NORM BASED APPROACHES FOR AUTOMATIC TUNING OF MODEL BASED PREDICTIVE CONTROL Pastora Vega, Mario Francisco, Eladio Sanz University of Salamanca – Spain European Congress of Chemical Engineering (Copenhaguen, September 2007)
Dec 20, 2015
NORM BASED APPROACHES FOR AUTOMATIC TUNING OF MODEL BASED PREDICTIVE CONTROL
Pastora Vega, Mario Francisco, Eladio Sanz
University of Salamanca – Spain
European Congress of Chemical Engineering (Copenhaguen, September 2007)
2
Index
1. Introduction and objectives
2. Description of the Model Predictive Controller
3. Optimal automatic tuning method
4. Results applied to the activated sludge process control
5. Conclusions
3
Introduction
Model based predictive control (MPC) is the most popular advanced controller for industrial applications, due to its simplicity for operators, the natural way of incorporating constraints and its easy application to multivariable systems.
MPC tuning parameters are real numbers (weights, etc.) and integer numbers (prediction and control horizons), determining closed loop system dynamics.
Usually these parameters are tuned by a trial and error procedure, taking into account general system behaviour and expert knowledge. There exist some optimization based methods for automatic tuning, but they are rather slow due to the simulations needed to evaluate dynamical indexes.
4
Objectives
Develop a method for optimal automatic tuning of Model Based Predictive Controllers (MPC) that considers both real and integer parameters, using norm based performance indexes, avoiding numerical simulations.
Validate this method using a simple reference model based on the activated sludge process of a wastewater treatment plant, particularly to minimize the output substrate variations considering typical process disturbances at the input.
Include this method in a further Integrated Design of wastewater treatment plants and their control systems.
5
Index
1. Introduction and objectives
2. Description of the Model Predictive Controller
3. Optimal automatic tuning method
4. Results applied to the activated sludge process control
5. Conclusions
6
General MPC controller structure
y1,y2 controlled (or constrained)
u1,u2 manipulated variables
12 2
0
ˆ ˆmin ( ) ( ( | ) ( | )) ( ( | ))Hp Hc
y uui Hw i
V k W y k i k r k i k W u k i k
Standard linear multivariable MPC controller, using state space model for prediction and state estimators (MPC Toolbox MATLAB)
MPC controller
index
Ref. y1
Ref. y2 Controller
u1 y1
y2
PROCESS u2
i i iylb y yub MPC constraints i i ilb u ub i i iulb u uub
7
Tuning parameters
Hp : Prediction horizonHc : Control horizonWu: Weights of the changes of manipulated variables
Integer parameters (Hp, Hc) Real parameters (Wu)
12 2
0
ˆ ˆmin ( ) ( ( | ) ( | )) ( ( | ))Hp Hc
y uui Hw i
V k W y k i k r k i k W u k i k
8
General MPC controller structure
MPC general structure for the linear case without constraints
Particular formulation: 2 1K K
-
r + y u
d
K1 G
Gd
K3
3
11dGK Gy
Sd GK
3 1
11dK K Gu
Md GK
Transfer functions used for Automatic Tuning: output sensitivity (S’), control sensitivity (M’)
Block diagrams (linear control system):
r y u
d
KMPC G
Gd
1 2 3 1 2 3
r
u K K K y K r K y K d
d
9
Index
1. Introduction and objectives
2. Description of the Model Predictive Controller
3. Optimal automatic tuning method
4. Results applied to the activated sludge process control
5. Conclusions
10
Optimal automatic tuning of MPC
Tuning procedure based on a H mixed sensitivity problem
where
are suitable weights
minx
f x f x N
p
esf
W SN
W s M , ,p c ux H H W
,p esfW W
1
pW S
max1M u
max1S y
Constraints:Over disturbance rejection and based on l1 norms to avoid actuator saturation
11
Optimization problem
1 2 3( ) ( ) ( ) ( )F x f x f x f x
Multiobjective approach
Objective function F:
2 1max
d possible
uf M
d
3 1max
d possible
yf S
d
x=(Wu, Hp, Hc)
1
pW S,
minx R
( )i i if x w f
where fi* is the desired value for each objective
function
1 f x N
S’= output sensitivity
M’= control sensitivity
N = mixed sensitivity
12
Step 2:F is minimized using “Goal
Attainment” method, keeping constant now the integer parameters (horizons)
with the values obtained in step 1
Step 1:F is minimized by a random search method keeping real
parameters constant
INTEGER PARAMETERS
TUNING Horizons
REAL PARAMETERS
TUNING
Wu Hp,Hc
The algorithm converges when changes in F are smaller than a certain bound
Algorithm developed
Method “Goal Attainment” (MATLAB)
Specific random search
An iterative two steps optimization algorithm has been proposed due to the existence
of real and integer parameters
13
Algorithm developed
Algorithm steps
Modified random search method for tuning MPC parameters
2. A random vector ξ(k) of Gaussian distribution is generated, with
integer elements.
