718 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 24, NO.

2, APRIL 2019

Run-Time Efficiency of Bilinear Model Predictive

Control Using Variational Methods, With
Applications to Hydronic Cooling
Michael B. Kane , Member, IEEE, Jerome P. Lynch, Member, IEEE, and Jeff Scruggs , Member, IEEE

Abstract—Effective real-time execution of nonlinear systems design paradigm [1]. Nonlinear model predictive con-
model predictive control (NMPC) on embedded systems trol (NMPC) has proven its usefulness for many of these
is significantly dependent on the controller formulation. constrained multi-physics control problems [2]. However, these
This paper studies the effect of model structure and cost
applications have been limited to systems with relatively slow
functions on the computation time of scalar bilinear NMPC
using variational methods, with hydronic cooling applica- dynamics and/or expensive computational hardware [3]. For
tions. Two algebraically equivalent nonlinear model struc- systems with fast dynamics and/or cost constraints, the NMPC’s
tures common in literature are primarily considered: a lin- online optimization may have excessive memory demands,
ear state equation with state-dependent control constraints slower than real-time computation, or limited adaptability to
and a bilinear state equation with time-invariant rectangular system parameter changes. This challenge is particularly appar-
control constraints. Additionally, the effects of three cost ent for nonlinear systems, such as the scalar bilinear systems
function formulations are also considered: minimum-time, (BLSs) considered here-in. This class of nonlinear dynamical
quadratic regulation, and efficient state constraints. High-
fidelity computer simulations, hardware-in-the-loop testing, systems, also known as control-affine or bi-affine systems [4],
and experiments on a bench-scale hydronic cooling sys- is characterized by a state equation
tem are used to study sources of computational complexity,
rates of convergence, initialization techniques, and overall ẋ (t) = ax (t) + (bx (t) + b0 ) u (t) + g (1)
effectiveness of the different models and costs. These re-
sults suggest that NMPC with bilinear state equations, min- where the state x(t) and control u(t) are constrained to subsets
imizing pump power and a one-sided quadratic state cost,
converges sufficiently fast and reliably. This presents an of R, a and b are strictly negative scalar constants, and b0 and g
attractive alternative to the traditionally constrained linear are scalar constants that may take any real value.
quadratic regulator-based NMPC on embedded systems. BLSs model a variety of real-world plants. In semi-active
vibration control, a bilinearity is produced by dampers with
Index Terms—Bilinear system, control design, control- controllable viscosity [5]. In biological systems, enzyme con-
affine system, microcontrollers, predictive control,
temperature control. centration is a bilinear control input to metabolic processes [6].
Forced-air heating, ventilation, and air-conditioning (HVAC)
I. INTRODUCTION systems are bilinear with an air-flow control input and room
temperature state [7]. This paper is concerned with hydronic
HE control of ever more complex and interconnected sys-
T tems, enabled by low-power microcontrollers and commu-
nication, is driving a shift toward an integrated cyber-physical
heat transfer BLSs; specifically, the cooling of hot objects with
chilled water. This hydronic cooling control problem appears in
many industrial plants and HVAC systems.
The optimization of output and control trajectories for a con-
Manuscript received May 30, 2017; revised April 14, 2018; accepted strained bilinear system is nonlinear. The solutions to which
December 30, 2018. Date of publication January 29, 2019; date of cur- are often nonanalytic, except for the minimum-time (MT) cost
rent version April 16, 2019. Recommended by Technical Editor Xiu-Tian
Yan. This work was supported in part by the US Office of Naval Research
formulation presented below. As such, linearization or other
under Contracts N00014-05-1-0596 and N00014-09-C0103 granted to suboptimal control strategies are often employed to improve
Jerome P. Lynch. (Corresponding author: Michael B. Kane.) tractability. Analytical stabilizing and optimal feedback control
M. B. Kane was with the University of Michigan, Ann Arbor MI laws exist for subclasses of bilinear systems, e.g., those without
48109 USA. He is now with the Northeastern University Department of
Civil and Environmental Engineering, Boston, MA 02115 USA (e-mail:, constraints [4], [8]. The passivity constraint and limited flow
[email protected]). in hydronic systems prevent the application of these analytical
J. P. Lynch is with the University of Michigan, Department of Civil control solutions to the problem at hand.
and Environmental Engineering and the Department of Electrical En- NMPC, capable of explicitly handling constraints and non-
gineering and Computer Science, Ann Arbor MI 48109 USA (e-mail:,
[email protected]). linearities, has become a powerful tool for designing bilin-
J. Scruggs is with the University of Michigan Department of Civil ear control systems. Research in this field has provided valu-
and Environmental Engineering, Ann Arbor MI 48109 USA (e-mail:, able theory on convergence [9], stability [10], [11, p. 200],
[email protected]).
Color versions of one or more of the figures in this paper are available
applications [8], feedback linearization [12], [13], and design
online at http://ieeexplore.ieee.org. tools [14]. The essential component of every NMPC is an algo-
Digital Object Identifier 10.1109/TMECH.2019.2896020 rithm for optimizing open-loop (OL) control trajectories. The
This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see http://creativecommons.org/licenses/by/3.0/
KANE et al.: RUN-TIME EFFICIENCY OF BILINEAR MPC USING VARIATIONAL METHODS, WITH APPLICATIONS TO HYDRONIC COOLING 719

