Chapter - 1

Introduction
There are dramatic changes in the power industry because of deregulation. One consequence
of this is that the demands for rapid changes in power generation is increasing. This leads to
more stringent requirements on the control systems for the processes. It is required to keep
the processes operating well for large changes in the operating conditions. One way to
achieve this is to incorporate more process knowledge into the systems. The goal is to
develop moderately complex nonlinear models that capture the key dynamical properties over
a wide operating range. The models are based on physical principles and have a small number
of parameters; most of which are determined from construction data. Particular attention has
been devoted to model drum level dynamics well. Drum level control is an important
problem for nuclear as well as conventional plants. In Parry, Petetrot and Vivien (1995) it is
stated that about 30% of the emergency shutdowns in French PWR plants are caused by poor
level control of the steam water level. One reason is that the control problem is difficult
because of the complicated shrink and swell dynamics. This creates a nonminimum phase
behaviour which changes significantly with the operating conditions. Since boilers are so
common there are many modelling efforts. There are complicated models in the form of large
simulation codes which are based on finite element approximations to partial differential
equations. Although such models are important for plant design, simulators, and
commissioning, they are of little interest for control design because of their complexity. The
model presented here is adapted from K.J Astrom and R.D. Bell (1998). A nonlinear dynamic
model for natural circulation drum-boilers is adapted. The model describes the complicated
dynamics of the drum, downcomer, and riser components. It is derived from first principles,
and is characterized by a few physical parameters. A strong effort has been made to strike a
balance between fidelity and simplicity. Since the model is derived from first principles it can
describe the system for a wide operating range.
Simulation of the model is done using MATLAB R2010a and results have been verified with
plant data presented in K.J Astrom and R.D. Bell (1998). The conventional PID controller is
applied when the boiler is operating at medium load for both servo and regulatory problems.
Advanced controllers such as Fuzzy logic controllers(FLC) and Neural Network(NN)
predictive controllers have been also tried with satisfactory results. The details of these
controllers are discussed in the report within.
The drum boiler is a simple boiler which consist of a drum, downcomer and riser
components. A simple schematic of the drum boiler have been shown in Figure 1. The heat
Q, supplied to the risers causes boiling. Gravity forces the saturated steam to rise causing a
circulation in the riser-drum-downcomer loop. Feedwater qf , is supplied to the drum and
saturated steam, qs, is taken from the drum to the superheaters and turbine. The presence of
steam below the liquid level in the drum causes the shrink-and-swell phenomenon which
makes level control difficult. In reality the system is much more complicated than shown in
the figure. The system has a complicated geometry and there are many downcomer and riser
tubes. The outflow from the risers passes through a separator to separate the steam from the
water. In spite of the complexity of the system it turns out that its gross behaviour is well
captured by global mass and energy balances.

Fig 1: Schematic of a drum boiler[1].

The shrink and swell phenomena in a drum boiler makes the level control very
difficult. Under boiling conditions, steam supporting field products such as bubbles
exist below the water/steam level interface. These bubbles have volume and therefore
displace water to create a misrepresentation of the true water level in the drum.
Another effect upon drum level is pressure in the drum. Because steam bubbles
compress under pressure the steam bubbles expand or contract respective to these
pressure changes. A higher steam demand will cause the drum pressure to drop, and
the steam bubbles to expand to give the appearance of a water level higher than it
truly is.This fictitious higher water level causes the feedwater input to be shut down at
a time when more water is really required. A surge in water level as a result of the
drum pressure decreasing is called 'swell'. A water level decrease due to drum
pressure increase is called 'shrink'.

Chapter - 2
Modeling and Simulation
2.1 Global mass and energy balance equations
The global mass balance equation is
Rate of accumulation of mass in the system = Rate input of mass Rate output of mass.
[

.(1)

where
S= density of saturated steam (kg/m3)
Vst= total volume of steam in the boiler system (m3)
W= density of saturated water (kg/m3)
Vwt= total volume of water in the boiler system (m3)
qf= feed flowrate (kg/s)
qs=steam flowrate(kg/s)
The global energy balance is
Rate of accumulation of energy in the system = Rate input of energy Rate output of energy.
[

..(2)

us= specific internal energy of steam (J/kg).

uw= specific internal energy of saturated water (J/kg).
mt= total mass of metal tubes and the drum.(kg)
Cp= specific heat of metal (J/kg 0C)
tm= temperature of the metal (oC)
Q= heat supplied to the risers (W)
hf= specific enthalpy of feedwater (J/kg)
hs= specific enthalpy of steam (J/kg)
From first law of thermodynamics
H=U+PV
3

which can be rearranged as

u = h - p/.
Now we write the global energy balance in terms of enthalpy instead of internal energy by
using the first law of thermodynamics.
*

.(3)

The total volume of the drum, downcomer and riser is

.(4)
Equation (1), (3) and (4) combined with saturated steam tables yields a simple boiler model.
We will however make manipulations of the model to obtain a state model. There are many
possible choices of state variables. Since all parts are in thermal equilibrium it is natural to
choose drum pressure ,p as one state variable. This variable is also easy to measure. Using
saturated steam tables, the variables S , W , hS and hW can be expressed as functions of
steam pressure. The second state variable can be chosen as the total volume of water in the
system VWt. Using equation (4) VSt can be eliminated and replaced by VWt.
Equation (1) can be written as
[

]
[

.
[

.
.

Equation (3) can be written as

[

.
.

.
4

.
.

So
[

(5)
.

Where
.
=

(6)
.
.

Steam tables are required to evaluate

results are based on approximations of steam tables with quadratic functions.

(

.
)

.
.

.
.
.
.

and

. The

The condensation flow rate is given by a simple energy balance done within the system
(

where

.
.

Here the term on the left hand side of the equation

accounts for the thermal energy change within the system and the term

accounts for the

pressure energy change within the system. The negative sign appears before

since

condensation is taking place.

