Features of practical PID controllers

Although real PID controllers use algorithms which are very similar to the
'theoretical' one I've been using up to now, there are several differences in
the way real controllers operate. The topics covered in this section are:
dimensionless gain; proportional bandwidth; integral
windup; implementation of derivative action; other practical considerations;
and the use of the StrathPID controller.

Dimensionless Controller Gain

The controller gain we've been using up to now has always had dimensions
of 'controller output units/controller input units'. Most (if not all) practical
controllers use dimensionless gains, where the input and output signals are
made dimensionless be expressing them as a percentage of their range. For
example, imagine a proportional temperature controller which moves its
control signal (an output to a valve) by 2mA for every 5 degC change in
temperature. Assume the temperature instrument has measurement limits of
-20 to 150 degC and that the controller output has limits of 4-20mA. Using
the gain definition we've been using up to now, the controller gain is
2mA/5degC = 0.4 mA/degC. Using the dimensionless gain expression, the
gain becomes:

To convert between from dimensioned to dimensionless gain you need to

multiply the dimensioned gain by 'Input range'/'Output range'.
The simulation work you'll be doing from now on will use a controller
which uses this dimensionless gain. All that you'll have to do is supply the
input and output limits (which are problem dependant) and then controller
will do the rest. The controller is described below.

Controller Proportional Bandwidth

Electrical and Electronic engineers like the term 'bandwidth' and when they
made up the bulk of control engineers, the term was carried over into control
engineering. The controller bandwidth is just an alternative way of
expressing the controller gain. The bandwidth is the range over which the
error must move to move the output of a proportional controller over its full
range. Low bandwidths are therefore equivalent to high gain. If
dimensionless controller gain is being used there is a simple conversion
between gain and bandwidth:

The bandwidth, or proportional band, is a rather old-fashioned way of

expressing controller gain, but can still be found on even modern control
systems - you need to be aware of what it means.

Controller Output Constraints and Integral Windup

The controllers we've been using up to now have had no constraints at all on
their output values (other than the numerical capacity of VisSim). Real
controllers obviously do, since a valve can't close any more than fully closed
and can't open any more than fully open. This doesn't cause any problems
for proportional and derivative actions, but does cause difficulties for
integral action. The problem that integral action suffers from is called
'integral windup' or 'reset windup'.
A movie demonstrating the effect of reset windup is here and the VisSim
model used in the demonstration is here.
There are a number of ways of protecting a controller from windup. The
method used in the StrathPID controller is to switch off the integration
whenever the output limits are exceeded, and to switch it back on when the
output returns to within its range. YOU NEVER NEED TO WORRY
ABOUT THE INTERNAL MECHANISM A CONTROLLER USES.
Windup protection is built into controllers and is invisible to users. The only
time you need to take special care about windup is when setting up cascade
controllers.

Practical Implementation of Derivative Action

It isn't possible to obtain the derivative of a signal directly from an
instantaneous measurement. As a result, most controllers obtain the
derivative of a signal by passing it through a first-order 'filter' (a first-order
differential equation):

where = a manufacturer's adjustable 'filter constant', with a value in the

range [0,1], but normally set between 0.1 and 0.2.
= Derivative time constant
y = the output of the derivative element.
This equation can be manipulated to obtain an expression which is
'physically realisable', i.e. it is possible to implement:

You saw in an earlier section how sensitive derivative action is to noise in

measurements. The filter in the derivative element provides a limited
amount of noise resistance, but extra filtering is usually required.
Another modification that is often made to the derivative action in practical
controllers is to take the derivative of only the measurement signal, and not
the error. In the absence of setpoint changes this gives exactly the same
result:

The reason for doing this is that it is quite possible to introduce sudden
changes in the setpoint (e.g. a step), that will produce very large error
derivatives (infinite for a step). If these are used for control they will cause
the control valve to suddenly move to its limit and then back (when the
derivative becomes reasonable). This isn't particularly desirable!

