Ench 617 Notes II (2021)
Ench 617 Notes II (2021)
ADVANCED CONTROL
The derivation of first-principles dynamic models (e.g. based upon material, energy and
momentum balances) for even relatively simple chemical processes can be a difficult and
time-consuming task. For this reason, the usual industrial practice is to construct
approximate models from available process data (Ogunnaike and Ray, 1994). For feedback
controller tuning purposes, the most commonly employed approximation is the “first-
order-plus-deadtime (FOPDT)” model.
This model may be obtained by many procedures, but in the interest of simplicity, a
particularly popular method utilizes open-loop step-response data. This involves making
a step change in the controller output signal U (t ) (e.g. steam valve stem position) with the
controller in manual operating mode in order to elicit a response in the measured variable
Y (t ) (e.g. exchanger outlet temperature). A record of this response with time is known as
the process reaction curve. Note that the observed response is due to the combined
dynamics of the final control element (steam valve), the process itself (heat exchanger),
and the measuring device (thermocouple). The approximate model will therefore describe
this combined process ensemble.
dY (t )
Y (t ) K U (t D ) (2.1)
dt
where
Yss K U ss (2.2)
Subtracting (2.2) from (2.1) and introducing deviation or perturbation variables (e.g. Y (t )
Y (t ) Yss ) leads to
dY (t )
Y (t ) K U (t D) (2.3)
dt
This linear ordinary differential equation is said to be in standard form because the
coefficient of the Y (t ) term is unity.
N.B. In this course, a variable appearing with the superscript ‘prime’ () will always denote
its deviation from its initial steady-state value and never its first derivative. Note also that
Let us pretend for now that the deadtime D 0. If we make a step change of size U in
the controller output at time t 0, then for t 0, U (t ) U (t ) U ss U and
dY (t ) 1
Y (t ) K U dY (t ) ( K U Y (t )) dt
dt
(2.5)
dY (t ) dt t
C ln ( K U Y (t )) C
( K U Y (t ))
But when t 0, Y (0) Y (0) Yss 0, and so the constant of integration is found to be
C ln ( K U ) (2.6)
It should be pointed out that it was not necessary to recast (2.1) into the form of (2.3) in
order to solve this simple differential equation. However, the use of deviation variables is
convenient because the boundary condition Y (0) becomes zero and any constant terms in
the original ODE will disappear when the steady-state version (e.g. (2.2)) is subtracted
from it.
2
Inserting (2.6) into (2.5) leads to
t K U Y (t ) t
ln ( K U Y (t )) ln ( K U ) ln
K U
K U Y (t )
et / Y (t ) K U (1 e t / )
K U
or
In actuality, the time delay is nonzero in many chemical process control applications. This
means that the measured variable Y cannot begin to respond to the bump in U until D
minutes have elapsed. Therefore, the solution to (2.1) is properly written as
Yss , t D
Y (t ) ( t D )/
(2.8)
Yss K U (1 e
), t D
i) In the limit as t , the measured variable (e.g. measured heat exchanger outlet
temperature) converges to the ultimate value
Finally, let Y Y denote the eventual change in Y. Then the physical significance
of the steady-state gain, K, is revealed:
3
ii) At time t D, the term in brackets in (2.8), namely 1 e (t D )/
1 e1 0.632, so from (2.8) we find that
Y ( D) 0.632 Y
It is apparent that after D minutes have elapsed, a FOPDT system will have made
63.2% of its eventual change, which will be observed when it has re-established
steady-state. Mathematically speaking, this new steady-state is only achieved as
t , but in practice we conclude that we have reached steady-state when no further
changes in Y can be observed outside of the sensor noise band. Similarly,
After 5 time constants (plus the deadtime) have elapsed, we are 99.3% of the way to
the asymptotic value Y . This is why many practitioners use the quantity 5 to
convey the approximate time-to-steady-state or settling time. The longer the settling
time, the more challenging the control problem is going to be; in other words, the more
difficult it will be for the controller to keep Y close to its setpoint, Y sp , once it has
been switched to automatic.
The simulated step response or process reaction curve of a FOPDT process is given in Fig.
2.1. Observations i) and ii) motivate a technique for estimating the parameters
K , and D of a FOPDT system which we shall refer to as Method I. Method I delivers
best results when the response looks truly first-order; that is, when the initial response of Y
is fastest. Method I is very easy to use. We first note the length of time it took Y to start
moving after the bump in U – this is the deadtime, D. Next, measure the eventual change
in Y, i.e. Y . The steady-state gain, K, is obtained through dividing Y by the size of the
bump you made in the controller output, i.e. U . Lastly, to estimate the time constant, ,
we calculate Y ( D) Yss 0.632 Y , then locate on the x-axis the time at which Y(t)
assumed that value. Our estimate of is yielded by subtracting the value of D from this
time. When this approach is applied to the system of Fig. 2.1, we arrive at the following
values of the unknown parameters: K 2 % / %, 5 min., D 5 min. Note that K
may be positive or negative, but it is always dimensionless.