1. An initial point for horizons, variances and centre of gaussians (for random numbers
generation) is chosen.
3. Two new points are obtained by adding and removing this vector to
the current point.
4. Cost function is evaluated at the original point and at new points, and the algorithm chooses the point with
smallest cost.
5. If some convergence criteria is satisfied, stop the algorithm, otherwise return to step
2. Variances are decreased.
( ) ,p cc k H H
( ) ( )c k k
( ) ( )c k k
14
Index
1. Introduction and objectives
2. Description of the Model Predictive Controller
3. Optimal automatic tuning method
4. Results applied to the activated sludge process control
5. Conclusions
15
Description of the process and control problem
EffluentSettlerBioreactorInfluent
Recycling
Ref. s1 MPC qr1
s1
x1
PROCESS
Non linear systemLarge disturbancesSubstrate control
problem
qr1 manipulated variable
s1 controlled x1
constrained
16
Process disturbances: input flow and substrate
Substrate concentration at
the plant input (si)
Flow rate at the plant input (qi)
Real data from a wastewater plant Benchmark disturbances
17
Tuning results (I)
Weights considered and parameters of the MPC tuned automatically
Substrate comparison for two weights(solid line – Wp1; dashed dotted line – Wp2)
MPC constraints
Wu=[0.0023] Hp=9, Hc=21
26.6 32
0.0001
sWp
s
2
8 9.6
0.0001
sWp
s
Comparison of sensitivity functions for tuning with weights (Wp1; Wp2)
Wu=[0.0118] Hp=8, Hc=3
Fixed plantV1=7668
A=2970.88
H mixed sensitivity problem considering objectives f1 and f2:Comparison of weights Wp
10 1000qr
1
1
0 125
400 3000
s
x
10 3500qr
Output variable: s1
2 1f M
18
Tuning results (II)
H mixed sensitivity problem considering objectives f1 and f2:Comparison of weights Wesf
Substrate comparison for two weights(dashed dotted line – Wesf2; solid line – Wesf3)
10 1000qr
1
1
0 125
400 3000
s
x
MPC constraints
Wu=[0.0011] Hp=6, Hc=2
10 3500qr
Comparison of sensitivity functions to the control efforts s*M’ for tuning with two weights
Wesf
Wu=[0.0118] Hp=8, Hc=32 2
0.2 0.02
5 0.0001
sWesf
s s
3 2
0.02 0.002
5 0.0001
sWesf
s s
Output variable: s1
Weights considered and parameters of the MPC tuned automatically
19
Tuning Results (III)
H mixed sensitivity problem considering objectives f1 and f3:Comparison of weights Wp
TABLE II
INDEX Case 4 Case 5
Wu 0.0019 0.0091
Hp 9 10
Hc 2 4
Max(qr1) 1185.5 1096.6
Max(s1) 64.17 65.33
0.95 0.97
Weights Wp1 Wp2
Computational time (min)
10.8 5.89
pW S
Comparison of substrate responses for two weights Wp1 and Wp2
Output variable: s1
Wp1 is more restrictive than Wp2
3 1f S
20
Index
1. Introduction and objectives
2. Description of the Model Predictive Controller
3. Optimal automatic tuning method
4. Results applied to the activated sludge process control
5. Conclusions
21
Conclusions and future work
– A new methodology has been develop to tune automatically all parameters of Model Based Predictive Controllers, considering simultaneously horizons and weights.
– This method has been tested for the MPC tuning of the activated sludge process in a wastewater treatment plant.
– The plant with the MPC tuned with this method is able to reject substrate disturbances in the influent.
– This method has been designed to be straightforward included within an Integrated Design scheme of wastewater treatment plants together with MPC controllers.
Future work:– Include some robust stability and robust performance indexes.