect of working in different operating regions. Here the neural
network is used as an associative memory that relates the parameters of the controller with the
state of the plant. The Hopfield network, and other connectionist architectures can be used for
this purpose.
1.4.3 Fuzzy Controllers
Two approaches are extensively followed for designing fuzzy controllers: Direct Fuzzy
control and Indirect Fuzzy control. In Direct Fuzz}' control the conventional controller is
completely replaced; while in
Indirect control the conventional control parameters
(Proportional, Integral and derivative gains) are adjusted with an inference method based on
Fuzz}' logic. The structure of the Fuzzy controller largely depends on the input and output
classifications. Some of the common and well understood models will be discussed here.
Fuzzy PID Controllers [1.18] (Fig. 1.15)
Here the error and the rate of change of error are fuzzified and then fed to the Fuzzy
decision block. The Fuzzy inference block refers to the rule-base to generate output Fuzzy sets
and their membership grades. Then a defuzzification scheme is used to derive the normalised
values of the output variables. It can be multiplied by a scaling factor to generate the actual
input to the plant. These are of two types i.e. linear and non-linear. The symmetrical rule table
is a characteristic of a linear Fuzz}' controller. However, very often the situations demand
asymmetrical configurations and therefore the rule table is asymmetrical leading, to non-linear
Fuzzy controllers. Also the method of defuzzification sometimes characterise a controller to be
linear or non-linear.
Fig. 1.15 Structure of the Fuzzy PID Controller
Hybrid Fuzzy Controller [1.19] (Fig. 1.16)
Often Fuzzy controllers have inputs generated by a conventional controller (e.g.
combinations of integrators and differentiators). Typically the error is the input to the
conventional controller. These types of controllers are robust and need a less complicated rule
base. The lead/lag times can also be adjusted by choosing proper time constants for the
conventional controller.
Fig. 1.16 An Hybrid Fuzzy Controller
Fuzzy Adaptive Controller [1.20] (Fig. 1.17)
Here the structure is same as that of Fuzzy PID controller. But the slope of the input
membership functions are adjusted to adapt with the error. In this case since the membership
function becomes adaptable, the controller becomes robust and insensitive to plant parameter
variations. In addition to the slope of the membership functions the following parameters of the
Fuzzy controller can also be adapted according to,
The set of rules and knowledge base
A finite set of values describing the universe of discourse.
Fig. 1.17 Block Diagram for the Adaptive Fuzz}-' Controller
Fuzzy Sliding Mode Controller [1.21]
Although Fuzzy control is very successful especially for non-linear systems, there is a
lack in design of such controllers with respect to performance and stability. The success behind
the Fuzzy controlled plants is that it is similar to the sliding model control (SMC) which is, for
a specific class of non-linear systems an appropriate robust control method. SMC can be
applied especially in the presence of model uncertainties, parameter fluctuations and
disturbances provided that the upper bounds of their absolute values are known. For the FSC
(Fuzz}' Sliding mode Controller) whose input is an error vector e and whose output is a scalar
control value u, the general design rules may be formulated as:
R\ :Tne normalised control value UN (controller output) should be negative above the
switching line and positive below it.
R2 '.Normalised states (Controller inputs) eN,eN.... that fall out of the phase plane should
produce maximum value
respective sign of UN .
i?3; jt/_y | should increase as the distance 'd' between the actual state and the line perpendicular
to the switching line grows, for the following reasons :
Normalised state eN,eN.... that are situated at the boundaries inside the phase
plane produce maximum values
with respective sign of UN so that
discontinuities at the boundaries of the phase plane can be avoided.
The central domain of the phase plane can be arrived at very quickly.
This controller has been successfully applied to small plants and the results obtained are better
than the sliding mode conu-oller. The sliding mode controller shows chattering effects whereas,
the Fuzzy sliding mode controller is smooth.
Fuzzy Model Following Controller [1.22] (Fig. 1.18)
Fig. 1.18 Fuzzy-model Following Control System
For a conventional Fuzzy control system, the resulting performance is difficult to
