Multi-Model Predictive Control (MMPC) for

Non-linear Systems with Time Delay: An

Experimental Investigation
G. Maruthi Prasad Vatsal Kedia A. Seshagiri Rao
Department of Chemical Engineering Department of Chemical Engineering Department of Chemical Engineering
National Institute of Technology National Institute of Technology National Institute of Technology
Warangal, India -506004 Warangal, India -506004 Warangal, India -506004
Email: [email protected] Email: [email protected] Email: [email protected]

Abstract—This paper depicts an experimental validation and an advanced control strategy is required. Model predictive
comparison of gap metric based weighting methods for nonlin- control (MPC) has been employed for better performance over
ear processes with delay using Multi-model predictive control conventional control strategy [5]. By using the each local linear
(MMPC). Controlling nonlinear processes is a difficult task and
the difficulty increases when there is time delay in the process. model, corresponding MPC has to be modeled known as a
Multi model technique is the simplest approach and is used to local controller. To form a global controller weighting methods
control the nonlinear process from decades. In this research, are used. In the weighting methods, finding proper weighting
model predictive control is developed in a multi model framework functions is essential. There are many weighting methods
(MMPC). In MMPC the global controller formation depends on available in the literature but gap metric based weighting
local weights and these weights are calculated based on gap
metric. From the literature, we found two popular weighing method of 1 − δ (δ is gap metric function) and 1/δ is more
functions based on gap metric. Here we present the compar- popular. These weighting functions are define the weights of
ative performance analysis of those weighing functions through each controller and combination forms the global controller
simulation and experimental studies. Level control in a conical [9] . Now the gap metric weighting method was employed to
tank process is cosnidered for experimental implementation of find the weights of each local controller and the weighted sum
the considered weighting methods in MMPC framework.
Index Terms—Model predictive control,multi model control
of local controllers(MPC) result to global controller. Hence the
scheme, gap metric. approach becomes multi-model predictive control (MMPC).
Gap metric based MMPC are very effecting in controlling
complicated nonlinear system(Proven in [8]), in order to test
I. I NTRODUCTION
the effectiveness of this method, a case-study on conical
Control of nonlinear processes is challenging and this tank (level control) process is considered. The method was
problem has been addressed by many researchers by using dif- employed on MATLAB simulation platform and was also
ferent types of controllers such as linear controllers, nonlinear validated experimentally. Comparison of both the weighting
control strategies, etc. When compared to single model control methods on both the scenarios was done to justify the need of
approach, multi-model control approaches are found to be advanced control strategy.
more effective for control of nonlinear systems. Multi-model The paper is arranged as: In section 2 detailed explanation
strategy depends on a problem breakdown approach where a on theoretical development of weighting function based on
non-linear system can be divided into many local linear models gap metric and global controller formation using respective
depending on their input region [1]. The local model enact weights. In section 3, multi-model predictive control strategy
the dynamic behavior of the system at particular region of is explained for linear models. A case-study on conical tank
the operating point, the gap metric approach has been used process is presented in section 4 to observe the effectiveness
widely for reducing the number of linear models in the recent of weighting functions based on gap metric using MMPC. In
years [2] [3]. The approach is very simple and mathematically Section 5, simulation and experimental results are presented
tractable. Moreover, when these local models are considered and followed by conclusion in section 6.
to be linear, then we can achieve the best performance linear
II. T HEORETICAL D EVELOPMENT OF W EIGHTING
model for control analysis. Multi-model controllers use linear
F UNCTION
control owing to their easy implementation.
The conventional PID multi model controller can control A. Gap Metric
a low level nonlinear system effectively, if the degree of The gap metric is an apt tool to analysis the uncertainty in
nonlinearity is increased these controllers are observed to give feedback control systems and is useful to measure the gap be-
some degradation in performance. To address such problems tween two linear systems which is better than any other metric

