2014 American Control Conference (ACC) June 4-6, 2014. Portland, Oregon, USA
Stochastic Nonlinear Model Predictive Control with Probabilistic Constraints Ali Mesbah1 , Stefan Streif1,2 , Rolf Findeisen2 , and Richard D. Braatz1 Abstract— Stochastic uncertainties are ubiquitous in complex dynamical systems and can lead to undesired variability of system outputs and, therefore, a notable degradation of closedloop performance. This paper investigates model predictive control of nonlinear dynamical systems subject to probabilistic parametric uncertainties. A nonlinear model predictive control framework is presented for control of the probability distribution of system states while ensuring the satisfaction of constraints with some desired probability levels. To obtain a computationally tractable formulation for real control applications, polynomial chaos expansions are utilized to propagate the probabilistic parametric uncertainties through the system model. The paper considers individual probabilistic constraints, which are converted explicitly into convex second-order cone constraints for a general class of probability distributions. An algorithm is presented for receding horizon implementation of the finite-horizon stochastic optimal control problem. The capability of the stochastic model predictive control approach in terms of shaping the probability distribution of system states and fulfilling state constraints in a stochastic setting is demonstrated for optimal control of polymorphic transformation in batch crystallization.
I. I NTRODUCTION Parametric uncertainties and exogenous disturbances are ubiquitous in complex dynamical systems and render the process stochastic. Deterministic robust model predictive control (MPC) approaches have been investigated to address the receding horizon optimal control problem for uncertain systems (e.g., see [1], [2] and the citations therein). If the uncertainties are considered to be in bounded sets, robust MPC formulations enable analyzing the stability and performance of the system against worst-case perturbations. This has led to a number of applications, where providing guarantees on robust stability and performance is the main objective [1], [2]. However, the worst-case perturbations may have a vanishingly small probability of occurrence in real practice [3], so control design based on worst-case uncertainties can be too conservative and, therefore, may lead to infeasible control problems. In addition, it is seldom practical to specify precisely the uncertainty bounds for complex dynamical systems. If the actual uncertainty realizations are larger than those assumed a priori, the guarantees on robust stability and performance will be diminished; conversely, exceedingly Financial support is acknowledged from Novartis Pharma AG, and from the Research Center Dynamic Systems, Saxony-Anhalt, Germany. 1 Massachusetts Institute of Technology, 77 Massachusetts Avenue, Cambridge, MA 02139, USA. 2 Otto-von-Guericke University Magdeburg, Universit¨ atsplatz 2, 39108 Magdeburg, Germany. Emails:
[email protected],
[email protected],
[email protected], and
[email protected].
978-1-4799-3274-0/$31.00 ©2014 AACC
large uncertainty bounds will lead to loss in the achievable closed-loop performance. In contrast to deterministic robust MPC, stochastic MPC (SMPC) with probabilistic constraints (aka chance constraints) takes a different approach to solve MPC problems under uncertainties. SMPC utilizes probabilistic descriptions of uncertainties and constraint violations and allows accounting for acceptable levels of risk during system operation. SMPC with probabilistic constraints is intended to control the predicted probability distributions of system states in some optimal manner over a finite prediction horizon, while ensuring the satisfaction of constraints with a desired probability level (e.g., see [4], [5], [6], [7], [8], [9], [10], [11], [12]). Probabilistic constraints enable the SMPC formulation to directly incorporate the tradeoffs between the satisfaction of constraints and closed-loop performance and, therefore, alleviate the inherent conservatism of worst-case MPC. In addition, the probabilistic framework of SMPC allows the shaping of the probability distributions of system states (e.g., rather than merely optimizing their expected value and/or variance), which can have significant safety and economic implications for the operation of complex dynamical processes. SMPC formulations are particularly advantageous when the control cost functions are asymmetric and the highperformance system operation is realized in the vicinity of constraints in a stochastic setting [13]. The propagation of probabilistic parameter uncertainties and exogenous disturbances through the system