through ¼” flexible tubing from a reservoir to a single thermal

load that is heated with a resistive heater. Additional cooling
results from natural convection with the air. This system is re-
alized as just one possible flow path of the University of Michi-
gan Chilled Water Demonstrator (UMCWD) [18]. Each of the
thermal loads consists of a 1” × 2” × 4” aluminum block with
two ½” holes drilled through for the flow. On each side of the
Fig. 1. Schematic of simplified hydronic plant. paired thermal loads is mounted a LM35DT temperature sen-
sor. These measure the blocks’ temperature transients because
of heat flow to the chilled water, to the air, and from a 70-Watt
optimality and computational and memory complexities are vi- resistive heater affixed beneath each paired thermal load. Each
tal factors to deploy this dynamic optimization. of these sensors and actuators is connected to a Martlet wireless
Indirect or variational approaches are based on the first-order control platform in which the control algorithms are executed
necessary conditions for optimality as defined by Pontryagin’s [19].
maximum principle. In this case, the optimal OL control tra-
jectory is a solution to a two-point boundary value problem
(TPBVP) containing the states and co-states, linked by a feed- B. Modeling hydronic cooling
back function. Alternatively, direct approaches solve the time- The change in temperature Tb of the thermal load with mass
coupled optimization directly with sequential or simultaneous m and specific heat capacity cp is a function of the rate of energy
approaches. Recent advances in algorithms and code generation transfer shown in Fig. 1 and modeled by
have brought these to the mainstream with computation times
on the order of milliseconds or less [9]. The speed advantage mcp Ṫb (t) = Q̇w (t) + Q̇a (t) + Q̇h (2)
of direct approaches comes at a quadratic demand for memory, where
as opposed to linear memory complexity with variational ap-
proaches [15, p. 375]. This is an important factor considering Q̇a (t) = (Ta − Tb (t)) ha (3)
the limited memory on low-cost micro-controllers. Elapsed time
to an optimal control trajectory can be further reduced with off- Q̇w (t) = (Tw − Tb (t)) hw (t) . (4)
line explicit NMPC methods. These perform the optimization
a priori, generating a lookup table of feedbacks within similar Thermal energy enters from the resistive heater, Q̇h , leaves
regimes [16]. Explicit NMPC methods, however, are slower to through natural convection, Q̇a , to the air with constant tempera-
adapt to system changes because the entire look-up table must ture Ta , and leaves through forced convection, Q̇w , to the water
be recomputed. See [17] for a thorough survey of these and other with constant temperature Tw . The heat transfer coefficients
dynamic optimization strategies. from the block to the air and water are ha and hw , respectively.
This paper presents a run-time efficiency study, which has Control of the blocks’ temperature dynamics is achieved by ad-
thus far been lacking in the literature, of variational approaches justing hw (t) through changes in flow rate q(t) as modeled by
to the real-time control of a hydronic cooling plant. Section II the Dittus–Boelter correlation [20].
derives a physics-based bilinear model of a hydronic cooling kw
testbed. The equivalency of this bilinear OL model with a lin- hw (t) = Nu (t)
ADH
ear state-equation based model is shown in Section III, along  0.8  
with three approaches to representing the engineering objec- kw DH ν 0.4 (5)
= (0.023) (q (t))0.8
tive. The variational calculus employed to optimize these six ADH νA α
cost-function—model pairs is derived in Section IV. Section V = α0 (q (t))α 1 , α0 ∈ R+ , α1 ∈ (0, 1] .
applies these algorithms to NMPC. Additionally, the embed-
ded codes for these algorithms are discussed. Tests are then In this relation, the thermal conductivity kw , pipe surface area
described in Section VI, showing both computer and hardware- A, hydraulic diameter DH , water viscosity ν, and water thermal
in-the-loop (HiL) simulations and experimental results that pro- diffusivity α are condensed to empirical parameters α0 and α1 .
vide insight into the run-time efficiency. Section VII concludes The monotonically increasing function hw (t) maps q → hw .
this paper, discussing the benefits and trade-offs of the different Therefore, an inverse exists, and the problem of choosing a flow
models and objective functions. rate can be algebraically abstracted to the problem of choosing
a heat transfer coefficient or vice versa.
II. BILINEAR MODEL OF HYDRONIC COOLING Using (2), the dynamics of the block temperature can be
expressed in the following bi-affine form with respect to control
Hydronic systems use liquid, e.g., water, as an efficient me- hw (t) and state Tb (t):
dia for moving thermal energy. As such, hydronic heating and
cooling systems are found in many industries including building mcp Ṫb (t) = − ha Tb (t) − hw (t) Tb (t) + hw (t) Tw
energy systems, process control, and naval ships.
+ (ha Ta + Q̇h ). (6)
A. University of Michigan Chilled Water Demonstrator The constraint imposed by the finite capacity of the pump adds
The simple hydronic system used in this analysis is schemat- an additional nonlinearity that bounds the flow q(t) ∈ [0, qm ax ]
ically shown in Fig. 1. A single pump delivers chilled water and thus heat transfer hw (t) ∈ [0, hw m a x ].
720 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 24, NO. 2, APRIL 2019

TABLE I where
MODEL PARAMETERS
aL := aB L < 0; bL := −bB L > 0 (11)
gL := gB L ∈ R (12)
 
b0B L
uL (t) := − xL (t) + uB L (t) (13)
bB L


UL (xL ) = uL (t) : uL (t) (bB L xL (t) + b0B L )uB L m a x
 
− bL uL (t) ≥ 0, ∀t . (14)

Defining a state as xB L (t) = Tb (t) − Tw , (6) is transformed

into the bilinear form (7) with the following parameters.
Fig. 2. Linear model state-dependent control constraints (figure is
shown assuming b0B L ≤ 0). uB L (t) = hw (t) = α0 (q (t))α 1 (15)
α1
um ax = hw m a x = α0 (qm ax ) (16)
−h a −1
Table I shows the empirically derived model parameters for aB L = m cp ; bB L = m cp ; (17)
the UMCWD used in the simulations and experiments presented  
in the following sections. b0B L = 0; gB L = 1
m cp ha Ta + Q̇h . (18)

Similarly, the parameters for (10) can be defined.

III. MODELS AND COSTS FOR OL CONTROL OF BLSS
−h a
Given the physical relations that describe a bilinear system, aL = ; bL = m1c p
m cp (19)
a controller may be designed such that the system performs 1  
as desired. But how does that controller achieve the desired gL = ha Ta + Q̇h . (20)
response? And what, specifically, is the desired response? Is mcp
there more than one way of achieving such performance?
In this section, methods for generating effective OL control B. Objective Functions
trajectories will be studied across multiple definitions of the In general, the engineering goal of the controller in this
control variable u(t) and objective functions. Section III-A will study is to maintain the state approximately within a safe set
discuss two possible model formulations: a linear state equation X ∗ := {x : x < xm ax }, subject to initial conditions x0 and ex-
with state-dependent constraints on control and a bilinear state ternal disturbances, while minimizing the power consumed by
equation with rectangular control constraints. bounded control u over a time horizon tf . This general objective
can be represented by a variety of exact and approximate cost
¯ u; t) to be minimized.
functions J(x,
A. Bilinear and Linear Model Equivalency
 tf
This paper studies the family of scalar bilinear control sys- ¯ u; t)dt.
min J(x, (21)
tems with state xB L (t) controlled by uB L (t) with time-invariant u (t)∈U (t) 0
bounds and subjects to detectable time-invariant disturbance ẋ(t)=f (x,u ;t)
gB L . x(0)=x 0

Three cost functions are studied herein. The MT and quadratic

ẋB L (t) = aB L xB L (t) + (bB L xB L (t)
regulator are common in industry. The third, efficient cooling
+ b0B L )uB L (t) + gB L (7) (EC), better captures the engineering objective.
The MT objective aims to drive the state to within the safe re-
where gion as quickly as possible, with no penalty added for exercising
control. The cost function is defined as
aB L < 0; bB L < 0; gB L ∈ R (8)
J¯M T (x, u; t) = 1 (22)
0 ≤ uB L (t) ≤ uB L m ax ⇔ UB L
   and the integration bound tf is the time at which the state enters
= uB L (t) : uB L (t) uB L m a x − uB L (t) ≥ 0 ∀ t . the safe set. Equation (21) must be augmented to include tf as
(9) a free variable in the minimization. If the system begins within
X ∗ , then no control is prescribed by this objective, and a constant
This system can be equivalently modeled by a linear state control should be used that maintains the state within X ∗ . Any
equation with state xL (t) and controlled by uL (t) with state- realization of a controller with this objective must include an
dependent bounds (graphically depicted in Fig. 2) and subject analysis of parameters that will result in tf → ∞.
to time-invariant disturbance gL . The quadratic regulator (Q) objective integrates the cost
ẋL (t) = aL xL (t) + bL uL (t) + gL (10) J¯Q (x, u; t) = (x (t) − xm ax )2 + ρu2 (t) (23)
KANE et al.: RUN-TIME EFFICIENCY OF BILINEAR MPC USING VARIATIONAL METHODS, WITH APPLICATIONS TO HYDRONIC COOLING 721

