2016 IEEE Conference on Control Applications (CCA)

Part of 2016 IEEE Multi-Conference on Systems and Control

September 19-22, 2016. Buenos Aires, Argentina

Nonlinear identification and control of Organic Rankine Cycle systems

using sparse polynomial models.
Andres Hernandez1,3 , Fredy Ruiz2 , Adriano Desideri3 , Clara Ionescu1 , Sylvain Quoilin3 ,
Vincent Lemort3 and Robin De Keyser1

Abstract— Development of a first principles model of a system mial nonlinear autoregressive exogenous (NARX) models are
is not only a time- and cost- consuming task, but often leads particular interesting, since they allow a direct interpretation
to model structures which are not directly usable to design a and analysis of the nonlinear input-output dependencies. Sev-
controller using current available methodologies. In this paper
we use a sparse identification procedure to obtain a nonlinear eral experimental studies have shown that many systems can
polynomial model. Since this is a NP-hard problem, a relaxed be accurately approximated by polynomial NARX models of
algorithm is employed to accelerate its convergence speed. The relatively small size (see e.g., [4], [5]).
obtained model is further used inside the nonlinear Extended In this contribution the objective is to identify a sparse
Prediction Self-Adaptive control (NEPSAC) approach to Non- nonlinear polynomial model using a modified version of
linear Model Predictive Control (NMPC), which replaces the
complex nonlinear optimization problem by a simpler iterative the Elastic Net regression algorithm originally proposed in
quadratic programming procedure. An organic Rankine cycle [6]. First it is essential to clarify what sparse means in this
system, characterized for presenting nonlinear time-varying context. It consists in approximating a function using a ‘few’
dynamics, is used as benchmark to illustrate the effectiveness basis functions properly selected within a ‘large’ set. More
of the proposed combined strategies. precisely, a sparse approximation is a linear combination of
fixed basis functions, where the vector of linear combination
I. INTRODUCTION
coefficients is sparse, i.e. it has only a ‘few’ non-zero
Today process industry is experiencing an increasing pro- elements. The pioneer work performed in [6], [7] shows
ductivity demand. This, together with tighter environmental that a sparse model can be obtained by making a basic
regulations, demanding economical considerations and the augmentation of the ordinary least squares solution, with the
need of operating in an energy-efficient manner, is forcing goal of obtaining a coefficient shrinkage and selection. The
industry to operate systems closer to the boundary of the augmentation often consists in including penalty functions to
admissible operating region, where productivity is often the least square problem, thus providing some regularization.
maximized. In this regard modeling and control design play In [6], [7] and the references therein, it is shown that for
an important role to achieve optimal operation of a system. sufficient amounts of regularization, solutions are sparse,
One of the control strategies which has been well-accepted i.e., some of the coefficients of the model are exactly zero,
by industry is Model Predictive Control (MPC) [1]. MPC leading to more compact models which are easier to interpret.
refers to a family of control approaches, which makes explicit The modification here proposed in section II, is used to
use of a model of the process to optimally obtain the control accelerate convergence speed by relaxing the optimization
signal by minimizing an objective function [2]. Due to this problem, whereas the goal is to minimize a combination
active use of a model, its performance is strongly linked to of the `1 - and `2 - norm of the parameter vector, with
the model’s ability to accurately predict the system dynamics. a constraint on the model accuracy. The relaxed Sparse
Real processes are inherently nonlinear, thus ideally a low- identification algorithm allows to obtain the structure of the
order nonlinear model and therefore a nonlinear MPC should model, then a second step is implemented by identifying
be used to control the system. However, linear predictive its coefficients using a least-squares criterion and applying
controllers have been predominantly used, this is due to pruning. Finally an input/output polynomial model of the
the easy to obtain linear parametric models compared to system with prescribed modeling error is obtained [8]. The
nonlinear ones [3]. methodology is similar to the one known as simulation error
Many model classes have been considered for nonlinear minimization with pruning (SEMP), which as reported in
identification purposes, e.g., Wiener or Hammerstein, neu- [9], can yield an increase in robustness for model selection
ronal networks, radial basis functions, support vector ma- compared to prediction error minimization (PEM) framework
chines and polynomial expansions. From those the polyno- with respect to low excitation identification conditions.
1 A.
Then the obtained polynomial model is used to construct
Hernandez, C. Ionescu and R. De Keyser are with
Faculty of Engineering and Architecture, Department of Electrical
a nonlinear MPC strategy, i.e., the Nonlinear Extended
energy, Systems and Automation, Ghent University, Belgium. Prediction Self-Adaptive Control (NEPSAC) approach to
[email protected] NMPC. This controller, fully described in [10], can deal
2 F. Ruiz is with Department of electronics engineering, Pontificia Uni-
with models in different formats (e.g. state-space, transfer
versidad Javeriana, Colombia. [email protected]
3 S. Quoilin and V. Lemort are with Thermodynamics Laboratory, Uni- functions, neuronal networks, polynomial parametric models,
versity of Liege, Belgium. [email protected] etc.) as it uses the nonlinear model for prediction directly,

