Sparse Identification of Polynomial Models

and its application to Nonlinear MPC ?

Andres Hernandez ∗ Fredy Ruiz ∗∗ Clara Ionescu ∗
Robin De Keyser ∗
∗
Faculty of Engineering and Architecture, Department of Electrical
energy, Systems and Automation, Ghent University, Belgium. (e-mail:
[email protected]).
∗∗
Department of electronics engineering, Pontificia Universidad
Javeriana, Colombia. (e-mail: [email protected])

Abstract: It is well-known that the performance of Nonlinear Model Predictive Control

(NMPC) strongly depends on the accuracy of the model used for prediction, which in an
industrial context means that one will face the challenge of building a nonlinear model of
the system. This is a time- and cost- consuming task, that could result on the development
of linear models and the subsequent implementation of linear MPC. In this paper we tackle
these challenges by using a nonlinear polynomial model obtained from a sparse identification
procedure. Since this is a hard NP-problem, we employ a relaxed algorithm to accelerate
convergence speed and to guarantee the solution is optimal and the sparsest one. The obtained
model is further used inside the nonlinear Extended Prediction Self-Adaptive control (NEPSAC)
approach to NMPC, which replaces the complex nonlinear optimization problem by a more
simple iterative quadratic programming procedure. A simulation example on a waste heat
recovery application using Organic Rankine Cycle (ORC) technology is presented to illustrate
the effectiveness of the proposed strategy.

Keywords: Nonlinear Systems Identification, Bounded Error Identification, Identification for

Control

1. INTRODUCTION fact that some of the nonlinear models and/or constraints

lead to non-convex nonlinear optimization problems that
Today process industry is experiencing an increasing pro- are relatively complex to solve Grune and Pannek (2011).
ductivity demand. This, together with tighter environmen- Interesting results have been achieved by researchers lead-
tal regulations, demanding economical considerations and ing to efficient algorithms for nonlinear MPC (NMPC)
the need of operating in an energy-efficient manner. Is as presented in the survey Diehl et al. (2009) and the
forcing industry to operate systems closer to the boundary references therein. Notice however that following the ap-
of the admissible operating region, where productivity is proaches there mentioned, one will face not just the diffi-
often maximized. culty of finding a nonlinear model, but also the restriction
that it must be in the state-space form, thus requiring of
Industry looked into the technological developments as a nonlinear observers such as moving horizon estimators Rao
solution to increase the production performance, resulting et al. (2003).
in the development and further implementation of ad-
vanced control algorithms. One of the control strategies Linear predictive controllers are often based on parametric
which has been well-accepted by industry is Model Predic- models. In this contribution we explore a logical extension
tive Control (MPC) Qin and Badgwell. (2003). MPC refers of these models by obtaining a nonlinear polynomial model
to a family of control approaches, which makes explicit use using a modified version of the sparse identification algo-
of a model of the process to optimally obtain the control rithm proposed in Novara (2011). Sparse identification can
signal by minimizing an objective function Camacho and be performed by looking for a coefficient vector of the basis
Bordons. (2004). function linear combination with a ‘small’ `0 quasi-norm,
that yields a ‘small’ prediction error on the measured data.
Although real processes are in general inherently nonlin- However, the `0 quasi-norm is a non-convex function and
ear, linear predictive controllers have been predominantly its minimization is in general an NP-hard problem. The
used, this is due to the easy to obtain linear models modification here proposed is used to accelerate conver-
compared to nonlinear ones Billings (2013). Besides the gence speed by relaxing the optimization problem using
? The results presented in this paper have been obtained within the as cost function the sum of the `1 norm of the parameter
frame of the IWT SBO-110006 project The Next Generation Organic vector and a weighted `2 norm of the prediction error. The
Rankine Cycles (www.orcnext.be), funded by the Institute for the relaxed Sparse identification algorithm allows to obtain the
Promotion and Innovation by Science and Technology in Flanders. structure of the model, then a second step is implemented
This financial support is gratefully acknowledged by identifying its coefficients using a least-squares crite-
rion. Finally an input/output polynomial nonlinear model 2.2 Sparse Identification algorithm
of the system with prescribed modeling error is obtained.
Further the obtained polynomial nonlinear model is used The sparsity of a vector is typically measured by the `0
to construct a NMPC strategy, i.e. the Nonlinear Ex- quasi-norm, defined as the number of its non-zero ele-
tended Prediction Self-Adaptive Control (NEPSAC) ap- ments. Sparse identification can be performed by look-
proach to NMPC. This controller, fully described in De ing for a coefficient vector of the basis function linear
Keyser (2003), can deal with models in different formats combination with a ‘small’ `0 quasi-norm, that yields a
(e.g. state-space, transfer functions, neuronal networks, ‘small’ prediction error  between the measured output ỹ
polynomial parametric models, etc.) as it uses the non- and the predicted output fa (x̃) = φ(x̃)a. Thus, a solution
linear model for prediction directly, taking full advantage to problem 1 could be found by solving the following
of the given nonlinear system dynamics to generate a high- optimization problem:
performance design without involving model linearization a0 = arg minn kak0
a∈R (5)
or gain scheduling for its implementation. Additionally,
the usually complex nonlinear optimization problem is in subject to kỹ − Φak2 ≤ 
the NEPSAC algorithm replaced by a more simple itera- with  the maximum desired prediction error and
tive quadratic programming procedure to solve problems ỹ =
˙ (ỹ1 , . . . , ỹL )
including constraints.  
φ1 (x̃1 ) . . . φn (x̃1 )
The content of this paper is as follows. The sparse iden- .. . . .. 
Φ= ˙  (6)

