Sie sind auf Seite 1von 11

IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 16, NO.

3, JUNE 2015 1199

Fast Online Computation of a Model Predictive


Controller and Its Application to Fuel
Economy–Oriented Adaptive Cruise Control
Shengbo Eben Li, Zhenzhong Jia, Keqiang Li, and Bo Cheng

Abstract—The recent progress of advanced vehicle control sys- performance maximization and cost minimization. Such a trend
tems presents a great opportunity for the application of model presents a great opportunity to use model predictive control
predictive control (MPC) in the automotive industry. However, (MPC) theory in automotive industry because of its exceptional
high computational complexity inherently associated with the
receding horizon optimization must be addressed to achieve capability in implementing optimal control while explicitly
real-time implementation. This paper presents a generic scale re- handling nonlinearities and constraints [1]–[4].
duction framework to reduce the online computational burden of A typical application of MPC in automotive field is the vehic-
MPC controllers. A lower dimensional MPC algorithm is formu- ular longitudinal automation, e.g. adaptive cruise control (ACC)
lated by combining an existing “move blocking ” strategy with a [2], [3]. Motivated by the growing concern on energy crisis,
“constraint-set compression” strategy, which is proposed to fur-
ther reduce the problem scale by partially relaxing inequality researchers started to explore fuel economy-oriented ACC tech-
constraints in the prediction horizon. The closed-loop stability is niques. This kind of design highly demands an optimal or quasi-
guaranteed by adding terminal zero-state constraint. The trade- optimal controller, which can possess desirable optimality of
off between control optimality and computational intensity is predefined cost function while still reserving the ability to
achieved by proper design of the blocking and compression ma- handle constraint and nonlinearity. Li et al. [5] have showed the
trices. The fast algorithm has been applied on intelligent vehicular
longitudinal automation, implemented as a fuel economy-oriented effectiveness of MPC-based ACC in improving fuel economy
adaptive cruise controller and experimentally evaluated by a series through simulation studies. However, the high computational
of real-time simulations and field tests. These results indicate that burden associated with the receding horizon optimization must
the proposed method significantly improves the computational be mitigated for real-time implementations.
speed while maintaining satisfactory control optimality without In MPC, a plant model is used to predict the state evolution
sacrificing the desired performance.
over a future horizon, whereas an optimization problem (min-
Index Terms—Adaptive cruise control (ACC), computation effi- imizing a cost function of state variables and control inputs
ciency, fuel economy, model predictive control (MPC). in the prediction horizon) is solved to determine the optimal
control sequence [6], [7]. One critical challenge for MPC is
I. I NTRODUCTION the computational burden in numerical optimization. This may
not be a problem in process control applications because of
I NTENDED to assist drivers during the driving process and
improve road safety, advanced vehicle control systems have
been investigated by both academia and industry for decades
relatively slow plant dynamics, infrequent control updates, and
abundant computing power [7]. It, however, becomes critical
to control some aspects of vehicle dynamics, e.g., antiskid for applications with fast dynamics (e.g., vehicles and robots)
traction/braking, desired yaw motion, and longitudinal/lateral and/or limited computing power (e.g., on-board single-chip
automation. Related technologies have also progressed from controllers).
previously basic functional realization to today’s focus on The computational issue of MPC motivates researchers to
explore more efficient computing techniques [7], some of which
are summarized as follows. An approach to achieve fast com-
Manuscript received May 21, 2013; revised October 31, 2013, January 22,
putation is to utilize the structural sparseness of predictive
2014, May 14, 2014, and August 10, 2014; accepted August 28, 2014. Date of controllers. By properly reordering the manipulated variables
publication September 16, 2014; date of current version May 29, 2015. This related to the sparse matrices, an interior-point method could
work was supported in part by the National Natural Science Foundation of
China under Grant 51205228 and in part by the Tsinghua University Initiative
become more efficient during directional search [8]. This type
Scientific Research Program under Grant 2012THZ0. The Associate Editor for of sparse formulation can be used to reduce computational time
this paper was A. Hegyi. of applications with large horizons without jeopardizing the
S. E. Li is with the State Key Laboratory of Automotive Safety and Energy,
Department of Automotive Engineering, Tsinghua University, Beijing 10084,
closed-loop stability. Another scheme is called explicit MPC,
China (e-mail: lisb04@gmail.com). which is often used for linear or piecewise affine linear plants
Z. Jia is with the Department of Naval Architecture and Marine Engineer- with 1-, 2- or ∞-norm-based costs and linear constraints [7]. In
ing, University of Michigan, Ann Arbor, MI 48109 USA (e-mail: zhenzjia@
umich.edu). this method, an explicit solution is generated offline by using
K. Li and B. Cheng are with the Department of Automotive Engineering, multiparametric programming [9]–[12]. The control input is
Tsinghua University, Beijing 10084, China. then calculated through a preprepared lookup table, thereby
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. greatly simplifying online computation. One major problem
Digital Object Identifier 10.1109/TITS.2014.2354052 associated with this approach is the large online data storage

