Documente Academic
Documente Profesional
Documente Cultură
Abstract—In this letter, a nonlinear control design is nonlinear controls need to be analytically designed and exam-
presented for systems whose input and output vectors ined. Nonlinear dynamics such as saturation arise naturally in
are saturated. The proposed synthesis includes an optimal practical applications due to safety and performance require-
tracker as the seed controller and explicitly imposes all the
relevant constraints on system input, on the rate of change ments as well as physical constraints. Control applications
in the input, and on system output by using the tools include hybrid vehicle control [1], underwater autonomous
of barrier functions, input–output feedback linearization, vehicles [2], magnetic levitation [3], and power systems [4]. In
comparison argument, Lyapunov argument, and a real-time particular, the problem of designing a real-time optimized con-
constrained optimization. The proposed design is applied trol for systems with output, input, and their rate constraints
to microgrid operation with high penetration of renew-
able generation, where net load and variable generation simultaneously has not been addressed.
must be balanced and the microgrid should be operated at Barrier functions provide an effective tool to deal with
an optimal or near-optimal performance level. Since tradi- constraints in dynamic systems. Existing results include open-
tional generation has a limited ramping rate, battery stor- loop techniques [5], barrier approximations [6], and Boolean
age devices, and demand responses become necessary to compositions [7]. Using input-output feedback linearization
effectively deal with variability of renewable generation and
to maintain frequency stability, but their capacities are also in conjunction with a comparison argument, a barrier func-
limited. It is shown that the proposed design is effective for tion method is presented in [8] to design feedback control
coordinated control of traditional generation, storage and for output constrained systems. While this result has salient
demand response so that the power system frequency is features of being analytical and using real-time optimization,
guaranteed to be within the required operational limits and it does not consider constraints on the input or its rate of
that renewable curtailment is eliminated or minimized.
change. It is well known that, as constraints on control input
Index Terms—Constrained control, optimal control, limit control authority, input constraints impede both stability
power systems. and performance and hence cannot simply be added with-
out adequate design or thorough analysis. Furthermore, none
of the existing results have considered tracking or optimal
performance.
I. I NTRODUCTION In this letter, the class of input-output saturated systems
PTIMAL control theory has been widely applied since
O the 1950s, and modern control units can perform real-
time computation of static optimization, but nonlinear optimal
with nominal linear dynamics is considered, and a nonlin-
ear optimized control design is presented. Nonlinear sta-
bility conditions are derived in closed form for the class
control problems involve dynamic optimization that are gen- of systems by using the Lyapunov direct method and by
erally too difficult to solve computationally online. Hence, utilizing the fact that the Jacobian of saturated linear dynam-
ics is piecewise constant, and the conditions can analyti-
Manuscript received June 2, 2019; accepted July 8, 2019. Date of cally be verified. Additionally, input-to-state stability is also
publication July 16, 2019; date of current version July 30, 2019. The
work of R. Harvey and Z. Qu was supported in part by the U.S. National established to ensure robustness of the closed loop system
Science Foundation under Grant ECCS-1308928, in part by the U.S. against exogenous input and modeling/prediction errors. The
Department of Energy’s Awards under Grant DE-EE0007998, Grant proposed control synthesis methodology embeds the lin-
DE-EE0007327, and Grant DE-EE0006340, and in part by the U.S.
Department of Transportation’s Award under Grant DTRT13-G-UTC51. ear quadratic optimal control into a point-wise constrained
Recommended by Senior Editor G. Cherubini. (Corresponding author: quadratic optimization formulation and uses the barrier func-
Roland Harvey.) tion method to derive algebraic state/input-dependent con-
R. Harvey and Z. Qu are with the Department of Electrical and
Computer Engineering, University of Central Florida, Orlando, FL 32816 straints for real-time optimization. Stability and optimal (or
USA (e-mail: rharvey2@knights.ucf.edu; qu@ucf.edu). near-optimal) performance is shown for the proposed control.