The term on the right hand side accounts for the physical changes taking place within the
system due to total change in stored energy within the system.
So it can be rewritten as
(

)..(8)

2.2 Distribution of steam in risers and drum

Saturated mixture quality in a heated tube
Consider a vertical tube with uniform heating. Let be the density of the steam-water
mixture. Furthermore let q be the mass flow rate, A be the area of cross section of the tube, V
the volume, h the specific enthalpy, and Q the heat supplied to the tube. All quantities are
distributed in time, t and space, z. Assume for simplicity that all quantities are same in a
cross-section of the tube.
The mass balance for heated section of tube can be derived from the mass continuity
equation. The continuity equation can be stated as
.
u,v and z are velocity in x,y and z directions respectively. Since mass is only entering and
leaving in z- direction. So we have
.
.
.(9)
6

The energy balance can be derived as follows. Let us consider a infinitesimally small block of
cube through which the fluid is flowing. The cube is a control volume. We will do the energy
balance around this cube. The cube has three coordinates x, y and z. So
Energy flow into the cube- Energy flow out of the cube= Rate of change of energy
Energy flow into the cube=
Energy flow out of the cube=
.

.
Dividing both sides by dxdydz we get.

.
.
.
.
(10)
The specific internal energy of mixture of steam and water is
.(11)
Where

denotes the mass fraction of steam in the flow i.e. the quality of the mixture.

In steady state we get

.
..from (10) and (11).
So
.
Let be the normalized length coordinate along the risers and
outlet. The steam fraction along the tube.
7

be the steam quality at riser

, 0 <= <= 1. ..(12)

The volume and mass fraction of steam are related through

) (

where

. .(13)
To model drum level it is essential to describe the total amount of steam in the risers. This is
governed by the average volume fraction in the risers.

) ..(14)

2.3 Mass and energy balance for the riser section

The global mass balance for riser section is

(15)

Where Vr = volume of the riser(m3).

qdc =flow rate in the downcomer section(kg/s)
qr=flow rate in the riser section(kg/s).
The global energy balance for the riser section is

.
.
.

.
(

.(16)
Circulation flow
The momentum balance for the downcomer riser loop is
8

Where Lr and Ldc are the length of riser and downcomer respectively, k is a dimensionless
friction coefficient and Adc is the area of the downcomer.
This is a first order system with time constant
.
With typical numerical values we find that the time constant is about a second. This is short
in comparison with the sampling period of the experimental data and we will therefore use
steady state relation.

.
.(17)

2.4 Distribution of steam in the drum

Vsd volume of steam under liquid level
Vwd volume of water under liquid level
qsd steam flowrate under the liquid surface
qcd - condensation of steam below the liquid level
The mass balance for steam under the liquid level is
.(18)
qcd can be given by (8)
(

..(19)
Flow qsd is driven by the density differences of water and steam, and momentum of flow
entering the drum.
Good fit to the experimental data have been obtained with the following empirical model.
.(20)
Here Vsd0 denotes volume of steam in the drum in the hypothetical situation where there is no
condensation of steam in the drum and Td is the residence time of steam in the drum.
Drum level
9

The volume of water in the drum is

.(21)

The deviation of the drum level l from normal operating level is

(22)
The term lw represents level variations due to change in water in the drum and ls represents
level variations caused by steam in the drum.
2.5 The Model
Selection of state variables
The accumulation of water is represented by the total water volume Vwt. The total energy is
represented by drum pressure p and the distribution of steam and water is captured by the
steam mass fraction in the risers r and the steam volume in the drum Vsd.
Pressure and water dynamics.
These equations are represented by equation (5).
Riser dynamics
Eliminating the flow rate out of the risers, qr , by multiplying eq(13) by (hw + rhc) and
adding to eq (14) gives

10

*(

*(

.
.(23)
Where
*(

+.

If the state variables p and r are known the riser flow rate qr can be computed from eq(13).

..(24)

Drum Dynamics
The dynamics for the steam in the drum is obtained from the mass balance (18). Introducing
(24) for qr , (19) for qcd and (20) for qsd we get
(

.
(
)

11

.
(25)

Where
*

(
)

+.

+.

.
2.6 Summary
The model can be written as
[

.
.
.
.

where
.
=
.
.

12

*(

+.

+.

+.

.
2.7 Steady state solution of the model
.
.
.
.
Where qdc is given by equation (15) i.e.

A convenient way to find the initial values is to first specify steam flow rate q s and steam
pressure p. The feedwater flowrate qf and input power Q are given by the first two equations
and the steam volume in the drum is given by the last equation. The steam quality is
obtained by solving the nonlinear equations.

.
) .

The steam volume in the drum then can be computed directly.

2.8 The Solution
The solution of the third order non linear model developed can be given by
(26)
13

Where

This is an ordinary differential equation which can be solved by MATLAB using the ODE
function. Basically we use ODE 45 since it is nonstiff differential equations. The ODE 45
implements the fourth or fifth order Runge-Kutta method.
2.9 Step Responses
Fuel Flow changes at medium load
A step change increase of 10 MW is given to the fuel flow rate and the results are discussed
below
As the steam flow out from the drum is constant ,this in turn keeps the pressure of the drum
constantly increasing at the same rate.The constantly increasing pressure in the drum causes
condensation to take place inside the drum which is the reason for an increase in the water
volume, Vwt within the drum.Initially the steam quality at riser outlet shows a rapid increase
and then as time value increases,the quality of steam tends to become more stable with
gradual increase. The volume of steam inside the drum initially shows an increases by a small
volume and eventually decreases as time value increases. As seen from Figure 2 we note a
sharp initial increase in the steam volume and then a steep fall in the curve, the sharp
increases could be atrributed with the increasing steam flowrate in the riser whereas the steep
fall occurs due to the increase in pressure within the drum causing condensation to take place.
At the onset of the step there is a rapid increase in the outlet flow rate of the steam from the
risers as the flowrate of steam from the outlet of the riser needs to match the downcomer
14

flowrate we observe a decrease in the outlet steam flow rate .These flow rates equalize
almost after 30 seconds. The condensation flow changes in an almost step like manner. A
combination of water and steam dynamics are responsible for the response in drum level
indication and is somewhat complicated .The initial increase in the water level is due to the
rapid initial response of steam. The response in level is a combination of two competing
mechanisms. The water volume in the drum increases due to increased condensation caused
by increasing pressure. The volume of steam in the drum first increases a little and then it
decreases because of the increasing pressure.