tification methodology is briefly introduced in section 2. . . .
Next the nonlinear Model Predictive Control methodology φ1 (x̃L ) . . . φn (x̃L )
is presented in section 3. In section 4 the NMPC strategy is =
˙ [φ1 (x̃) . . . φn (x̃)]
tested on a waste heat recovery application using Organic
Rankine Cycle (ORC) technology, and its performance However, the `0 is a non-convex function and therefore
compared to the one achieved using linear MPC. Finally minimization of (5) is in general an NP-hard problem. Sim-
a conclusion section summarizes the main outcome of this ilarly as proposed by Novara (2011), we propose a convex
investigation. relaxation, where an optimization problem similar to (5)
is solved, by replacing the `0 quasi-norm by a combination
2. SPARSE IDENTIFICATION OF NONLINEAR of a `1 and weighted `2 norms. The optimization problem
FUNCTIONS (5) is thus redefined as a two-step algorithm:
Identifying a sparse approximation of a function from Step 1: solve the optimization problem
a set of data, possibly corrupted by noise, is what is a1 = arg minn L−1 kak1 + β minn kak2
called sparse identification. It consists in approximating a∈R a∈R (7)
a function using a ‘few’ basis functions properly selected subject to kỹ − Φak2 ≤ 
within a ‘large’ set. More precisely, a sparse approximation with L the length of the data set, β a weighting coefficient
is a linear combination of ‘many’ basis functions, where the and  the maximum expected prediction error.
vector of linear combination coefficients is sparse, i.e. it has
only a ‘few’ non-zero elements. Step 2 : define a sub-space with the active basis of of a1
and compute the estimate a∗ of a0 using least-squares of
2.1 Problem formulation prediction error
a∗ = arg minn kỹ − Φai k2
Consider a nonlinear function f0 defined by ai ∈R
(8)
y = f0 (x) (1) subject to ai = 0, ∀i |a1i | < γ
nx
where x ∈ R , y ∈ R. Suppose that f0 is not known but with γ an additional tolerance value used for pruning, thus
a set of noise-corrupted data D = {x̃, ỹ}L k=1 is available, enhancing sparsity.
described by
ỹ = f0 (x̃k ) + nk , k = 1, 2, . . . , L (2)
3. NMPC-NEPSAC ALGORITHM
where nk is noise. Define the following parametrized func-
tion: A brief introduction to NEPSAC algorithm is presented
Xn
fa (x) = ai φi (x) = φ(x)a (3) in this section. For a detailed description the reader
i=1 is referred to De Keyser (2003) and De Keyser and
where φ(x) = [φ1 (x), φ2 (x), . . . , φn (x)], φi : X → Y are Hernandez (2014).
known basis functions, and a = (a1 , a2 , . . . , an ) ∈ Rn is an
unknown coefficient vector. 3.1 Computing the Predictions
Problem 1 : from the data set D, identify a coefficient
vector a such that Using NEPSAC algorithm, the measured process output
can be represented as:
(i) a is ‘sparse’ y(t) = x(t) + n(t) (9)
(ii) the identification error
where x(t) is the model output which represents the effect
˙ 0 − fa kp
e(fa )=kf (4) of the control input u(t) and n(t) represents the effect of
is ‘small’. the disturbances and modeling errors, all at discrete-time