978-1-5090-0754-7/16/$31.00 ©2016 IEEE 1012

taking full advantage of the given nonlinear system dynamics A. Problem formulation
to generate a high-performance design without involving Consider a nonlinear function f0 defined by
model linearization or gain scheduling for its implementa-
tion. Additionally, the usually complex nonlinear optimiza- y = f0 (x) (1)
tion problem, is in the NEPSAC algorithm replaced by a
simpler iterative quadratic programming procedure, to solve where x ∈ Rnx , y ∈ R. Suppose that f0 is not known but a set
problems including constraints as briefly described in section of noise-corrupted data D = {x̃, ỹ}Lk=1 is available, described
III. by
The effectiveness of the proposed combined strategy is y˜k = f0 (x̃k ) + nk , k = 1, 2, . . . , L (2)
tested in simulation using an organic Rankine cycle (ORC)
where nk is noise. Define the following parametrized func-
system for waste heat recovery applications (WHR). This
tion:
system is particularly challenging, since its dynamics are n
highly nonlinear and time-varying because they change as fa (x) = ∑ ai φi (x) = φ (x)a (3)
function of the waste-heat source [11]. Their dynamic com- i=1

ponents are mainly allocated in the evaporator, therefore where φ (x) = [φ1 (x), φ2 (x), . . . , φn (x)], φi : X → Y are known
research in modeling has been focused on finding a trade-off fixed basis functions, and a = [a1 , a2 , . . . , an ]T ∈ Rn is an
between model accuracy and simulation speed as evaluated in unknown coefficient vector.
[12], where finite volume and moving boundary approaches Problem 1: from the data set D, identify a coefficient
to evaporator modeling are proposed. vector a such that
Regarding the control design for ORC systems, it is (i) is sparse and
essential to guarantee a vapor quality at the inlet of the (ii) the identification error
expander, since liquid droplets can damage the machine. This
requirement can be achieved by ensuring a certain level of ˙ f 0 − f a k2
e( fa )=k (4)
superheating [13], [14]. For example, in [15] a feed-forward is small and bounded.
and gain-scheduled PI controller is proposed. In [16] an Sparse solutions may be preferred to full counterparts if
explicit multi-model MPC is applied to control superheating. the model can be assumed to be sparse or when interpretation
In conclusion, identifying a nonlinear model for superheating of the results can be important.
and achieving a proper regulation of this signal despite waste-
heat variations represents the target for the controller. B. Sparse Identification algorithm
The performance of the nonlinear identification and con- The sparsity of a vector is typically measured by the `0
trol algorithms is evaluated by developing a superheating quasi-norm, defined as the number of its non-zero elements.
control of an ORC system as described in section IV. Finally, Sparse identification can be performed by looking for a
the main outcome of this investigation is summarized in coefficient vector of the basis function linear combination
section V. with a ‘small’ `0 quasi-norm, that yields a ‘small’ prediction
error ε between the measured output ỹ and the predicted
output fa (x̃) = φ (x̃)a. Thus, a solution to problem 1 could
II. S PARSE I DENTIFICATION OF NONLINEAR FUNCTIONS be found by solving the following optimization problem:

In cases where an accurate first principle model is not a0 = arg minn kak0
a∈R (5)
available for a complex system, models deduced from ex- sub ject to kỹ − Φak2 ≤ ε
perimental data by means of system identification techniques
provide reliable alternatives. Whilst existing system identifi- with ε the maximum desired prediction error and
cation methods for linear systems can be considered to be at a
ỹ =
˙ (ỹ1 , . . . , ỹL )
mature stage, nonlinear system identification techniques still  
reveal several fundamental issues to be resolved to further φ1 (x̃1 ) . . . φn (x̃1 )
improve general model performance. In this section a sparse Φ= ˙
 .. . . ..  (6)
. . . 
identification algorithm of nonlinear functions is presented. φ1 (x̃L ) . . . φn (x̃L )
Identifying a sparse approximation of a function from a set =
˙ [φ1 (x̃) . . . φn (x̃)]
of data, possibly corrupted by noise, is what is called sparse
identification. It consists in constructing a sparse model However, the `0 is a non-convex function and therefore
which involves a minimum model variables and terms but can minimization of (5) is in general an NP-hard problem.
generate a maximum of model generalization performance. Similarly as proposed by [8], we propose a convex relaxation,
Therefore, a sparse approximation is a linear combination of where an optimization problem similar to (5) is solved, by
fixed basis functions, where the vector of linear combination replacing the `0 quasi-norm by a combination of a `1 and
coefficients is sparse, i.e. it has only a ‘few’ non-zero weighted `2 norms. The optimization problem (5) is thus
elements. redefined as a two-step algorithm:

1013
 Step 1: solve the optimization problem x(t + k|t) and n(t + k|t) can be predicted by recursion of
1 −1 the process model (10) and by using filtering techniques on
a = arg minn L kak1 + β kak2
a∈R (7) the noise model (11), respectively [10].
s.t. kỹ − Φak2 ≤ ε
B. Optimization Procedure
with L the length of the data set, β a weighting coefficient A key element in linear MPC is the use of base (or
and ε the maximum expected prediction error. free) and optimizing (or forced) response concepts [2]. This
This structure of the optimization problem in equation (7) is however no longer valid for nonlinear processes, since
is chosen, as the `2 penalty ensures an unique solution also the superposition principle does not hold in that case. In
when the number of basis functions is larger than the data NEPSAC this is solved as follows:
length, and the `1 penalty offers variable selection via a The future response can then be expressed as:
sparse vector of coefficients a as demonstrated in [6].
Step 2: define a sub-space with the active basis of a1 y(t + k|t) = ybase (t + k|t) + yoptimize (t + k|t) (13)
and compute the estimate a∗ of a0 using least-squares of
The two contributing factors have the following origin:
prediction error
• ybase (t + k|t) is the effect of the past inputs, the apriori
a∗ = arg minn kỹ − Φai k2 defined future base control sequence ubase (t + k|t) and
ai ∈R
(8) the predicted disturbance n(t + k|t).
s.t. ai = 0, ∀i |a1i | < γ
• yoptimize (t + k|t) is the effect of the additions δ u(t + k|t)
with γ an additional tolerance value used for pruning, thus that are optimized and added to ubase (t + k|t), according
enhancing sparsity. to δ u(t + k|t) = u(t + k|t) − ubase (t + k|t). The effect of
Note that the major problem with regularization methods these additions is the discrete time convolution of ∆U =
is the optimal selection of the regularization parameters, in {δ u(t|t), . . . , δ u(t +Nu −1|t)} with the impulse response
this case γ, β and ε, which correspond to a complexity/bias coefficients of the system (G matrix), where Nu is the
vs. variance trade-off [6]. chosen control horizon.
III. NMPC-NEPSAC ALGORITHM The control ∆U is the solution to the following constrained
optimization problem:
A brief introduction to NEPSAC algorithm is presented in
N2
this section. For a detailed description the reader is referred
∆U =arg min ∑ [r(t + k|t) − y(t + k|t)]2
to [10] and [17]. ∆U∈ RNu k=N (14)
1