1524-9050 © 2014 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.
1200 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 16, NO. 3, JUNE 2015

requirements for high-order controllers due to exponentially function with respect to the pair (x, u) and satisfies l(0, 0) = 0.
increasing complexity. Other constraints on x and u can also exist in the aforemen-
In engineering practices, an effective approach to reduce tioned problem but omitted here for narrative simplicity. For
the computational intensity is to reformulate the original MPC the sake of convenience, we introduce three column vectors,
problem as a lower order optimization problem by reducing i.e., Uk ∈ RP m , Umax ∈ RP m , and Vmax u
∈ RP m , defined as
the dimension of manipulated variables. Two examples are  T
the parameterization method [13] and the “move blocking Uk = uT0|k , . . . , uTP −1|k
(MB)” method [14]–[16]. The former often assumes a priori  T
Umax = uTmax , . . . , uTmax
knowledge of the control law, which can be approximated by  T
a parameterized function (with less unknowns to reduce the u
Vmax = (vmaxu
)T , . . . , (vmax
u
)T . (4)
order of the original problem), e.g., polynomial [13]. However,
this method may change the structure of the original problem Similar definition can be found for state vector Xk ∈ RP l .
due to specific parameterized functions used, thereby resulting Plugging (1) and (4) into (2) and (3), the MPC problem is
in pseudo-optimal solutions and losing the degree of control transformed into a nonlinear programming (NLP) problem [6]
optimality. In the MB method, the number of free variables
is reduced by fixing the input or its derivatives to be constant min J = L(Xk , Uk ) + ρε2k (5)
over several time steps by using a “blocking matrix” [14]. This
straightforward method can be easily integrated with other fast- subject to set Θ
 

computing schemes.  Uk
The purpose of this paper is to extend the MB scale re- Θ = Uk , εk [IU − Vmax
u
] ≤ Umax (6)
εk
duction strategy by introducing an additional “constraint-set
compression (CSC)” strategy capable of further reducing the where L(·) is assumed to be a convex function, i.e., IU = IP m
computational intensity of MPC for real-time implementation, (unity matrix).
particularly in the fields of engineering practice, e.g., auto- For an NLP problem, its computational intensity depends on
motive control techniques. The remainder of this paper is three aspects: 1) problem style; 2) problem scale; and 3) opti-
organized as follows. Section II reviews a simplified MPC mization algorithm to be used. The change of problem style and
formulation, followed by a brief analysis of its problem scale the improvement of optimization algorithm are not the focus
and computational complexity. Section III presents a generic of this paper. How to reduce the problem scale is more pre-
scale reduction framework for the MPC optimization problem, ferred in engineering practice due to its easy-to-use and in-field
together with the aforementioned MB strategy and the proposed flexibility. For example, the scale of a commonly constructed
CSC strategy. The MPC-based vehicular adaptive cruise con- optimization problem relies on two critical factors: the number
troller is presented in Section IV. Application of the proposed of manipulated variables NA and the number of inequality con-
fast MPC algorithm to ACC is evaluated through real-time straints NB [17]. In the Dantzig–Wolfe algorithm introduced in
simulation and field tests in Sections V and VI, respectively. [18], the required iteration is often no less than Niter , which is
The last section concludes this paper. equal to the sum of NA and NB . Hence, the scale reduction is
expected to reduce the computing time effectively.
II. P ROBLEM S TATEMENT
Consider a nonlinear (but affine with respect to the input) III. F RAMEWORK OF P ROBLEM S CALE R EDUCTION
system in the discrete-time domain
The scale reduction framework consists of two parts, i.e.,
xk+1 = f (xk ) + g(xk ) · uk (1) “MB” and “CSC.” The former aims to reduce the number of
manipulated input variables in the predictive horizon, whereas
where u ∈ Rm is the control input, x ∈ Rl is the system state, the latter aims to reduce the number of inequality constraints.
and f (·) and g(·) are nonlinear functions of x. The pair (x, u) =
(0, 0) is the exclusive equilibrium point of (1).The receding
horizon optimization problem has the following form: A. MB Strategy


