This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3057674, IEEE Access

Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2017.DOI

Impact of MPC Embedded Performance

Index on Control Quality
PAWEł D. DOMAŃSKI1 and MACIEJ ŁAWRYŃCZUK2
1
Warsaw University of Technology, Institute of Control and Computation Engineering, Faculty of Electronics and Information Technology, ul. Nowowiejska
15/19, 00-665 Warsaw, Poland, (e-mail: [email protected])
2
Warsaw University of Technology, Institute of Control and Computation Engineering, Faculty of Electronics and Information Technology, ul. Nowowiejska
15/19, 00-665 Warsaw, Poland, (e-mail: [email protected])
Corresponding author: Maciej Ławryńczuk (e-mail: [email protected]).

ABSTRACT Model Predictive Control (MPC) is a well established advanced process control technology.
There are many successful implementations of different predictive strategies in process industry. There may
be found various modifications of the MPC, however, one aspect remains fixed. MPC performance index
is in quadratic form. Nonetheless, statistical analysis frequently points out that the quadratic regression
formulation has some drawbacks. It is sensitive against the outliers. This work analyzes alternative and
robust formulations of the MPC embedded performance index. It is shown that the quadratic formulation
is not an optimal one, while the linear `1 weight improves control. Classical `2 norm together with robust
Cauchy and Dynamic Covariance Scaling gives worse results.

INDEX TERMS MPC, performance index, robust regression, `1 norm, Dynamic Covariance Scaling.

I. INTRODUCTION mation of the system dynamics and quadratic approximation

Model Predictive Control (MPC) constitutes a major compo- of the objective functions. The MPPI control framework has
nent of the Advanced Process Control (APC). Its story starts been successfully applied to a variety of robotic systems for
with the first formulation of the linear quadratic regulator tasks such as aggressive autonomous driving and autonomous
(LQR) by Kalman. Within the next ten years the ongoing flying through 2D/3D cluttered environments [15].
research enabled successful industrial applications of MPC Though the approach may vary, one single element re-
in industry. Model Predictive Heuristic Control (known as mains unchanged in the majority of variants. It is a quadratic
Model Algorithmic Control (MAC)) has been presented fol- formulation of the embedded performance index. Excep-
lowed by Dynamic Matrix Control (DMC) [1]. Next, Gen- tions from this rule are very rare. In the beginning, authors
eralized Predictive Control (GPC) using minimum variance improved classical raw quadratic error adding weighting.
control ideas [2] has been proposed. The backbone of the Davidson [16] has proposed an alternate “cheap control” per-
MPC framework is established. formance index transferred from discrete time servo control
Further research investigated in details baseline formula- context. The quadratic formulation is named as a `2 -MPC.
tion extending its basic idea towards various scenarios and There are also alternate formulations of MPC performance
configurations. The researchers have investigated multivari- index in terms of mean and variance [17], however, they
ate, nonlinear constrained MPC configurations [3], [4]. As are equivalent. Interesting discussion about applicability and
a result, MPC algorithms have been applied to numerous deficiency of the MPC quadratic cost function may be found
processes, ranging from relatively slow process control plants in recent work of [18]. Authors show that typical choice of
such as chemical reactors [5], distillation columns [6], NOx `2 -MPC is not always suitable, in particular, a sufficiently
control [7] and coal mills [8] to very systems such as fast long horizon satisfying MPC asymptotically stable control.
robots [9], micro grids [10], electric drives [11], spark- Limitations and drawbacks of the quadratic performance
ignition gasoline engines [12] and autonomous vehicles [13]. index might also be observed in other than MPC contexts,
Recently, Williams et al. [14] proposed a sampling-based and as for instance statistics and regression.
derivative-free MPC algorithm, known as Model Predictive The research focuses on outliers, which pose a serious
Path Integral (MPPI) control framework, that can be easily challenge for any control. The classical quadratic perfor-
utilized without requiring the first- or second-order approxi- mance index is highly sensitive to any kind of the outliers

VOLUME 4, 2016 1

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3057674, IEEE Access

Paweł D. Domański and Maciej Ławryńczuk: Impact of MPC Embedded Performance Index on Control Quality