A. Computing the Predictions s.t. A.∆U ≤ B

Using NEPSAC algorithm, the measured process output where N1 and N2 are the minimum and maximum prediction
can be represented as:
horizons, Nu is the control horizon, r(t + k|t) is a future
y(t) = x(t) + n(t) (9) setpoint or reference sequence. The various process inputs
where x(t) is the model output which represents the effect and output constraints can all be expressed in terms of ∆U,
of the control input u(t) and n(t) represents the effect of the resulting in matrices A, B.
disturbances and modeling errors, all at discrete-time index As the cost function (14) with decision variables ∆U
t. Model output x(t) can be described by the generic system is quadratic with linear constraints, then the minimization
dynamic model:
problem can be solved by a QP algorithm.
x(t) = f [x(t − 1), x(t − 2), . . . , u(t − 1), u(t − 2), . . .] (10)
C. NEPSAC iterative procedure
Notice that x(t) represents here the model output, not the The procedure described until now can be used to im-
state vector. Also important is the fact that f can be either plement a linear MPC (EPSAC), if f [·] in (10) is linear, as
a linear or a nonlinear function. superposition principle holds in (13). Instead, if a nonlinear
Furthermore, the disturbance n(t) can be modeled as
colored noise through a filter with the transfer function: model is employed, the condition is that the term yoptimize (t +
k|t) is small enough compared to ybase (t + k|t), in order to
C(q−1 ) make the superposition principle still valid. Such condition
n(t) = e(t) (11)
D(q−1 ) can be satisfied if δ u(t + k|t) is close to zero, making thus
with e(t) uncorrelated (white) noise with zero-mean and ubase (t + k|t) close to the optimal u∗ (t + k|t). To address
C, D monic polynomials in the backward shift operator q−1 . this issue, the idea is to recursively compute δ u(t + k|t)
The disturbance model must be designed to achieve robust- solving (14), within the same sample instant, until δ u(t +k|t)
ness of the control loop against unmeasured disturbances and converges to 0. Inside the recursion ubase (t + k|t) is updated
modeling errors [2]. each time to ubase (t + k|t) + δ u(t + k|t).
A fundamental step in the MPC methodology consists Also important is the fact that in NEPSAC the G matrix
of the prediction. Using the generic process model (9), the has to be recomputed at every sampling instant [17]. The
predicted values of the output are: number of required iterations depends on how far the op-
y(t + k|t) = x(t + k|t) + n(t + k|t) (12) timal u∗ (t + k|t) is away with respect to u∗ (t + k|t − 1). In

1014
quasi-steady-state situations, the number of iterations is low speed is kept constant at 3000 rpm to emulate an installation
(1 . . . 2) [17]. On the other hand, during transients the number directly connected to the grid.
of iterations might raise to 10 [17]. Regarding the optimal operation of the ORC unit, the
superheating and the evaporating temperature are the most
IV. C ASE STUDY: WASTE HEAT RECOVERY APPLICATION relevant variables to be controlled [19]. The superheating is
This section describes the architecture and main charac- defined as:
teristics of the Organic Rankine Cycle (ORC) system used
∆Tsh = Texp,su − Tsat,ev (15)
for evaluating the performance of the developed control
strategies.
where Texp,su is the temperature measured at the inlet of the
A. The Organic Rankine Cycle system expander and Tsat,ev the evaporating temperature, correspond-
ing to the temperature at which the fluid undergoes the phase
The ORC power unit investigated in this work is based transition from saturated liquid to saturated vapor at the fixed
on a sub-critical 11 kWel unit for stationary low temperature evaporating pressure Psat,ev .
waste heat recovery. The system is based on a regenerative
Research performed on ORC technology (e.g., [13], [14])
cycle and Solkatherm (SES36) as working fluid. The 11 kWel
has already established that to optimally operate the ORC
expander is originally a single screw compressor adapted to
power unit, two conditions need to be satisfied: 1) A ‘high’
run in expander mode.
efficiency and a safe operation of the ORC unit is achieved
if a minimum amount of superheating at expander inlet is
guaranteed. 2) For each heat source condition there exists
an optimal evaporating temperature which maximizes the
output power. In this study we will focus on fulfilling the
first condition, by designing a controller for superheating.
In order to assess the overall performance of the system the
net-output power Ẇel,net is computed:

Ẇel,net = Ẇexp − Ẇpump (16)

where Ẇexp is the expander electrical power, Ẇpump is the

pump electrical power.