model and the reformulation of probabilistic constraints to computationally tractable expressions are key issues in SMPC for real applications. A common approach for probabilistic analysis and control of uncertain systems is to use samplingbased techniques, which solve convex chance constrained optimization problems (e.g., [14], [15], [16], [17]). However, these approaches can be prohibitively expensive for realtime control due to the large number of samples required for uncertainty propagation. Additional complexity in SMPC arises from probabilistic constraints, which involve the computation of multivariate integrals. In [4], [5], [11], [18], the probabilistic constrains are transformed into deterministic expressions to facilitate solution of the optimal control problem. However, these analytic approximations are restrictive, as they hold only for certain classes of probability distributions and are limited to convex problems. The restriction with respect to special distributions is alleviated in sampling-based methods, where so-called scenarios are used to formulate an optimization problem that replaces the probabilistic constrained control
2413
problem [3], [19], [20], [21], [22]. These methods can, however, result in conservative formulations unless special sampling techniques [9], which may require a prohibitive number of samples, are employed. Recently, [23] proposed an adaptive SMPC scheme that starts with a conservative probabilistic constrained formulation and, subsequently, adapts the constraints formulation online based on the experienced constraint violation frequencies. Guarantees of convergence to the desired level of constraints violation are established for a special class of linear systems. In this paper, the generalized polynomial chaos (PC) theory is employed to devise a stochastic MPC formulation for a general class of nonlinear systems that are subject to parametric uncertainties. The PC theory [24], [25], [26] provides computationally efficient spectral tools to replace or accelerate Monte Carlo-based uncertainty analysis methods. In the PC approach, the implicit mappings between the uncertain variables/parameters and the states (defined in terms of differential equations) are replaced with explicit functions in the form of a series of orthogonal polynomials, whose statistical moments can be readily computed from the expansion coefficients. This approach is distinct from past works on the use of PC expansions for probabilistic simulation and controller synthesis of uncertain systems [10], [11], [13], [27], [28], [29] in terms of the general framework proposed for stochastic nonlinear MPC with probabilistic constraints. The remainder of the paper is structured as follows. In Sec. II, the stochastic nonlinear optimal control problem is formulated to design the probability distributions of the system states, while satisfying probabilistic constraints on states. The PC expansion is introduced in Sec. III. An algorithm is presented in Sec. IV to facilitate the receding horizon implementation of the SMPC by updating the coefficients of PC expansions online at discrete time instants. To obtain a computationally tractable formulation, individual probabilistic constraints are converted explicitly into convex second-order cone constraints for a general class of probability distributions with known mean and covariance [30]. In Sec. V, the stochastic nonlinear MPC approach with probabilistic constraints is applied for optimal control of polymorphic transformation in batch crystallization, which is a process that most commonly occurs in the pharmaceutical industry. II. P ROBLEM F ORMULATION Consider an uncertain, discrete-time nonlinear system xt+1 = f (xt , ut , θ)
(1)
where t is the discrete time, xt := [x1 (t), · · · , xnx (t)]> ∈ Rnx denotes the system states, ut ∈ Rnu denotes the control inputs, θ ∈ Rn denotes the uncertain system parameters, and f : Rnx × Rnu × Rn → Rnx denotes the nonlinear system dynamics. A probability triple (Ω, F, P) is defined on the basis of sample space Ω, σ-algebra F, and probability measure P on (Ω, F). The parameter vector θ is composed of independent distributed random variables θi , with known
probability distribution functions (pdfs) fθi , such that θi ∈ L2 (Ω, F, P) ∀i ∈ {1, · · · , n}. Here, L2 (Ω, F, P) is the Hilbert space of all random variables θi , whose L2 -norm is finite. The expected value (first-order moment) of a stochastic R variable ψ : Ω → R is denoted as E[ψ] := Ω ψdfψ , where fψ is the pdf of ψ over its support Ω. The variance (central second-order moment) of a stochastic variable ψ is represented by Var[ψ] := E[(ψ − E[ψ])2 ]. The solution of (1) is assumed to exist and to be unique with probability one for any initial condition x0 and any feasible input. Under the uncertainties in the system parameters, the solution trajectories of system (1) are stochastic processes. The goal of controller synthesis is to shape the probability distributions of x to possess desirable statistics. Assuming that the system states can be measured at all times, the finite-horizon SMPC problem with probabilistic constraints is stated as follows. Problem 1