Fig 2: Response due to step change in fuel flow at medium load.

Steam flow changes at medium load

15

A step increase of 10 kg/s in steam flow at medium load is given and the results are
discussed below.
As observed in figure 3 there is a linear decrease in the pressure of the steam because
of an increased steam flow rate. As we note a decrease in the pressure of steam , this
causes an increase in the evaporation rate of the water within the drum, in turn
causing an overall decrease in the water volume within the drum boiler system. As
mentioned above the steam quality at the riser outlet first increases rapidly due to
pressure decrease and it then decreases due to an increase in the water circulation flow
rate. The drop in steam pressure within the boiler system causes an increase in the
steam volume .There is very rapid increase of flow out of steam from the riser as a
reason of high pressure drop within the system. After this initial transient response of
the riser flow rate it decreases to match the downcomer flow rate.The decreased steam
pressure causes both the downcomer as well as riser flow rates to increase. The
constant decrease in the pressure happens to create an almost step like decrease in the
condensation flow rate response.
The shrink and swell effect is clearly visible. The contributions from the volumes of
steam and water have initially the same sign. There would however be a decrease in
the water volume because of steam flow rate alterations.

16

Fig 2: Response due to step change in steam flow rate at medium load.

17

Chapter 3
Control Schemes

The feedback control scheme will be implemented to control the water level with the
conventional PID controller and advanced controller fuzzy logic controller(FLC).
However we will try only feedwater flow rate as the manipulated variable and both
servo and regulatory testing will be done.
The feedback control scheme
3.1 PID controllers
The control implemented by the PID controller can be given as

(27)

Where u(t) control action taken by the PID controller

e(t) error at time t .
Kc proportional gain of controller
I integral time constant (s)
D derivative time constant (s)

Fig 4: feedback PID control scheme.( r(t) disturbance)(14).

The salient features of the PID controllers are
18

Eliminates offset due to integral action

Better stability criteria due to derivative action
If error is constant then no derivative action.
For a very noisy response with almost zero error the derivative term leads to
large control action which is not needed and hence a drawback.

3.1.1 Tuning methods

Tuning of a controller is a very tedious task. Tuning basically means to set the
parameters (proportional band/gain, integral gain/reset, derivative gain/rate) to get its
optimum values for some desired response. The most basic requirement is that both
response of the system should be stable i.e. the system should not give unbounded
oscillations. It seems even though there are only three parameters in PID controller
tuning is a tedious problem. There are various online and offline methods for tuning a
PID controller. Some of the methods with their advantages and disadvantages are
tabulated below.
Method

Advantages

Disadvantages

Manual Tuning

No math required; online.

Requires
personnel.

Zeigler-Nichols

Proven method; online.

Process upset, some trialand-error,

very
aggressive tuning

Software tools

Consistent tuning; online or offline - Some cost or training

can employ computer-automated involved.
control system design (CAutoD)
techniques; may include valve and
sensor analysis; allows simulation
before downloading; can support
non-steady-state (NSS) tuning.

Cohen-Coon

Good process models.

experienced

Some math; offline; only

good
for
first-order
processes.

Table 1: Choosing a tuning method(14).

Manual Tuning
Manual tuning is a online tuning method and it basically requires an experienced
personnel. Even though the tuning method is manual some rules can be followed for
improved response. First of all set KI and KD values to zero. Increase KP until the loop
19

starts to oscillate. For a quarter amplitude decay type response the KP value should
be set as half of the value at which the response starts to oscillate. Next we have to
increase KI until the offset is eliminated in sufficient time of the process and at last
increase KD, if required until loop is quick to reach its reference after a load
disturbance. But, too much increase in KD can cause excessive response and
overshoot. A fast PID loop usually overshoots a little bit to reach its response
reference point more quickly; however in some system overshoot is not acceptable, in
which case we require an overdamped response.

Parameter

Rise time

Overshoot

Settling
time

Steadystate error

Stability

Kp

Decrease

Increase

Small
change

Decrease

Degrade

KI

Decrease

Increase

Increase

Eliminate

Degrade

KD

Minor
change

Decrease

Decrease

No effect in Improve in
theory
Kd small

Table 2: Effects of increasing a parameter independently(14).

In our process we will try the manual tuning method only to tune the PID controller.

Fig 5: Simulink model to control the level in feedback control scheme using the PID
controller.
20

3.2 Fuzzy Logic Controller

The basic and key concept of fuzziness is that it allows a gradual and continuous
transition rather than a sudden and abrupt change in values. For example in a fuzzy set
a member does not either or neither belongs to a group. Each and every member have
some membership in that group whereas in a normal set it either belongs to a group or
it neither belongs to a group[ either 0 or 1]. In fuzzy logic a member can take any
value between 0 to 1 depending on its membership of the group i.e. how much it
belongs to the group.The membership function of a set maps each element to its
degree i.e. it gives the degree of membership for each element under consideration.
Fuzzy control is a control technique based on fuzzy logic. The basic idea of fuzzy
control is to apply fuzzy inference to control problems. In fuzzy control, the control
box includes fuzzification, fuzzy inference using fuzzy if-then rules, and
defuzzification procedures. Fuzzy rules can include human descriptive judgements.
Commonly used fuzzy variables and their membership functions
We define fuzzy variables that can represent values of the input and output variables.
A commonly used set of seven fuzzy variables follows:
NB = Negative Big
NM = Negative Medium
NS = Negative Small
ZO = Zero
PS = Positive Small
PM = Positive Medium
PB = Positive Big
Or, the two mediums, NM and PM, may be omitted, resulting in the following set of
five fuzzy variables. This smaller set of fuzzy variables is simpler, but it would result
less fine or delicate control. For simplicity, we will use this five fuzzy variable
version hereafter.
NB = Negative Big
NS = Negative Small
ZO = Zero
PS = Positive Small
PB = Positive Big
The next step is to define membership functions for these fuzzy variables. Defining a
membership function is up to us, and the selection of membership
functions affects the control performance. What membership function we choose
depends on many factors, such as the type of application, how much fine control is
required, how fast the control must be performed, and so on. A rule of thumb is that
simpler membership function causes lesser computation time but reduces fine control.
There are two categories of membership functions. One is continuous and the other
discrete