B. Low-order model suitable for prediction

Fig. 1. Schematic layout of the organic Rankine cycle power unit.
The ORC unit considered has one manipulated variable
A schematic layout of the ORC system is presented in (the pump speed N pp ), and two outputs (the evaporating
Fig. 1. Starting from the bottom of the scheme, it is possible temperature Tsat,ev and the superheating ∆Tsh ). Notice that
to recognize the liquid receiver (b) installed at the outlet the temperature Th f and mass flow variations mh f in the
of the condenser (a) where the fluid is collected in saturated heat source also influence Tsat,ev and ∆Tsh , thus becoming
liquid condition. From the receiver outlet, the fluid is pumped two measured disturbances. As result we are interested on
(c) through the regenerator (d) cold side, and the evaporator identifying a system consisting of 3 inputs (one manipulated
(e), where it is heated up to superheated vapor, reaching its and two measured disturbances) and 2 outputs.
maximum temperature at the evaporator outlet. The fluid, The following input signals have been considered to
after being expanded in the volumetric machine (f), enters perform the identification: the heat source conditions Th f
the regenerator hot side, and then it flows into the condenser and mh f have been simulated as white noise filtered to a
(a) to close the cycle. maximum band of 0.5 rad/s and 0.63 rad/s with amplitude
In order to assess the performance of the different devel- variations of ± 20 ◦C around 110 ◦C and ± 0.5 kg/s around
oped control strategies, a validated dynamic model of the 1.0 kg/s, respectively. The pump speed has been taken as the
ORC system presented in Fig. 1 has been developed in the sum of 10 sinusoids spread over the band [0.005, 0.63] rad/s,
Modelica language [12], using existent components from the taking values between 1320 rpm and 2100 rpm. A set of
ThermoCycle library [18]. The developed Modelica model is L = 2000 samples has been generated from the ‘true’ system,
then exported into Simulink/Matlab environment by means considering a sampling time of Ts = 1 s.
of the Functional Mock-Up Interface (FMI) open standard, A linear parametric identification is performed, using the
using a model exchange format. This simulation approach prediction error minimization method in the data collected
takes advantage of the strengths of each platform: Modelica from multisine excitation signals [20]. The identified linear
for modeling and Simulink/Matlab for control design. For the model in the form of discrete-time transfer functions for a
simulations performed in this paper, the generator rotational sampling time of Ts = 1 s, is presented in (17):

1015
 30
25 Data Linear Nonlinear
∆Tsh (q) −0.063q−1 +0.059q−2
= 1−2.44q−1 +1.955q−2 −0.51q−3
(17a) 20

∆ Tsh [°C]
N pp (q) 15
∆Tsh (q) 0.47q−1 10
= 1−0.51q−1
(17b)
Th f (q) 5

∆Tsh (q) −2.98q−1 +4.29q−2 −1.31q−3 0

= 1−1.35q−1 −0.11q−2 +0.46q−3
(17c) 0 500 1000 1500 2000
mh f (q)
10
Tsat,ev (q) 0.066q−1 −0.063q−2
= 1−2.42q−1 +1.91q−2 −0.49q−3
(17d)
N pp (q) 0

Error
Tsat,ev (q) −11
0.0017q −0.0017q −12
= 1−3.6q−1 +4.88q−2 −2.95q−3 +0.67q−4
(17e) −10
Th f (q)
Tsat,ev (q) −1 −2
2.43q −6.16q +5.33q −1.6q −3 −4
= 1−2.93q−1 +3.12q−2 −1.42q−3 +0.23q−4
(17f) −20
0 500 1000 1500 2000
mh f (q) Time [s]

Next, we are interested in identifying a nonlinear poly-

nomial model of the system. In order to perform the sparse Fig. 2. Prediction of superheating using the identified linear and non-linear
identification, a set of n = 57 polynomial basis functions models as simulators.
has been considered and the corresponding matrix Φ =
(Φ1 (x̃), . . . , ΦL (x̃)) has been obtained according to:
140 2
Φk (x̃) =[φ1 (x̃), . . . , φn (x̃))] Thf
m
=[1, y(k − r), N pp (k − r), Th f (k − r), mh f (k − r), hf

y(k − r)2 , y(k − r) ∗ N pp (k − r), y(k − r) ∗ Th f (k − r), 125 1.625

2 (18)
y(k − r) ∗ mh f (k − r), N pp (k − r) ,

mhf [kg/s]
Thf [ºC]