P −1 A common strategy to reduce the computational complexity
min J = l(xi+1|k , ui|k ) + ρ · ε2k . (2) of optimal control is to reduce degrees of freedom by fixing
u
i=0 the control input (or the control increment) to be constant over
several steps. This policy is referred to as “MB” [14]. Instead
Subject to (1) and constraint
of solving for the optimal Uk ∈ RP m , problem (5) is restated
ui|k ≤ umax + εk · vmax
u
, i=0:P −1 (3) in terms of solving for a lower order vector
 T
where P is the length of the prediction horizon, xi|k is the Zk = z0|kT T
, . . . , zQ−1|k ∈ RQm (Q < P ) (7)
predicted state using the measured (or estimated) state xk , ε ∈
R+ is the slack variable, ρ ∈ R+ is the weighting coefficient, so that
umax ∈ Rm is the upper bound of u, and vmax u
∈ Rm is the re-
laxing coefficient [6]. The function l(x, u) is a positive definite U k = MT · Z k , MT = T ⊗ I m . (8)
LI et al.: COMPUTATION OF A MODEL PREDICTIVE CONTROLLER AND ITS APPLICATION TO FUEL ECONOMY–ORIENTED ACC 1201

Fig. 2. Illustration of “CSC” strategy (take m = 1 as an example).


Fig. 1. Illustration of “MB” strategy (take m = 1 as an example).
single-step computational time), which must be limited within
In (8), ⊗ denotes the Kronecker product and T ∈ R Q×P
is a an acceptable range (< 100 ms in this paper). However, the
so-called “blocking matrix” with Q < P [14]. Matrix T is often numerical computation of MPC is more likely to take a longer
assumed to be fully column ranked and consisted of only ones time when constraints become active. Hence, a more effective
and zeros with each row containing exactly one nonzero element. strategy to reduce the computational complexity is to reduce the
Recall that only the first entry uk+0|k in Uk will be applied to number of active constraints.
the plant, whereas all other entries (i.e., uk+i|k , i = 1, . . . , P − In addition, it is also not necessary in practice to impose
1) are not used for direct control [6]. Thus, it is not necessary restrictions on every data point in the prediction horizon, since
to calculate the exact value of every entry in Uk except for the only the constraints at the first time step directly affect the
first element uk+0|k . Hence, a blocking matrix (see Fig. 1) can actual control inputs. Hence, constraints at some selected steps
be formulated as can be deactivated in the prediction horizon to reduce the
⎡ ⎤ dimension of the constraint set, thus benefiting the computa-
1 0 0 0 tional intensity. This method is referred to as “CSC” strategy
⎢ 0 1q1 ×1 0 0 ⎥ (i.e., compress or squeeze the constraint set). The concept is
T =⎣ ⎦ (9)
0 0 1q2 ×1 0 illustrated in Fig. 2.
0 0 0 1qQ−1 ×1 P ×Q Analogous to (8), a “compression matrix” is formulated to be
where 1qj×1 is a qj -dimensional column vector of ones only, ⎡ ⎤
1 0 0 0
and qj ∈ N + (j = 1, . . . , Q − 1) is the length (i.e., number of ⎢ 0 π1 0 0 ⎥
steps) of the (j + 1)th interval, in which the input is hold as Π=⎣ ⎦ (Ω < P ) (12)
0 0 π2 0
constant, and 0 0 0 πΩ−1 Ω×P

Q−1
with πj defined as
qj = P − 1. (10)
j=1
πj = [1, 0, . . . , 0] (j = 1, . . . , Ω − 1) (13)
  
In Fig. 1, let Uk∗ ∈ RP m and Zk∗ ∈ RQm denote solutions to ωj

the optimization problem (5) before and after the MB method where ωj is the length of interval, in which all the constraints
(8) being applied, respectively. Consider vector U Z defined as except those of the first step are deactivated (see Fig. 2), and Ω
U Z = MT · Zk∗ ∈ RQm . (11) is the number of steps, in which the constraints are enabled in
the prediction horizon. Note that Π is a full-row rank matrix.
We see that U Z is actually a staircase approximation of Uk∗ . Define a linear compressing mapping from the original con-
The length (i.e., qj ) of each interval can be different from each straint set Θ in (6) to a lower dimensional constraint set Φ as
other, i.e., they do not need to be equally distributed. Smaller qj 
indicates a more accurate approximation and accordingly less φ(MΠ ) : Θ → Φ
(14)
loss of optimality after the MB transformation. In engineering MΠ = Π ⊗ I m .
practice, we can select small qj ’s for points that are close to the A lower dimensional problem can be formulated by applying
current step and large qj ’s for those relatively far away. the compressing mapping (14) to deactivate some inequality
constraints. In practice, a properly selected compression matrix
B. CSC Strategy Π can significantly reduce the computational intensity (par-
Under some circumstances, the computational intensity is ticularly the maximum single-step computational time) with
still challenging even after the MB strategy being applied, negligible effect on the optimality of solutions.
as shown in the following real-time simulation in Section V.
One major reason is that the dimensional reduction of ma- C. Formulation of Fast MPC Algorithm
nipulated variables should be not too excessive when using To construct a fast MPC algorithm, we apply both the MB
MB in case of large loss of control optimality. In a plant with and CSC strategies into (5) and (6), yielding
fixed controlling frequency, e.g., ACC, a critical factor for real-
time implementation is the peak CPU time (i.e., the maximum min J = L(Xk , MT · Zk ) + ρε2k (15)
1202 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 16, NO. 3, JUNE 2015