21

Fig 6: Graphical representations of continuous triangular membership functions for seven

fuzzy variables.
The membership functions can be of various shapes such as triangular, trapezoidal
Gaussian etc.
Typical fuzzy control setup
At each time interval our fuzzy control system receives specific values for two inputs,
E and E and yields one output, W.
Where En = ln-lsp : the level difference at time period n
E = En En-1 : changing rate of E at time period n
Just as E is the difference between the current and target values, rather than the
current control value itself, W is a deviation from the current output value. For
example, suppose that level is controlled by a feedwater flow rate, and the amount of
feedwater at time period n is qfn; then
qf n+1 = qf n +W.
Fuzzy if-then rules that derive W from E and E
The fuzzy if then rules described for our system for level control is
Rule 1: If E is ZO and E is NB then W is PB.
Rule 2: If E is ZO and E is NS then W is PS
.
.
.
.
Rule 25: If E is PB and E is PB then W is NB.

22

The rules are tabulated below.

NB

NS

ZO

PS

NB

PB

PB

PM

PS

ZO

NS

PB

PB

PS

ZO

NS

ZO

PB

PS

ZO

NS

NB

PS

PM

ZO

NS

NB

NB

PB

ZO

NS

NB

NB

NB

PB

Table 3: If then rules for our system for level control.

3.2.1 Optimization of the membership functions
The optimization of the membership functions to obtain the best response is done by
using a non-traditional optimization technique called the genetic algorithm (GA).
Genetic Algorithm(GA)
Genetic algorithm is a non-traditional optimization technique developed by Prof. John
Holland and his colleagues and students at University of Michigan at around 1975
based on process of evolution. It is basically a search algorithm based on mechanics
of natural selection and natural genetics. It is based on the survival of the fittest
concept (darwanian theory) and simulates the process of evolution.
The key idea behind GA is that evolution is an optimizing process. The central theme
of research in GA is robustness.
Why GA?

23

Fig 7: A comparison between GA and traditional techniques.

As we can see from the figure 7 that F(x) is a function with of x over a region of x
with one global optimum and three local optimum for min(F(x)).

The table below clearly describes why we are using GA.

Initial guess

Traditional optimization

Genetic Algorithm

Region I

Converges to 1

Converges to 2

Region II

Converges to 2

Converges to 2

Region III

Converges to 3

Converges to 2

Table 4: A comparison between GA and traditional techniques

As we can see that in traditional optimization the solution approaches the local
optimum value whereas in GA the solution will always approach the global optimum
value.
However there is a disadvantage that
optimization technique.

GA is slow compared to other general

The differences between GA and traditional methods of optimization can be stated as:

GA work with coding of parameter set ,not the parameter themselves

GA search for population of point not a single point.

24

GA use the objective function information and not the derivative or second
derivative.
GA use stochastic transition rates , not deterministic rates.

In our case the objective function is to minimize the integral of square error(ISE) and
the variables are k1, k2 and k3. k1 , k2 , k3 are the values which need to be multiplied
with the ranges of the functions E, E and W. The optimum values of k1,k2 and k3
are obtained by using GA inthe optimization toolbox in MATLAB.

Fig 8:Simulink model for feedback fuzzy logic controller(FLC)

3.3 Neural Network (NN) predictive control
Neural Networks as function approximators

As shown in Figure 9, we have some unknown function that we wish to approximate.We

want to adjust the parameters of the network so that it will produce the same response as the
unknown function, if the same input is applied to both systems.

Fig 9: Neural networks as function approximator(7).

25

Neuron Model
The scalar input p is multiplied by the scalar weight w to form wp , one of the
terms that is sent to the summer. The other input,1 , is multiplied by a bias b and then passed
to the summer. The summer input often referred to as net input , goes into a transfer function,
f which produces the scalar neuron output a.

Fig 10: Single Input neuron(7).

The neuron output is calculated as
.(28)
w and b are both adjustable scalar quantities of the neuron.
The transfer function may be linear or non-linear function of n. One of the most commonly
used is the log-sigmoid transfer function.

Fig 11: log sigmoid transfer function(7).

This transfer function takes the input and squashes the output in the range of 0 to 1.
..(29)

26

Fig 12: A multiple input neuron[7].

Typically a neuron has more than one input . A multiple input neuron is shown in figure 12.
The neuron has a bias b, which is summed with the weighted inputs to form the net input n.
.
which can be written in matrix form as
..(30)
Where the matrix W from the single neuron case has only one row.

Fig 13: Neuron with R inputs, Matrix notation[7].

Commonly one neuron, even with many inputs, is not sufficient. We might need five or ten,
operating in parallel, in what is called a layer. A single-layer network of S neurons.

27

Fig 14: Layer of S neurons[7].

The S- neuron R-input layer can also be shown in matrix notation as

Fig 15: Layer of S- neurons. Matrix notation [7]

Consider a network with several layers . Each layer has its own weight matrix W, its own
bias vector b a net input vector n, and an output vector a. We need to introduce some
additional notation to distinguish between these layers. We will use superscripts to identify
the layers. Thus, the weight matrix for the first layer is written as W1, and the weight matrix
for the second layer is written as W2. This notation is used in the three-layer network shown
in Figure 16. As shown, there are R inputs, S1 neurons in first layer, S2 neurons in second
layer etc. A layer whose output is the network output is called an output layer. The other
layers are called hidden layers.

28

Fig 16: Three- Layer Network[7].

The first stage of model predictive control is to train a neural network to represent the
forward dynamics of the plant. The prediction error between the plant output and the neural
network output is used as the neural network training signal. The process is represented by
Figure 17.
One standard model that has been used for nonlinear identification is the Nonlinear
Autoregressive-Moving Average (NARMA) model:
[

]..(31)

where u(k) is the system input , y(k) is the system output and d is the system delay (we will
use a delay of 1 for the predictive controller).

Fig 17: Plant Identification[7].

For the identification phase, we train a neural network to approximate the nonlinear function .
The structure of the neural network plant model is given in Figure 14, where the blocks
labelled TDL are tapped delay lines that store previous values of the input signal.

29

Fig 18: Neural network plant model[7].