[19], as it is characterized by the 0% breakdown point. Sim- •The paper investigates and compares the impact of var-
ilar breakdown value appears for normal standard deviation. ious robust norm formulations on an overall nonlinear
Integral of absolute error is only a little bit better, but still MPC control performance.
holds 0% breakdown point. Robust statistics [20] proposes In the following sections, we will introduce robust re-
M-estimators that exhibit 50% breakdown. gression estimators and the way how they are incorporated
An outlier is a strange data observation [21], [22]. A single into MPC performance index. Then, the simulations of MPC
outlier may originate from an erroneous observation (ex- controller using four selected index formulations are per-
ogenous) or can be an intrinsic symptom of some unknown formed using nonlinear control problem in the form of the
underlying mechanism [23] (endogenous). Outliers impede pH neutralization reactor. The simulations are summarized
analysis as they increase the variance and reduce the power with the results and their discussion. Advantages and possible
of statistical tests [24], destroy Gaussianity, introduce tails areas for further research are highlighted in conclusions.
[25], and bias data regression [19]. An α-stable distribution
is a common approach [26] to model them. II. NONLINEAR MODEL PREDICTIVE CONTROL
Following the above indications, statistics suggest to use Description of the robust MPC formulations consists of two
mean absolute error, which is formulated as a `1 norm. following sections. First, the general layout of applied MPC
Statistically, this index is only slightly better, as it is robust is presented, which is followed by the specific performance
to an outlier in y-direction, but is sensitive to outliers in x- index formulations addressed in the simulation analysis.
direction (leverage points) and still holds a 0% breakdown
[19]. Formulation and an analysis of the `1 -MPC can be A. MPC ALGORITHM
found in [27]. It was also quite natural that the researchers The process input, i.e. Manipulated Variable (MV) is denoted
also considered other norms as for instance `∞ -MPC [28]. by u and the output, i.e. Controlled Variable (CV), is denoted
Some other approach to improve MPC quadratic perfor- by y. The vector of decision variables determined on-line
mance index has been proposed [29]. Classical least squares at each discrete sampling instant (k = 0, 1, 2, . . .) by MPC
approach has been regularized with `1 component. algorithm [3] is
Investigation of `1 and `∞ MPCs is mostly driven by 4u(k) = [4u(k|k) 4u(k + 1|k) . . . 4u(k + Nu − 1|k)] ,
T
two factors: simplification of the MPC evaluation due to the (1)
computationally less complex problem linear programming where Nu is the control horizon, i.e. the number of calculated
formulation. On the other hand, there have been observed future control increments defined as backward differences,
some issues with the performance and suitability of `1 and i.e. 4u(k|k) = u(k|k) − u(k − 1) and 4u(k + p|k) =
`∞ criteria. The main observed consequence was that they u(k+p|k)−u(k+p−1|k) for p = 1, . . . , Nu −1. For p ≥ Nu
may yield either dead-beat or idle control performance [30]. it is assumed that the manipulated variable is constant, i.e.
It must be noted that the outlier robustness aspect is not as u(k + p|k) = u(k + Nu − 1|k). The decision variables of
such addressed. On the other hand, robust indexes are suc- MPC (1) are calculated from an optimization problem. Its
cessfully used in the MPC control performance assessment typical form is
[31]–[33], which is independent of the controller operation, 
however, should reflect predictive operation. min Jy (k) + Ju (k) ,
4u(k)
In contrary, robust regression estimators are successfully
used in other engineering contexts, as for instance camera- subject to (2)
based localization. Robust function in Iteratively Reweighted umin ≤ u(k + p|k) ≤ umax , p = 0, . . . , Nu − 1,
Least-Squares (IRLS) problem allows to deal with frequent − 4umax ≤ 4u(k + p|k) ≤ 4umax , p = 0, . . . , Nu − 1,
correspondence outliers present in visual navigation [34]. y min ≤ ŷ(k + p|k) ≤ y max , p = 1, . . . , N,
These results have stimulated the presented work. Quadratic
`2 estimator inside of MPC performance index formulation where two components of the minimized cost-function are
is compared with three robust estimators: `1 , M-estimator N
X
with Cauchy function [35] and Dynamic Covariance Scaling Jy (k) = `(e(k + p|k)), (3)
(DCS) [36], which is not present in the literature as a robust p=1
cost function, but can be expressed as such. u −1
NX
Concluding, the paper addresses the following subjects: Ju (k) = λ (4u(k + p|k))2 . (4)
p=0
• The most popular formulations of the MPC control
incorporate quadratic internal performance index norm The role of the first part of MPC cost-function is to minimize
`2 . However quadratic index is sensitive to the outliers predicted control errors over prediction horizon N
as 0% breakdown point.
e(k + p|k) = sc(y sp (k + p|k) − ŷ(k + p|k)). (5)
• Robust regression research delivers other estimators
with a larger breakdown, like linear `1 , robust Cauchy Setpoint and predicted values of the process output for future
`Cauchy and Dynamic Covariance Scaling `DCS . sampling instant k + p known/calculated for current instant
2 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3057674, IEEE Access

Paweł D. Domański and Maciej Ławryńczuk: Impact of MPC Embedded Performance Index on Control Quality