subject to although it might cause some additional computation burden.


 
An improvement of zero-state constraint is called dual-mode
 MT Z k

Φ = Zk , εk [MΠ −MΠ Vmax ]
u
≤ MΠ Umax . (16) method, which replaces the equality constraints by inequalities
εk for the terminal states [23]. In the dual-mode control, the MPC
switches to a stable linear control, e.g., LQ, once the terminal
Matrix MT has a full column rank, and (15) maintains its states in the prediction horizon enter a suitable neighborhood
original convexity after transformation. The MB transformation of the equilibrium points. This method reduces the possibility
reduces the number of manipulated variables from P m + 1 of infeasibility at the expense of frequent switching on the
to Qm + 1, whereas the compressing mapping reduces the bounds of the neighborhood. Another stabilizing method is
number of inequality constraints from P m to Ωm. Hence, to use the terminal contractive constraints, which require that
the overall scale reduction ratio is approximately P 2 /QΩ the norm of the terminal states is sufficiently smaller than
(analogous to the gear ratio). Moreover, the MB matrix and that of the initial states in the prediction horizon [24]. The
CSC matrix can be independently designed. The MB strategy soft constraint is widely used in practical implementations of
is inherently to reduce the dimension of optimized variables. MPC; however, its stability analysis has received relatively little
In the receding horizon optimization, the peak, rather than attention. In [25], MPC with mixed hard and soft constraints
average, on-board computing time determines the feasibility of is considered and stability is proven by properly extending the
real-time implementation. In this sense, the MB is less effective stability proof in normal MPC. In [26], a soft constrained MPC
than the CSC strategy. The reason is that the computing time approach with guaranteed stability is proposed by enlarging
critically depends on the number of active constraints, which the terminal constraint and relaxing the state constraints. The
could be significantly reduced by directly relaxing some of approach can ensure the feasibility in a large region, and the
the constraints in the prediction horizon. Both the MB and optimal performance is preserved whenever all state constraints
CSC strategies are easy to use for engineers and are relatively can be enforced.
flexible in parameter tuning. Their performance also relies on In the following, we will show that the zero-state terminal
the design of the blocking and compression matrices, thereby constraint can stabilize the problem (15) even when the MB and
presenting customers with some challenges. A trial-and-error CSC strategies are applied. To add the terminal constraint, let
method can be used in practical engineering design to balance
the computational efficiency and other demands, e.g., stability xi+1|k = 0, i = P − 1. (17)
and optimality.
The keystone of the proof is to establish the monotonicity
property [24], which will be shown in (22).
D. Discussion on Optimality and Stability Proof of Stability: Define a Lyapunov function for the MPC
The MB and CSC strategies will lead to the loss of optimality problem after using MB and CSC strategies
because they actually change the original predictive problem. A
V ∗ (x∗k , zk∗ , ε∗k , k) = min L(Xk , MT · Zk ) + ρε2k . (18)
desirable balance between the computational speed and the loss zk+i|k,,i=1:Q ,εk
of optimality can be achieved by properly tuning the MB and
CSC matrices. Another, but more important, topic is the issue This is the Lyapunov function at the kth step, in which
of closed-loop stability after using MB and CSC strategies. “∗”denotes that V is calculated from optimized solutions. To
As illustrated in many examples/papers, MPC cannot guaran- calculate the Lyapunov function for the (k + 1)th step, we can
tee the closed-loop stability for finite-horizon cost functions, construct a virtual control sequence as
particularly for short prediction horizon. Designers often add  ∗
zi+1|k , i = 0, . . . , Q − 2
additional terms into the cost function and/or constraints to zi|k+1 =
0, i=Q−1
stabilize the closed-loop system by somehow reformulating εk+1 = ε∗k . (19)
the MPC problems. Key mechanisms to guarantee stability
fall into two categories: 1) extending the prediction horizon; It is easy to check that (19) is a feasible solution for the
2) incorporating proper additional conditions, i.e., penalty (k + 1)th step, which satisfies (16). Note that (19) is equivalent
and/or terminal constraints [16], [19], [20]. Here, we focus to the case that ui|k+1 = 0 when i = P − qQ−1 : P − 1 by
on the latter case because the former has disadvantages in considering the MB transformation. By using the zero-state
computational efficiency. terminal constraint (17), and the fact that the pair (x, u) =
Using additional conditions to ensure the stability of linear (0, 0) is the exclusive equilibrium point of (1), we have
MPC is often easier than the nonlinear cases. Commonly used  ∗
approaches include complementing the cost function with a xi+qQ−1 |k , i = 0 : P − qQ−1 − 1
xi|k+1 = . (20)
zero-state terminal constraint [20], adding a terminal penalty 0, i = P − qQ−1 : P − 1
which is equal to the infinite-horizon cost [21]. Although
Then
there is well-established theory for linear cases, its importance
diminishes because MPC with nonlinearities and constraints is V ∗ (x∗k , zk∗ , ε∗k , k) = V (xk+1 , zk+1 , εk+1 , k + 1)
hardly better than infinite-horizon linear quadratic (LQ) control. qQ−1 −1   −1
 