index t. Model output x(t) can be described by the generic problem can be solved by a QP algorithm. In this work
system dynamic model: we make use of the Hildreth QP algorithm, to ensure
x(t) = f [x(t − 1), x(t − 2), . . . , u(t − 1), u(t − 2), . . .] (10) numerical robustness and stability of the solution as it
is particularly suitable to recover from ill-conditioned
Notice that x(t) represents here the model output, not the problems Wang (2010).
state vector. Also important is the fact that f can be either
a linear or a nonlinear function.
Furthermore, the disturbance n(t) can be modeled as 3.3 NEPSAC iterative procedure
colored noise through a filter with the transfer function:
C(q −1 ) The procedure described until now can be used to imple-
n(t) = e(t) (11) ment a linear MPC (EPSAC), if f [·] in (10) is linear, as
D(q −1 )
superposition principle holds in (13). Instead, if a non-
with e(t) uncorrelated (white) noise with zero-mean and linear model is employed, the condition is that the term
C, D monic polynomials in the backward shift operator yoptimize (t + k|t) is small enough compared to ybase (t +
q −1 . The disturbance model must be designed to achieve k|t), in order to make the superposition principle still
robustness of the control loop against unmeasured distur- valid. Such condition can be satisfied if δu(t + k|t) is close
bances and modeling errors Maciejowski. (2002). to zero, making thus ubase (t + k|t) close to the optimal
u∗ (t + k|t). To address this issue, the idea is to recursively
A fundamental step in the MPC methodology consists of compute δu(t + k|t) solving (14), within the same sample
the prediction. Using the generic process model (9), the instant, until δu(t+k|t) converges to 0. Inside the recursion
predicted values of the output are:
ubase (t+k|t) is updated each time to ubase (t+k|t)+δu(t+
y(t + k|t) = x(t + k|t) + n(t + k|t) (12) k|t).
x(t + k|t) and n(t + k|t) can be predicted by recursion of Also important is the fact that in NEPSAC the G matrix
the process model (10) and by using filtering techniques has to be recomputed at every sampling instant De Keyser
on the noise model (11), respectively De Keyser (2003). and Hernandez (2014). This algorithm results after con-
vergence to the optimal solution for the underlying non-
3.2 Optimization Procedure linear predictive control problem. The number of required
iterations depends on how far the optimal u∗ (t + k|t) is
A key element in linear MPC is to use of base (or free) away with respect to u∗ (t + k|t − 1). In quasi-steady-state
and optimizing (or forced) response concepts Camacho situations, the number of iterations is low (1 . . . 2) De
and Bordons. (2004). This is however no longer valid for Keyser and Hernandez (2014). On the other hand, during
nonlinear processes, since the superposition principle does transients the number of iterations might raise to 10.
not hold in that case. In NEPSAC this is solved as follows:
The future response can then be expressed as: 4. EXAMPLE 2: WASTE HEAT RECOVERY
APPLICATION
y(t + k|t) = ybase (t + k|t) + yoptimize (t + k|t) (13)