k are denoted by y sp (k + p|k) and ŷ(k + p|k), respectively,

α-stable disturbance Gaussian
sc is a scaling factor. Predicted values of the process output q2(t)
noise
are calculated on-line using a mathematical model of the
setpoint
pH + + pH(t)
controlled process. The role of the second part of the cost- + q1(t) neutralization
function is to eliminate excessive changes of the manipulated
MPC
-
variable, λ > 0. In general, the constraints may be imposed
on: (i) future values of the manipulated variable (over the
control horizon), the minimal and maximal allowed values FIGURE 1. Simulation environment of pH neutralization reactor
are umin and umax , respectively, (ii) future changes of that
variable, the maximal value is 4umax , (iii) predicted values
of the controlled variable (over the prediction horizon), the 4) `DCS -MPC
minimal and maximal values are y min and y max , respectively. In this approach [36],
Although at each sampling instant whole sequence of deci- (
sion variable (1), is calculated, only its first entry is applied to 0.5(e(k + p|k))2 if e(k + p|k) ≤ φ
the process. At the next sampling instant, k +1, measurement Jy (k) = 2φ(e(k+p|k))2 .
φ+(e(k+p|k))2 − 0.5e(k + p|k) if e(k + p|k) ≥ φ
of the process output is updated and the procedure is re- (9)
peated. In this work, a nonlinear dynamical model is used for The parameter φ is a threshold. Its value must be chosen
prediction calculation. It means that the MPC optimization taking into account properties of a particular application. In
problem (2) is a nonlinear task. It is solved by the Sequential all simulations discussed in this work, φ = 1.
Quadratic Programming algorithm.
III. SIMULATIONS
B. PERFORMANCE INDEX FORMULATION
Simulation layout has been prepared for the analysis. The
Four performance index formulations used in the analysis nonlinear process has been selected in the form of the pH
are described in the following paragraphs. Although different neutralization reactor [37]. It has one controlled input and
functions ` are used to measure the influence of predicted one considered as the unmeasured disturbance. Simulation
control errors reflected in the first part of the cost-function layout is sketched in Fig. 1.
(3), the classical quadratic second part of the cost-function
Two different simulation scenarios have been tested. The
(4) is used.
main set of results has been obtained using the undisturbed
reactor, i.e. unmeasured disturbance buffer inlet q2 (t) re-
1) `2 -MPC
mains unchanged. In the second part q2 (t) is simulated as
The classical approach is to use a sum of squared predicted
a non-Gaussian process with α-stable distribution and sta-
control errors [3]
bility factor α = 1.70. Such a selection enables to analyze
N
X industrial-like aspects with sufficient process excitement.
Jy (k) = (e(k + p|k))2 . (6) Another aspect that requires testing is the range of applied
p=1
cost function denoted as sc. The following values have been
When a linear model is used for prediction, the both parts tested to reflect possible shapes of the M-estimator costs
of the optimized MPC cost-function are quadratic in terms function: sc = {1, 2, 5, 10, 20, 50}.
of the calculated decision variables (1) and the whole MPC We use control performance measures, called KPIs (Key
optimization problem (2) becomes a quadratic optimization Performance Indicators) to evaluate the quality of a con-
problem (a quadratic cost-function and linear constraints). trol system. In general, in Control Performance Assessment
Such a problem may be very efficiently solved on-line. (CPA), there are two classes of methods: model-driven and
data-driven. Model-driven techniques need some assumed
2) `1 -MPC process knowledge that is used to identify the process model
When the sum of absolute control errors is minimized, the further used to calculate the KPI. Data-driven methods re-
first part of the MPC cost-function is [27] quire only raw process data without any further assumptions.
N
X Classical KPIs are not always sufficient and industry searches
Jy (k) = |e(k + p|k)|. (7) for robust measures that would be suitable in the broader
p=1 sense. This analysis uses measures that use control error
((k) = y sp (k) − y(k)) time series of length Np as the basis
3) `Cauchy -MPC for index calculation [32].
In this approach [35], 1) MSE: mean square error – commonly used, but not
N
X 
(e(k + p|k))2
 robust against outliers, exhibits 0% breakdown
Jy (k) = 0.5c2 log 1 + . (8)
c Np
p=1 1 X 2
MSE =  (k), (10)
In all simulations, c = 1. Np
k=1
VOLUME 4, 2016 3

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3057674, IEEE Access

Paweł D. Domański and Maciej Ławryńczuk: Impact of MPC Embedded Performance Index on Control Quality

2) IAE: integral absolute error – not robust against out- and one algebraic output equation
liers, exhibits 0% breakdown
Wa (t) + 10pH(t)−14 − 10−pH(t)
Np
1 X 1 + 2 × 10pH(t)−K2
IAE = |(k)| , (11) + Wb (t) = 0. (20)
Np
k=1
1 + 10K1 −pH(t) + 10pH(t)−K2
State variables Wa and Wb are reaction invariants. Parame-
3) LMS: least median square – robust against outliers and
ters of the first-principle model are given in Table 1 and the
exhibits 50% breakdown [19]
values of process variables for the nominal operating point in
LMS = med (k)2 , (12) Table 2. Buffer inflow q2 (t) is the disturbance, while the acid
k
stream q3 (t) is constant.
4) σG : Gaussian standard deviation calculated as
buffer q 2(t)
s
PN p 2
k=1 ((k) − 0 )
DV base q 1(t)
σG = , (13)
Np − 1 acid q 3(t)

where 0 is mean value of the control error,

5) γ: stable scaling factor evaluated for the fitted α-stable
PDF MV
MPC
stab α
Fα,β,δ,γ () = exp {iδ − |γ| (1 − iβl ())} , (14) PV STP

where SP
(
sgn () tan πα


2 for α 6= 1
l () = 2
(15) pH
sgn () π ln || for α = 1.
V
The coefficient 0 < α ≤ 2 is called a stability index
or a characteristic exponent, |β| ≤ 1 is a skewness
coefficient, δ ∈ R is distribution location and γ > 0 FIGURE 2. Schematic diagram of the pH neutralization reactor

is called a scale or dispersion [38].