N pp (k − r) ∗ Th f (k − r), N pp (k − r) ∗ mh f (k − r), 110 1.25

Th f (k − r)2 , Th f (k − r) ∗ mh f (k − r), mh f (k − r)2 ]

where r is the regressor which defines how many variables
95 0.875
we look in the past, for example r = 2 represents y(k − r) ≡
[y(k − 1), y(k − 2)]; in this example r = 4 and k = 1, 2, . . . , L.
Using the regularization parameters β = 20, γ = 1e − 6 80 0.5
and ε = 1 ◦ C, the identified model for ∆Tsh requires of 17 200 400 600 800 1000 1200
Time [s]
1400 1600 1800 2000

active basis functions. Simulations using the validation data

are performed for both linear and nonlinear models (Fig. 2). Fig. 3. Temperature and mass flow rate of the defined heat source.
An accurate identification of low superheating values is still
a challenging task, modeling errors can be attributed to the
choice of the fixed basis functions, these results suggest that
other type of nonlinearities should be considered. The nor- 100 rpm) and constraints at the process output (∆Tsh,min =
malized root-mean-square error (RMSE,[20]) of each model 10 ◦ C). Both linear and nonlinear controllers are tuned for
to the validation data is computed, reporting that a fitting horizons N1 = 1 and N2 = 10.
percentage of 48% and 72% for the linear and nonlinear
model, respectively. Nevertheless, from control point of view, The performance of the control strategies is depicted in
this is not a problem as still a robust controller can be tuned Fig. 4, both linear and nonlinear controllers are able to
to deal with modeling errors or unmodeled dynamics. track the reference, however the linear controller presents
oscillations at low superheating values due to large modeling
C. Application to constrained predictive control errors. The Integral Absolute Error (IAE) is used as criteria to
Two predictive controllers are implemented, a EPSAC- quantify the control performance for tracking the reference,
MPC based on the linear parametric model (17) and a reporting as result 972.76 and 750.84, for the linear and
NEPSAC-NMPC based on the nonlinear polynomial model. nonlinear control, respectively. The control effort is similar
The controllers are tested in simulation using a heat source in magnitude for both controllers, except for the observed
profile which could be typically observed in industrial waste oscillations, thus suggesting that either the region/conditions
heat [13], as for example from exhaust gases from a reheat where the controller has been tested presents a weak nonlin-
furnace, as depicted in Fig. 3. earity or that the nonlinear model is still not capturing the
The performance of the controllers is assessed during main nonlinearity present in the system. From a practical
a tracking test, where it is expected to have the fastest point of view, the nonlinear controller achieves a smoother
response with minimum overshoot, while satisfying actuator control effort which might be beneficial for the actuator (i.e.,
constraints (N pp,min = 1320 rpm; N pp,max = 2100 rpm; ∆N pp = enlarging the pump life).

1016
 25
Next Generation Organic Rankine Cycles (www.orcnext.be),
20 funded by the Institute for the Promotion and Innovation by
Science and Technology in Flanders. This financial support
∆ T (°C)
15
10 is gratefully acknowledged
5 MPC NMPC SetPoint R EFERENCES
0
0 500 1000 1500 2000 [1] S. Qin and T. Badgwell., “A survey of industrial model predictive
Time (s)
control technology.” Control engineering practice, vol. 11, pp. 733–
2200 764, 2003.
2000
[2] E. F. Camacho and C. Bordons., Model Predictive Control, 2nd ed.
Npp (rpm)

Springer-Verlag, London,, 2004, vol. 405 pages.