The two contributing factors have the following origin: This section describes the architecture and main charac-
teristics of the Organic Rankine Cycle (ORC) system used
• ybase (t+k|t) is the effect of the past inputs, the apriori for evaluating the performance of the developed control
defined future base control sequence ubase (t+k|t) and strategies. A schematic layout of the ORC system is pre-
the predicted disturbance n(t + k|t). sented in Fig. 1, where also position of the sensors are
• yoptimize (t + k|t) is the effect of the additions δu(t + included.
k|t) that are optimized and added to ubase (t + k|t),
according to δu(t + k|t) = u(t + k|t) − ubase (t + k|t).
The effect of these additions is the discrete time
convolution of ∆U = {δu(t|t), . . . , δu(t + Nu − 1|t)}
with the impulse response coefficients of the system
(G matrix), where Nu is the chosen control horizon.
The control ∆U is the solution to the following constrained
optimization problem:
N2
X 2
∆U =arg min [r(t + k|t) − y(t + k|t)]
∆U ∈ RNu
k=N
(14)
1

subject to A.∆U ≤ B
where N1 and N2 are the minimum and maximum pre-
diction horizons, Nu is the control horizon, r(t + k|t) is a
future setpoint or reference sequence. The various process
inputs and output constraints can all be expressed in terms
of ∆U , resulting in matrices A, B. Fig. 1. Schematic layout of the pilot plant available at
As the cost function (14) is decision variables ∆U is Ghent University, campus Kortrijk (Belgium)
quadratic with linear constraints, then the minimization
4.1 The Organic Rankine Cycle system where Ẇexp is the expander electrical power, Ẇpump is the
pump electrical power and Q̇in,ORC is the thermal power
The system considered in this study is the pilot plant supplied to the ORC system in the evaporator.
available at Ghent University campus Kortrijk (Belgium).
The system based on a regenerative cycle and Solkatherm 4.2 Low-order model suitable for prediction
(SES36) as working fluid, has a nominal power of 11 kWe.
The expander is originally a single screw compressor The ORC unit considered in this study has one ma-
adapted to run in expander mode. It drives an asyn- nipulated variable (the pump speed), and two outputs
chronous generator connected to the electric grid through (the evaporating temperature Tsat,ev and the superheating
a four-quadrant inverter, which allows varying the gen- ∆Tsh ). Notice that the temperature Thf and mass flow
erator rotational speed (Nexp ). During the experiments variations mhf in the heat source also influence Tsat,ev and
performed in this paper, the generator rotating speed ∆Tsh , thus becoming two measured disturbances. As result
is kept constant at 3000 rpm to simulate an installation we are interested on identifying a system consisting of 3
directly connected to the grid. The circulating pump Npp inputs (one manipulated and two measured disturbances)
is a vertical variable speed 14-stage centrifugal pump with and 2 outputs.
a maximum pressure of 14 bar and 2.2 kW nominal power.
A more detailed description of the setup and acquisition A linear parametric identification is performed in the
system can be found in Gusev et al. (2014). system using the prediction error minimization method
in the data collected from a multisine excitation signal
Starting from the bottom of the scheme (figure 1) it is Ljung (2007). The identified model presented in (19) is in
possible to recognize the liquid receiver installed at the the form of discrete-time transfer functions for a sampling
outlet of the condenser, where the fluid is collected in time of Ts = 1 s .
saturated liquid condition. From the receiver outlet, the
∆Tsh (q) −0.063q −1 + 0.059q −2
fluid is pumped through the regenerator cold side, and the = (19a)
Npp (q) 1 − 2.44q −1 + 1.955q −2 − 0.51q −3
evaporator, where it is heated up to superheated vapor,
reaching its maximum temperature at the evaporator ∆Tsh (q) 0.47q −1
= (19b)
outlet. The fluid, after being expanded in the volumetric Thf (q) 1 − 0.51q −1
machine, enters the regenerator hot side, and then it flows ∆Tsh (q) −2.98q −1 + 4.29q −2 − 1.31q −3
= (19c)
into the condenser to close the cycle. mhf (q) 1 − 1.35q −1 − 0.11q −2 + 0.46q −3