6) σH : robust scale M-estimator with logistic ψ function
[39], TABLE 1. Parameters of pH neutralization model

7) Hdiff : differential entropy [40]

Z ∞ Wa1 = −3.05 × 10−3 mol Wb1 = 5 × 10−5 mol
Hdiff = − γ () ln γ () d, (16) Wa2 = −3 × 10−2 mol Wb2 = 3 × 10−2 mol
−∞ Wa3 = 3 × 10−3 mol Wb3 = 0 mol
K1 = 6.35 V = 2900 ml
8) Hrat : rational entropy [41] K2 = 10.25
Z ∞  
γ ()
Hrat = − γ () log d. (17)
−∞ 1 + γ ()
TABLE 2. Nominal operating point of pH reactor

A. PROCESS DESCRIPTION
In the considered neutralization reactor a base (NaOH) q̄1 = 15.55 ml/s W a = −4.32 × 10−4 mol
stream q1 , a buffer (NaHCO3 ) stream q2 and an acid (HNO3 ) q̄2 = 0.55 ml/s W b = 5.28 × 10−4 mol
q̄3 = 16.60 ml/s pH = 7
stream q3 are mixed in a constant volume tank. The process
has one input (manipulated) variable which is the base flow
rate q1 [ml/s] and one output (controlled) variable which
B. SIMULATION RESULTS
is the value of pH (Fig. 2). Continuous-time fundamental
The parameters of all compared MPC variants are the same:
model of the process comprises of two ordinary differential
the prediction and control horizons are N = 10 and Nu = 3,
equations
respectively, the weighting coefficient is λ = 0.5, the mag-
dWa (t) q1 (t)(Wa1 − Wa (t)) q2 (Wa2 − Wa (t)) nitude constraints imposed on the manipulated variable are
= +
dt V V defined by q1min = 0 ml/s and q1max = 30 ml/s, respectively.
q3 (Wa3 − Wa (t)) At first no disturbances are taken into account during
+ (18)
V simulations. Fig. 3 shows the control error for one step set-
dWb (t) q1 (t)(Wb1 − Wb (t)) q2 (Wb2 − Wb (t)) point change for four MPC controllers in case of sc = 1.
= +
dt V V The `1 -MPC algorithm is the fastest approach, the `2 -MPC
q3 (Wb3 − Wb (t)) one is significantly slower, both `Cauchy -MPC and `DCS -
+ (19)
V MPC ones are the worst ones (they both give practically the
4 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3057674, IEEE Access

Paweł D. Domański and Maciej Ławryńczuk: Impact of MPC Embedded Performance Index on Control Quality

same results). Next, Fig. 4 presents also undisturbed time 0.1

trends showing an effect of scaling sc changes for `2 -MPC.
Of course, the higher the scaling factor, the faster the control. 0.05
Finally, Fig. 5 presents disturbed time trends (all MPCs share
the same disturbance realization). Similarly to Fig. 3, The `1 -
0
MPC algorithm gives the fastest response, the `2 -MPC one is
significantly slower, both `Cauchy -MPC and `DCS -MPC ones
are the worst ones . -0.05

-0.1
0.6 L1-MPC
L1-MPC
L2-MPC
L2-MPC
0.5
Cauchy-MPC -0.15
Cauchy-MPC
DCS-MPC
DCS-MPC
0.4 -0.2
3720 3740 3760 3780 3800 3820 3840

0.3
FIGURE 5. Control error time trends for MPC controllers simulated with
disturbances
0.2

TABLE 3. Impact of a controller on step response

0.1
`1 -MPC `2 -MPC `Cauchy -MPC `DCS -MPC
0 κ 0.598% 0.850% 0.805% 0.785%
Tset 100 160 200 200
4245 4250 4255 4260 4265 4270

FIGURE 3. Control error time trends for MPC controllers with coefficient MPCs. It is well seen that increasing the scaling improves
sc = 1 simulated without disturbances
control, both in the settling time and the overshoot.

TABLE 4. Impact of scaling coefficient sc on step response performance

0.6 sc=1 indexes
sc=2
sc 1 2 5
0.5
sc=5
sc=10 κ 0.850% 0.795% 0.680%
Tset 160 120 100
sc=20
0.4 sc=50 sc 10 20 50
κ 0.660% 0.640% 0.638%
0.3 Tset 90 90 90

0.2 Further comparison of the obtained results is presented in

a set of the figures showing the relationship between scaling
0.1 factor sc and CPA index simultaneously for four considered
robust estimators of MPC performance index: `2 -MPC, `1 -
0 MPC, `Cauchy -MPC and `DCS -MPC.
4245 4250 4255 4260 4265 4270 The comparison starts with Fig. 6, which shows the re-
lationship reflected with the mean square error. It shows that
`1 -MPC controller reaches the best performance according to
FIGURE 4. Control error time trends for various scaling using `2 -MPC
simulated without disturbances the MSE. It is also seen that three of the controllers saturate
on the same performance `1 -MPC, `2 -MPC and `Cauchy -
Numerical results of settling time Tset and overshoot κ are MPC, except `DCS -MPC as scaling factor reaches sc = 20.
presented in a tabular way. Table 3 shows values of Tset and It is due to the fact that then the shapes of the performance
κ for MPCs having the same scaling sc = 1. It is well seen index function converge to the same function. It should be
that `1 -MPC has the shortest settling time. Though overshoot noticed that for large scaling sc = 20 the `DCS -MPC loses
is small in all cases, it reaches the minimum for `1 -MPC as stability.
well. Similar behavior appears for IAE and σG CPA measures.
Table 4 presents an effect of performance scaling sc for Three robust measures: LMS (Fig. 9), σH (Fig. 10) and γ
classical `2 -MPC, however the same effect appears for other (Fig. 11) show slightly different pattern, however still `1 -
VOLUME 4, 2016 5