T. Namerikawa is with the Department of System Design The rest of this letter is organized as follows. In Section II,
Engineering, Keio University, Yokohama 223-8522, Japan (e-mail:
namerikawa@sd.keio.ac.jp). the class of saturated systems is introduced, and the control
Digital Object Identifier 10.1109/LCSYS.2019.2929159 problem is defined. In Section III, stability and robustness
2475-1456 c 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
368 IEEE CONTROL SYSTEMS LETTERS, VOL. 4, NO. 2, APRIL 2020
conditions under a general nonlinear control law are derived, • An optimal tracking control is developed for system (2)
the linear optimal tracker is introduced as the seed con- with respect to performance index
trol, the barrier function method is applied to derive the
1 1 tf T
algebraic state/input-dependent constraints, and the proposed J = ỹTf Sf ỹf + [ỹ Qỹ + uT Ru]dt, (5)
nonlinear optimized control is solved in real-time from the 2 2 t0
corresponding point-wise constrained optimization problem. In where ỹ = y − yd , yf = Cη(tf ), Sf = S(tf ) ≥ 0, with
Section IV, the proposed control is applied to the microgrid yd signifying the desired output. Square penalty matrices
control problem with high penetration of renewables. Then, Q ≥ 0 and R > 0 are symmetric and of appropriate
conclusions are drawn in Section V. dimensions. Although the nonlinear optimal control may
not be solvable in general, a Lyapunov-based design is
II. P ROBLEM F ORMULATION presented to force nonlinear system (2) to move in such
Consider the following class of dynamic systems with satu- a way that η gets into set and remains there, hence
ration on partial state (i.e., output), input and the rate of change the resulting control converges to the closed-form linear
of input: optimal solution.
⎧ • For any given set , the constraints on η are handled by
⎨ ż = Az z + Bz SAT[uz ] + Gz r, yz = Cz z, a barrier function design into which the aforementioned
u̇z = SAT[u], (1) analytical optimal tracking control law is embedded.
⎩
ηz = SAT[yz ],
where z is the original state vector, yz is the unconstrained III. N ONLINEAR AND O PTIMAL C ONTROLS
output, ηz is the constrained output, SAT[ · ] denotes a vec- This section begins with a general nonlinear control design
tor of component-wise saturation functions, uz is the control for system (2). The corresponding stability condition is derived
input vector subject to both magnitude and rate saturations, using a combination of [12, Lemma 1] and the standard
and u ∈ m is the actual dynamic control to be synthesized. Lyapunov argument [13].
Matrices Az , Bz , Cz and Gz are of appropriate dimensions.
Exogenous input r ∈ l is a function of time, and its his- A. Nonlinear Control
torical value is recorded. Although the future value of r(t) is
While saturation functions are nonlinear, their Jacobians are
typically unknown, its estimate r̂(t) could be generated using
piecewise constant. This observation leads us to use lemma 1
such technique as stochastic forecast [9], [10], [11].
whose Lyapunov stability conditions are in terms of the
By introducing the augmented state x ∈ n and the overall
Jacobian properties rather than the original nonlinear system.
constrained vector η ∈ p+m as
For the case that V is quadratic, lemma 1 can be used to
x [zT uTz ]T , η [ηzT SATT [uz ] SATT [u]]T , conclude theorem 1.
Lemma 1 [12]: Consider an error system in the form of
we can rewrite dynamics of system (1) as ė = F(e, t), and assume that its Lyapunov function V(e, t) is
positive definite and decrescent. If Jacobian matrix
e F has
ẋ = F(x, u) + Ax + Bu + Gr, y = Cη, η = H(x, u), (2) the property that
subject to η ∈ , where
t V + [
e V]T
w F(w, t)|w=x−δe ≤ −ρ(e)
= ci ≤ Hi (x, u) ≤ ci , i = 1, . . . , (p + m) (3) for some class-K function ρ(·), for all x ∈ n and for all con-
stants δ ∈ (0, 1), then the system is uniformly asymptotically
ci , ci are known lower and upper limits, stable.
