Professional Documents
Culture Documents
Ovation Advanced Control Algorithms: Predictor
Ovation Advanced Control Algorithms: Predictor
Data Sheet
High-level functions such as dynamic feed forward, Smith predictor, fuzzy logic control, Neuralnet and
multivariable control are available from the Ovation system through the use of algorithms running in an
Ovation controller.
Predictor
The PREDICTOR algorithm is used to compensate for pure time delay. This algorithm is designed to be
fitted into the standard Smith Predictor control system structure. The advantage of using a Smith
Predictor is that the controller (typically a PID) tuning can be performed as if the process exists with no
dead time, and hence a relatively high gain can be applied to expedite the response. The output of the
PREDICTOR is the result of the summation of the process output and no-delay model output minus the
with-delay model output. The process input to the algorithm is gain and biased. The algorithm output
directly connects to the downstream PID algorithm process input. When no time delay is present in the
process (and the model), the process variable input to the PREDICTOR directly passes through to the
downstream PID process input. The internal model used is a second order dynamics with pure time delay
(dead time).
CTRL
PVAR(IN1)
DIN
Since local memory size for each algorithm is limited, a long dead time will be characterized by an internal
time delay plus a chain of cascaded external delays (implemented by the TRANSPORT algorithm). The
actual number of time delay steps is determined by the dead time divided by the sampling time, with the
result rounded off to an integer.
M G
TCT1
P R E D IC T O R
OUT
DOUT
TCT2
Emerson Process Management Power & Water Solutions. All rights reserved.
OV-SYS-XX-106 [7] Proprietary
Page - 1
r(s)
u (s)
+
y (s)
P ro c e ss
G (s )
P ID (C (s))
_
x (s)
G 0 ( (s)
(1 -e -ts )
where y(s) ---- process variable, u(s) ---- control output, r(s) --- setpoint, and t --- dead time.
In the ideal case G(s) = G0(s)e-ts =
K e-ts
, and the closed loop transfer function is:
T2s + T1s + 1
2
P R E D IC T O R
(O p tio n a l)
T R A N S P O R T
T R A N S P O R T
S e tp o in t
P ID
M /A
NEURALNET
The Neuralnet algorithm calculates the output of a neural network model. Specifically, it calculates the
output response of a two-layer feed forward neural network model based on a given set of input variables.
Emerson Process Management Power & Water Solutions. All rights reserved.
OV-SYS-XX-106 [7] Proprietary
Page - 2
Algorithm inputs are usually selected as process input variables, and algorithm outputs are predicted
nominal process responses. The network training program resides in an Ovation engineer workstation.
All weights and biases obtained for the model are passed to the controller through the system database
IN1
T1
NEURALNET
T3
OUT3
OUT1
IN2
Algorithm Tuning
Function of NEURALNET
The network consists of two layer neurons. The first layer neurons are tangent hyperoblic sigmoids, i.e.
he function for each neuron can be represented by:
f(x) = 1 - e-2x
1 + e-2x
The second layer neurons are all linear functions.
In the following illustrated diagram, the number of inputs is assumed to be m, the number of neurons in
the first layer is assumed to be n, and the number of output (same as the number of neurons in the output
layer) is assumed to be p. Weights and biases in the first layer are represented by W1 and b1, and the
weights and biases in the second layer are represented by W2 and b2. The output for the first layer is
denoted by Z. By these assumptions, the network output at each layer can be calculated as:
Layer 1: Zi = f(Xi) = 1 - e-2x
1 + e-2x
m
where Xi = b1i + W1i,k Uk
k=1
n
Layer 2: where Yi = b2i +
W2i,k Zk,
(I = 1,.n)
k=1
Emerson Process Management Power & Water Solutions. All rights reserved.
OV-SYS-XX-106 [7] Proprietary
Page - 3
W 2 11
W 1 11
b11
b21
U1
Y1
Um
Yp
b2p
W 2 PN
W 1NM
b1n
Future enhancements to the NEURALNET algorithm include an option for a radial-basis network
structure, accommodation for up to three layers of neurons, and flexibility for select transfer function for
each individual neuron.