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3057674, IEEE Access

Paweł D. Domański and Maciej Ławryńczuk: Impact of MPC Embedded Performance Index on Control Quality

L1 L2 Cauchy DCS L1 L2 Cauchy DCS

0.16 8.0E-06
0.15 7.0E-06
0.14
6.0E-06
0.13
MSE index value

LMS index value

0.12 5.0E-06

0.11 4.0E-06
0.1 3.0E-06
0.09
2.0E-06
0.08
0.07 1.0E-06

0.06 0.0E+00
1 2 5 10 20 50 1 2 5 10 20 50
SCALING COEFFICIENT sc SCALING COEFFICIENT sc

FIGURE 6. Performance index impact measured with MSE FIGURE 9. Performance index impact measured with LMS

L1 L2 Cauchy DCS L1 L2 Cauchy DCS

0.12 0.0045

0.11 0.004

0.1
0.0035
IAE index value

0.09
σH index value

0.003
0.08
0.0025
0.07
0.002
0.06

0.05 0.0015

0.04 0.001
1 2 5 10 20 50 1 2 5 10 20 50
SCALING COEFFICIENT sc SCALING COEFFICIENT sc

FIGURE 7. Performance index impact measured with IAE FIGURE 10. Performance index impact measured with σH

L1 L2 Cauchy DCS L1 L2 Cauchy DCS

0.39 0.002

0.37 0.0018

0.0016
0.35
σG index value

0.0014
γ index value

0.33
0.0012
0.31
0.001
0.29
0.0008

0.27 0.0006

0.25 0.0004
1 2 5 10 20 50 1 2 5 10 20 50
SCALING COEFFICIENT sc SCALING COEFFICIENT sc

FIGURE 8. Performance index impact measured with σG FIGURE 11. Performance index impact measured with γ

MPC exhibits the best values. It confirms outlier robustness controllers, while the next Table 6 shows similar compari-
properties observed in the CPA analysis [32]. son for the simulations disturbed with α-stable disturbance
Differential entropy Hdiff sketched in Fig. 12 detects per- variable. Tabular results confirm earlier observations. In an
formance in a similar way to MSE, while rational entropy undisturbed case the `1 -MPC reaches the best performance
Hrat presented in Fig. 13 exhibits the highest scattering of according to all used CPA measures. The results for disturbed
the curves. simulations show a different pattern. Outlier sensitive CPA
Table 5 presents CPA indexes for four undisturbed MPC indexes (MSE, IAE, σG and Hrat ) select `2 -MPC, while the
6 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3057674, IEEE Access

Paweł D. Domański and Maciej Ławryńczuk: Impact of MPC Embedded Performance Index on Control Quality

TABLE 5. CPA indexes in undisturbed simulations (best value highlighted)

MPC type MSE IAE σG σH γ Hdiff Hrat LMS

`1 -MPC 7.457 × 10−2 5.065 × 10−2 2.731 × 10−1 1.123 × 10−3 4.681 × 10−4 1.413 × 103 1.342 4.114 × 10−7
`2 -MPC 7.487 × 10−2 5.163 × 10−2 2.736 × 10−1 1.218 × 10−3 5.128 × 10−4 1.406 × 103 1.354 4.864 × 10−7
`Cauchy -MPC 7.606 × 10−2 5.328 × 10−2 2.758 × 10−1 1.278 × 10−3 5.399 × 10−4 1.399 × 103 1.391 5.479 × 10−7
`DCS -MPC 7.668 × 10−2 5.319 × 10−2 2.769 × 10−1 1.273 × 10−3 5.371 × 10−4 1.400 × 103 1.373 5.377 × 10−7

TABLE 6. CPA indexes in disturbed simulations (best value highlighted)

MPC type MSE IAE σG σH γ Hdiff Hrat LMS

`1 -MPC 7.975 × 10−2 6.772 × 10−2 2.824 × 10−1 1.016 × 10−2 5.278 × 10−3 1.272 × 103 1.422 3.959 × 10−5
`2 -MPC 7.797 × 10−2 6.687 × 10−2 2.792 × 10−1 1.123 × 10−2 5.878 × 10−3 1.244 × 103 1.392 4.944 × 10−5
`Cauchy -MPC 7.938 × 10−2 6.858 × 10−2 2.818 × 10−1 1.178 × 10−2 6.175 × 10−3 1.238 × 103 1.400 5.393 × 10−5
`DCS -MPC 7.957 × 10−2 6.836 × 10−2 2.821 × 10−1 1.178 × 10−2 6.153 × 10−3 1.234 × 103 1.441 5.387 × 10−5

L1 L2 Cauchy DCS facilitates its application.