1800 [3] S. A. Billings, Nonlinear System Identification: NARMAX Methods in
1600 the Time, Frequency, and Spatio-Temporal Domains. Wiley, 2013.
1400 [4] N. Chiras, C. Evans, and D. Rees, “Nonlinear gas turbine modeling
1200
using narmax structures,” IEEE Tans. Instrum. Meas., vol. 50, pp.
0 500 1000 1500 2000 893–898, 2001.
Time (s) [5] O. Boaghe, S. Billings, L. Li, P. Fleming, and J. Liu, “Time and
frequency domain identification and analysis of a gas turbine engine,”
Control engineering practice, vol. 10, pp. 1347–1356, 2002.
Fig. 4. Control performance of MPC and NMPC, for waste heat recovery [6] H. Zou and T. Hastie, “Regularization and variable selection via the
application using an Organic Rankine Cycle. elastic net,” Journal of the Royal Statistical Society, Series B, vol. 67,
pp. 301–320, 2005.
[7] R. Tibshirani, “Regression shrinkage and selection via the lasso,”
V. C ONCLUSIONS Journal of the Royal Statistical Society, Series B, vol. 58, pp. 267–288,
1994.
In the present contribution, a relaxed sparse identification [8] C. Novara, “Sparse identification of nonlinear functions and paramet-
ric set membership optimal analysis.” in in proc. American control
algorithm is presented in order to identify a parametric model conference, San Francisco, USA, 2011.
of a nonlinear system, with prescribed modeling error. The [9] M. Bonin, V. Seghezza, and L. Piroddi, “Narx model selection based
augmentation of the basic ordinary least square solution on simulation error minimisation and lasso,” IET control Theory and
Applications, vol. 4(7), pp. 1157–1168, 2010.
with a `1 penalty allows to identify the structure of the [10] R. De Keyser, Model based Predictive Control for Linear Systems.
model and to enhance sparsity, while the `2 penalty ensures Invited chapter in UNESCO EoLSS. Oxford (6.43.16.1), 2003.
obtaining an unique solution. Sparse polynomial models [11] C. Sprouse and C. Depcik, “Review of organic Rankine cycles for
internal combustion engine exhaust waste heat recovery,” Applied
compared to other black-box identification techniques, for Thermal Engineering, vol. 51, no. 1-2, pp. 711–722, mar 2013.
example Neuronal Networks, offer the benefit that a certain [12] A. Desideri, B. Dechesne, J. Wronski, M. Van Den Broek, S. Gu-
degree of understanding about the type (e.g., square, cubic, sev, V. Lemort, and S. Quoilin, “Comparison of moving boundary
and finite-volume heat exchanger models in the modelica language,”
etc) and location (e.g., input, output, etc) of nonlinearity submitted to energies, pp. 1–17, 2016.
present in the system can be obtained. The simulation [13] S. Quoilin, R. Aumann, A. Grill, A. Schuster, and V. Lemort., “Dy-
example provided in this paper is of industrial relevance and namic modeling and optimal control strategy for waste heat recovery
organic rankine cycles,” Applied Energy, vol. Vol. 88, pp. 2183–2190,
represents a challenging for modeling and control design. 2011.
The optimal performance for these systems is often achieved, [14] A. Hernandez, A. Desideri, C. Ionescu, S. Quoilin, V. Lemort, and
by proper regulation of the quality of vapor entering the R. De Keyser., “Experimental study of predictive control strategies for
optimal operation of organic rankine cycle systems.” in Proceedings
expander. The latter can be observed by computing the level of the European Control Conference (ECC15), Linz, Austria, 2015.
of superheating, which presents highly nonlinear and time- [15] J. Peralez, P. Tona, O. Lepreux, A. Sciarretta, L. Voise, P. Dufour, and
varying dynamics. Results show that it is possible to identify M. Nadri., “Improving the control performance of an organic rankine
cycle system for waste heat recovery from a heavy-duty diesel engine
a sparse polynomial model that describes the dynamics over using a model-based approach,” in In IEEE Conference on Decision
a wide operating range. and Control, Florence, Italy, 2013.
The obtained input/output polynomial model is further [16] V. Grelet, P. Dufour, M. Nadri, V. Lemort, and T. Reichel, “Explicit
multi-model predictive control of a waste heat rankine based system
used in order to construct a constrained Nonlinear Model for heavy duty trucks,” in In IEEE Conference on Decision and
Predictive Control strategy using the NEPSAC approach. Control, Osaka, Japan, 2015.
The performance of the proposed controller is tested and [17] R. De Keyser and A. Hernandez, “Evaluation of the NEPSAC non-
linear predictive control on a thermal process.” in Proceedings of the
compared to the one achieved using a linear model predictive 13th European Control Conference (ECC 2014), 2014.
controller for the studied case. The obtained results suggest [18] S. Quoilin, A. Desideri, J. Wronski, I. Bell, and V. Lemort., “Thermo-
that the NMPC strategy leads to a smoother, safer and cycle: A modelica library for the simulation of thermodynamic sys-
tems,” in Proceedings of the 10th International Modelica Conference,
more efficient operation; where similar or better tracking Lund, Sweden, 2014.
performance is obtained at a lower control effort. [19] V. Lemort, A. Zoughaib, and S. Quoilin., “Comparison of control
Future work includes identification and control of multiple strategies for waste heat recovery organic rankine cycle systems,” In
Sustainable Thermal Energy Management in the Process Industries
input multiple output systems and using the knowledge of the International Conference (SusTEM2011), 2011.
prescribed modeling error to build a robust NMPC controller. [20] L. Ljung, System identification: theory for the user. Prentice-Hall,
2007.
ACKNOWLEDGMENT
The results presented in this paper have been obtained
within the frame of the IWT SBO-110006 project The

1017