⎡ ⎤ Theorem 1: Consider system (2) under control
I
A Bz 0 Gz
A= z , B= , CT = ⎣0⎦, G = , (4) u(x, t) = −R−1 BT (k(x, t) − v(r̂, t)), (6)
0 0 I 0
0
where v is a uniformly bounded function of r̂ and time. Under
y is the (unconstrained) output. Vector η consists of all vari- control (6), denote F(x, t) F(x, u(x, t)). If matrices S(t) and
ables subject to saturation, and function F(x, u) : n+m → n
contains all the saturation-induced nonlinearities that vanish in (w, t) Ṡ + S
w [F(w, t) + Aw − BR−1 BT k(w, t)]
set . Specifically, elements of F(x, u) are the nonlinear terms + {
w [F(w, t) + Aw − BR−1 BT k(w, t)]}T S
of saturated variables minus their unsaturated versions, i.e.,
F(x, u) ≡ 0 if η ∈ . It can be assumed in most applications are decrescent and positive definite for all w ∈ n , then
that ⊂ p is a compact set with 0 ∈ . system (2) under control (6) has the following properties:
The control problem solved in this letter consists of the • If r̂ = r, the state is asymptotically convergent to
following aspects: equilibrium state xe described by
• Analytical stability conditions are derived for nonlinear
control system (2), and its robustness is established in the ẋe = [F(xe , ue ) + Axe − BR−1 BT k(xe , t)]
presence of forecasting error (r − r̂). + BR−1 BT v(r, t) + Gr. (7)
HARVEY et al.: OPTIMIZED INPUT/OUTPUT-CONSTRAINED CONTROL DESIGN WITH APPLICATION TO MICROGRID OPERATION 369
• If r̂ = r, the tracking error is input-to-state stable with −1 ∂F(x, u) T
u = −R B+ λ
respect to forecast error (r − r̂) ∂u
where ue = u(xe , t). with x0 given. Comparing the above expression with con-
Proof: Under control (6), system (2) becomes trol (6) of theorem 1, we know that, as long as SAT[u] = u,
ẋ = [F(x, t) + Ax − BR−1 BT k(x, t)] + BR−1 BT v(r̂, t) + Gr. λ = k(x, t) − v(r̂, t), and control (6) is also optimal if k(x, t)
and v(r̂, t) are the solutions to the following equations:
If r̂ = r, the trajectory of equilibrium state xe is governed
by (7). Conversely, if r̂ = r, then for the state error e = x − xe , ∂k(x, t) ∂k(x, t) T
=− [F(x, u) + Ax − BR−1 BT k(x, t)]
we have the error system ∂t ∂x
T
∂y(x) T ∂F(x, u)
ė = [F(x, t) − F(xe , t) + Ae − Qy(x) − + A k(x, t), (12)
∂x ∂x
+ BR−1 BT (−k(x, t) + k(xe , t))]
dv(r̂, t) ∂k(x, t) ∂F(x, u) T
+ BR−1 BT (v(r̂, t) − v(r, t)). = A − BR−1 BT + v(r̂, t)
dt ∂x ∂x
For positive definite matrix S(t), we choose the Lyapunov ∂k(x, t) ∂y(x) T d
function V = 12 eT S(t)e. It follows that, for some δ ∈ (0, 1), − Gr + Qy ; (13)
∂x ∂x
1 T 1 1
V̇ = e Sė + ėT Se + eT Ṡe with terminal conditions k(xf , tf ) = Sf yf , v(r̂, tf ) = Sf yd .
2 2 2 Although equations (12) and (13) can only be solved numer-
1
= eT SBR−1 BT [v(r̂, t) − v(r, t)] + eT |w=x−δe e ically in the general setting, they can be solved analytically
2 whenever state η stays within set as defined in (3). By
≤ λmax (SBR−1 BT )ev(r̂, t) − v(r, t) parameterizing the solutions as k(x, t) = S(t)x(t) and recalling
1 F(x, u) = 0 for η ∈ , equations (12) and (13) reduce to the
− λmin ()e2 , (8)
2 standard linear quadratic tracker:
where λmax (·) and λmin (·) denote the maximum and minimum
eigenvalues, respectively, and from which asymptotic stability Ṡ = −SA − AT S + SBR−1 BT S − CT QC (14)
−1 T
is seen from lemma 1 and input-to-state stability can also be v̇ = [A − BR B S] v − SGr̂ + C Qy
T T d
(15)
concluded using [13]. u = −R−1 BT (Sx − v),
∗
(16)
In general, there are many choices of k(x, t) in feedback con-
trol (6). In the next subsection, an optimal control is presented where (14) and (15) are solved backwards. The closed-form
as the seed control in the proposed design. control (16) is only optimal for η ∈ . For η ∈ , a sta-
bilizing control can be designed according to theorem 1 and
B. Optimal Control for system (2) whose nonlinearities consist of saturation func-
tions. Specifically, linear optimal tracker (16) can always be
In what follows, an optimal tracking control is derived for embedded into control (6) with v(r̂, t) given by (15) together
system (2) using standard techniques [14]. With respect to cost with (14), and the nonlinear feedback control part k(x, t) will
function (5), the Hamiltonian with costate λ(t) is: be chosen to force η into set and hence achieve both stability
1 T 1 and optimality. This design, which utilizes barrier functions,
H= ỹ Qỹ + uT Ru + λT (F(x, u) + Ax + Bu + Gr̂). (9)
2 2 is the topic of the next subsection.