4
75 100
Y∞ 90
70
Measured Variable, Y(t) <%>
70
Yss
60
60
55 U
50
D
Steady-State Gain, K ≡ Y/ U
50 40
-10 0 10 20 30 40
Time, t <min>
Alternatively, plants which exhibit an ‘S-shape’ in their process reaction curves (Fig. 2.2),
characteristic of higher-order dynamic systems, are better-modelled using the method
originally proposed by Ziegler and Nichols (1942). It is widely referred to as the Ziegler-
Nichols ‘open-loop method’, but in this course we’ll call it Method II. It is also very easy
to use Method II, but its theoretical justification is more difficult to wade through than that
of Method I. The first step in using Method II to fit a FOPDT model to an S-shaped reaction
curve is to locate its point of inflection, i.e. where the rate-of-change of Y(t) is greatest in
magnitude. We then draw a tangent to the curve at this point and extend it in both directions
until it intersects with the horizontal lines f (t ) Yss and f (t ) Y . The times
corresponding to the two points of intersection give us our estimates of the effective
deadtime and time constant. (The time of the second one is t D, so we must again
subtract our estimate of D to obtain the effective time constant.) The steady-state gain
K Y / U as in Method I. Using Method II on the S-shaped reaction curve of Fig. 2.2,
we would obtain K 0.86 % / %, 36 min. and D 18 min. as the parameters of
our FOPDT approximation.
5
55 90
Y∞
54 85
53 80
52 Y 75
51 70
Yss
50 65
49 60
48 55
U
47 50
D
46 45
Steady-State Gain, K Y / U
45 40
-10 10 30 50 70 90 110
Time, t <min>
We will now explain how this technique can be derived from the step response of a true
FOPDT system, Eq. (2.8). For t D,
dY (t ) d K U ( t D )/
( K U e (t D )/ ) e (2.12)
dt dt
The function e (t D )/ decreases monotonically on the interval t [ D, ), beginning with
a value of 1 at t = D and converging to 0 as t . We introduce a new variable ,
defined as the absolute value of the slope of Y:
dY (t ) | K U | (t D )/
(t ) e (2.13)
dt
The exponential term and the time constant move outside the absolute value sign since they
can never be negative. Clearly, (t ) is maximized when t D, since that’s when e (t D )/
assumes its maximum value of unity. The maximum value of (t ) will be designated
6
dY (t ) | K U | | Y |
max (2.14)
dt tD
recalling (2.11). Simple rearrangement gives an expression for the time constant,
(Ogunnaike and Ray, 1994):
| Y |
(2.15)
max
To summarize our argument thus far, we have established that the time constant of an open-
loop FOPDT process can be estimated by measuring the eventual change in Y(t) as well as
its slope at its point of infection, i.e. when the magnitude of its slope is largest. Put another
way,
dY (t )
max slope of tangent to Y (t ) at its inflection point
dt inflection pt .
It next falls to us to derive the equation of this tangent. Remember that the equation of a
straight line can be obtained from knowledge of its slope and one of its points. As just
shown, the inflection point of a true FOPDT system occurs at t = D. The value of Y(t) at
this time can be obtained from (2.8):
hence the point (D, Yss) lies on the tangent. With respect to the slope, there are two cases
to consider, according to whether the measured variable Y rises or falls in response to the
step change in U.
X (t ) Y ( D) X (t ) Yss max , Y 0
Tangent slope (2.18)
t D t D max , Y 0
(We’ve used the symbol X(t) to refer to the tangent line to prevent confusion with the
process reaction curve, Y(t).)
Case I: Y 0
Now that we have identified the equation of the tangent, we substitute (2.15) to obtain
7
| Y | Y Y Yss
X (t ) (t D) Yss (t D ) Yss (t D ) Yss (2.20)
Case II: Y 0
| Y | Y Y
X (t ) (t D ) Yss (t D ) Yss (t D ) Yss
Y Yss
X (t ) (t D ) Yss (2.22)
It should be noted that all of the foregoing results were derived for the true first-order-plus-
deadtime process (2.8), and not the higher-order S-shaped response sketched in Fig. 2.2.
These arguments are worth taking the time to understand because they give insight into the
thinking of Ziegler and Nichols as they were developing their ‘open-loop’ modelling
technique, called Method II here. Method II enables us to approximate higher-order
dynamic systems by fitting the three parameters (steady-state gain, effective deadtime and
effective time constant) of a simple FOPDT model. It is not recommended for application
to response curves such as that of Fig. 2.1 since the point of inflection occurs immediately
(e.g. at t D ). The slope of the tangent drawn to this point must be nearly perfect in order
to avoid making large errors in your estimate of the time constant after extending the
tangent to Y .
8
Once we have fit a FOPDT model to our plant step response data, we’ll be in a position to
choose amongst hundreds of FOPDT model-based controller tuning algorithms (Foley et
al., 2005a, b; O’Dwyer, 2009). These rules can greatly simplify selection of tuning
parameters for the ‘industry default’ Proportional-Integral-Derivative (PID) controller.
2.3 References
Foley, M.W., Ramharack, N.R. and B.R. Copeland (2005a). Comparison of PI controller
tuning methods. Ind. Eng. Chem. Res., 44, 6741-6750.
Foley, M.W., Julien, R.H. and B.R. Copeland (2005b). A comparison of PID controller
tuning methods. Can. J. Chem. Eng., 83, 712-722.
O’Dwyer, A. (2009). Handbook of PI and PID Controller Tuning Rules (3rd Edition).
Imperial College Press, London.
Ogunnaike, B.A. and W.H. Ray (1994). Process Dynamics, Modeling and Control. Oxford
University Press, New York.
Ziegler, J.G. and N.B. Nichols (1942). Optimum settings for automatic controllers.
Transactions of the ASME, 42, 94-100.