predict. In order to have the advantages of a Fuzzy logic controller with desired performance, a
Fuzzy adaptive controller is introduced to a model-following control system. Fig. 1.18 shows
the basic structure of the model following controller. The error between the plant output and the
reference model output is used to adjust the membership functions of the Fuzzy controller. It
has been demonstrated [1.22] that the proposed controller displays good performance.
Hierarchical Fuzzy Controller [1.23] (Fig. 1.19)
In an hierarchical Fuzzy controller the structure is divided into different levels. Similar
to tuning a radio set by coarse and fine adjustments the hierarchical controller gives an
approximate output at the first level which is then modified by the second level rule set. This
process is repeated in succeeding level of hierarchy.
Fig. 1.19 The Configuration of Adaptive Hierarchial Fuzzy Control System
The rules in the first level are of following form
if ( x{ is ax, x2 is a12,....*„i ls a\,n\) then (output yl is bx)
The rules in the ith (i>l) level are of the form.
if ( *y.+I is aNji.... xN.+n. is aN_) and(yM is A_,)then(output y-is b()
The number of rules in such kind of controllers are drastically reduced. In addition the
controller action is more human like and therefore is easy to formulate the rule base. Such type
of controllers can also be made adaptable by changing the membership functions according the
performance index of the plant.
1.4.4 Self Organising Controllers
These controllers are more or less similar to adaptive Fuzzy controllers. But in this case
the rule base, and membership functions are tuned with the plant operating conditions. These
controllers have the self-learning capability. Therefore the rule base is automatically formulated
with little or no intervention of the human expert. By using certain strategy the controller can
be made to start from an empty rule base. Control rules are generated and modified by the
learning algorithm evaluating the performance enhancement matrix. And after several runs (on
several changes in the reference signal) the controller can give a good performance.
1.4.5 Indirect Fuzzy Controllers
Sometimes it is convenient to improve the existing conventional PI strategy rather than
replacing the whole controller itself. From the stability point of view, there is not any particular
method to design a direct Fuzzy controller which will be stable under all circumstances.
Therefore, for plants where the operating conditions exhibit a wide range of variations, the
design of a direct Fuzzy controller becomes difficult. However, the proportional and integral
gains of the conventicnal controllers can be adjusted with the help of Fuzzy logic. °
1.4.6 Neuro-Fuzzy Controllers
In order to get a mapping with substantial flexibility and non-linearity Neural networks
are used. To represent the abstract knowledge and to infer it like human brains Fuzzy logic is
the best approach. However, it suffers from the absence, of learning and mapping. To get the
advantage of the Fuzzy logic as well as Neural networks a number of network architectures
employing the principle of Fuzzy logic have been proposed. These are termed as the Fuzzy
Neural Networks(FNN).
FNN1 [1.27] (Fig1.20)
As shown in the Fig. 1.20 the neural network consists of two layers. The input layer
(7j, /2_) fuzzifies the inputs into crisp sets. The connections from the inputs to outputs are
made through the weights. At the output neurons the bias is combined (logically, by Zadeh
AND, OR, Leukaswiecz OR) to produce outputs.
if, V = max A = min
Vk = (**!»...... ,vhn)
tk =
.....>'**) desired outputs
0k = {0k i,.....,Okn) outputs of the network.
given input Vk the output neurons first combine the incoming signals with their corresponding
weights using
V = max and A = min as follows
Hi - VjAK,*)...... .{A{Kto>r„,}} fbrlS<S»
Output neuron Oi now combines Wkl together with bias 0( to produce its output
Oki =
for 1 <i <n, all k.
The difference of this network with the traditional one is that, here the min, max operations are
used for multiplication and addition. A modified delta rule [1.27] has been proposed to adjust
the weights and 0{.
FNN2 [1.28] (Fig. 1.21)
This type of network uses a typical Fuzzy neuron. The neuron sends a crisp signal xt to
N, 1 < i < n . The weights for N are Fuzzy sets A-t, 1 </</?. The membership function of
input xi is evaluated through At- (xt) and the results are combined into one number y inside
N . N then transforms y into (Crisp) output 0 which is then sent to some other (Fuzzy)
neurons in the network.
FNN3 [1.28] (Fig. 1.22)
This uses another type of neuron which sends Fuzzy signal Xx towards N , 1 <i<n.