We have modified our previous notation here, to allow more than one input into the network.
IWi,j is a weight matrix from input number j to network number i.
The equation for the plant model is given by

].(32)

where [ ] is a function implemented by the neural network, and x is a vector containing

all the network weight and biases.
The model predictive control method is based on the receding horizon technique. The neural
network model predicts the plant response over a specified time horizon. The predictions are
used by a numerical optimization program to determine the control signal that minimizes the
following performance criterion over the specified horizon.

..(33)

where N1 , N2 and Nu define the horizons over which the tracking error and the control
increments are evaluated. The u variable is the tentative control signal , yr is the desired
response and ym is the network model response. The value determines the contribution that
the sum of the squares of the control increments has on the performance index..
The following block diagram illustrates the model predictive control process. The controller
consists of the neural network plant model and the optimization block. The optimization
block determines the values of u that minimize J and then the optimal u is the output to the
plant.

30

Fig 19: Neural network predictive control[7].

Fig 20: The Simulink model for neural network predictive control.

31

Chapter 4
Results and Discussions
4.1 PID controller
4.1.1 Regulatory Problem
In the first regulatory problem we are giving a step change in fuel flow as 10 MW.
The PID parameters obtained by manual tuning are
KC = 500
I = 1 , D = 1.
The response of the system can be seen as follows

Fig 20: PID response for a step change in fuel flow.

The response obtained is a under damped response and the charecteristic of the
response are
Rise time, tr = 0 s
Peak time , tp= 5 s

32

Time period, T= 30 s
Response time , ts = 100 s
Decay ratio = 0.33.
Overshoot = 0.58
The manipulated variable i.e. feedwater flow rate changes are given by

Fig 21 : Changes in feed water flow rate to control the level.

The changes in the feed water flow rate at the beginning drops to almost 26 kg/s and
then rises to almost 55 kg/s and after about 160 s it settles to a constant value of 51
kg/s.
In the second regulatory problem we are giving a step change in steam flow as 10
kg/s.
The PID parameters obtained by manual tuning are
KC = 500
I = 1 , D = 1.
The response of the system can be seen as follows

33

Fig 22: PID response for a step change in steam flow.

The response obtained is a under damped response and the charecteristic of the
response are
Rise time, tr = 0 s
Peak time , tp= 4 s
Time period, T= 30 s
Response time , ts = 110 s
Decay ratio = 0.5.
Overshoot = 0.707
The manipulated variable i.e. feedwater flow rate changes are given by

34

Fig 23 : Changes in feed water flow rate to control the level.

The changes in the feed water flow rate at the beginning drops to almost 40 kg/s and
then rises to almost 65 kg/s and after about 200 s it settles to a constant value of 59
kg/s. The water level increases at last to compensate to the losses in steam due to
increase in steam flow rate.
4.1.2 Servo problem
In the servo problem we are giving a step change of 0.1 m in the level set point.
The PID parameters obtained by manual tuning are
KC = 700
I = 3 , D = 1.
The response of the system can be seen as follows

35

Fig 24: PID response for a step change in steam flow.

The response obtained is a under damped response and the charecteristic of the
response are
Rise time, tr = 11 s
Peak time , tp= 21 s
Time period, T=62 s
Response time , ts = 90 s
Decay ratio = 0.146.
Overshoot = 0.38
The manipulated variable i.e. feedwater flow rate changes are given by

36

Fig 25 : Changes in feed water flow rate to control the level.(servo problem)
The changes in the feed water flow rate at the beginning rises to almost 180 kg/s and
then drops to almost 15 kg/s and after about 150 s it settles to a constant value of 50
kg/s.
4.2 Fuzzy Logic Controller
4.2.1 Regulatory Problem
Step Change in fuel flow rate by 10 MW
The membership function of the fuzzy logic controller before optimization are shown
below

37

Fig 26 : membership function for input variable E. Range [-0.1 0.1]

Fig 27: membership function for input variable E. Range [-0.1 0.1]

Fig 28: membership function for output variable W . Range [-20 20]
The responses before and after optimization of the membership functions are
compared below. As can be seen from the figure that the integral of square error (ISE)
before and after optimization have a difference of almost 26. The three values k1,k2
and k3 which are multiplied with the membership functions of E,E and W
respectively to minimize ISE are 0.109,0.814 and 0.302 respectively.

38

Fig 27: response before optimization of membership functions (change in fuel flow)
ISE=148.1217

Fig 28: results after optimization of membership functions (change in fuel flow)
ISE=122.0842

39

Final results for step change in fuel flow

As we see from the response that in fuzzy control the response obtained is more
steady as obtained in PID controller. The response rises a peak of about 0.1 m above
set point and then eventually decreases to the set point. The response obtained after 90
seconds.

Fig 29: response of FLC due to step change in fuel flow rate.

Fig 30: changes in feedwater flow rate to control the level for FLC for a step in fuel
flow.

40

As we can see from figure 20 that the changes in feedwater flow rate can be limited
within a range ,in or case it is +- 20 kg/s but in PID controller it is not possible if the
disturbance by somehow increases to a great extent.
4.2.2 Servo Problem
The membership function of the fuzzy logic controller before optimization are shown
below

Fig 31 : membership function for input variable E. Range [-0.1 0.1]

Fig 32: membership function for input variable E. Range [-0.1 0.1]

Fig 33: membership function for output variable W . Range [-20 20]

41

The responses before and after optimization of the membership functions are
compared below. As can be seen from the figure that the integral of square error (ISE)
before and after optimization have a difference of almost 0.3 which is low but the
response time decreases by an extent of almost 5 s. The three values k1,k2 and k3
which are multiplied with the membership functions of E,E and W respectively to
minimize ISE are 1.981,1.593 and 1.34 respectively.

Fig 34: response before optimization of membership functions (change in set point )
ISE=156.6541

Fig 35: response after optimization of membership functions (change in set point )
ISE=156.333

42

Final results for step change in set point.

Fig 36: response of FLC due to step change in set point.