Regarding the optimal operation of the ORC unit, the Tsat,ev (q) 0.066q −1 − 0.063q −2
= (19d)
superheating and the evaporating temperature are the Npp (q) 1 − 2.42q −1 + 1.91q −2 − 0.49q −3
most relevant variables to be controlled Lemort et al. Tsat,ev (q) 0.0017q −11 − 0.0017q −12
= (19e)
(2011). The superheating is defined as: Thf (q) 1 − 3.6q −1 + 4.88q −2 − 2.95q −3 + 0.67q −4

∆Tsh = Texp,su − Tsat,ev (15) Tsat,ev (q) 2.43q −1 − 6.16q −2 + 5.33q −3 − 1.6q −4
= (19f)
mhf (q) 1 − 2.93q −1 + 3.12q −2 − 1.42q −3 + 0.23q −4
where Texp,su is the temperature measured at the inlet
of the expander and Tsat,ev the evaporating temperature, Next, we are interested in identifying a nonlinear polyno-
corresponding to the temperature at which the fluid un- mial model of the system. The following input signals have
dergoes the phase transition from saturated liquid to sat- been considered to perform the sparse identification: the
urated vapor at the fixed evaporating pressure Psat,ev . heat source conditions Thf and mhf have been simulated
as white noise filtered to a maximum band of 0.5 rad/s and
Tsat,ev = f (Psat,ev ) (16)
0.63 rad/s with amplitude variations of ± 20 ◦ C around
where f corresponds to a function that correlates the 110 ◦ C and ± 0.5 kg/s around 1.0 kg/s, respectively. The
pressure for the refrigerant SES36. pump speed has been taken as the sum of 10 sinusoids
spread over the band [0.005, 0.63] rad/s, taking values be-
Research performed on ORC technology has already es- tween 1320 rpm and 2100 rpm. A set of L = 2000 samples
tablished that to optimally operate the ORC unit, two has been generated from the ‘true’ system, considering a
conditions need to be satisfied Quoilin et al. (2011): sampling time of Ts = 1 s.
i A ‘high’ efficiency and a safe operation of the ORC In order to perform the sparse identification, a set of
unit is achieved if a minimum amount of superheating n = 57 polynomial basis functions has been considered
at expander inlet is guaranteed. and the corresponding matrix Φ = (Φ1 (x̃), . . . , ΦL (x̃)) has
ii For each heat source condition there exists an optimal been obtained according to:
evaporating temperature which maximizes the output Φk (x̃) =[φ1 (x̃), . . . , φn (x̃))]
power. =[1, y(k − r), Npp (k − r), Thf (k − r), mhf (k − r),
The main terms to assess the performance of the ORC y(k − r)2 , y(k − r) ∗ Npp (k − r), y(k − r) ∗ Thf (k − r),
system are the net output power and the cycle efficiency y(k − r) ∗ mhf (k − r), Npp (k − r)2 ,
which are defined in equation (17) and (18), respectively. Npp (k − r) ∗ Thf (k − r), Npp (k − r) ∗ mhf (k − r),

Ẇel,net = Ẇexp − Ẇpump (17) Thf (k − r)2 , Thf (k − r) ∗ mhf (k − r), mhf (k − r)2 ]
(20)
where r is the regresor which defines how many variables
Ẇel,net we look in the past, for example r = 2 represents y(k −
ηcycle = (18)
Q̇in,ORC r) ≡ [y(k − 1), y(k − 2)]; in this example r = 4 and k =
1, 2, . . . , L. Also note that in this case two multiple-input a NEPSAC-NMPC based on the nonlinear polynomial
single-output (MISO) systems have been identified, one model. The controller are tested in simulation using a
for superheating and other for evaporating temperature. heat source profile which could be typically observed in
Therefore, in (20) y(k − r) has to be replaced by each of industrial waste heat Quoilin et al. (2011), as for example
our outputs, i.e. ∆Tsh (k − r) and Tsat,ev (k − r). from exhaust gases from a reheat furnace, as depicted in
figure 4.
For the case of ∆Tsh the basis function set contains 17
functions defining the ‘true’ system equation, while for
Tsat,ev only 8 functions were required. Meaning that su- 140 2
Thf
perheating has more complex dynamics than the one of mhf
evaporating temperature. Simulations using the validation
data are performed for both linear and nonlinear models. 125 1.625