978-1-7281-3069-9/20/$31.00 2020
c IEEE

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY BOMBAY. Downloaded on May 03,2023 at 10:31:01 UTC from IEEE Xplore. Restrictions apply.
 based norms [6]. Let Li is a linear model defined in the H2 III. M ULTI -M ODEL P REDICTIVE C ONTROL (MMPC)
space. The left and rightco-prime factorizations of the
transfer
  Ni MPC is an advanced control strategy that is used in most
function Li are Ǧi , Gi , Ǧi = Ni Mi , Gi = . The of the process industries. It uses a dynamical model of
Mi
gap metric can be calculated as the process to predict its likely future response and then
choosing the best control action possible while satisfying set
δg (L1 , L2 ) = max(~δg (L1 , L2 ), ~δg (L2 , L1 )) (1) of constraints. Nowadays, it finds application in aerospace,
automotive, smart electricity grids, etc. Because of the
Where, advantage associated with MPC over conventional control
   
inf strategy, it has been employed. MPC inherently has feed-
N1 − N2 Q

δg (L1 , L2 ) = (2) forward nature as it takes measured disturbances as input and
QH∞ M1
M2
∞
it negates the effect of the disturbance beforehand making it
The gap metric properties: very popular in highly disturbed plants as well [4]. The future
1) Zero to one is the range of gap metric. control signal is computed in such a way that minimizes the
2) The metric value is nearly zero, then a single tuned quadratic objective cost function defined as, Minmize J
controller based any one of the model can stabilize both
NP
systems at an operating range; else if metric value is X 2
J= kΓy (y (Ki + l|Ki ) − r (Ki + l))k
nearly one, it is little difficult to stabilizing the system
l=1
in both operating region using a single controller [7].
NC
X 2
B. 1/δ and 1 − δ weighting function + kΓ∆u (∆u (Ki + l − 1))k (7)
l=1
At time t, the nonlinear system is nLt . Then L(θt ) is the
linearized model of nLt at steady state conditions. The gap The 1st term denotes the objective of minimization of error
metric function defines the distance between two local linear between predicted outputs and set-point signal and the 2nd
model of Li and L(θt ), denoted as γi (θt ): term denotes the objective to find optimal ∆u values such
that error is reduced. Γy denotes the penalty on tracking
γi (θt ) = δ (Li , L (θt )) , i = 1, ...., Nm (3) error known as output weighting, Γ∆u denotes the penalty
on the actuation known as input-rate weighting, y (Ki + l|Ki )
C. 1/δ weighting function represents the predicted value of output at Ki + l instant given
The weight of the ith local linear controller using 1/δ information up to ki instant. Tuning parameters of the MPC
weighting function is defined as: are prediction horizon (NP ), control horizon (NC ), control
  interval (∆t), rate weight on MVs (Γ∆u ), weight on CV
1 (Γy ). Few distinguishing features of MPC from conventional
γi (θt ) kw
ϕi = P   (4) control strategy is, it has ability to forecast, optimize and
Nm 1
j=1 γj (θt ) kw good constraint handling capability. The disadvantages are it
requires simple linear state-space model, too many degrees of
D. 1 − δ weighting function freedom (horizons, weights, constraints, etc.), requires real-
The weight of the ith local linear controller using 1 − δ time optimization, etc.
weighting function is defined as: If we have Nm number of local linear models then we
need to design Nm numbers of local MPC controllers and
(1 − γi (θt )) kw the response of all local controllers are merged together to
ϕi = PNm (5)
j=1 (1 − γj (θt )) kw
form an exhaustive controller according to (6).

Where kw is greater P than or equal to one and is a tuning IV. CASE STUDY
Nm
parameter. ϕi satisfies i=1 ϕ (θt ) = 1, kw = 1 as consider
for this work. Therefore, the output of the global controller of Conical tank process has been studied for examining the
multi-model is: performance of MMPC with measurement delay of 7 secs.
Nm
X A. Identification of Conical tank process
u (t) = ϕ (θt ) ui (t) (6)
i=1
Fig 1 shows the conical tank process, where the input flow
rate(qi ) is varied to control the level (h) in the tank [9].The
Where ui (t) is the ith local controller for particular region mathematical model is given as.
and modeled using ith linear model. According to (4) and (5),
a gap metric is close to one will be smaller weight, and vice dv
ρqi + ρqo = ρ (8)
versa. The weighting function formed by (4) and (5) is used dt
in the 1/δ and 1 − δ weighting methods.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY BOMBAY. Downloaded on May 03,2023 at 10:31:01 UTC from IEEE Xplore. Restrictions apply.
 TABLE I
R EDUCED N UMBER OF M ODELS U SING G AP M ETRIC M ETHOD