Again we can change from the response that it is more smooth than the response
obtained by PID controller and nonoscillatory. The response time is about 105
seconds.
The feed water flow rate is shown below

43

Fig 37: changes in feedwater flowrate to control the level.(servo problem)

4.3 Comparison between PID and FLC
Step
Change
in

PID
ISE

FLC
ISE

PID
Response
time

FLC
Response
time

PID
Type of
response

Fuel flow
Steam
Flow
Set Point

147.57
147.55

122.08
--------

100 s
110 s

90 s
--------

Oscillatory Smooth
Oscillatory Smooth

160.95

156.33

90 s

105 s

Oscillatory Smooth

Table 4: Comparison between PID and FLC response

44

FLC
Type of
response

Chapter 5
Conclusion
After analyzing the results we can conclude as:

The model obtained clearly takes the shrink and swell effect in the drum
The water level contribution(lw) to the water level(l) actually is the water in
the drum and hence we can distinguish it from the model.
Both servo and regulatory problems are taken into account to control the water
level.
The feed water flow rate is very efficient in controlling the level of water.
PID gives an oscillatory response whereas FLC gives a smooth response.
We can restrict the manipulated variable, in our case the feed water flow rate
to a certain range in FLC but it is not possible in PID controllers.

45

References:
1) K.J.Astrom, R.D.Bell, Drum-boiler dynamics, Automatica,2000,vol 36,pg 363-378.
2) A.Parry, J.F. Petotrot and W.J.Vivier. Recent progress in sg level control in French
pwr plants. In Proc. Int. Conf on Boiler Dynamics and Control in Nuclear
Powerstations, pg 81-88. British Nuclear Energy Society ,1995
3) H.H. Hacene Habbi, M.Zelmat, B.O. Bouamama,A dynamic fuzzy model for drum
boiler turbine system,Automatica,2002,vol 39,pg 1213-1219.
4) E.J.Adam, J.L.Maarchetti (1997) ,Dynamic Simulation of large boilers with natural
circulation, Computers and Chemical Engineering,1999,vol 23,pg 1031-1040.
5) M.E. Flynn, M.J.O.Malley(1999), A drum boiler model for long term power system
dynamic simulation, IEEE transactions on Power Systems,1999, Vol. 14, pg .
6) H.Kim, S.Choi(2005), A model on water level dynamics in natural circulation drumtype boilers, International Communications in Heat and Mass Transfer,2005,vol
32,pg 786-796.
7) M.T.Hagan,H.B.Demuth,O.D.Jesus, An introduction to use of neural networks in
control system, International journal of robust and non linear control,2002, Vol 12
,pg 959-985.
8) M. Hudson Beale, M. T. Hagan, H. B. Demuth,Neural Network Toolbox 7,
Users Guide MATLAB.
9) D.M. Himmelblau,Applications of Artificial Neural Networks in Chemical
Engineering, Korean Journal. Chemical Engineering.2000,vol 17,pg 373-392.
10) Toshinori Munakata, Fundamentals of the New Artificial Intelligence, SpringerVerlag London Limited, 2008, ISBN- 978-1-84628-839-5.
11) Fuzzy Logic Toolbox, Users Guide, MATLAB ,R2014a.
12) Basics of S-functions retrieved march 5, 2014 from the world wide web
http://www.mathworks.in/help/simulink/s-function-basics.html
13) Fuzzy Logic toolbox retrieved april 5, 2014 from the world wide web
http://www.mathworks.in/products/fuzzy-logic/
14) PID Controller retrieved may 9, 2014
http://en.wikipedia.org/wiki/PID_controller.

46

from

the

world

wide

web

15) Neural Network toolbox retrieved may 5, 2014 from the world wide web
http://www.mathworks.in/products/neural-network/

47

Appendix A
MATLAB CODE

function[sys,x0,str,ts]=sfd302_sfcn(t,x,u,flag,x_init)
%sfd302 S-function version of the four state model
%compatible with simulink block diagrams.
global qr qdc qct;
%plant parameters------------------------------mt=300000;mr=100000;md=20000;Cp=650;
Cfw=4.18;Ad=23;Vd=37;Vr=37;Vdc=11;
Vt=Vd+Vr+Vdc;
ke=200;Vsd0=8;beta=0.3;
%properties of steam and water in saturated table
a01=2.7254E6;a11=-1.8992E4;a21=-1160.0;
a02=53.1402;a12=7.673;a22=0.36;
a03=1.4035E6;a13=4.9339E4;a23=-880.0;
a04=691.35;a14=-18.672;a24=-0.0603;
a05=310.6;a15=8.523;a25=-0.33;
%------------------------------------------------if abs(flag) ==1, %compute state derivatives
%state initial conditions
Vwt=x(1);
p=x(2);
ar=x(3);
Vsd=x(4);
%controls or inputs
Q=u(1)*1e6; %conversion to watts
qf=u(2);
tf=u(3);
qs=u(4)+4.8*(p-8.5); %correction for steam transducer
% properties of steam and water in saturated state
hs=a01+(a11+a21*(p-10))*(p-10);
dhsdp=a11+2*a21*(p-10);
rs=a02+(a12+a22*(p-10))*(p-10);
drsdp=a12+2*a22*(p-10);
hw=a03+(a13+a23*(p-10))*(p-10);
dhwdp=a13+2*a23*(p-10);
rw=a04+(a14+a24*(p-10))*(p-10);
drwdp=a14+2*a24*(p-10);
ts=a05+(a15+a25*(p-10))*(p-10);
dtsdp=a15+2*a25*(p-10);
%properties of water in subcritical state
hf=(Cfw*tf+p*1e3/rw)*1e3;
%the model equations-------------------hc=hs-hw;
hr=ar*hs+(1-ar)*hw;
%average steam quality volume rati and partial derivatives
z=ar*(rw-rs)/rs;
48