1430.0
Tables 7 and 8 present how the MPC control performance
1410.0
is assessed by various performance measures. Tables show
1390.0
relative improvement measured versus the best case. Such
1370.0 data presentation informs about the scale of the improvement
Hdiff index value

1350.0 exhibited by the best measure.

1330.0 Above observations are consistent with the ones noticed
1310.0 in the context of the controller tuning according to the mean
1290.0 square error and absolute error cost function. Research shows
1270.0 that tuning minimizing the MSE punishes large deviations
1250.0
and causes aggressive control [42]. The absolute error is less
1 2 5 10 20 50 conservative and it is often used for an on-line controller
SCALING COEFFICIENT sc
tuning. The IAE has the closest relation with economic
aspects [43]. It penalizes continued cycling.
FIGURE 12. Performance index impact measured with Hdiff
Finally, let us compare the calculation time of all four con-
sidered MPC algorithms. It turns out that the `1 -MPC method
L1 L2 Cauchy DCS
1.9
is almost twice more computationally demanding than the
other MPC schemes. Let the scaled computational time for
1.8 the `1 -MPC algorithm equals to 100%. The obtained times
are: 55.37%, 54.72% and 53.31% for `2 -MPC, `Cauchy -MPC
1.7
and `DCS -MPC approaches.
Hrat index value

1.6

IV. CONCLUSIONS AND FURTHER RESEARCH

1.5
Performed research focuses on the possibility to improve
1.4 MPC control through the use of other, outlier robust, perfor-
mance index formulation. Classical `2 cost function is com-
1.3 pared against `1 -MPC and two other formulations originating
1 2 5 10 20 50
SCALING COEFFICIENT sc from other contexts, i.e. `Cauchy -MPC and `DCS -MPC.
The results show consistent observations. `1 -MPC always
FIGURE 13. Performance index impact measured with Hrat exhibits the best performance, independently on the applied
CPA measure. Furthermore, it is less sensitive to scaling
sc. At the same time, it obtains the shortest step response
robust ones (LMS, σH , γ and Hdiff ) point out the `1 -MPC. settling time and the smallest overshoot. It shows that linear
It is due to the fact that the applied disturbance is fat-tailed. introduction of the cost function is more realistic than the
In such situations, the robust indexes are more appropriate. quadratic one. Finally, obtained results show that two se-
In all cases two other controllers, i.e. `Cauchy and `DCS -MPC lected robust estimators using Cauchy function and Dynamic
exhibit worse performance. Covariance Scaling are not improving MPC operation. As
Moreover, the use of the `1 -cost function seems to be positive properties of the `1 norm are well known, it is
less sensitive to scaling sc in view of obtained results. That expected that similar results might be observed in other MPC
seems to be a key advantage of the `1 -MPC as it significantly applications as well.
VOLUME 4, 2016 7

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3057674, IEEE Access

Paweł D. Domański and Maciej Ławryńczuk: Impact of MPC Embedded Performance Index on Control Quality

TABLE 7. CPA index relative change versus the best one in undisturbed case (best value highlighted)

MPC type MSE IAE σG σH γ Hdiff Hrat LMS

`1 -MPC 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00%
`2 -MPC 0.40% 1.92% 0.20% 8.39% 9.55% 0.52% 0.91% 18.25%
`Cauchy -MPC 1.99% 5.19% 0.99% 13.78% 15.33% 0.96% 3.61% 33.19%
`DCS -MPC 2.82% 5.01% 1.40% 13.33% 14.74% 0.91% 2.33% 30.71%

TABLE 8. CPA index relative change versus the best one in disturbed case (best value highlighted)

MPC type MSE IAE σG σH γ Hdiff Hrat LMS

`1 -MPC 2.28% 1.27% 1.15% 0.00% 0.00% 0.00% 2.16% 0.00%
`2 -MPC 0.00% 0.00% 0.00% 10.53% 11.37% 2.20% 0.00% 24.88%
`Cauchy -MPC 1.81% 2.56% 0.93% 15.94% 17.00% 2.67% 0.57% 36.22%
`DCS -MPC 2.05% 2.23% 1.04% 15.94% 16.58% 2.99% 3.52% 36.07%