In the process Xf is transformed, by some transformation 7), into X) = [(X)), 1 < i < n . N
has Fuzzy weights A.: which are first combined with transformed signal into Wt = At * X] for
some operator *, l <i<n, and these Wt- are combined into one Fuzzy set W inside N .
Finally N transforms W to its output O . The training algorithm for such networks have been
proposed in [1.28].
Fig. 1.22 Fuzzy Neural Net-Ill
FNN4 [1.28] (Fig. 1.23)
This type Neural Network uses neuron which sends Fuzzy signal Xi to N . There are
no weights and N has a Fuzzy transfer relation R , so that its output 0 is (Xj,......,Xn)oR
for some type of composition operator "o". The network built-up from such type of Fuzzy
neuron appears ideal for rule extraction from training data, in a Fuzzy expert system.
ANFIS (Adaptive Network-based Fuzzy Inference Systems) [1.29] (Figl.24)
An adaptive network is a multilayer feedforward network in which each node performs
a particular function (node function) on incoming signals as well as a set of parameters
pertaining to this node. The formulas for the node functions may vary from node to node, and
the choice of each node function depends on the overall input-output function which the
adaptive network is required to carry out. To reflect different adaptive capabilities, both circle
and square nodes have been used. A square node (adaptive node) has parameters while a circle
node (fixed node) has none.
In order to achieve a desired input-output mapping, the parameters of all the square
nodes are updated according to given training data and a gradient-based learning procedure can
be used. As described in [1.29] a Fuzzy controller can be easily made to fit into this. Thus it is
equivalent to a typical Fuzzy inference system with an adaptive structure.
A Self Tuning Neuro-Fuzzy Controller [1.30] (Fig. 1.25)
Fig. 1.25 A Self Tuning Neuro-Fuzzy Controller
This network is similar to CMAC ( Cerebellum Model Articulation Controller). Such
type of controllers are very much similar to a Simple Fuzzy controller with an adaptive
defuzzification procedure. As shown in the Fig. 1.25 the CMAC architecture represents the
connectionist model cf a Fuzzy controller. The error between the reference and the plant output
is trained to adjust the weights. This architecture is simple and well understood. The training
also needs the least number of calculations.
This architecture is discussed in
subsequent Chapters.
In the literature a number of architectures have been proposed. Out of them only a few
are applicable to the HVDC link control. The HVDC link is a very complicated and large
system. Therefore to qualify for such systems the controller need to have the following
essential properties.
It should be robust over a wide range of operating conditions.
It should adapt to all the transient and dynamic conditions
It should be easier to choose the parameters of the controllers on-line
mathematical computations should be less such that it does not create any
computational burden
Out of all the above described structures only a few satisfy these conditions. Therefore,
only some of them have been tested for their feasibility and effectiveness in controlling HVDC
systems. They will be discussed in detail in the following chapters.
1.5 Motivation and Problem Statement
The motivations behind this thesis is work can be attributed to the following reasons:
1. The intelligent control techniques have found themselves successful in small plants where
the number of control variables are less.
2. For large dynamic systems these techniques need extensive off line study for their
feasibility and effectiveness.
3. The HVDC links are highly dynamic and non-linear. The plant conditions demand more
efficient and robust control schemes for a wide range of operating points.
4. Extensive research need to be carried out in the application of Fuzzy and Neural controllers
to the HVDC links.
5. Furthermore, in the study of AC-DC system interactions the HVDC control schemes need
thorough investigations, for these controllers very often regulate some of the vital
parameters of the AC system. Therefore, along with the applications of Neural and Fuzzy
controllers some of the more advanced control schemes have been tested to establish the
advantages of the proposed controllers in AC-DC system interaction.
6. In addition to the Fuzzy and Neural controllers some other advanced schemes should be
tried to evaluated the relative merits and demerits. In the literature input-output linearising
controllers have been applied successfully to control some of the practical systems. The