Then, the corresponding unconstrained optimization problem
is to minimize the following cost function: C. Barrier Function Design
tf
1 T Based on input-output feedback linearization and a com-
J = ỹf Sf ỹf + [H − λT ẋ]dt. (10) parison argument, the so-called barrier function translates
2 t0
algebraic constraints on the output (including both the states
It follows from calculus of variation [14], that the first-order variables and control u) into algebraic state-dependent inequal-
variation of the cost function is given by ities in terms of the control. By satisfying such inequalities,
tf T the control can force output vector η into set . By nature,
∂H
δJ = (Sf ỹf − λf )δx +
T
+ λ̇ δx such a control design ensures Lyapunov stability and, by mak-
t0 ∂x
T T ing an invariant set, control (6) becomes linear control (16)
∂H ∂H that achieves optimal performance within set .
+ δu + − ẋ δλ dt, (11)
∂u ∂λ Definition 1: Consider system (2) with u satisfying its con-
straints and with r = r̂ being a time-frozen value, expressed as
∂ ỹ
with λf = Sf ∂x |y=yf . Upon δJ in (11) being set to be zero, the ẋ = f (x, r̂)+Bu, with barrier function ξi (x). Then, an lth-order
following necessary conditions are obtained: Lie derivative of ξi (x) is recursively defined as
ẋ = F(x, u) + Ax + Bu + Gr̂ (1) ∂ξi (x)
T Lf ξi (x) = f (x, r̂)
∂y(x) T ∂F(x, u) ∂x
λ̇ = − Q[y(x) − y ] +
d
+A λ ..
∂x ∂x .
370 IEEE CONTROL SYSTEMS LETTERS, VOL. 4, NO. 2, APRIL 2020
(l) ∂ (l−1)
Lf ξi (x) = [L ξi (x)]f (x, r̂). The specific steps of designing an optimized nonlinear
∂x f control using barrier functions are as follows.
The barrier function ξi (x) is said to have relative degree μi if Step 1: It follows from (3) that there are a total of
∂ξi (x) ∂ (l−2) 2p state-dependent constraints. Choose their corresponding
B = 0, ··· , [L ξi (x)]B = 0, barrier functions as:
∂x ∂x f
and ξ2i−1 = Hi (x) − ci , ξ2i = ci − Hi (x), i = 1, . . . , p;
∂ (l−1)
[L ξi (x)]B = 0. and determine the relative degree μi for each of ξi .
∂x f
Step 2: Choose positive constants γij for j = 1, . . . , μi and
Using the concept of relative degree, we can write high- calculate the corresponding characteristic polynomial:
order time derivatives of ξi (x) as [13]: using differential
operator s, μi
μ
i −1
(j) (j)
i (s) (s + γij ) = sμi + βj sj . (23)
ξi (x) sj ξi (x) = Lf ξi (x), j = 1, . . . , μi − 1; (17) j=1 j=0
and It follows from the above expression and equation (18) that
(μ ) (μ ) ∂
ξi i (x) sμi ξi (x) = Lf i ξi (x) + [Lf(l−1) ξi (x)]Bu. (18) ∂ (μi −1) (μi )
μ
i −1
(j)
∂x αμi = [L ξi (x)]Bu + Lf ξi (x) + βj Lf ξi (x).
The following lemma based on the comparison theorem pro- ∂x f
j=0
vides the mechanism of employing a set of barrier functions,
We know from lemma 2 that barrier ξi (x) ≤ 0 can be ensured
denoted by ξi (x) ≤ 0 for i = 1, . . . , 2p, into a nonlinear
either all the time or asymptotically by choosing u such that
control design. In the remainder of the section, we denote
the control as
ξi (t) ξi (x(t)) to simplify the notation.
Lemma 2: Consider the following differential inequality: μ
i −1
∂ (μi −1) (μ ) (j)
μi [Lf ξi (x)]Bu ≤ −Lf i ξi (x) − βj Lf ξi (x). (24)
∂x
(s + γij )ξi (t) ≤ 0, (19) j=0
V. C ONCLUSION
In this letter, a nonlinear control design is presented for
systems that are subject to saturation constraints on the
system output, the system input, and the input rate of change.