(Finite-horizon stochastic MPC with probabilistic constraints):
min J(x(k), uN ) uN
¯ t+1 = f (¯ subject to: x xt , ut , θ), h(¯ xt , ut , θ) ≤ 0, Pr[g(¯ xt ) ≤ 0] ≥ β, ut ∈ U, ¯ 0 = x(k) x
∀t = 0, . . . , N ∀t = 0, . . . , N ∀t = 0, . . . , N ∀t = 0, . . . , N
(2)
where uN := [u0 , . . . , uN ]> denotes the decision variables ¯ t denotes the states predicted by the (control policy), x nonlinear system model, x(k) denotes the measured states at discrete time instant k, h : Rnx × Rnu × Rn → Rnh denotes the hard inequality constraints, Pr denotes probability, g : Rnx → Rng denotes the joint probabilistic constraints on the states, β ∈ (0, 1) ⊂ R denotes the lower bound of the desired joint probability that the state constraints should satisfy under uncertainties, and U ⊂ Rnu denotes the convex compact set of input constraints. For the nonlinear system (1), Problem 1 considers joint probabilistic constraints on the states, hard constraints on the inputs, and general inequality constraints. In (2), the cost function J(x(t), uN ) can be defined in terms of some statistics or moments of the pdfs of the states x or, alternatively, in terms of some metrics (e.g., Wasserstein distance [31], [32]) on the states’ pdfs when explicitly available. Using the latter in the cost function, the difference between the shapes of states’ pdfs and their desired pdfs can be quantified, which enables the shaping of the probability distributions of x. Sampling from the uncertain parameter distributions and, subsequently, simulating the system (1) can be used to derive the pdfs of states. To render the online solution of (2) computationally feasible, it is proposed to use PC expansions to propagate the parametric uncertainties through (1). Next, the main principles of PC expansions are introduced. III. P OLYNOMIAL C HAOS E XPANSIONS Polynomial chaos expansions provide means for the approximation of stochastic variables with finite second-
2414
order moments [24], [33]. A stochastic variable ψ(θ) ∈ L2 (Ω, F, P) has the L2 -convergent expansion [26] ψ(θ) =
∞ X
ak Φαk (θ)
(3)
k=0
where Qn ak denotes the expansion coefficients and Φαk (θ) := i=1 Φαi,k (θi ) denotes the multivariate polynomials with Φαi,k being univariate polynomials in θi of degree αi,k . The choice of the univariate polynomials Φαi,k is made in accordance with the distribution of random variables θi (e.g., Hermite polynomials are utilized for Gaussian random variables) [25]. R The polynomials should satisfy the orthogonality property Ω Φαi Φαj dfθ = h2i δij , where δij is theRKronecker delta and hi is a constant term that corresponds to Ω Φ2αi dfθ . For practical reasons, the PC expansion (3) needs to be truncated. The total number L of terms in the truncated expansion depends on P the number of uncertain parameters n and the order m (i.e., i=1 αi,k ≤ m ∀αk ) of expansion (e.g., see [10]) (n + m)! . (4) L= n!m! Hence, the truncated PC expansion takes the form ˆ := ψ(θ)
L−1 X
ak Φαk (θ) = a> Λ(θ)
(5)
k=0
:= := with Λ(θ) [Φα0 , · · · , ΦαL−1 ]> and a > [a0 , · · · , aL−1 ] . The PC expansion is convergent in the mean-square sense [25]. Probabilistic collocation methods are commonly used to compute the coefficients a for a nonlinear system (e.g., see [10], [13], [34], [35]). These methods essentially consist of the least-squares estimation of the coefficients from a finite data set (i.e., a set of collocation points), requiring the residuals between the PC expansion and the nonlinear model predictions to be zero at the collocation points. The selection of the collocation points can greatly affect the accuracy of the PC approximation. A common choice for collocation points is the roots of the orthogonal polynomial of one degree higher than the order of the PC expansion [36]. Once the coefficients a in (5) are computed, the orthogonality property of multivariate polynomials can be exploited to efficiently compute the statistics of the stochastic variable ˆ ψ(θ). For instance, the first- and second-order moments of ˆ ψ(θ) are defined by ˆ E[ψ(θ)] = a0 ˆ Var[ψ(θ)] =
L−1 X
a2k E[Φ2αk (θ)]
(6) (7)
k=1
where the terms E[Φ2αk (θ)] ∀k = 1, · · · , L−1 are calculated only once and offline. These moments, along with the higher order moments [28], can be used to approximate the pdf of ˆ the stochastic variable ψ(θ). Next, PC expansions are used to present a tractable formulation for Problem 1.