As depicted in figure 2, superheating exhibits a more com-

plex dynamics compared to the evaporating temperature

mhf [kg/s]
Thf [ºC]
for which both models present an acceptable performance 110 1.25

(figure 3). An accurate identification of low superheating

values is still a challenging task, modeling errors can be
attributed to the choice of the function basis, these results 95 0.875

suggest that other type of nonlinearities should be consid-

ered (e.g. sine, cos, etc). Nevertheless, form control point
of view, this is not a problem as still a robust controller 80
200 400 600 800 1000 1200 1400 1600 1800
0.5
2000
Time [s]
can be tuned to deal with modeling errors or unmodeled
dynamics.
Fig. 4. Temperature and mass flow rate of the defined heat
30 source.
25 Data Linear Nonlinear

20
We assume that there exists an optimizer which computes
∆ Tsh [°C]

15
the optimal evaporating temperature as a function of the
10
5
heat source conditions (Thf and mhf ). The control objec-
0
tive is thus to follow the optimal setpoint which maximizes
0 500 1000 1500 2000 the output power, while keeping the superheating above
10
a desired threshold value, in order to guarantee a safe
operation.
0
Error

The control strategy must satisfy these conditions us-

−10
ing only one degree of freedom (i.e. pump speed Npp ),
−20
while satisfying actuator constraints (Npp,min = 1320 rpm;
0 500 1000
Time [s]
1500 2000 Npp,max = 2100 rpm; ∆Npp = 100 rpm) and constraints at
the process output (∆Tsh,min = 10 ◦ C). Both controller
linear and nonlinear controllers are tuned for horizons
Fig. 2. Prediction of superheating using the identified N1 = 1 and N2 = 20.
linear and non-linear models as simulators.
110
Data Linear Nonlinear 25
105
20
∆ T (°C)

100 15
Tsat [°C]

95 10
5 MPC NMPC
90
0
85 200 400 600 800 1000 1200 1400 1600 1800 2000

80 120
0 500 1000 1500 2000
5 110
Tsat (°C)

100

0 90
Error

80

−5 70
200 400 600 800 1000 1200 1400 1600 1800 2000

−10 2000
0 500 1000 1500 2000
Npp (rpm)

Time [s] 1800

1600
Fig. 3. Prediction of evaporating temperature using the 1400
identified linear and non-linear models as simulators. 200 400 600 800 1000 1200 1400 1600 1800 2000
Time (s)

4.3 Application to constrained predictive control

Fig. 5. Control performance of MPC and NMPC, for waste
Two predictive controllers are implemented, a EPSAC- heat recovery application using an Organic Rankine
MPC based on the linear parametric model (19) and Cycle.
The performance of the control strategies is depicted in fig- REFERENCES
ure 5, where due to the large variations in the heat source
Billings, S.A. (2013). Nonlinear System Identification:
the superheating tends to undergo the imposed minimum
NARMAX Methods in the Time, Frequency, and Spatio-
superheating threshold of ∆Tsh = 10 ◦ C. Because the
Temporal Domains. Wiley.
NMPC uses a more accurate model for prediction, it is
Camacho, E.F. and Bordons., C. (2004). Model Predictive
able to detect that superheating will drop, reacting on time
Control, volume 405 pages. Springer-Verlag, London,,
ensuring that ∆Tsh will not go below 10 ◦ C. By making
2nd edition.
a zooming around 1400 s one observe that the NMPC
De Keyser, R. (2003). Model based Predictive Control for
strategy remains always above the threshold value. Instead
Linear Systems. Invited chapter in UNESCO EoLSS.
the MPC slightly crosses the threshold value, forcing it to
Oxford (6.43.16.1).
make a sudden decrease in the pump speed.
De Keyser, R. and Hernandez, A. (2014). Evaluation of
the nepsac nonlinear predictive control on a thermal
25
process. In Proceedings of the 13th European Control
20 Conference (ECC 2014).
Diehl, M., Ferreau, H., and Haverbeke, N. (2009). Efficient
∆ T (°C)