over a fixed horizon tf defined a priori. This objective adds a TABLE II

penalty both when the state is above and below xm ax , which FEASIBLE PERMUTATIONS OF xe 0 , xe m a x , x0 , xf , AND xs
causes tuning difficulties in certain circumstances shown below.
The control cost ρu2 represents the pumping energy cost; how-
ever, this approximation depends upon the model formulation
used, i.e., the physical definition of u. As such, even with an
identical ρ weighting term, the optimal trajectories will differ
if a linear or bilinear model is used.
The EC objective function is an accurate mathematical rep-
resentation of the design goal. This is the sum of two costs. A
quadratic penalty is placed on states outside the safe set, other-
wise the state cost is zero. The second cost is proportional to the
power consumed P(t) by the controller over a fixed horizon tf [0, tf ] are defined by (28)–(30).
defined as a priori. H (x, u, p; t) = 1 + pT (t) f (x, u; t) (27)
 2 ∗ ∗ ∗ ∗ ∗
1 H (x , u , p ; t) ≤ H (x , u, p ; t)
J¯E C (x, u; t) = ((x (t) − xm ax ) + |x (t) − xm ax |)
2 ∀t ∈ [0, tf ] and ∀u ∈ U (x, u; t) (28)
+ ρP (x, u; t) . (24) H (x∗ , u∗ , p∗ ; t) = 0∀t ∈ [0, tf ] (29)
The function P(x, u; t) shall be monotonic with respect to ∂H
u ∈ U and maps the state and control to the power consumed by = Hx = −ṗT = pT fx . (30)
∂x
the system. The resulting objective function is independent of
the definition of u in the system model. The soft-constraint may 1) Bilinear State Equation Model: Solving (28) for u∗ with
allow the state to rise above the set point, but only in-so-much-as Hamiltonian (27) using the bilinear state (7) and the monotonic
it offsets the additional positive ρ-weighted cost incurred by the co-state equation asymptotically approaching the origin (30)
additional power consumed. yield an optimal control that switches between two modes, i.e.,
a bang-bang type control, and satisfies (29).
⎧  
IV. DERIVING THE OPTIMIZATION APPROACHES ⎪ b0B L
⎪
⎪ 0, (p (t) ≥ 0) ∧ x (t) ≤ xs = −
The following three sections will derive the control policies ⎪
⎪ bB L
⎪
⎪
that minimizes each of these objective functions, for both linear ⎨u (p (t) ≥ 0) ∧ (x (t) > xs )
m ax ,
and bilinear system models. u∗ (t) = .
⎪
⎪
⎪
⎪ u m ax , (p (t) < 0) ∧ (x (t) ≤ x )
⎪
⎪
s
A. MT Cost ⎪
⎩
0, (p (t) < 0) ∧ (x (t) > xs )
/ X ∗,
Given the following system1 with x0 ∈ (31)
ẋ (t) = f (x, u; t) ; x (0) = x0 ∈ R . (25) Switching between the two modes (A and B) occurs when
the state touches the guard, G(A, B) = xs . In the time domain,
The goal of the MT controller is to drive the final state of the
mode A is modeled by (35) when it enters with state x0 and
system x(tf ) into X ∗ as quickly as possible given the constraints
monotonically approaches xe 0 . Likewise, the state in mode B
on the control u(t) ∈ U(x, u; t)∀t ∈ [0, tf ]. The scalar problem
(36) monotonically approaches xe m a x . Thus, not all values of
is reduced to driving the state to the boundary of X ∗ , i.e., xm ax =
x(t) are reachable from all initial states x0 if xe 0 = xe m a x .
x(tf ) = xf . This problem can be described by the following
minimization aA = aB L ; gA = gB L (32)
 tf
min dt. (26) aB = aB L + bB L uB L m a x (33)
ẋ(t)=f (x,u ;t)
u (t)∈U (x,u ;t)
0
gB = b0B L uB L m a x + gB L (34)
/X ∗
x(t 0 )=x 0 ∈ −gA
x(t f )∈X ∗ x (t) = (x0 − xe 0 ) ea A t + xe 0 , xe 0 = (35)
0≤t f aA
−gB
From this point, an analytical feedback control law is derived x (t) = (x0 − xe m a x ) ea B t + xe m a x , xe m a x = . (36)
from Pontryagin’s maximum principle [15]. Using the calculus aB
of variations [21] and a Hamiltonian defined by (27), the three The automaton’s analytical time domain solution results in a
necessary conditions for optimality of the state, co-state, and piece-wise function switching at-most one time before tf . Only
control trajectories, x∗ , p∗ , and u∗ respectively, over horizon nine permutations of xe 0 , xe m a x , x0 , xf , and xs shown in Table II
result in feasible solutions given the following conditions:
1) max {xe 0 , xe m a x } < xf
1 Within this paper, the notation f (g, h; t) is an abbreviated notation for
2) x0 < xf
f (g(t), h(t)). It describes a function f of two time-varying inputs g(t) and
h(t), but the function f is time invariant, i.e., the variable t does not explicitly 3) a < 0; b < 0
show up in f . 4) sign (p (t)) = const.
722 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 24, NO. 2, APRIL 2019

Once the final time has been reached, another control algo- The Hamiltonian was simplified to H̄ by removing terms in
rithm must take over to keep the state within the safe set. For which ∂duH = 0, resulting in no effect on (45). When the solu-
example, hysteresis could be built into the bang–bang controller tion to (45) violates the control constraint, the control saturates
or a constant control with an equilibrium state equal to xm ax according to
could be used. ⎧
2) Linear State Equation Model: Solving (27)–(30) with the ⎪
⎪ uL ∗u c (p∗L ) , if u∗L u c (p∗L ) ∈ UL (x∗L )
⎨  
linear state, (10) yields a control exactly equivalent to (32)–(36). u∗L = u∗L m ax (x∗L ) , else if HL p∗L , u∗L m a x < 0 (48)
⎪
⎪
⎩
B. Quadratic Regulation 0, otherwise