Operating Range (cm) Transfer function models

0.667
0-15 M2 = 12.253s+1 e−7s
1.364
15-40 M6 = 225.682s+1 e−7s
1.978
40-60 M10 = 908.856s+1 e−7s

Fig. 1. Conical Tank

Fig. 2. Weights for different set-point range using 1 − δ method

Writing differential equation of height (h) in terms of inlet

flow rate (qi ), B. Controller design
dh qi CD Identified models are specified in Table I, MPC controller
= 2 − 3/2 (9)
dt 3h (πD/12H) 3h (πD/12H) has to be designed corresponding to each model. Three MPC
Controller has been designed and their parameters are shown
Now linearizing (9) by applying Taylors series expansion up
in the Table II. After designing the individual controllers,
to 1st order terms around the steady-state operating point then
based on gap metric weighting approach weights for each
taking Laplace transform and rearranging the terms to get
controller is selected for 1 − δ and 1/δ weighting method
transfer function between height in the tank and inlet flow
respectively as explained in section 2 and the weights are
rate, we get,
shown in the Fig 2 and Fig.3 respectively.
√
s 2 h
H (s) K V. S IMULATION AND E XPERIMENTAL S TUDIES
=  2 2  Cd√  = (10)
Qi (s) πD hs
2
2 hs
s+1 τ s +1 In order to test the effectiveness of these methods on
4H Cd
nonlinear system plus delay, a simulation and experimental
Equation (10) represents the approximated first order transfer analysis was carried out on conical tank process. The two
function of conical tank process. Since the process is non- weighting function, 1/δ and 1 − δ was implemented for
linear, the parameters of the model has to be obtained at multi set-point change to the level of the process and the
different steady-state heights making it multi-model transfer responses are presented below. For testing the controller
function. The maximum height (H) and diameter (D) of the performance, two performance metrics were used namely
conical tank is 62 cm and 30 cm respectively. The transfer Integral Square Error (ISE) and Integral Absolute Error (IAE).
function is obtained for 12 different steady-state heights and
calculated using (10). Since all real time process have some From the simulation response i.e. Fig 4. and from Table III,
delay associated with it. Therefore we need to incorporate it it can be deduced that in overall sense 1−δ weighting function
to our model for more realistic controller performance. The
delay was calculated by giving sudden change in inlet-flow
rate from 0% to 100% of valve opening and the time taken
from actuating the signal till it reached 100% was recorded.
The delay-time approximately came out to be 7 seconds. Now
in order to design the controllers for multi-model transfer
functions the number of models are lessened owing to the
practical limitation of implementing many controllers. Hence,
the number of models is reduced by using gap-metric method
as explained in section II. A gap-value of 0.25 is selected and
the following three models were retained as shown in the Table
I. The model M2 is used to take-care of the process dynamics
in range of heights between 0-15 cm, similarly M6 for 15-40
cm and M10 for 40-60cm. Fig. 3. Weights for different set-point range using 1/δ method

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY BOMBAY. Downloaded on May 03,2023 at 10:31:01 UTC from IEEE Xplore. Restrictions apply.
 TABLE II
M ULTI -M ODEL MPC C ONTROLLER T UNING PARAMETERS F OR D IFFERENT S ET-P OINT R ANGE

Control Prediction Control

Operating Name of
interval horizon horizon
Range (cm) Controller
(secs) (interval) (interval)
0-15 MPC-1 5 40 2
15-40 MPC-2 10 150 2
40-60 MPC-3 10 600 2

Fig. 4. Closed loop response for different weighting methods to multi set- Fig. 5. Closed loop response for different weighting methods to multi set-
point change (Simulation result) point change (Experimental result)