It is worth to evaluate the performance of other possible [14] G. Williams, A. Aldrich, and E. A. Theodorou, “Model predictive path in-
robust regression functions. As in the present research, the tegral control: From theory to parallel computation,” Journal of Guidance,
Control, and Dynamics, vol. 40, pp. 344–357, 2017.
control cost element in the MPC performance index is still [15] I. S. Mohamed, G. Allibert, and P. Martinet, “Model predictive path inte-
quadratic, one might be interested in verifying whether its gral control framework for partially observable navigation: A quadrotor
formulation has any impact. The other opportunity would be case study,” in 16th International Conference on Control, Automation,
Robotics and Vision (ICARCV), Shenzhen, China, 2020, pp. 197–203.
to exchange mean operator in MPC performance index with [16] D. E. Davison, R. Milman, and E. J. Davison, “Optimal transient response
a median. shaping in model predictive control,” IFAC Proceedings Volumes, vol. 38,
no. 1, pp. 147–152, 2005, 16th IFAC World Congress.
[17] L. Hewing and M. N. Z. J. Kabzan, “Cautious model predictive control
REFERENCES using gaussian process regression,” IEEE Transactions on Control Systems
[1] R. Cutler and B. Ramaker, “Dynamic matrix control – a computer control Technology, vol. 28, pp. 2736 – 2743, 2019.
algorithm,” in Proc. AIChE National Meeting, Houston, TX, US, 1979. [18] M. A. Müller and K. Worthmann, “Quadratic costs do not always work in
[2] W. Clarke, C. Mohtadi, and P. S. Tuffs, “Generalized predictive control - mpc,” Automatica, vol. 82, pp. 269–277, 2017.
I. the basic algorithm,” Automatica, vol. 23, no. 2, pp. 137–148, 1987. [19] P. J. Rousseeuw and A. M. Leroy, Robust Regression and Outlier Detec-
[3] P. Tatjewski, Advanced control of industrial processes, structures and tion. New York, NY, USA: John Wiley & Sons, Inc., 1987.
algorithms. London, UK: Springer-Verlag, 2007. [20] P. J. Huber and E. M. Ronchetti, Robust Statistics, 2nd Edition. Wiley,
[4] M. Ławryńczuk, Computationally efficient model predictive control algo- 2009.
rithms: a neural network approach, ser. Studies in Systems, Decision and [21] H. Wainer, “Robust statistics: A survey and some prescriptions,” Journal
Control, vol. 3. Cham, Switzerland: Springer International Publishing, of Educational Statistics, vol. 1, no. 4, pp. 285–312, 1976.
2014. [22] D. M. Hawkins, Identification of outliers. London; New York: Chapman
[5] F. Tahir, E. Mercer, I. Lowdon, and D. Lovett, “Advanced process control and Hall, 1980.
and monitoring of a continuous flow micro-reactor,” Control Engineering [23] L. B. Klebanov and I. Volchenkova, “Outliers and the ostensibly heavy
Practice, vol. 77, pp. 225–234, 2018. tails,” Mathematical Methods of Statistics, vol. 28, pp. 74–81, 2019.
[6] B. Huyck, J. De Brabanter, B. De Moor, J. F. Van Impe, and F. Logist, [24] J. W. Osborne and A. Overbay, “The power of outliers (and why re-
“Online model predictive control of industrial processes using low level searchers should always check for them),” Practical Assessment, Research,
control hardware: A pilot-scale distillation column case study,” Control and Evaluation, vol. 9, no. 6, pp. 1–8, 03 2004.
Engineering Practice, vol. 28, pp. 34–48, 2014. [25] N. N. Taleb, “Statistical consequences of fat tails: Real world preasymp-
[7] J. Gabor, D. Pakulski, P. D. Domański, and K. Świrski, “Closed loop nox totics, epistemology, and applications,” 2020, arXiv:2001.10488, Cornell
control and optimization using neural networks,” in IFAC Symposium on Univesity Library.
Power Plants and Power Systems Control 2000, Brussels, Belgium, 2000, [26] M. Shao and C. L. Nikias, “Signal processing with fractional lower order
pp. 141–146. moments: Stable processes and their applications,” Proceedings of the
[8] R. Jankowski, P. D. Domański, and K. Świrski, “Optimization of a coal IEEE, vol. 81, no. 7, pp. 986–1010, 1993.
mill using a mpc type controller,” in Proceedings of IMECE’03, ASME [27] A. Dötlinger and R. M. Kennel, “Near time-optimal model predictive
International Mechanical Engineering Congress, Washington, DC, 2003, control using an l1-norm based cost functional,” in 2014 IEEE Energy
pp. 233–243. Conversion Congress and Exposition (ECCE), 2014, pp. 3504–3511.
[9] J. Huang, H. An, Y. Yang, C. Wu, Q. Wei, and H. Ma, “Model predictive [28] A. Bemporad, F. Borrelli, and M. Morari, “Model predictive control based
trajectory tracking control of electro-hydraulic actuator in legged robot on linear programming - the explicit solution,” IEEE Transactions on
with multi-scale online estimator,” IEEE Access, vol. 8, pp. 95 918–95 933, Automatic Control, vol. 47, no. 12, pp. 1974–1985, 2002.
2020. [29] M. Gallieri, `asso -MPC - Predictive Control with `1 -Regularised Least
[10] S. Kayalvizhi and D. M. Vinod Kumar, “Load frequency control of an Squares, ser. Springer Theses. Switzerland: Springer International Pub-
isolated micro grid using fuzzy adaptive model predictive control,” IEEE lishing, 2016.
Access, vol. 5, pp. 16 241–16 251, 2017. [30] C. V. Rao and J. B. Rawlings, “Linear programming and model predictive
[11] C. Jia, X. Wang, Y. Liang, and K. Zhou, “Robust current controller for control,” Journal of Process Control, vol. 10, no. 2, pp. 283–289, 2000.
IPMSM drives based on explicit model predictive control with online [31] P. D. Domański and M. Ławryńczuk, “Control quality assessment of
disturbance observer,” IEEE Access, vol. 7, pp. 45 898–45 910, 2019. nonlinear model predictive control using fractal and entropy measures,”
[12] P. Wang, C. Zhu, and G. J., “Feedforward model predictive control of fuel- in Nonlinear Dynamics and Control, W. Lacarbonara, B. Balachandran,
air ratio for lean-burn spark-ignition gasoline engines of passenger cars,” J. Ma, J. A. Tenreiro Machado, and G. Stepan, Eds. Cham: Springer
IEEE Access, vol. 7, pp. 73 961–73 969, 2019. International Publishing, 2020, pp. 147–156.
[13] S. Li, Z. Li, Z. Yu, B. Zhang, and N. Zhang, “Dynamic trajectory planning [32] P. D. Domański, Control Performance Assessment: Theoretical Analyses
and tracking for autonomous vehicle with obstacle avoidance based on and Industrial Practice. Cham: Springer Nature Switzerland AG, 2020.
model predictive control,” IEEE Access, vol. 7, pp. 132 074–132 086, [33] P. D. Domański, “Performance assessment of predictive control—a sur-
2019. vey,” Algorithms, vol. 13, 2020, article ID 97.