Given (25), a trajectory u∗ is to be found that solves (37). where

Without loss of generality, the system has been defined and/or −bL p∗L
shifted such that xm ax = 0 and ρ ≥ 0. The feedback law is uL ∗u c (p∗L ) = (49)
derived using the same calculus of variations and Pontryagin’s 2RL
maximum principle used to derive the control law for the MT (bB L x∗L + b0B L ) uB L m a x
case. First, a Hamiltonian is defined by (38) with respect to uL ∗m a x (x∗L ) = . (50)
bL
the state, co-state, and control at each moment. (For simplicity,
variation with respect to time will no longer be shown and will The forward state dynamics (43), the backward co-state dy-
be implied from here forward for x, u, and p.) namics (44) and feedback law (48) form a TPBVP. The TPBVP
 tf solution, x∗L , u∗L , and p∗L meet the necessary optimality condi-
tions.
min (x2 (t) + u2 (t)ρ)dt (37)
ẋ(t)=f (x,u ;t) 0
ẋL = aL xL + bL uL + gL , xL (0) = xL 0 (51)
u (t)∈U (x,u ;t)
x(0)=x 0 ṗL = −aL pL − 2xL , pL (tf ) = 0 (52)
⎧
H(x, u, p) = (x + u ρ) + pf (x, u).
2 2
⎪ uL u c (pL ) , if uL u c (pL ) ∈ UL (xL )
(38) ⎪
⎨
The four necessary conditions for optimality then become uL (xL , pL ) = uL m a x (xL ) , else if HL (pL , uL m a x ) < 0
⎪
⎪
⎩
ẋ∗ = f (x∗ , u∗ ) 0, otherwise
(39) (53)
∀t ∈ [0, tf ] , x∗ (0) = x0
ṗ∗ = − ∂∂H ∗ ∗ ∗ where
x (x , u , p )
(40)
∀t ∈ [0, tf ] , p∗ (tf ) = 0 H̄L (pL , uL m a x ) = ρL u2L m a x + bL p∗L uL m a x (54)

H (x∗ , u∗ , p∗ ) ≤ H (x∗ , u, p∗ ) −(bL pL )2

(41) uL u c (pL ) = (55)
4RL
∀ u ∈ U (x, u) , t ∈ [0, tf ]
(bB L xL + b0B L ) uB L m a x
Ḣ (x∗ , u∗ , p∗ ) = 0 ∀ t ∈ [0, tf ] . (42) uL m a x (xL ) = . (56)
bL
The first three conditions lead to an optimization problem in 2) Bilinear State Equation Model (QBL): Similarly, the nec-
which u(k) is decoupled from u(k) at all other time steps. This essary conditions for the optimal control of the bilinear plant
property results in a more computationally tractable optimiza- described by (7) become (57) to (60) with the Hamiltonian sim-
tion compared to directly optimizing (37). The fourth condition plified to (61).
should be automatically satisfied by the optimized trajectories,
as verified numerically. ẋ∗B L = aB L x∗B L + (bB L x∗B L + b0B L ) u∗B L + gB L
(57)
1) Linear State Equation Model (QL): The linear model im- ∀t ∈ [0, tf ] ; xB L (0) = xB L 0
plementation of these conditions yields
ṗ∗B L = −aB L pB L − bB L uB L pB L − 2xB L
ẋ∗L = aL x∗L + bL u∗L + gL (58)
(43) ∀t ∈ [0, tf ] ; pB L (tf ) = 0
∀t ∈ [0, tf ] , xL (0) = xL 0
min H̄B L (x∗B L , p∗B L , uB L ) ∀t ∈ [0, tf ] (59)
ṗ∗L = −aL p∗L − 2x∗L 0≤u B L ≤u B L m a x
(44)  
∀t ∈ [0, tf ] , pL (tf ) = 0 0 = (x∗B L )2 + (u∗B L )2 ρB L + (aB L x∗B L
min H̄L (p∗L , uL ) ∀ t ∈ [0, tf ] (45) + (bB L x∗B L + b0B L ) u∗B L + gB L )p∗B L ∀t ∈ [0, tf ] (60)
u L ∈g L (x ∗L ,u L )
  H̄B L (x∗B L , p∗B L , uB L ) = ρB L u2B L
(x∗L )2 + (u∗L )2 ρ + (aL x∗L + bL u∗L + gL ) p∗L = 0
(46) + (bB L x∗B L + b0B L ) p∗B L uB L . (61)
∀t ∈ [0, tf ]
Following the same process as for the linear system, a TP-
H̄L (p∗L , uL ) = ρu2L + bL p∗L uL . (47) BVP is defined that yields a solution that satisfies the necessary
KANE et al.: RUN-TIME EFFICIENCY OF BILINEAR MPC USING VARIATIONAL METHODS, WITH APPLICATIONS TO HYDRONIC COOLING 723

conditions for optimality. (77) as a function of the (78), (79), P(x∗ , u∗ ), and (P(x∗ , u∗ ))x .
ẋB L = aB L xB L + (bB L xB L + b0B L ) uB L + gB L , xB L (0) ẋ∗L = aL x∗L + bL u∗L + gL ∀t ∈ [0, tf ] ; x∗L (0) = x0 (74)
 