IV. S TOCHASTIC M ODEL P REDICTIVE C ONTROL The most computationally efficient way to solve the probabilistic control Problem 1 for online control applications is to recast (2) in terms of a deterministic optimal control problem. Theorem 1 (Distributionally robust probabilistic constraints [30]): Consider a single generic probabilistic constraint of the form Pr[l> ν ≤ 0] ≥ 1 − ,
∈ (0, 1)
(8)
where l ∈ Rnl denotes some random quantities and ν ∈ Rnl denotes some variables. Denote with L the family of all distributions with known mean ˜l and covariance Γ. For any ∈ (0, 1), the distributionally robust probabilistic constraint inf Pr[l> ν ≤ 0] ≥ 1 −
l∼L
(where l ∼ L denotes that the distribution of l belongs to the family L) is equivalent to the convex second-order cone constraint p κ Var[l> ν] + E[l> ν] ≤ 0, κ = (1 − )/ > where E[l> ν] = ˜l ν and Var[l> ν] = ν > Γν. Thm. 1 provides an explicit expression for the probabilistic constraint (8), which is enforced robustly with respect to an entire family of probability distributions on random quantities l. Assumption 1: In Problem 1, the probabilistic constraints on states are assumed to consist of merely individual constraints, each of which is a linear function of one state variable. Hence, the probabilistic constraints in (2) take the form Pr[ci x ¯i (t) + di ≤ 0] ≥ β¯i , ∀i ∈ I (9)
¯t where {¯ xi (t)}, ∀i ∈ I ⊆ {1, . . . , nx } is a subset of x ¯ and βi ∈ (0, 1) denotes user-specified lower bounds for the individual probabilistic constraints (note that the choice of β¯i is independent of β in (2)). By approximating {¯ xi (t)}, i ∈ I in terms of PC expansions (5) (i.e., x ¯i (t) ≈ x ˆi (t)), (9) is rewritten as ¯ Pr[ci a> t,i Λ(θ) + di ≤ 0] ≥ βi ,
∀i ∈ I
(10)
where at,i denotes the coefficients of PC expansions corresponding to x ¯i (t) at different times. Thm. 1 can now be used to convert the individual probabilistic constraints (10) into deterministic constraints ci κ1−β¯i Var x ˆi (t) + E x ˆi (t) + di ≤ 0, ∀i ∈ I (11) where E[ˆ xi ] and Var[ˆ xi ], ∀i ∈ I are computed using (6) and (7), respectively. Using PC expansions to propagate the parametric uncertainties through the nonlinear model (1) and explicit expressions (11) to replace the probabilistic constraints, a deterministic formulation for the finite-horizon stochastic MPC Problem 1 is presented as follows.
2415
Problem 2
TABLE I K INETIC EXPRESSIONS FOR BATCH POLYMORPHIC CRYSTALLIZATION OF L- GLUTAMIC ACID [39]
(Deterministic surrogate for stochastic MPC with probabilistic constraints):
min J(x(k), uN )
α-form crystal nucleation rate
uN
subject to: x ˆi (t) = a> t,i Λ(θ), h(ˆ xnom (t), ut , θ) ≤ 0, i ˆi (t) + ci κ1−β¯i Var x E x ˆi (t) + di ≤ 0, ut ∈ U, ¯ 0 = x(k). x
∀t = 0, · · · , N ∀i = 1, . . . , nx ∀t = 0, · · · , N ∀i = 1, . . . , nx (12) ∀t = 0, · · · , N ∀i ∈ I ∀t = 0, · · · , N
In the optimal control problem (12), the coefficients of the PC expansions (i.e., at,i ) depend on the control inputs. To solve (12) online, the coefficients at,i , ∀i = 1, . . . , nx are recomputed at each discrete time instant using the new state measurements x(k) and the decision variables uN (see Algorithm 1). Note that the hard inequality constraints on states are enforced only at the nominal values of the states x ˆnom (t). For online implementations, Problem 2 is i embedded in a receding horizon algorithm. Algorithm 1. (Receding horizon implementation of PCbased MPC). Input:
Initial time step k Feasible initial control inputs uk Initial states x(k) Uncertainty description of parameter vector θ 5) Type and order of orthogonal polynomials 6) Number of collocation points s
1) 2) 3) 4)
At discrete time instant k: ¯ 0 = x(k) and uk to carry out 1) Use x s simulations of the system (1) using s samples drawn from the known pdfs of θ 2) Use simulation data from Step 1 to estimate the coefficients of the PC expansions along the prediction horizon 3) Compute the statistics of the ˆt PC-approximated states x 4) Solve the deterministic optimal control problem (12) to determine the optimal control policy. For each optimization iteration, this requires repeated simulation of the PC expansion and of the statistics as in Steps 1 to 3 5) Apply the first element u0 to the system (start from Step 1 at each time instant k).