15
10 numerical methods for nonlinear mpc and moving hori-
MPC NMPC
5
zon estimation. In Nonlinear Model Predictive Control,
0
1300 1350 1400 1450 1500 volume 384, 391–417. Springer Berlin Heidelberg.
120 Grune, L. and Pannek, J. (2011). Nonlinear Model Pre-
110 dictive Control. Theory and algorithms. Springer.
Tsat (°C)

100
Gusev, S., Ziviani, D., Bell, I., De Paepe, M., and Van den
90
Broek, M. (2014). Experimental comparison of work-
80
70
ing fluids for organic rankine cycle with single-screw
1300 1350 1400 1450 1500 expander. In In 15th International Refrigeration and
2000 Air Conditioning Conference at Purdue, Proceedings.
Lemort, V., Zoughaib, A., and Quoilin., S. (2011). Com-
Npp (rpm)

1800
parison of control strategies for waste heat recovery
1600
organic rankine cycle systems. In Sustainable Thermal
1400
Energy Management in the Process Industries Interna-
1300 1350 1400 1450 1500
Time (s) tional Conference (SusTEM2011).
Ljung, L. (2007). System identification: theory for the user.
Fig. 6. Zoom on Control performance for MPC and Prentice-Hall.
NMPC. Maciejowski., J. (2002). Predictive Control: With Con-
straints. Pearson Education. Prentice Hall.
Novara, C. (2011). Sparse identification of nonlinear
Although the linear MPC represents a very good option to functions and parametric set membership optoptimal
achieve optimal and safe operation of the ORC unit. The analysis. In in proc. American control conference, San
performance can be enhanced by means of the nonlinear Francisco, USA.
MPC controller, which is able to keep superheating above Qin, S. and Badgwell., T. (2003). A survey of industrial
the threshold value by making a more accurate prediction model predictive control technology. Control engineer-
(avoiding any liquid drop in the expander) and to present ing practice, 11, 733–764.
a smoother control effort (enlarging the pump life). Quoilin, S., Aumann, R., Grill, A., Schuster, A., and
Lemort., V. (2011). Dynamic modeling and optimal
control strategy for waste heat recovery organic rankine
5. CONCLUSIONS
cycles. Applied Energy, Vol. 88, 2183–2190.
Rao, C., Rawlings, J., and Mayne, D. (2003). Constrained
In the present contribution, a relaxed sparse identification state estimation for nonlinear discrete-time systems: sta-
algorithm is presented in order to identify a parametric bility and moving horizon approximations. Automatic
model of a nonlinear system, with prescribed modeling Control, IEEE Transactions on, 48(2), 246–258. doi:
error. The obtained input/output polynomial model is 10.1109/TAC.2002.808470.
further used in order to construct a constrained Non- Wang, L. (2010). Model Predictive Control System Design
linear Model Predictive Control strategy using NEPSAC and Implementation Using MATLAB . R Advances in
approach. The performance of the proposed controller is Industrial Control. Springer.
tested and compared to the one achieved using a linear
model predictive controller for a waste heat recovery appli-
cation using ORC technology. The obtained results suggest
that the NMPC strategy leads to a smoother and safer
operation of the system, allowing to operate closer to the
boundary conditions where production is maximized.
Future work includes identification and control of multiple
input multiple output systems and using the knowledge of
the prescribed modeling error to build a robust NMPC
controller.