Other practical aspects

Although we won't be covering these any further in the course you should be
aware that they are a consideration if you are applying a controller for real.
1. Is the input reasonable? Most controllers include some checks on the
input. The normal method for analogue communication in process
plants is a 4-20mA current loop. If the current at a controllers input is
less than 4mA, or more than 20mA, then something is seriously
wrong. In situations like this, controllers should sound an alarm and
NOT carry out a control calculation. Input checking needs to be built
into a controller by the manufacturer.
2. Does the input need to be linearised? PID controllers are linear
controllers and produce less than optimal results when applied to
significantly non-linear processes. Often a signal will be passed
through a 'linearising' element to try to improve the controlled
systems performance (e.g. the signal from an orifice plate, the
differential pressure, isn't linearly related to the flow being measured,
but is proportional to the square of the flow. If the signal is passed
through an element which takes its square root, then the controller
input will then be linearly related to the flow, and flow control should
be improved)
3. Is the measurement noisy? Noise can enter a measurement signal
from a number of sources. Process noise is random variations arising
from a variety of minor process disturbances and tends to be
relatively low frequency. Measurement noise arises during the
measurement process and tends to be a high frequency signal (flow
measurements from orifice plates are notoriously noisy and are
affected by bubbles, suspended particles, vortices, etc,etc). Electrical
noise arises from induction from nearby power cables or electric
motors. Noise from power cables is always at 50Hz, the frequency of
the AC supply (except in countries where it is 60Hz!). Electrical noise
can usually be suppressed by shielding the data cables and keeping
them away from obvious sources of interference. High frequency
noise filtered relatively easily using low-pass analogue (i.e. NOT
digital) filters. Noise which has a frequency similar to the process
dynamics is very difficult to remove. The filtering of measurement
signals is not covered in this course, but it is important!
4. Are limits and alarms required? Most controllers allow alarm limits to
be set on measurement signals. These limits trigger an alarm to warn
operators when a variable is moving out of a desirable range. Many
controllers allow several alarm limits to be set ranging from: 'I think
you should have a look at this', to 'Panic and run'! As well as absolute
output limits, many controllers allow the setting of rate of change

limits on the output. Limits on the setpoint are also often applied to
prevent operators requesting process conditions which are outside of
the normal operating range of the process.

Using the 'StrathPID' controller

In the remainder of the simulation work for this course you will be using a
more realistic controller than the simple arrangements we've been using up
to now. The controller is contained in the file 'StrathPID.vsm' which you
will find in the 'StrathModels' subdirectory or click here. To use the
StrathPID controller in your simulations you'll need to follow the following
procedure:
1. Open a new instance of the VisSim program (just press the Windows
'Start' button and start-up a new copy of VisSim). When you've done
this you'll have two copies of VisSim running: one containing your
model; and the other currently empty.
2. Open the StrathPID file into the empty instance of VisSim. You will
see the following:

the controller is containing in the compound block in the centre. It has three
input signals: the setpoint; the measurement; and a special anti-windup input
for cascade control loops. The controller has two outputs: the controller
output for connection to the manipulated variable; and an anti-windup
output.

3. Select the controller block and 'copy' from the VisSim 'Edit' menu. It
isn't necessary to copy the variable blocks attached to the inputs and
outputs - these are only present for documentation.
4. Connect the controller inputs and outputs to the process. Unless you
are setting-up a cascade control system you won't need either the antiwindup input or output, and you can leave them unconnected.
Everything else MUST be connected or you will get very strange
results. The measurement input needs to be connected to the
controlled variable (process output) and the controller output needs to
be connected to the manipulation (process input). The setpoint needs
to be connected to whatever you are using as a setpoint source.
Usually this will be a constant block.
5. Double click on the controller block to open the configuration level:

The bits you are interested in this diagram are those with brown labels
against them - these allow you to set-up the various controller parameters,
they are:
Controller gain. This is the dimensionless controller gain
Integral time constant: Just what it says. The dimensions of this
time constant are the same as the time dimension you used in
building your simulation model. In real controllers the time
dimensions are usually minutes. To 'switch-off' integral action
just set this time constant to a very large value (e.g. 1e+6).
Derivative time constant. Again this has the same time
dimensions as your simulation model, but in real controllers is
usually in minutes. To switch off derivative action set this to
zero.
Derivative filter constant. This is the 'alpha' factor used in
creating a realisable method of calculating the error derivative
- I would recommend that you don't touch this.
Direction of control action. Rather than switch signs on the
formation of the error, to change the direction of control action
you need to change the sign on this variable. A value of 1
produces a direct acting controller, and a value of -1 produces a
reverse acting controller. Any other values will muck up the
controller gain!
The Bias term lets you set the initial controller output. If you
leave this at zero and connect the controller to a simulation
input that should be non-zero at steady-state then you'll end up
getting a huge start-up transient when you start the simulation
as the controller tries to find the correct output. You should
always start simulations with a short period of 'steady-state'
operation prior to introducing any disturbance or setpoint
changes. If you see the outputs moving significantly during
this period it means that the simulation isn't starting at steadystate and that any results you get will be incorrect.
The output minimum and maximum values do two things: they
limit the output signal; and they scale the dimensionless gain.
You should set these limits to the limits specified in your
simulation brief.

 The input minimum and maximum limits are used to scale the
dimensionless gain. You should be provided with values for
these in your simulation brief ( in a real situation you would
use the real instrument limits obtained from the instrument
engineers).