Next, stochastic nonlinear model predictive control of polymorphic transformation in a batch crystallization system is investigated. V. S TOCHASTIC C ONTROL OF P OLYMORPHIC T RANSFORMATIONS IN BATCH C RYSTALLIZATION Polymorphism, in which multiple crystal forms exist for a chemical compound, is of paramount importance in the spe-
α-form crystal growth/dissolution rate
β-form crystal nucleation rate
Bα = kbα (Sα − 1)µα,3 kgα (Sα − 1)gα if Sα ≥ 1 Gα = kdα (Sα − 1) otherwise −E kgα = kgα,0 exp 8.314(Tgα +273) Bβ = kbβ,1 (Sβ − 1)µα,3 + kbβ,2 (Sβ − 1)µβ,3 −kgβ,2
Gβ = kgβ,1 (Sβ − 1)gβ exp
β-form crystal growth rate
kgβ,1 = kgβ,0 exp
Sβ −1 −Egβ 8.314(T +273)
cialty chemical and pharmaceutical industries [37]. Polymorphic transformations lead to variations in physical properties (e.g., shape, solubility, and chemical reactivity) of crystals, and, therefore, can be detrimental to their performance. The problem of controlling polymorphic transformations consists of ensuring consistent production of the desired polymorph in a stochastic environment [38]. The batch polymorphic crystallization of L-glutamic acid is investigated using the kinetic model developed in [39] for polymorphic transformations of metastable α-form and stable β-form L-glutamic acid crystals. Under operating conditions that promote nucleation and growth as the prevalent kinetic phenomena, the batch crystallization dynamics are described by a population balance equation [40] ∂(Gi ni ) ∂ni + = Bi δ(L − L0 ), i = α, β (13) ∂t ∂L where ni is the crystal size distribution (CSD) of the iform crystals (#/m4 ) with # representing the number of crystals, Bi and Gi are the nucleation (#/m3 s) and growth (m/s) rates, respectively, L and L0 are the characteristic size of crystals and nuclei, respectively (m), and δ(·) is a Dirac delta function. Equation (13) is coupled with the solute mass balance 3000 dC ρα kvα Gα µα,2 + ρβ kv,β Gβ µβ,2 =− dt ρs where C is the solute concentration (g/kg), ρs is the solvent density (kg/m3 ), ρi is the density of the i-form crystals (kg/m3 ), kvi is the volumetric shape factor (dimensionless), and µi,2 is the second moment (#/m) of the crystal size distribution defined by Z ∞ Lj ni dL, i = α, β j = 2, 3. µi,j = 0
The kinetic expressions for nucleation and growth/dissolution rates of the α- and β-form crystals are listed in Table I, where Si = C/Csat,i is the supersaturation, Csat,i (T ) is the saturation concentration (g/kg), and T is the solution temperature (◦ C). The initial conditions (i.e., seed properties), the values of physical properties, and the kinetic parameters and their respective 95% credible intervals are taken from [39]. The method of moments [40] is applied to the population balance equation (13), yielding a set of nonlinear differential algebraic equations to describe the system dynamics. The kinetic parameters gα , Egα , kbβ,2 , kgβ,2 , and Egβ (see Table I) are considered to be Gaussian random variables, whose
2416
Solute Concentration (g/kg)
35 30 25 20 PCE−20 collocation points PCE−40 collocation points PCE−60 collocation points Nonlinear model
15 10 0
2000
4000
6000 Time (s)
8000
10000
12000
Fig. 1. Comparison between solute concentration profiles predicted by the nonlinear model and third-order PC expansions.