potential of this method needs to be verified in case of the DC links relative to the
conventional and Intelligent control.
As a whole the thesis is devoted to the study of Intelligent controllers for HVDC
converters. However, in the study of AC-DC system interaction, some of the advanced
controllers along with the Intelligent controller has been tested.
1.6 Thesis Outline
The thesis has been primarily devoted to evaluate the performance of some of the Fuzzy
and Neuro-Fuzzy controllers in point-to-point and Multi-terminal HVDC links. Before they are
tested on bench-mark models, the Fuzzy controller has been tested as an auxiliary control in an
AC-DC system with a quasi-steady-state model.
Chapter-II discusses the application of Fuzzy controller to the auxiliary control loop of
the PI current regultator. Here the DC link has been used to damp out the transient oscillations
in the connected AC system. The firing angle to the rectifier has been modulated by an
auxiliary signal derived from the system frequency deviation. Along with the Fuzzy controller,
some of the advanced controllers have also been tested.
The self-tuning and the variable structure sliding mode controllers are some of the very
advanced non-linear controllers which in recent years have been proven to be very effective.
These two controllers have been tested as the auxiliary controllers for the AC-DC system in
Chapter-II. Finally, a comparative study has been taken up to prove the worth of the Fuzzy
In Chapter-Ill some of the Fuzzy and Neuro-Fuzzy controllers have been tested on a
two-terminal point-to-point HVDC benchmark model. The transformers, transmission lines and
converters have been represented in detail in the simulation program. The simulation have been
carried out in an EMTP type program known as EMTDC. Here the Fuzzy and the Neuro-Fuzzy
controllers act as robust and self tuning controllers. Unlike Chapter-II here the main objective is
to damp out the DC link transients. Therefore, the generators at either end are represented by
simple voltage sources . There is no auxiliary control provided here. As robust controllers the
conventional rectifier current regulator has been completely replaced by Fuzzy and NeuroFuzzy blocks. As self-tuning controller the proportional and the integral gains of the
conventional controller have been tuned on-line by methods based on Fuzzy Logic. Five
different robust and self-tuning Fuzzy and Neuro-Fuzzy controllers have been tested by
subjecting the system to different normal and abnormal situations.
In Chapter-IV the same benchmark model of Chapter-Ill has been used to evaluate the
performance of an Input-Output Linearising Controller, The Linearising Controllers are best
known for their application to non-linear plants. The philosophy is to cancel out the plant nonlinearities thus making the design task easier. Substantial literature available tells the success of
such controllers in large number of control problems. The simulation results pertaining to the
proposed linearising controller has been compared with the Fuzzy and Neuro-Fuzzy controllers
*of Chapter-Ill.
Out of all the proposed controllers the better ones are chosen to be implemented as the
rectifier current regulators in a Multi-terminal-DC system. The performance and design of these
controllers for a 3-terminal DC system has been presented in Chapter-V. The study includes the
transient response of the system under various kinds of faults.
Much remains to be explored in designing suitable Fuzzy controllers for non-linear and
uncertain plants. As these controllers are also highly non-linear, it leads to difficulty in the
analysis and design of appropriate structures for a given plant. Chapter-VI gives an overall
assessment and tells the future scope of this research work especially in the' area of design and
analysis of these controllers.
1.7 Conclusion
The thesis has been primarily devoted to the study of Intelligent controls for large
systems. AC-DC systems in fact come under such systems. The wide spread non-linearity and
high degree of uncertainty makes it difficult for the designer to choose the proper control
strategy. Therefore while undertaking this work it was anticipated that the AI (Artificial
Intelligent) based controllers would give a better solution.
From the coincised literature review it is found that, a number of controllers based on
Neural network and Fuzzy logic have proved themselves successful in controlling small-non­
linear plants. However, for large-non-linear plants it is yet to prove its worth. To investigate the
possibility of their application in such plants this thesis work has been taken up. Along with the