av=rw/(rw-rs)*(1-(1/z)*log(1+z));
davdar=rw/rs*((1/z)*log(1+z)-1/(z+1))/z;
zp=(rs*drwdp-rw*drsdp)*ar/rs/rs;
z1=(rw*drsdp-rs*drwdp)/(rw-rs)/(rw-rs);
z2=rw/(rw-rs)*((1/z)*log(1+z)-1/(1+z))/z*zp;
davdp=z1*(1-(1/z)*log(1+z))+z2;
%circulation flow
s1=ke*(rw-rs)*Vr*av;
qdc=sqrt(s1);
%equations for coefficents of derivatives of state variables
Td=600/qs;
Vst=Vt-Vwt;
Vwd=Vwt-Vdc-(1-av)*Vr;
e11=rw-rs;
e12=Vst*drsdp+Vwt*drwdp;
e21=hw*rw-hs*rs;
e22x=-Vt*1e6+mt*Cp*dtsdp;
e22=Vst*(hs*drsdp+rs*dhsdp)+Vwt*(hw*drwdp+rw*dhwdp)+e22x;
e3w=(rw*dhwdp-ar*hc*drwdp)*(1-av)*Vr;
e3x=(rs+(rw-rs)*ar)*hc*Vr*davdp-Vr*1e6+mr*Cp*dtsdp;
e32=((1-ar)*hc*drsdp+rs*dhsdp)*av*Vr+e3w+e3x;
e33=(rs+(rw-rs)*ar)*hc*Vr*davdar;
e42y=(1+beta)*ar*Vr;
e42x=e42y*((1-av)*drwdp+av*drsdp+(rs-rw)*davdp)+Vsd*drsdp;
e42=(rs*Vsd*dhsdp+rw*Vwd*dhwdp-Vsd*1e6+md*Cp*dtsdp)/hc+e42x;
e43=(1+beta)*(ar*Vr*(rs-rw)*davdar);
e44=rs;
e=[e11,e12,0,0
e21,e22,0,0
0,e32,e33,0
0,e42,e43,e44];
%the right hand side of state equations
b4=(hf-hw)*qf/hc+rs*(Vsd0-Vsd)/Td;
b=[qf-qs; Q+qf*hf-qs*hs; Q-qdc*ar*hc; b4];
%solve linear equation for derivatives
dx=e\b;
qx=(rw-rs)*Vr*davdp*dx(2)+(rw-rs)*Vr*davdar*dx(3);
%two important flows for understanding behaviour
qr=qdc-(av*drsdp+(1-av)*drwdp)*Vr*dx(2)+qx;
qctx=rs*Vst*dhsdp+rw*Vwt*dhwdp-Vt*1e6+mt*Cp*dtsdp;
qct=((hw-hf)*qf+qctx*dx(2))/hc;
%step the derivatives, dx is a 4X1 vector
sys = [dx];
%------------------------------------------------------------elseif flag==3, %Compute outputs
%extract state variables
Vwt=x(1);
p=x(2);
ar=x(3);
Vsd=x(4);
49

%drum level
rs=a02+(a12+a22*(p-10))*(p-10);
rw=a04+(a14+a24*(p-10))*(p-10);
z=ar*(rw-rs)/rs;
av=rw/(rw-rs)*(1-(1/z)*log(1+z));
Vwd=Vwt-Vdc-(1-av)*Vr;
ls=Vsd/Ad;
lw=Vwd/Ad;
l=lw+ls;
sys=[l,p,lw,ls,qr,av,qdc,ar,Vwt,Vsd,qct];
%-------------------------------------------------------------elseif flag==0, %INITIALIZATIONS
str=[];
sizes(1)=4; %4 continious states
sizes(2)=0; %0 discrete states
sizes(3)=11; % 11 outputs
sizes(4)=4; %4 inputs
sizes(5)=0; % no discontinious roots
sizes(6)=0; % no direct term
sizes(7)=1; ts=[0,0]; %efficency fix(see[1])
sys=sizes';
x0(1)=x_init(1);
x0(2)=x_init(2);
x0(3)=x_init(3);
x0(4)=x_init(4);
else
sys=[]; %flags 2 or 4 , not considered since we have no discrete states.
end;
% Also flag 5 is of no interest.

50

Appendix B
Optimization Code
The code for the fuzzy logic file is:
d=newfis('boiler','mamdani');
d=addvar(d,'input','err',[-0.1 0.1]*k1);
d=addmf(d,'input',1,'nb','trimf',[-0.19 -0.1 -0.01]*k1);
d=addmf(d,'input',1,'ns','trimf',[-0.03 -0.015 0]*k1);
d=addmf(d,'input',1,'z0','trimf',[-0.0004 0 0.0004]*k1);
d=addmf(d,'input',1,'ps','trimf',[0 0.015 0.03]*k1);
d=addmf(d,'input',1,'pb','trimf',[0.01 0.1 0.19]*k1);
d=addvar(d,'input','dele',[-0.1 0.1]*k2);
d=addmf(d,'input',2,'nb','trimf',[-0.15 -0.1 -0.05]*k2);
d=addmf(d,'input',2,'ns','trimf',[-0.1 -0.05 0]*k2);
d=addmf(d,'input',2,'z0','trimf',[-0.0004 0 0.0004]*k2);
d=addmf(d,'input',2,'ps','trimf',[0 0.05 0.1]*k2);
d=addmf(d,'input',2,'pb','trimf',[0.05 0.1 0.15]*k2);
d=addvar(d,'output','w',[-20 20]*k3);
d=addmf(d,'output',1,'nb','trimf',[-27.5 -20 -12.5]*k3);
d=addmf(d,'output',1,'nm','trimf',[-15 -10 -5]*k3);
d=addmf(d,'output',1,'ns','trimf',[-10 -5 0]*k3);
d=addmf(d,'output',1,'z0','trimf',[-2.5 0 2.5]*k3);
d=addmf(d,'output',1,'ps','trimf',[0 5 10]*k3);
d=addmf(d,'output',1,'pm','trimf',[5 10 15]*k3);
d=addmf(d,'output',1,'pb','trimf',[12.5 20 27.5]*k3);
ruleList=[ ...
11711
12711
13611
14511
15411
21711
22711
23511
24411
25311
31711
32511
33411
34311
35111
41611
42411
43311
44111
45111
51411
51

52311
53111
54111
5 5 1 1 1];
d=addrule(d,ruleList);