statistics are characterized using the parameter estimation results in [39]. The crystallization control objective is to minimize the ratio of the nucleated crystal mass to the seed crystal mass of β-form crystals at the end of the batch ζ(tf ) =
µβ,3 (tf ) − µseeds β,3 µseeds β,3
(14)
is the thirdwhere tf is the end time of the batch and µseeds β,3 moment of the seed size distribution. The optimal control problem should prevent the nucleation and growth of α-form crystals and the dissolution of β-form crystals during polymorphic transformation. The control objective should be realized in a stochastic environment due to the kinetic parametric uncertainties. A stochastic nonlinear model predictive control approach is designed to solve a finite-horizon stochastic control problem at every discrete time instant
PC expansions at the discrete time instants (i.e., every 200 s). The SMPC is implemented in Matlab based on the sequential optimization strategy, in which a stiff ODE integrator is exploited in combination with the optimization subroutine fmincon. Fig. 1 shows the solute concentration profiles predicted by the nonlinear model and PC expansions, whose coefficients are determined using different numbers of collocation points. Using 20 and 40 collocation points resulted in inadequate estimation of the coefficients of the PC expansions and poor approximation of the solute concentration profile. On the other hand, when 60 points are used to estimate a in (5), the PC expansions provide accurate approximations of the system dynamics. To evaluate the performance of the control approach, Monte Carlo simulations of the closed-loop system are performed using 100 random parameter vectors generated from the Gaussian pdfs of the kinetic parameters (the parameters are used for simulating the process model to which the optimal control inputs are applied). The performance of the SMPC approach is compared to that of a nonlinear model predictive control (NMPC) approach with the same performance index (14). Fig. 2 shows the histograms of the performance index based on 100 closed-loop simulations of control approaches. The NMPC and SMPC approaches result in comparable mean performance indexes of 30.87 and 29.71, respectively. However, the SMPC approach leads to a significant reduction (i.e., 26%) in the variance of the performance index in the presence of parametric uncertainties, as compared to the NMPC approach, due to the probabilistic formulation of the optimal control problem (15) that enables
min E[ζ(tf )] + wVar[ζ(tf )] T (t)
Frequency
8 6 4 2 0 28
29
30
31 (tf)
32
33
34
33
34
(a) Nonlinear model predictive control 20 15 Frequency
nonlinear model equations C(tf ) ≤ Cmax (tf ) Pr[Csat,β (t) < C(t) < Csat,α (t)] ≥ 0.95 Tmin ≤ T (t) ≤ Tmax (15) where w is a weighting coefficient on the variance of the performance index, and Tmin and Tmax are the minimum and maximum temperatures, respectively. In (15), the first inequality constraint ensures that the minimum crystallization yield required by economic considerations will be achieved; the probabilistic constraint guarantees that the solute concentration stays within the admissible region in the presence of parametric uncertainties (to avoid the formation and dissolution of α- and β-form crystals, respectively); and the hard input constraint ensures that the optimal temperature profile remains within the operating range of the crystallizer. The stochastic optimal control problem (15) is converted into a deterministic optimal control problem using polynomial chaos expansions (see Problem 2). Third-order expansions of Hermite polynomials are utilized to describe the parametric uncertainties, resulting in PC expansions with 56 terms (see expression (4)). Algorithm 1 facilitates the receding horizon implementation of the stochastic MPC, where 60 collocation points are used to compute the coefficients of the
10
subject to:
10 5 0 28
29
30
31 (tf)
32
(b) Stochastic nonlinear model predictive control Fig. 2. Histograms of the performance index ζ(tf ) based on 100 closedloop simulations.
2417
Solute Concentration (g/kg)
30
propagation of parametric uncertainties through the system model and the reformulation of probabilistic constraints to computationally tractable expressions. To obtain a deterministic surrogate for the stochastic MPC problem, polynomial chaos expansions are used to propagate the parametric uncertainties and the individual probabilistic constraints are transformed into convex second-order cone constraints that are valid for a general class of probability distributions with known mean and covariance. The application of the stochastic control approach to batch polymorphic crystallization reveals its ability in shaping the probability distribution of system states, as well as ensuring the fulfillment of state constraints in a stochastic setting. It is known that the prediction quality of PC expansions worsens for large times [28]. In the presented receding horizon stochastic MPC approach, the PC expansions are recomputed at discrete time instants (using measurements), which improves the prediction accuracy. A future research topic is to include adaptive polynomial chaos approaches to further enhance the long-time prediction quality of PC expansions.
25
20
Solubility Curve of −form Crystals
15
Solubility Curve of −form Crystals
27.5 27 26.5 26
10 36
38
40
49.8
50
42 44 46 Temperature (C)
50.2
48
50
52
(a) Nonlinear model predictive control
Solute Concentration (g/kg)
30
25
20
Solubility Curve of −form Crystals
15
Solubility Curve of −form Crystals
26.5
10 36
26 25.5 25 48.4
38
40
48.6
42 44 46 Temperature (C)
48.8
49
48
49.2
50
52
(b) Stochastic nonlinear model predictive control Fig. 3. Solute concentration trajectories for 100 closed-loop simulations. The insets show a magnification of the α-crystals solubility curve to illustrate constraint violations.