various Fuzzy, Neural and Hybrid Controllers, a number of advanced control schemes have
been applied to the HVDC links. As will be seen , the performance of these controllers are far
superior to the conventional ones under transient and dynamic conditions.
Padiyar, K.R., 'HVDC Power Transmission Systems", Wiley Eastern Limited.
Tarrawkey, M.Z., "HVDC Transmission Control Schemes", Proceedings of Manitoba
Power Conference on EHVDC, 1971, pp.598-714.
Alexandridis A.T. and Galanos G.D., "Design of an optimal current regulator for Weak
AC/DC Systems using Kalman Filtering in the Presence of Unknown Inputs", 1989,
March Proc. IEE, Vol.136, pt.c, No.2, pp.57-63.
Rostamkolai N., Wegner C.A., Piwko R.J., et ah, "Control Design of Santo Tome Backto-back HVDC Link", Aug, 1992, Trans. IEEE on Power Systems, Vol.8, No.3,
To K.W.V., David, A.K., "Using an HVDC Link to Control the Behaviour of
Generators", 1993, March, Proc. of International Power Engineering conference,
Singapore, pp.205-209.
Alvis J.E.R., Pilotto L.A.S., and Waterbe E.H., "An Adaptive Digital Control Applied
to HVDC Transmission", IEEE Trans., No.93, WM 067-9PWRD.
To K.W.V., David A.K., Hammad A.E., "A Robust Co-ordinated Control Scheme for
HVDC Transmission with Parallel AC-DC Systems", IEEE Trans., No.94 WM 0612PWRD.
Sood V.K., Kandil N., Patel R.V., and Khorasani K., Comparative Evaluation of Neural
Network-Based and PI Controllers for HVDC Transmission", May 1994, No.2, Vol.9,
IEEE Trans, on Power Electronics, pp.208-295.
Grant E., and B.Zhang, "A Neural Net Approach to Supervised Learning of Pole
Balancing", IEEE Int. Symposium on Intelligent Control, 1989, pp.123-129.
[1.10] Miller W.T., Glanz F.H. and Knaf L.G., "Application of a General Algorithm to Control
of Robotic Manipulators", The Int.-J. of Robotic Research, 1987, pp.84-98.
[1.11] Narendra K.S., and Parthsarathy K., "Identification and Control for Dynamic Systems
Using Neural Networks", IEEE Trans, on Neural Network ,1, 1990, pp.4-27.
[1.12] Hint K.J., and Subarbaro D., "Neural Networks for Non-linear Internal Model Control",
1991, Proc. IEE, Pt.D, 138, pp.431-438.
[1.13] Mayne D.Q., and Michalska H., "Receding Horizon Control of Non-linear Systems",
1990, IEEE Trans, on Aut. Control. 35, pp.814-824.
[1.14] Fu.K.S.,
"Learning Control Systems-review an Outlook", IEEE Trans, on Auto.
Control, 1970,16, pp.210-221.
[1.15] Sak S.H., "Robust Tracking Control of Dynamic Systems with Neural Networks", 1990,
IEEE Int. Joint Conf. on Neural Networks, IJCNN'90, pp.563-566.
[1.16] Barto A.G., "Neural Networks for Control 1990, Chapter-1, pp.5-58, MIT Press,
Cambridge, MA.
[1.17] Guez A., J.L.Elibert and M.Kaon (1988), "Neural Network Architecture for Control",
IEEE Control Systems Magazine, 8, pp.22-25.
[1.18] Chen Chieh-li, Chen P.C., and Chen C.K., "Analysis and design of Fuzzy Control
Systems", Fuzzy Sets and Systems, 57 (1993), pp.125-140.
[1.19] Passino K.M., and Stephen Yurkovich, "Fuzzy Control Theory and Applications", Proc.
of the CCA Workshop, Sept. 1992, Dayton.
[1.20] Batur C., and Karparian V., "Adaptive Expert Control", INT.J. Control, 1991, Vol.54,
No.4, pp.867-881.
[1.21] Palm Rainer, "Robust Control by Fuzzy Sliding Mode", Automatica, Vol.30, No.9,
pp. 1429-1437,1994.
[1.22] Che Cheih-li, Chen P.C., and Cheng C.K., "A Pneumatic Model-following Control
System using a Fuzzy Adaptive Controller" 0003:1098/93, Pergaman Press Ltd.
[1.23] G.V.S. Raju, Jun Zhao,"Adaptive Hierarchial Fuzzy Controller", IEEE Trans, on System
Man & Cybematics, vol.23,No.4,Jan 1993,pp973-980.
[1.24] A.D. Carli, P.Ligousi, A.Marroni," A Fuzzy PI Control Strategy" Control Engg.
Practice,Vol.2,No. 1,1994,pp 147-153.
[1.25] Z.Y. Zhao, M.Tomizuka, S.Isaka, " Fuzzy Gain Sheduling PI Cntrollers", IEEE Trans,
on System Man & Cybematics,Vol.23,No.5,Sept./Oct. 1993, ppl392-1398.
[1.26] H.Tseng, V.H. Hwang, " Servo Controller Tuning using Fuzzy Logic",IEEE Trans, on
Control System Technology, Vol.l, No.4,Dec 1993,pp262-269.
[1.27] Y. Hayashi, JJ. Buckley, E.Czogola," Fuzzy Neural Controllers", 0-7803-0236-2/92
1992 IEEE.
[1.28] Y. Hayashi, J.J. Buckley, E.Czogola," System Engineering Application of Fuzzy Neural
Networks", II-413-II416 0-7803-0559-0/92 1992 IEEE.
[1.29] J.Shing, R.Jang, "ANFIS: Adaptive Network based Fuzzy Inference System", IEEE
Trans, on System Man & Cybematics", Vol-23 No.3 May/June 1993, pp665-690
[1.30] G.Calcev," A Self Tuning Neuro-Fuzzy Controller", Proc. of.Int. Symp. on Intelligent
Control, Chicago, Illinois, USA, Aug 1993, pp-577-581.
Related flashcards


28 cards

Electrical engineering

45 cards

Signal processing

22 cards

Electrical components

22 cards


21 cards

Create Flashcards