The code for the process file is:

mt=300000;mr=100000;md=20000;Cp=650;
Cfw=4.18;Ad=23;Vd=37;Vr=37;Vdc=11;
Vt=Vd+Vr+Vdc;
ke=200;Vsd0=8;beta=0.3;
%properties of steam and water in saturated table
a01=2.7254E6;a11=-1.8992E4;a21=-1160.0;
a02=53.1402;a12=7.673;a22=0.36;
a03=1.4035E6;a13=4.9339E4;a23=-880.0;
a04=691.35;a14=-18.672;a24=-0.0603;
a05=310.6;a15=8.523;a25=-0.33;
dt=1;
% t(1)=0;
ise=zeros(1,200);
err=zeros(1,200);
qf=zeros(1,200);
p=zeros(1,200);
ar=zeros(1,200);
t=zeros(1,200);t=1:1:200;
Vwt=zeros(1,200);
Vsd=zeros(1,200);
l=zeros(1,200);lw=zeros(1,200);ls=zeros(1,200);Q=zeros(1,200);qs=zeros(1,200);
for i=1:200
Vwt(1)=57.2;
p(1)=8.5;
ar(1)=0.0511;
Vsd(1)=4.8971;
Q(1:50)=86.813*1e6;%conversion to watts
Q(51:200)=96.813*1e6;
qf(1)=50;
tf=240;
qs(1:200)=50+4.8*(p(i)-8.5);
%qs(51:200)=60+4.8*(p(i)-8.5);
% properties of steam and water in saturated state
hs=a01+(a11+a21*(p(i)-10))*(p(i)-10);
dhsdp=a11+2*a21*(p(i)-10);
rs=a02+(a12+a22*(p(i)-10))*(p(i)-10);
drsdp=a12+2*a22*(p(i)-10);
52

hw=a03+(a13+a23*(p(i)-10))*(p(i)-10);
dhwdp=a13+2*a23*(p(i)-10);
rw=a04+(a14+a24*(p(i)-10))*(p(i)-10);
drwdp=a14+2*a24*(p(i)-10);
ts=a05+(a15+a25*(p(i)-10))*(p(i)-10);
dtsdp=a15+2*a25*(p(i)-10);
%properties of water in subcritical state
hf=(Cfw*tf+p(i)*1e3/rw)*1e3;
%the model equations-------------------hc=hs-hw;
hr=ar(i)*hs+(1-ar(i))*hw;
%average steam quality volume rati and partial derivatives
z=ar(i)*(rw-rs)/rs;
av=rw/(rw-rs)*(1-(1/z)*log(1+z));
davdar=rw/rs*((1/z)*log(1+z)-1/(z+1))/z;
zp=(rs*drwdp-rw*drsdp)*ar(i)/rs/rs;
z1=(rw*drsdp-rs*drwdp)/(rw-rs)/(rw-rs);
z2=rw/(rw-rs)*((1/z)*log(1+z)-1/(1+z))/z*zp;
davdp=z1*(1-(1/z)*log(1+z))+z2;
%circulation flow
s1=ke*(rw-rs)*Vr*av;
qdc=sqrt(s1);
%equations for coefficents of derivatives of state variables
Td=600/qs(i);
Vst=Vt-Vwt(i);
Vwd=Vwt(i)-Vdc-(1-av)*Vr;
e11=rw-rs;
e12=Vst*drsdp+Vwt(i)*drwdp;
e21=hw*rw-hs*rs;
e22x=-Vt*1e6+mt*Cp*dtsdp;
e22=Vst*(hs*drsdp+rs*dhsdp)+Vwt(i)*(hw*drwdp+rw*dhwdp)+e22x;
e3w=(rw*dhwdp-ar(i)*hc*drwdp)*(1-av)*Vr;
e3x=(rs+(rw-rs)*ar(i))*hc*Vr*davdp-Vr*1e6+mr*Cp*dtsdp;
e32=((1-ar(i))*hc*drsdp+rs*dhsdp)*av*Vr+e3w+e3x;
e33=(rs+(rw-rs)*ar(i))*hc*Vr*davdar;
e42y=(1+beta)*ar(i)*Vr;
e42x=e42y*((1-av)*drwdp+av*drsdp+(rs-rw)*davdp)+Vsd(i)*drsdp;
e42=(rs*Vsd(i)*dhsdp+rw*Vwd*dhwdp-Vsd(i)*1e6+md*Cp*dtsdp)/hc+e42x;
e43=(1+beta)*(ar(i)*Vr*(rs-rw)*davdar);
e44=rs;
e=[e11,e12,0,0
e21,e22,0,0
0,e32,e33,0
0,e42,e43,e44];
%the right hand side of state equations
b4=(hf-hw)*qf(i)/hc+rs*(Vsd0-Vsd(i))/Td;
b=[qf(i)-qs(i); Q(i)+qf(i)*hf-qs(i)*hs; Q(i)-qdc*ar(i)*hc; b4];
%solve linear equation for derivatives
dx=e\b;
qx=(rw-rs)*Vr*davdp*dx(2)+(rw-rs)*Vr*davdar*dx(3);
53

%two important flows for understanding behaviour

qr=qdc-(av*drsdp+(1-av)*drwdp)*Vr*dx(2)+qx;
qctx=rs*Vst*dhsdp+rw*Vwt(i)*dhwdp-Vt*1e6+mt*Cp*dtsdp;
qct=((hw-hf)*qf(i)+qctx*dx(2))/hc;
Vwt(i+1)=(Vwt(i)+dx(1)*dt);
p(i+1)=(p(i)+dx(2)*dt);
ar(i+1)=ar(i)+dx(3)*dt;
Vsd(i+1)=Vsd(i)+dx(4)*dt;
ls(i)=Vsd(i)/Ad;
lw(i)=Vwd/Ad;
l(i)=lw(i)+ls(i);
lwsp(1:50)=0.859;
%lwsp(51:200)=0.909;
err(i)=lw(i)-lwsp(i);
if(err(i)>=0.00001 || err(i)<=-0.00001)
if i==1;
ise(i)=err(i)*(err(i));
else
ise(i)=ise(i-1)+err(i)*(err(i));
end
end

if i==1
dele=0;
else
dele=err(i)-err(i-1);
end
w=evalfis([err(i) dele], d);
qf(i+1)=50+w;
end
iseplot=ise(200);

The objective function is

:
function [iseplot]=objective(par)
k1=par(1);
k2=par(2);
k3=par(3);
fuzzyboils;
drumboilerdynamics;
end

54

55