= xB L 0 (62) ṗ∗L = − RηL (x∗L ) x − ρPL x (x∗L , u∗L ) − aL p
(75)
ṗB L = −aB L pB L − bB L uB L pB L − 2xB L , pB L (tf ) = 0 ∀t ∈ [0, tf ] ; p∗L (tf ) = 0
(63) u∗L = argmin HL (x∗L , uL , p∗L ) ∀t ∈ [0, tf ] (76)
u L ∈UL (x L )
uB L (xB L , pB L )
⎧ ḢL (x∗L , u∗L , p∗L ) = 0∀t ∈ [0, tf ] (77)
⎪ uB L u c (xB L , pB L ) ,
⎪
⎪
⎪
⎪ HL (xL , uL , pL ) = RηL (xL ) + ρPL (xL , uL ) + fL (xL , uL ) p
⎪
⎪ if 0 ≤ uB L u c (xB L , pB L ) ≤ uB L m a x
⎨ (78)
= uB L m a x , (64)  η 
⎪
⎪ RL (xL ) x = ηxηL−1 1 (xL ) (79)
⎪
⎪ else if H̄ (xB L , pB L , uB L m a x ) ≤ 0
⎪
⎪
⎪
⎩ Derivation of the feedback control law shows that the optimal
0, otherwise control is the solution to the convex program (80) that can
where: be solved numerically using efficient computer programs (e.g.,
gradient descent or simplex algorithms).
H̄B L (xB L , pB L , uB L ) = ρB L u2B L
u∗L = argmin (ρPL (x∗L , uL ) + BL uL p∗L ) (80)
+ (bB L xB L + b0B L ) pB L uB L u L ∈UL (x ∗L )
(65)
As in the previous section the state dynamics, co-state dy-
− (bB L xB L + b0B L ) pB L namics, and feedback are combined into a TPBVP.
uB L u c (xB L , pB L ) = (66)
2ρB L
ẋL = aL xL + bL uL + gL , xL (0) = xL 0 (81)
Methods for solving these TPBVPs are presented  
Section IV.D of this paper. ṗL = − ηxηL−1 1 (xL ) −ρPL x (xL , uL ) − aL pL , pL (tf ) = 0
(82)
C. Efficient Control with Soft-constraint
Given (25), a trajectory u∗ is to be found that balances vio- uL = argmin (ρPL (xL , uL )+bL uL pL ) (83)
u L ∈UL (x L )
lation of a soft constraint with an explicit accounting of power
consumption, resulting in the minimization problem 2) Bilinear State Equation Model (BLEC): For the bilinear
 tf model (7), the same processes can be followed to derive the
min R2 (x) + ρP(x, u)dt (67) TPBVP which solves for a solution that meets the necessary
ẋ(t)=f (x,u ;t) 0 conditions for optimality of (67).
u (t)∈U (x;t)
x(t 0 )=x 0 ẋB L = aB L xB L + (bB L xB L + b0B L ) uB L
where, + gB L , xB L (0) = xB L 0 (84)
 2  
R2 (x) =
1
(x + |x|) = (x1 (x))2 (68) ṗB L = − η xηB−1 L 1 (xB L ) − ρPB L x (uB L )
2
− aB L pB L , pB L (tf ) = 0 (85)
∂P (x, u)
P (x, 0) = 0 and ≥ 0 ∀ (u ∈ U) (69) uB L = argmin (ρPB L (uB L )+(bB L xB L +b0B L ) uB L pB L )
du
u B L ∈UB L
As with the quadratic regulator, and without loss of generality, (86)
the system has been defined and/or shifted such that xm ax = 0
and ρ ≥ 0. The same calculus of variations used in the previ- The run-time efficiency of EC control algorithms will be
ous section applies here to derive the following four necessary greater than that of the quadratic regulation algorithms due to the
conditions for an optimal solution to (67). secondary optimization required in (83) and (86), as opposed to
the analytic feedback law of (53) and (64). The advantage of the
ẋ∗ = f (x∗ , u∗ ) ∀t ∈ [0, tf ] ; x∗ (0) = x0 ; (70) EC algorithms may be in the ease of tuning due to the one-sided
∗ ∗ ∗ ∗ ∗
ṗ = −Hx (x , u , p ) ∀t ∈ [0, tf ] ; p (tf ) = 0 (71) thermal cost that closely represents the engineering objective (as
opposed to the two-sided regulation behavior), accurate model
H (x∗ , u∗ , p∗ ) ≤ H (x∗ , u, p∗ ) of pump power, and the equivalency of linear and bilinear state
(72) equation formulations.
∀u ∈ U (x, u) , t ∈ [0, tf ]
Ḣ (x∗ , u∗ , p∗ ) = 0∀t ∈ [0, tf ] (73) D. Solving the TPBVP
1) Linear State Equation Model (LEC): Applying these the General methods for solving TPBVPs fall into three cate-
four necessary conditions to the linear model (10) yields (74) to gories: gradient projections, shooting methods, and first-order
724 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 24, NO. 2, APRIL 2019

gradient methods [15]. The gradient projection method itera- TABLE III
tively optimizes the entire control trajectory in a single gradient. CONTROLLER PARAMETERS FOR EMBEDDED TESTS (AND SIMULATIONS)
Shooting methods iteratively estimate p0 and forward integrate
the system to match pf . First-order gradient methods integrate
state and co-state trajectories with estimated control trajectories,
then update the control to better satisfy the feedback law with
the updated states and co-states.
Synthesizing on-line embedded gradient projection con-
trollers demand potentially prohibitive amounts of memory. culations [19]. The first-order gradient procedure for solving
Shooting methods have minimal memory requirements, an ideal the TPBVPs was codified in the C language, integrated into the
feature for embedded applications. However, numerical stabil- Martlet firmware, compiled using Code Composer Studio v5
ity issues arise when the control can bang between the lim- [24] and embedded into the Martlet wireless controller.
its, causing large fluctuations in the co-state final value due The GNU Scientific Library (GSL) [25] was ported to the
to minor changes in the initial values. The first-order gradient Martlet in its native single floating point precision. The GSL
method balances numerical stability with a memory require- Runge-Kutta 4-5 ODE solver forward integrates the state and
ment of O(M K + N K) where M , N , and K are the number backwards integrates the co-state. The GSL one-dimensional
of control inputs, the number of states, and the number of time minimization routine, using the Brent minimization algorithm,
steps in the control horizon, respectively. solves the convex optimization required at each step along the
The first-order gradient algorithm used in this work is adapted prediction trajectory for the LEC and BLEC optimizations. The
from [15], pp. 335–337] to systems with constraints. The itera- TPBVP solver extends the GSL architecture, i.e., passes func-
tive algorithm consists of the following steps: tion pointers for the application specific functions. Execution
1) Forward integrate the state equations from x0 along the time was minimized by avoiding nested functions, passing of
horizon using u(i − 1) and p(i − 1) from the previous iter- large variables (instead pointers to structures are used), and re-
ation. Store the updated state trajectory estimate x(i) . casting of structures. The NMPC step was extended to ensure
2) Backwards integrate the co-state equation from pt f along convergence. However, “fast NMPC” was also tested.
the horizon using u(i − 1) and x(i) .
3) Apply the feedback law to compute û(i) from x(i) B. Controller Tuning
and p(i) .
The goal of this work is to compare the four controllers (i.e.,
4) Stop if |u(i) − û(i) | < ε. QL, QBL, LEC, and BLEC). Each require tuning of eight pa-
5) Perturb each value of u(i) (tk ) towards the value of rameters: ρ, τm in , τ0 , τm ax , K, dt, im ax , and εu . As such, a
û(i) (tk ) by some percentage τ . tuning procedure was developed to achieve a fair comparison
6) Go to step 1. among the controllers with different objectives. The parameters
The speed of decent parameter τ should be selected to balance τm in , τm ax , K, dt, im ax , and εu essentially have the same effect
potential oscillations (and lack of convergence) if too large, on all four controllers, so they were set consistently across all
and slow convergence if too small. Kirk proposes adjusting τ the controllers. A step size of 5s was defined approximately
after each iteration: slightly increasing (e.g., 1%) if the cost an order of magnitude faster than the system dynamics, while
J descreased from the previous iteration, or more significantly the prediction horizon of K · dt = 500 s was selected conser-
decreasing (e.g., 5%) if the cost J increased from the previous vatively as two orders of magnitude longer than the system
iteration [15], pp. 335–337]. dynamics. Convergence is defined as an RMS change in the
control trajectory of less than εu = 1E6, or when im ax = 250
V. NMPC OF BLSS iterations are reached. Values of τm in = 0.005 and τm ax = 0.3
The adverse effect of modeling errors, sensor errors, and were selected such that convergence was achieved robustly, yet
unknown disturbances on control performance are mitigated by quickly in all controllers and all cases. The linear controllers
recalculating the optimal OL trajectories at fixed intervals in were particularly sensitive to smaller τm in values or larger τm ax
time, i.e., NMPC [22]. The optimizations are initialized with values which prevented convergence either due to oscillation or
the solution at the previous NMPC step, appropriately shifted slow descent. The embedded GSL ODE solver and minimizer
forward in time. If full convergence is slower than real-time, a also have ε parameters for stopping criteria. These were set to
“fast NMPC” approach [23] stops the iterations at the end of the the largest values that produced results comparable to the re-
time interval, and applies the suboptimal results, converging as sults generated by the equivalent MATLAB functions with the
the system evolves. default parameters.
Thus, only two parameters are left to be tuned for each con-
troller, ρ and τ0 . Values of τ0 were selected such that convergence
A. Embedded NMPC of BLSs
was achieved quickly, yet robustly in all cases. For the LEC and
The experiments below compare the performance of NMPC BLEC controllers, ρ was tuned to subjectively achieve the de-
realizations synthesized using the two model formulations and sired performance. Then, the QL and QBL controllers were
three cost functions for the simple hydronic cooling system. each tuned by a line-search on ρ. This search minimized the
The synthesized digital controls were deployed using a Mart- aggregate retroactively computed EC cost of the simulated tra-
let wireless control platform which contains an 80 MHz 16-bit jectories from test cases A–D presented below. Table III shows
microcontroller capable of 32-bit hardware floating point cal- the values of ρ and τ from the tuning procedure.
KANE et al.: RUN-TIME EFFICIENCY OF BILINEAR MPC USING VARIATIONAL METHODS, WITH APPLICATIONS TO HYDRONIC COOLING 725