FUZZYCTR
IN2(PV)
IN1(STPT)
The FUZZYCTR algorithm calculates the control output of a fuzzy controller. The control philosophy of
this method is very similar to a variable gain PID. It is typically used when the PID controller tuning is
difficult for the operators to effectively control the process in manual mode. The algorithm inputs are the
setpoint and process output; the algorithm output is the control output. Tracking and bumpless transfer
are handled the same way as in the PID algorithms.
OUT
FUZZYCTR
Emerson Process Management Power & Water Solutions. All rights reserved.
OV-SYS-XX-106 [7] Proprietary
Page - 4
Function
Commonly defined and quantified linguistic sets:
Large positive
Medium positive
Small positive
Zero
Small negative
Medium negative
Large negative
(LP)
(MP)
(SP)
(ZE)
(SN)
(MN)
(LN)
The user has the option to select any one of the two commonly used membership function types, notably,
the trapezoid-shape, and the bell-shape (triangle is a special case of trapezoid). User selection can be
indicated by BIT03 of the X5 field (BYTE).
For trapezoid membership functions, each of them can be uniquely represented by four parameters
center x0, two end-points in the baseline xmin and xmax, and width w of the top line. Therefore, the function
can be basically determined by two line equations:
y = 1 + k1 (x-(x0-w/2)) and y = 1 + k2(x-(x0+w/2))
where the slopes k1 =
1
x0-0.5w-xmin
y
and k2 =
1
x0+0.5w-xmax
Normalized membership function
k1
Xmin
k2
x0
Xmax
Presumably, the center (x0), two end points xmin and xmax, and top width ware tunable either on or off line.
Note that when w 0, the trapezoid becomes a triangle.
For bell-shaped membership functions, a Gaussian distribution can be used to describe the curve shape.
In this case, only two parameters are needed (mean and variance). For example, for a bell centered at
(x0,1), the membership function curve can be described by:
y = e -((x-x0)/)2
Emerson Process Management Power & Water Solutions. All rights reserved.
OV-SYS-XX-106 [7] Proprietary
Page - 5
x0
The basic definitions and fuzzy logic formulation are briefly described as the following:
error: e(k) = setpoint r(k) - process output y(k)
error change e(k) = e(k) - e(k-1)
control output = u(k) (in the case of a stepper type of control output, this term is replace by control output
change u(k))
The following are some sample control rules:
1. If e is LP and e is any, then u is LP
2. If e is SP and e is SP or ZE, then u is SP
3. If e is ZE and e is SP or ZE, then u is ZE
4. If e is ZE and e is SN or ZE, then u is ZE
5. If e is SN and e is SN or ZE, then u is SN
6. If e is LN and e is any, then u is LN
For this setup, a mixed use of membership function shape (trapezoid and bell-curve) in one single
algorithm is not recommended. According to the current setup which involves 7 membership functions
(LP,MP,SP,ZE,SN,MN,LN), we need 4x7=28 record fields (type: REAL32) for trapezoids and 2x7=14
algorithm record fields (type: REAL32) for bell-curves.
Defuzzification (approximate centroid): The control output is calculated as:
n
n
u = ( x Ui)/( i)
i=1
i=1
where i is the number of rules and Ui is an associated crisp control output.
The rule base (table) is adjustable online providing operators the option of adding or deleting rules as
needed. Each rule occupies one algorithm record field (type: INT32). Each bit in the field performs the
following rule:
15
14
13
12
Rule Number
BIT12-15:
BIT7:
11
10
9
Error
7
O
5
4
3
Error Change
1
0
Control
Emerson Process Management Power & Water Solutions. All rights reserved.
OV-SYS-XX-106 [7] Proprietary
Page - 6
FEEDFWD
The Ovation PIDFF algorithm is designed to accept an externally generated feed-forward analog signal.
This analog signal is summed with the actual PID output (sum of the proportional, integral and derivative
terms) to become the final control output. The advantage to having the feed-forward input incorporated in
the PID algorithm is that saturation conditions, caused by either the final control output exceeding the
algorithm limits or raise/lower inhibit signals generated by downstream algorithms, are handled internally
to the algorithm. This can greatly improve controller response times to sudden changes in the algebraic
sign of the error between the process variable and the setpoint when the controller is in a saturation
condition.