TABLE IV
shows better performance as compared to 1/δ weighting C OMPARATIVE A NALYSIS OF C ONTROLLER P ERFORMANCE FOR
function. From the experimental response i.e. Fig 5. and from E XPERIMENTAL C ASE
Table IV, it can be deduced that 1−δ weighting function shows Set Point 1/δ 1−δ
better performance as compared to 1/δ weighting function IAE(104 ) ISE(105 ) IAE(104 ) ISE(105 )
at higher level while 1 − δ weighting function shows better 15 6.2208 4.2027 5.7348 3.9169
performance as compared to 1/δ weighting function at lower 35 18.116 21.249 14.121 15.456
55 37.706 41.209 38.358 42.029
level.It can also be seen from Fig 5. at lower levels there is 40 15.806 13.583 12.379 10.227
much oscillations because of the effect of delay and large gap 20 20.837 16.958 16.729 15.060
value between respective models.

TABLE III experimental studies. From the analysis, it can be concluded

C OMPARATIVE A NALYSIS OF C ONTROLLER P ERFORMANCE FOR that 1 − δ weighting function provides better performance as
S IMULATION C ASE
compared to 1/δ weighting function. In most of the practical
Set Point 1/δ 1−δ cases, 1 − δ weighting function is recommended because it
IAE ISE IAE ISE gives lower IAE value that leads to good controller perfor-
15 282.14 2863.4 110.98 4075.4
35 963.59 6276.9 410.92 5853.1
mance. When lower levels in the tank are considered as set-
55 2036.9 14436 1163.9 13949 points, more oscillations are observed across set point for both
40 2143.7 21104 2077.5 21998 the weighting methods. This issue will be considered as future
20 2795.3 34261 2741.5 34230 work.

VI. C ONCLUSION R EFERENCES

The gap metric based weighting methods were evaluated for [1] Adeniran, A. A., and El Ferik, S. ’Modeling and identification of
nonlinear systems: A review of the multimodel approachPart 1’, IEEE
control of conical tank process with delay using MMPC. The Transactions on Systems, Man, and Cybernetics: Systems, Vol. 47 No.
effectiveness of the method was justified using simulation and 7, pp.1149-1159, May 2016.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY BOMBAY. Downloaded on May 03,2023 at 10:31:01 UTC from IEEE Xplore. Restrictions apply.
 [2] Galan, O., Romagnoli, J. A., Palazoglu, A., and Arkun, Y. ’Gap metric
concept and implications for multilinear model-based controller design’,
Industrial and engineering chemistry research, Vol. 42, pp.2189-2197,
May 2003.
[3] Du, J., Song, C., and Li, P. ’Multimodel control of nonlinear systems: an
integrated design procedure based on gap metric and H loop shaping’,
Industrial and engineering chemistry research, Vol. 51, pp.3722-3731,
March 2012.
[4] Dougherty, D., and Cooper, D. ’A practical multiple model adaptive
strategy for single-loop MPC’, Control engineering practice, Vol. 11,
pp.141-159, February 2003.
[5] Chi, Q., and Liang, J. ’A multiple model predictive control strategy
in the PLS framework’, Journal of Process Control, 25, pp.129-141,
January 2015.
[6] Tan, G. T. ’On measuring closed-loop nonlinearity: a topological ap-
proach using the v-gap metric’, Doctoral dissertation, University of
British Columbia, 2003.
[7] Hosseini, S., Fatehi, A., Johansen, T. A., and Sedigh, A. K. ’Multiple
model bank selection based on nonlinearity measure and H-gap metric’,
Journal of Process Control, Vol 22, pp.1732-1742, October 2012.
[8] Du, J., and Johansen, T. A. ’A gap metric based weighting method for
multimodel predictive control of MIMO nonlinear systems’, Journal of
Process Control, Vol 24, pp.1346-1357, September 2014.
[9] Maruthi Prasad, G., and Seshagiri Rao, A. ’Evaluation of gap metric
based multi model control schemes for nonlinear systems: An experi-
mental study’, ISA Transactions, Vol 94 pp.246-254, November 2019.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY BOMBAY. Downloaded on May 03,2023 at 10:31:01 UTC from IEEE Xplore. Restrictions apply.