shaping the distribution of the performance index (in terms of the first- and second-order moments) in a stochastic environment. Fig. 3 depicts the solute concentration trajectories, which should be constrained between the solubility curves of α- and β-form crystals. Fig. 3a indicates that in the NMPC approach the solute concentration may violate its upper bound (the solubility curve of α-form crystals) with a likelihood of 39% in the presence of parametric uncertainties. As shown in the inset figure, the constraint violation occurs in the initial phase of the batch process, where the abrupt dissolution of the α-form seeds upon introduction into the crystallizer leads to rapid supersaturation generation that is subsequently consumed by the growth of the β-form crystals. On the other hand, the SMPC approach ensures that the solute concentration remains within its admissible region in an uncertain setting by inclusion of the probabilistic constraint in (15). Fig. 3b shows that the probabilistic constraint is fulfilled in 99% of closed-loop simulations, which is greater than the pre-specified probability of 95%. The simulation results of batch polymorphic crystallization clearly indicate the capability of the SMPC approach with probabilistic constraints to control the pdfs (in terms of their moments) of system states, while guaranteeing the fulfillment of state constraints in the presence of stochastic uncertainties. VI. C ONCLUSIONS AND F UTURE W ORK A framework for stochastic MPC of nonlinear systems in the presence of parametric uncertainties is presented. The key challenges in stochastic MPC of real applications are the
R EFERENCES [1] A. Bemporad and M. Morari, “Robust model predictive control: A survey,” in Robustness in Identification and Control (A. Garulli and A. Tesi, eds.), pp. 207–226, Springer Berlin, 1999. [2] D. Limon, T. Alamo, D. M. Raimondo, J. M. Bravo, D. M. de la Pena, A. Ferramosca, and E. F. Camacho, “Input-to-state stability: A unifying framework for robust model predictive control,” in Nonlinear Model Predictive Control, pp. 1–26, Springer-Verlag, Berlin, 2009. [3] S. Garatti and M. C. Campi, “Modulating robustness in control design: Principles and algorithms,” IEEE Control Systems, vol. 33, pp. 36–51, 2013. [4] A. Schwarm and M. Nikolaou, “Chance-constrained model predictive control,” AIChE Journal, vol. 45, pp. 1743–1752, 1999. [5] D. H. V. Hessem and O. H. Bosgra, “A full solution to the constrained stochastic closed-loop MPC problem via state and innovations feedback and its receding horizon implementation,” in Proceedings of the 42nd IEEE Conference on Decision and Control, pp. 929–934, Maui, 2003. [6] J. Primbs and C. Sung, “Stochastic receding horizon control of constrained linear systems with state and control multiplicative noise,” IEEE Transactions on Automatic Control, vol. 54, pp. 221–230, 2009. [7] D. Bernardini and A. Bemporad, “Scenario-based model predictive control of stochastic constrained linear systems,” in Proceedings of the 48th IEEE Conference on Decision and Control, pp. 6333–6338, Shanghai, 2009. [8] M. Cannon, B. Kouvaritakis, S. Rakovi´c, and Q. Cheng, “Stochastic tubes in model predictive control with probabilistic constraints,” IEEE Transactions on Automatic Control, vol. 56, pp. 194–200, 2011. [9] L. Blackmore, M. Ono, A. Bektassov, and B. C. Williams, “A probabilistic particle-control approximation of chance-constrained stochastic predictive control,” IEEE Transactions on Robotics, vol. 26, pp. 502– 517, 2010. [10] L. Fagiano and M. Khammash, “Nonlinear stochastic model predictive control via regularized polynomial chaos expansions,” in Proceedings of the 51st IEEE Conference on Decision and Control, pp. 142–147, Maui, 2012. [11] K. K. Kim and R. D. Braatz, “Generalized polynomial chaos expansion approaches to approximate stochastic receding horizon control with applications to probabilistic collision checking and avoidance,” in Proceedings of the IEEE International Conference on Control Applications, pp. 350–355, Dubrovnik, 2012. [12] T. Huschto and S. Sager, “Stochastic optimal control in the perspective of the Wiener chaos,” in Proceedings of the European Control Conference, pp. 3059–3064, Zurich, 2013.