TABLE IV
TEST CASE PARAMETERS

VI. EXPERIMENTAL AND SIMULATED TESTING

A. Test Scenarios
Five test procedures are used to study the four controllers in
Table III and the three benchmark controllers (i.e., zero-flow,
max-flow, and min.-time). These cases capture the following
primary characteristics of standard test procedures for thermal
control (i.e., neglecting humidity as a control variable) of single-
zone HVAC system control algorithms (e.g., [26]):
1) Case A: Maintain a safe temperature under heating Fig. 3. Computation time of each iteration measured during HiL testing
on the Martlet for all four test Cases A-D.
2) Case B: Get to a safe temperature
3) Case C: Get to a safe temperature under heating
4) Case D: Trivially maintain a safe temperature
5) Case E: Maintain safe temperature under changing heat
Case A and Case D are identical, as are Case B and Case
C, except for the heater state (off or on). Cases A–D are pri-
marily used for tuning and studying OL behavior. Case E com-
bines much of the same characteristics of Cases A–D into a
time-varying system that will demonstrate the CL behavior of
the proposed NMPC. Specifically, the heater switches on for
10 minutes, off for 5 minutes, on for 30s, off for 30s, on for
1 minute, off for 1 minute, on for two minutes, and then off.

B. Results and Analysis of OL Testing

Simulations and HiL tests were conducted with Cases A
through D to provide new understanding into the sources of
computational complexity, rates of convergence, appropriate
initialization techniques, and overall effectiveness of the four
controllers (QL, QBL, LEC, and BLEC) and three benchmarks
controls (zero flow, max flow, and MT). Fig. 4. Retroactively calculated EC cost function value versus HiL com-
Timing data were collected from HiL tests conducted on the putation time for four controllers and four test case. “Init” results are ini-
Martlet for each of the four cases and four controllers, where tialized with the OL trajectories calculated with the system having the
sensor data were emulated according to Table IV. Three of the opposite heater state.
Martlet’s GPIO pins were programmed to mark the start and
end of code execution for the state forward integration, co- and (64), the control update execution time for QL and QBL
state backwards integration, and the feedback calculation and is negligible. However, control updates for LEC and BLEC are
update for each iteration. The timing of these pins was measured lengthy due to the iterative optimization required in (83) and
with a digital oscilloscope connected to the Matlab based HiL (86). This calculation for BLEC can take a significantly longer
simulation environment. time. The total computation time per iteration is driven by the
The boxplot in Fig. 3 shows statistics on computation times complexity of the co-state reverse integration, and the feedback
collected for every iteration of every case for each of the four optimization for LEC and BLEC. The BLEC converges to the
controllers. The data show that the forward integration time is same result as LEC, yet computes faster per iteration.
consistently short among all the controllers. However, there are The total calculation time for each optimal OL trajectory is a
large timing variations in the co-state reverse integration, i.e., factor of both the time per iteration, and the number of iterations
due to small adaptive-step-sizes in the GSL rkf45 algorithm, required to converge. Fig. 4 shows these rates of convergence.
particularly for the LEC. This stems from the ill-conditioned For a fair comparison, the EC cost, which most closely repre-
reverse integration as Tb → Tw due to the presence of (Tb − Tw ) sents the engineering goal, was retroactively calculated for each
in the denominator of P(x, u) for LEC. As expected from (53) of the four controllers and each of the cases. The “init” results
726 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 24, NO. 2, APRIL 2019

TABLE V
RETROACTIVELY CALCULATED EC COSTS

Subscripts ∗0 indicate the optimization was initialized with zeros, while ∗i n i t indicate Fig. 5. OL temperature and flow trajectories for Case C, simulated in
initialization with the results of the same parameters, but with opposite heater state. Matlab and HiL testing on the Martlet. “Init” results are initialized with the
OL trajectories calculated from Case B.

(solid line) are initialized with the OL trajectories calculated

with the system having the opposite heater state (emulating
NMPC where the heater state just changed), while the other
results (dashed line) are initialized with all control values equal
to zero (emulating a fresh start of the controller).
The QL and QBL controllers initially converge near the op-
timal solutions faster than LEC and BLEC; a surprising result
considering the QL and QBL controllers are optimizing a dif-
ferent objective function, albeit tuned to emulate EC. Case D
demonstrates the undesirable behavior of the QL and QBL reg-
ulators; i.e., even though the block temperatures are safe, the
controllers raise the temperature closer to the set point, possible
because the Tw > Tb . On the other hand, the LEC and BLEC
controller behave as expected for Case D, producing zero flow.
Fig. 6. Case E BLEC convergent NMPC experimental response.
The convergence per iteration of LEC and BLEC approach that
of the QL and QBL controllers; however, the convergence per
second (the important measure for real-time performance) lags In Fig. 5, the single-precision floating-point HiL trajectories
significantly due to the increased calculation time per iteration are compared with the double-precision Matlab trajectories for
(Fig. 3). Upon close inspection, oscillations of varying ampli- Case C. The Matlab results demonstrate that the LEC and BLEC
tudes are observed in the final solutions for all the controllers yield nearly the same trajectory, regardless of initialization, as
and initializations in Case B and Case C. expected from their mathematical equivalency. However, some
The final values of the retroactively calculated EC costs for Martlet results differ significantly due to the lower-precision and
Case A–D are shown in Table V. As expected, the best LEC ill-conditioned reverse integration, among other factors. HiL and
or BLEC implementations outperform the best costs from the Matlab results most closely match for BLEC control. On the
QL and QBL controllers, since the QL and QBL controllers other hand, QL has the largest discrepancy between precision
are optimizing a different objective function. The discrepancies and initialization, illustrating the difficulty of finding values for
between simulation environments and initialization procedures τ that converge well across all cases and precisions.
demonstrate that full convergence is not always achieved, es-
pecially for the HiL calculations. Poor final convergence is a
C. Results and Analysis of Closed-Loop NMPC Testing
known behavior of the gradient method [15], and future work
should study these controllers and cases for different approaches The BLEC controller, with reliable convergence properties,
to solving the OL optimal control problem. The discrepancy is moderate computational complexity, and straight forward tun-
particularly bad for trajectories with sharp edges, e.g., those that ing, was selected for further study. Closed-loop NMPC experi-
exhibit bang-bang behavior, since the iterations converge slowly ments using test case E were run on the UMCWD controlled by
to fill into the corners. This may pose a problem for NMPC using a Martlet running the embedded NMPC algorithm. Two exper-
these methods if, even though most of the trajectory has con- imental results are shown in Figs. 6 and 7. The green shadows
verged, the first point (which is the only point applied during indicate when the heater is on (producing 35 W of heat). A “+”
NMPC) is slow to converge due to the proximity to a switch. on the middle axis indicates convergence of the TPBVP in that
KANE et al.: RUN-TIME EFFICIENCY OF BILINEAR MPC USING VARIATIONAL METHODS, WITH APPLICATIONS TO HYDRONIC COOLING 727