The proposed control employs barrier functions to explic-
itly take into account and enforce all the constraints, and it
embeds a linear optimal control as the seed controller into
a real-time optimization with dynamic and state-dependent
constraints. Analytical stability conditions in terms of the
piecewise-constant Jacobian matrix are obtained for systems
with saturations, and the proposed control is shown to be sta-
bilizing and ensure near-optimal or optimal performance with
respect to a standard performance index. The proposed control
Fig. 2. Trajectory of ug and its limiting values. is applied to the microgrid control problem in which traditional
generation, battery storage, and demand response are subject
to their perspective ramping rate and capacity constraints. The
proposed control effectively coordinates the assets under their
constraints and achieves frequency stability in the presence of
rapid changes in renewable generation.
R EFERENCES
[1] K. van Berkel, R. Titulaer, T. Hofman, B. Vroemen, and M. Steinbuch,
“From optimal to real-time control of a mechanical hybrid power-
train,” IEEE Trans. Control Syst. Technol., vol. 23, no. 2, pp. 670–678,
Mar. 2015.
[2] J. Biggs and W. Holderbaum, “Optimal kinematic control of an
autonomous underwater vehicle,” IEEE Trans. Autom. Control, vol. 54,
no. 7, pp. 1623–1626, Jul. 2009.
[3] A. T. Tran, S. Suzuki, and N. Sakamoto, “Nonlinear optimal control
design considering a class of system constraints with validation on a
magnetic levitation system,” IEEE Control Syst. Lett., vol. 1, no. 2,
Fig. 3. Trajectory of system frequency. pp. 418–423, Oct. 2017.
[4] H. Dagdougui, A. Ouammi, and R. Sacile, “Optimal control of a
network of power microgrids using the Pontryagin’s minimum princi-
ple,” IEEE Trans. Control Syst. Technol., vol. 22, no. 5, pp. 1942–1948,
Sep. 2014.
[5] C. Feller and C. Ebenbauer, “A barrier function based continuous-
time algorithm for linear model predictive control,” in Proc. IEEE Eur.
Control Conf. (ECC), 2013, pp. 19–26.
[6] J. Hauser and A. Saccon, “A barrier function method for the optimization
of trajectory functionals with constraints,” in Proc. 45th IEEE Conf.
Decis. Control, 2006, pp. 864–869.
[7] P. Glotfelter, J. Cortés, and M. Egerstedt, “Nonsmooth barrier func-
tions with applications to multi-robot systems,” IEEE Control Syst. Lett.,
vol. 1, no. 2, pp. 310–315, Oct. 2017.
[8] X. Xu, “Constrained control of input–output linearizable systems using
control sharing barrier functions,” Automatica, vol. 87, pp. 195–201,
Jan. 2018.
[9] H. V. Haghi and Z. Qu, “Stochastic distributed optimization of reactive
power operations using conditional ensembles of V2G capacity,” in Proc.
IEEE Amer. Control Conf., 2015, pp. 3292–3297.
[10] K. G. Boroojeni, M. H. Amini, S. Bahrami, S. S. Iyengar, A. I. Sarwat,
and O. Karabasoglu, “A novel multi-time-scale modeling for electric
power demand forecasting: From short-term to medium-term horizon,”
Elect. Power Syst. Res., vol. 142, pp. 58–73, Jan. 2017.
Fig. 4. Trajectory of battery and DR responses. [11] H. V. Haghi and Z. Qu, “A Kernel-based predictive model of EV capacity
for distributed voltage control and demand response,” IEEE Trans. Smart
Grid, vol. 9, no. 4, pp. 3180–3190, Jul. 2018.
[12] Z. Qu, “Robust state observer and control design using command-to-state
between 6-9am, and DR becomes the only asset to be con- mapping,” Automatica, vol. 41, no. 8, pp. 1323–1333, Aug. 2005.
trolled, which is due to the relative weights used in the optimal [13] H. K. Khalil, Nonlinear Systems. 3rd ed. Upper Saddle River, NJ, USA:
control design. Between 8:30am-2pm, both battery and DR Prentice-Hall, 2003.
[14] F. L. Lewis, D. L. Vrabie, and V. L. Syrmos, Optimal Control. Hoboken,
compensate for the large swings of PV generation without let- NJ, USA: Wiley, 2012.
ting the frequency drift out of its operational range. In short, [15] R. Harvey, Z. Qu, and T. Namerikawa, “Coordinated optimal control of
the three control inputs are coordinated optimally such that all constrained DERs,” in Proc. IEEE Conf. Control Technol. Appl. (CCTA),
2018, pp. 224–229.
the constraints are satisfied at all times, including frequency [16] Hourly Load Data Archives. Accessed: Apr. 18, 2018. [Online].
stability. Available: http://www.ercot.com/gridinfo/load/load_hist/