2418
[13] Z. K. Nagy and R. D. Braatz, “Distributional uncertainty analysis using power series and polynomial chaos expansions,” Journal of Process Control, vol. 17, pp. 229–240, 2007. [14] M. Vidyasagar, “Randomized algorithms for robust controller synthesis using statistical learning theory,” Automatica, vol. 37, pp. 1515– 1528, 2000. [15] A. Shapiro, “Stochastic programming approach to optimization under uncertainty,” Mathematical Programming, vol. 112, pp. 183–220, 2008. [16] G. C. Calafiore and L. Fagiano, “Robust model predictive control via scenario optimization,” IEEE Transactions on Automatic Control, vol. 58, pp. 219–224, 2013. [17] X. Zhang, K. Margellos, P. Goulart, and J. Lygeros, “Stochastic model predictive control using a combination of randomized and robust optimization,” in Proceedings of the 52nd IEEE Conference on Decision and Control, pp. 7740–7745, Florence, 2013. [18] P. Li, M. Wendt, and G. Wozny, “A probabilistically constrained model predictive controller,” Automatica, vol. 38, pp. 1171–1176, 2002. [19] G. Calafiore and M. Campi, “The scenario approach to robust control design,” IEEE Transactions on Automatic Control, vol. 51, pp. 742– 753, 2006. [20] M. C. Campi and S. Garatti, “The exact feasibility of randomized solutions of uncertain convex programs,” SIAM Journal on Optimization, vol. 19, pp. 1211–1230, 2008. [21] G. Calafiore, “Random convex programs,” SIAM Journal on Optimization, vol. 20, pp. 3427–3464, 2010. [22] G. Schildbach, G. C. Calafiore, L. Fagiano, and M. Morari, “Randomized model predictive control for stochastic linear systems,” in Proceedings of the American Control Conference, pp. 417–422, Montreal, 2012. [23] F. Oldewurtel, D. Sturzenegger, P. M. Esfahani, G. Andersson, M. Morari, and J. Lygeros, “Adaptively constrined stochastic model predictive control for closed-loop constraint satisfaction,” in Proceedings of the American Control Conference, pp. 4681–4688, Washington, 2013. [24] N. Wiener, “The homogeneous chaos,” American Journal of Mathematics, vol. 60, pp. 897–936, 1938. [25] R. Ghanem and P. Spanos, Stochastic finite elements - A spectral approach. Springer-Verlag, New York, 1991. [26] D. Xiu and G. E. Karniadakis, “The Wiener-Askey polynomial chaos for stochastic differential equations,” SIAM Journal of Scientific Computation, vol. 24, pp. 619–644, 2002. [27] F. S. Hover and M. S. Triantafyllou, “Application of polynomial chaos in stability and control,” Automatica, vol. 42, pp. 789–795, 2006. [28] J. Fisher and R. Bhattacharya, “Linear quadratic regulation of systems with stochastic parameter uncertainties,” Automatica, vol. 45, pp. 2831–2841, 2011. [29] J. Fisher and R. Bhattacharya, “Optimal trajectory generation with probabilistic system uncertainty using polynomial chaos,” Journal of Dynamic Systems, Measurement, and Control, vol. 133, pp. 014501–1 to 6, 2011. [30] G. C. Calafiore and L. E. Ghaoui, “On distributionally robust chanceconstrained linear programs,” Journal of Optimization Theory and Application, vol. 130, pp. 1–22, 2006. [31] S. S. Vallander, “Calculation of the Wasserstein distance between distributions on the line,” Theory of Probability and its Applications, vol. 18, pp. 784–786, 1973. [32] A. Halder and R. Bhattacharya, “Model validation: A probabilistic formulation,” in Proceedings of the 50th IEEE Conference on Decision and Control, pp. 1692–1697, Orlando, 2011. [33] R. Cameron and W. Martin, “The orthogonal development of nonlinear functionals in series of Fourier-Hermite functionals,” Annals of Mathematics, vol. 48, pp. 385–392, 1947. [34] D. Xiu and J. S. Hesthaven, “High-order collocation methods for differential equations with random inputs,” SIAM Journal on Scientific Computing, vol. 27, pp. 1118–1139, 2005. [35] A. Mesbah, S. Streif, R. Findeisen, and R. D. Braatz, “Active fault diagnosis for nonlinear systems with probabilistic uncertainties,” in Proceedings of the IFAC World Congress, Cape Town, 2014. [36] M. A. Tatang, W. Pan, R. G. Prinn, and G. J. McRae, “An efficient method for parametric uncertainty analysis of numerical geophysical models,” Journal of Geophysical Research, vol. 102, pp. 925–932, 1997. [37] N. Blagden and R. Davey, “Polymorphs take shape,” Chemistry in Britain, vol. 35, pp. 44–47, 1999.
[38] M. W. Hermanto, M. S. Chiu, X. Woo, , and R. D. Braatz, “Robust optimal control of polymorphic transformation in batch crystallization,” AIChE Journal, vol. 53, pp. 2643–2650, 2007. [39] M. W. Hermanto, N. C. Kee, R. B. H. Tan, M. Chiu, and R. D. Braatz, “Robust Bayesian estimation of kinetics for the polymorphic transformation of L-Glutamic acid crystals,” AIChE Journal, vol. 54, pp. 3248–3259, 2008. [40] H. M. Hulburt and S. Katz, “Some problems in particle technology: A statistical mechanical formulation,” Chemical Engineering Science, vol. 19, pp. 555–574, 1964.
2419