For example, consider the case where the output limits on the PIDFF algorithm are 0-100%, the external
feed-forward input has a normalized value of 75%, the actual PID output is 50% and the error between
process variable and set-point is such as to cause the PID output to increase. This would require that the
final output be 125%, which violates the algorithms high limit of 100%. The final PIDFF output will be
clipped at 100%. If the error between the process variable and set-point was to suddenly change direction
(i.e. the PID output now wants to decrease) , assuming that the feed-forward input remains constant, the
PID output would have to move 25% before it had any effect on the final control output and thus the
process variable. Depending on the tuning parameters, the amount of time that this takes can be
significant.
By incorporating the feed-forward input into the PIDFF algorithm, a sign change in the error between the
process variable and set-point will have an immediate effect on the final control output. In the case of the
previous example, this is accomplished mathematically as follows:
The integral term is adjusted to make-up the 25%.
Iterm = 100 - (pTerm + dTerm + ffInput)
Where :
Iterm = Integral term
Pterm = Proportional term
dTerm = Derivative term
ffInput = Feed forward input
Emerson Process Management Power & Water Solutions. All rights reserved.
OV-SYS-XX-106 [7] Proprietary
Page - 7
The PIDFF now begins integrating from the high limit of 100% and thus has immediate influence on the
final control output and subsequently on the process variable.
DTIN
The FEEDFWD algorithm is used together with any other feedback controller (PID, lead-lag, etc.).
Although disturbances can be eliminated by feedback, it is necessary for there to be an error before the
controller can take actions to eliminate disturbances. In some situations, it is possible to measure
disturbances before they have influenced the process. In other situations, the disturbance is just another
interacting control input. It is then natural to try and eliminate the effects of the disturbances before errors
are created. The input variable to the FEEDFWD algorithm is the disturbance; the output variable is the
feed forward action. Usually, the summation of FFED-FWD output and PID output is the total control
input to the process. This scheme requires a mathematical model of the process. Two models are used,
one for the process and the other for the disturbance. All models are represented by a 1st order dynamics
with pure time delay (dead time).
OUT
FEED FW D
and
G2(s) = K2e-t1s
T2s + 1
In order to eliminate the disturbance effect, the feed forward compensation is selected as:
ff (s) = G2(s) v(s)
G1(s)
By performing a bi-linear transformation to discretize the continuous model, the final feed forward output
ff at any specific sampling k is given by:
ff (k) = A1ff (k-1) + A2v (k-d) + A3v (k-1-d)
Emerson Process Management Power & Water Solutions. All rights reserved.
OV-SYS-XX-106 [7] Proprietary
Page - 8
where d = max(int(t2-t1 )/sampling time), 0) and Ais are the parameters calculated internally by the
algorithm.
A typical configuration example in the control system can be shown as follows:
Disturbance
Feedback
Controller
FEEDFWD
SUM
Process G
1 (s)
Disturbance G
2 (s)
SUM
MIMOCTR
The MIMOCTR algorithm handles multivariable systems with up to 3 input-output pairs. Also, up to 3
disturbance variables can be taken into account. This algorithm is based on the model predictive control
principle, and hence requires a mathematical model for each input-output pair (also for each disturbance
output pair if any disturbance appears). A first order model (which consists of a gain, time constant and
deadtime) is used to characterize each dynamics. For the current version an unconstrained online
optimization is employed to calculate the control outputs.
The algorithm has the following features:
Easy design and tuning. Once the process model is obtained, controller design is performed
automatically by the algorithm. Users have to specify just ONE tuning parameter to set the closed
loop response speed.
Easy Manual/Auto transfer. When any of the control output (MV) is set to Manual, the other control
loops keep running with the overall closed loop performance not being affected.
Emerson Process Management Power & Water Solutions. All rights reserved.
OV-SYS-XX-106 [7] Proprietary
Page - 9
PV3
PV2
PV1
ST3
ST2
ST1
Data Sheet
DT1
DT2
MIMOCTR
MV3
MV2
MV1
DT3
Emerson Process Management Power & Water Solutions. All rights reserved.
OV-SYS-XX-106 [7] Proprietary
Page - 10