ear state equations with state-dependent control constraints, and

bilinear state equations with rectangular control constraints. Op-
timization occurred over a receding horizon with MT, quadratic,
and EC cost functions.
The MT objective, or thermostatic controller is most com-
mon in practice due to computational simplicity; however, per-
formance can be poor when the engineering objective requires
balancing constraint violation with control energy use.
Quadratic objectives can be tuned to closely match the engi-
neering objective (see Table V); however, the tuning parame-
ters lack intuitive physical meaning. NMPC using bilinear state
equations with quadratic costs (QBL) more closely represent the
engineering objective and converge more reliably than NMPC
with linear state equations in the system model (QL). However,
Fig. 7. Case E BLEC “fast NMPC” experimental response. the QBL TPBVP contains a poorly conditioned reverse integra-
tion of the co-state, resulting in computation times per iteration
TABLE VI that are twice that of QL; although slightly faster convergence
RETROACTIVELY COMPUTED COSTS FOR EXPERIMENTAL RESULTS (CASE E) per second.
The engineering objective can be accurately represented with
a one-sided quadratic cost for violation of the safety thresh-
old, and a 3rd order polynomial of pump power with respect
to flowrate. Synthesizing an NMPC with such a cost function
requires solving the standard TPBVP, with an additional scalar
NMPC step, while an “·” denotes the TPBVP did not converge. optimization required at each control step along the prediction
The trajectory resulting from each online optimization is shown horizon. Here the linear state equation model (LEC) and bilinear
as a dotted line that starts dark blue and slowly fades to red. state equation model (BLEC) yield mathematically equivalent
The trajectories in Fig. 6 show an NMPC implementation controllers, yet with different computational complexity and
where each TPBVP ran until completion, even if that was longer convergence properties: Compared with LEC control, the BLEC
than the allotted time window. The convergence slows when the control demonstrated faster computations per iteration, due to a
heater switches on or off, as seen by the spacing of the crosses. better conditioned co-state reverse integration, and faster con-
This delay resulted in an undesirable overshoot of the set point. vergence per iteration. Compared to QL and QBL, the BLEC
The “fast” NMPC implementation is intended to mitigate this control converges about half as fast, yet still well within the 5s
delay. The experimental results shown in Fig. 7 closely resemble control period, except when system properties change. In such
the trajectories in Fig. 6, without the overshoot. When the costs a case, a “fast NMPC” was successfully employed.
are retroactively computed for the two NMPC implementations, These variational approaches employed have linear memory
shown in Table VI, the “fast” NMPC implementation is closer complexity in the length of the prediction horizon, making it
to optimal: the optimality loss in non-convergence is made up ideal for implementation on devices like the Martlet. This is in
for in a faster dynamic response. comparison to direct approaches with quadratic memory com-
Inaccuracies in the online model used for NMPC (primarily, plexity [27], or worse [28], for bilinear NMPC. Unlike implicit
unmodeled nonlinearities in the voltage to flow relationship) re- NMPC approaches, the methods shown herein are easily cou-
sult in discrepancies between prediction trajectories (the dotted pled with system ID algorithms. Thus, the controller can adapt
lines in Figs. 6 and 7) and measured trajectories (the thick black online to system parameter changes.
line). However, due to the closed-loop constantly iterating na- These properties and the promising experimental outcomes
ture of NMPC, the measured temperature trajectories are robust present designers with new insight into effective design of model
to these divergences (a well-known feature of MPC [3]) and predictive controllers, especially those run as part of wirelessly
generally maintain a safe temperature (e.g., the “state cost” is networked control systems. For example, the BLEC controller
less than 1% of the “control cost”). presented herein has been extended from the simple pump and
The drop in flow and temperature flare at the end of the pre- load scenario to a redundant network of pumps, valves, and
diction trajectories is an artifact of the finite prediction horizon loads [29].
that could be correct with a terminal penalty in the optimization
[3]. However, such a correction is not strictly necessary since
the artifact is sufficiently down the horizon that the zero flow REFERENCES
and high temperature are never implemented. [1] J. Sztipanovits et al., “Toward a science of cyber–physical system integra-
tion,” Proc. IEEE, vol. 100, no. 1, pp. 29–44, Jan. 2012.
[2] C. E. Garcı́a, D. M. Prett, and M. Morari, “Model predictive control:
VII. SUMMARY AND CONCLUSIONS Theory and practice—A survey,” Automatica, vol. 25, no. 3, pp. 335–348,
This paper provides new insight into the computational May 1989.
[3] E. F. Camacho and C. Bordons, Model Predictive Control in the Process
complexity and convergence characteristics of variational ap- Industry. New York, NY, USA: Springer-Verlag, 1995, p. 239.
proaches to model predictive control of scalar bilinear systems. [4] D. L. Elliott, “Bilinear systems,” in Wiley Encyclopedia of Electrical and
Two mathematically equivalent models were considered: lin- Electronics Engineering. Hoboken, NJ, USA: Wiley, 2007.
728 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 24, NO. 2, APRIL 2019