8 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3057674, IEEE Access

Paweł D. Domański and Maciej Ławryńczuk: Impact of MPC Embedded Performance Index on Control Quality

[34] K. MacTavish and T. D. Barfoot, “At all costs: A comparison of robust cost
functions for camera correspondence outliers,” in 2015 12th Conference on
Computer and Robot Vision, 2015, pp. 62–69.
[35] G. Hu, K. Khosoussi, and S. Huang, “Towards a reliable slam back-end,”
in 2013 IEEE/RSJ International Conference on Intelligent Robots and
Systems, 2013, pp. 37–43.
[36] P. Agarwal, G. D. Tipaldi, L. Spinello, C. Stachniss, and W. Burgard,
“Robust map optimization using dynamic covariance scaling,” in 2013
IEEE International Conference on Robotics and Automation, 2013, pp.
62–69.
[37] J. Gómez, A. Jutan, and E. Baeyens, “Wiener model identification and
predictive control of a pH neutralisation process,” Proc. IEE, Part D,
Control Theory Applications, vol. 151, pp. 329–338, 2004.
[38] S. Borak, A. Misiorek, and R. Weron, “Models for heavy-tailed asset
returns,” in Statistical tools for finance and insurance, 2nd ed., P. Cizek,
K. Härdle, W., and R. Weron, Eds. Springer, New York, 2011, pp. 21–56.
[39] S. Verboven and M. Hubert, “LIBRA: a Matlab library for robust analysis,”
Chemometrics and Intelligent Laboratory Systems, vol. 75, pp. 127–136,
02 2005.
[40] H. Yue and H. Wang, “Minimum entropy control of closed-loop tracking
errors for dynamic stochastic systems,” IEEE Transactions on Automatic
Control, vol. 48, no. 1, pp. 118–122, 2003.
[41] J. Zhang, M. Jiang, and J. Chen, “Minimum entropy-based performance
assessment of feedback control loops subjected to non-Gaussian distur-
bances,” Journal of Process Control, vol. 24, no. 11, pp. 1660–1670, 2015.
[42] D. E. Seborg, D. A. Mellichamp, T. F. Edgar, and F. J. Doyle, Process
dynamics and control. Wiley, 2010.
[43] F. G. Shinskey, “Process control: As taught vs as practiced,” Industrial and
Engineering Chemistry Research, vol. 41, pp. 3745–3750, 2002.

PAWEŁ D. DOMAŃSKI was born in Warsaw,

Poland in 1967. He received the M.S. degree in
1991, the Ph.D. degree in 1996 and the D.Sc.
degree in 2018, all in control engineering from
the Warsaw University of Technology, Faculty of
Electronics and Information Technology.
He works in the Institute of Control and Compu-
tational Engineering, Warsaw University of Tech-
nology from 1991. Apart from scientific research,
he participated in dozens of industrial implemen-
tations of APC and optimization in power and chemical industries. He is
the author of one book and more than 100 publications. His main research
interest is with industrial APC applications, control performance quality
assessment and optimization.

MACIEJ ŁAWRYŃCZUK was born in Warsaw,

Poland, in 1972. He obtained his M.Sc. in 1998,
Ph.D. in 2003, D.Sc. in 2013, all in automatic
control, from Warsaw University of Technology,
Faculty of Electronics and Information Technol-
ogy.
He has been with the Institute of Control and
Computation Engineering, Warsaw University of
Technology, since 1998, where he is currently
employed as an associate professor. He is the
author or a co-author of 6 books and more than 100 other publications,
including more than 40 journal articles. His research interests include:
advanced control algorithms, in particular, Model Predictive Control (MPC)
algorithms, set-point optimization algorithms, soft computing methods, in
particular neural networks, modeling, and simulation.

VOLUME 4, 2016 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/