[5] P. S. J. Harvey, H. P. Gavin, J. T. Scruggs, and J. M. Rinker, “Determining [28] A. Bemporad and M. Morari, “Control of systems integrating logic, dy-
the physical limits on semi-active control performance: A tutorial,” Struct. namics, and constraints,” Automatica, vol. 35, no. 3, pp. 407–427, 1999.
Control Health Monit., vol. 21, no. 5, pp. 803–816, May 2014. [29] M. B. Kane, L. P. Jerome, and J. Scruggs, “Development of a scalable dis-
[6] R. Mohler, “Natural bilinear control processes,” IEEE Trans. Sci. Cybern., tributed model predictive control system for hydronic networks with bilin-
vol. 6, no. 3, pp. 192–197, Jul. 1970. ear and hybrid dynamics,” J. Comput. Civ. Eng., vol. 32, no. 5, Sep. 2018,
[7] A. Kelman and F. Borrelli, “Bilinear model predictive control of a HVAC Art. no. 04018038.
system using sequential quadratic programming,” in Proc. 18th IFAC
World Congr., Jan. 2011, vol. 44, no. 1, pp. 9869–9874.
[8] P. M. Pardalos and V. Yatsenko, Optimization and Control of Bilinear
Systems: Theory, Algorithms, and Applications. New York, NY, USA: Michael Kane (Member, IEEE) received the
Springer-Verlag, 2008. B.S. degree in architectural engineering and the
[9] H. H. J. Bloemen, T. J. J. van den Boom, and H. B. Verbruggen, “An M.S. degree in civil engineering from Drexel Uni-
optimization algorithm dedicated to a MPC problem for discrete time versity, Philadelphia, PA, USA, in 2009, and the
bilinear models,” in Proc. Amer. Control Conf., Arlington, VA, USA, M.S. degree in electrical engineering and the
2001, pp. 2376–2381. Ph.D. degree in civil engineering from the Uni-
[10] A. B. Fontes, C. E. T. Dorea, and M. R. da S. Garcia, “An iterative versity of Michigan, Ann Arbor, MI, USA, in 2011
algorithm for constrained MPC with stability of bilinear systems,” in Proc. and 2014, respectively.
2008 16th Mediterranean Conf. Control Automat., 2008, pp. 1526–1531. He is an Assistant Professor of Civil and En-
[11] D. Q. Mayne, J. B. Rawlings, C. V. Rao, and P. O. M. Scokaert, “Con- vironmental Engineering with Northeastern Uni-
strained model predictive control: Stability and optimality,” Automatica, versity, Boston, MA, USA, prior to which he was
vol. 36, no. 6, pp. 789–814, 2000. a Fellow at the United States Department of Energy Advanced Research
[12] L. Del Re, J. Chapuis, and V. Nevistik, “Predictive control with embedded Project Agency-Energy, Washington, DC, USA. He is currently an Engi-
feedback linearization for bilinear plants with input constraints,” in Proc. neer in Training in Pennsylvania. His current research interests include in
Decision Control, 1993., Proc. 32nd IEEE Conf. on, San Antonio, TX, the areas of bi-linear and hybrid control, and human-in-the-loop control
1993, pp. 2984–2989, vol. 4. of infrastructure systems.
[13] M. Bacic, M. Cannon, and B. Kouvaritakis, “Constrained control of SISO Dr. Kane is a member of the American Society of Civil Engineers.
bilinear systems,” IEEE Trans. Autom. Control, vol. 48, no. 8, pp. 1443–
1447, Aug. 2003.
[14] J. Lunze and F. Lamnabhi-Lagarrigue, “Handbook of hybrid systems con-
trol: theory, tools, applications.” Cambridge, UK: Cambridge Univ. Press, Jerome P. Lynch (Member, IEEE) received the
2009, p. 565. B.E. degree in civil and environmental engineer-
[15] D. Kirk, Optimal Control Theory: An Introduction. New York, NY, USA: ing from Cooper Union, New York, NY, USA, and
Dover, 2004, p. 452. the M.S. and Ph.D. degrees in civil and envi-
[16] R. Zhang and S. Wang, “Predictive control of a class of bilinear systems ronmental engineering and the second M.S. de-
based on global off-line models,” J. Zhejiang Univ. Sci. A, vol. 7, no. 12, gree in electrical engineering from Stanford Uni-
pp. 1984–1988, Dec. 2006. versity, Stanford, CA, USA, in 1998, 2002, and
[17] L. T. Biegler, “An overview of simultaneous strategies for dynamic opti- 2003, respectively.
mization,” Chem. Eng. Process. Process Intensif., vol. 46, no. 11, pp. 1043– He is a Professor of Civil and Environmental
1053, Nov. 2007. Engineering at the University of Michigan, Ann
[18] M. B. Kane and J. P. Lynch, “An agent-based model-predictive controller Arbor, MI, USA. He is also a Professor of Elec-
for chilled water plants using wireless sensor and actuator networks,” in trical Engineering and Computer Science by courtesy. His current re-
Proc. Amer. Control Conf., Montreal, Canada, 2012, pp. 1192–1198. search interests include the areas of wireless cyber-physical systems,
[19] M. B. Kane et al., “Development of an extensible dual-core wireless cyberinfrastructure tools for management of structural monitoring data
sensing node for cyber-physical systems,” in Proc. SPIE, Nondestruc- sets, and nanoengineered thin-film sensors for damage detection and
tive Characterization Composite Materials, Aerospace Eng., Civil In- structural health monitoring.
frastructure, Homeland Security 2014, San Diego, CA, 2014, vol. 9061, Dr. Lynch is the recipient of the 2005 Office of Naval Research (ONR)
pp. 90611U1–11. Young Investigator Award, the 2009 National Science Foundation (NSF)
[20] T. L. Bergman, Fundamentals of Heat and Mass Transfer, 7th ed. Hobo- CAREER Award, the 2009 Presidential Early Career Award for Scien-
ken, NJ, USA: Wiley, 2011, p. 1048. tists and Engineers (PECASE), the 2012 ASCE EMI Leonardo da Vinci
[21] A. E. Bryson and Y.-C. Ho, Applied Optimal Control: Optimization, Esti- Award, and the 2013 ASCE Huber Award.
mation, and Control. Bristol, PA, USA: Hemisphere, 1975.
[22] E. F. Camacho and C. Bordons, Model Predictive Control. New York, NY,
USA: Springer-Verlag, 1999, p. 280. Jeff Scruggs (Member, IEEE) received the B.S.
[23] Y. Wang and S. Boyd, “Fast model predictive control using online opti- and M.S. degrees in electrical engineering from
mization,” IEEE Trans. Control Syst. Technol., vol. 18, no. 2, pp. 267–278, Virginia Tech, Blacksburg, VA, USA, in 1997 and
May 2010. 1999, respectively, and the Ph.D. in applied me-
[24] “Code Composer Studio.” Texas Instruments Incorporated, Dallas, TX, chanics from the Caltech, Pasadena, CA, USA,
USA, 2013. in 2004.
[25] M. Galassi and J. Theiler, GNU Scientific Library Reference Manual, 3rd He is an Associate Professor with the De-
ed. Network Theory Ltd., 2013. [Online]. Available: http://www.gnu.org/ partment of Civil & Environmental Engineering,
software/gsl/ the University of Michigan, which he joined in
[26] B. Tashtoush, M. Molhim, and M. Al-Rousan, “Dynamic model of an 2011. Prior to joining the University of Michigan,
HVAC system for control analysis,” Energy, vol. 30, no. 10, pp. 1729– he held postdoctoral positions at Caltech and
1745, Jul. 2005. the University of California, San Diego, CA, USA, and from 2007 to 11,
[27] N. Giorgetti, a. Bemporad, H. E. Tseng, and D. Hrovat, “Hybrid model he was on the faculty at Duke University. His research interests include
predictive control application towards optimal semi-active suspension,” the areas of mechanics, vibration, energy, and control. His research is
Int. J. Control, vol. 79, no. 5, pp. 521–533, May 2006. supported by NSF, ONR, and DOE.