P
For nonlinear cases, it has been proven in [22] and [23] that the + l x∗i|k , zi|k

− l(xi|k+1 , zi|k+1 ). (21)
zero-state terminal constraint guarantees closed-loop stability, i=0 i=P −qQ−1
LI et al.: COMPUTATION OF A MODEL PREDICTIVE CONTROLLER AND ITS APPLICATION TO FUEL ECONOMY–ORIENTED ACC 1203

From (19) and (20), and the fact that l(·, ·) is a positive
definite function with respect to pair (x, u), we have
V ∗ (x∗k , zk∗ , ε∗k , k) ≥ V (xk+1 , zk+1 , εk+1 , k + 1). (22)
Using the optimality property [24], we always have
 
V (xk+1 , zk+1 ,εk+1 ,k+1) ≥ V ∗ x∗k+1 , zk+1

,ε∗k+1 ,k+1 . (23)
Therefore, the defined Lyapunov function monotonically de-
creases when increasing the control step
 
V ∗ (x∗k , zk∗ , ε∗k , k) ≥ V ∗ x∗k+1 , zk+1

, ε∗k+1 , k + 1 . (24)
Hence, the closed-loop stability is ensured (end of proof).
Note that the proof of stability uses the so-called feasibility Fig. 3. Hierarchical control architecture, consisting of an upper layer con-
property: a feasible control consequence at the (k + 1)th step troller and a lower layer controller.
is obtained by appending zeros to previous optimal control and state constraints. The lower level controller is used to deal
sequence. An interesting property is that once the feasibility with nonlinearities of vehicle dynamics and errors in modeling.
property holds, the existence of inequality constraints will Interested readers can refer to [5] and [27] for more details. The
not affect the closed-loop stability. In addition, the proposed emphasis in this paper is to implement the MPC-based upper
CSC strategy always relaxes inequality constraints, instead lever controller in a real-time manner.
of strengthening them. Therefore, the introduction of CSC
actually does not affect the closed-loop stability when using A. Continuous-Time Model for Controller Design
zero-state terminal constraint. It should be noted that other The vehicle to be controlled is a passenger car with 2.0-L
aforementioned approaches can also guarantee closed-loop sta- gasoline engine, five-speed automatic transmission and hy-
bility. In practical applications, however, those approaches are draulic braking system. Its longitudinal dynamics includes
not always used because they enforce unnecessary constraints static nonlinearity of the engine, discontinuous gear ratio, and
and/or penalty to the original problem. In this case, a posteriori quadratic aerodynamic drag. These nonlinearities are compen-
check of stability via simulations can be used [7], [14]. The sated by using an inverse model in the lower level controller by
closed-loop stability can be achieved if not using rather worse neglecting the powertrain dynamics. In addition, a switching
blocking and compression matrices. logic with hysteresis is built to avoid simultaneous actions of
driving and braking. The outputs of the lower level controller
IV. M ULTIOBJECTIVE -BASED ACC are the acceleration pedal position aaccl and the brake pressure
This section summarizes the design of a MPC-based multiple Pbrk . A proportional–integral (PI)-type feedback controller
objective ACC system, with special interest in reducing fuel is incorporated to track the desired acceleration af des . The
consumption, which is expected to have a large penetration into combination of vehicle longitudinal dynamics, inverse model,
the markets in the near future. It should be noted that the term and PI-type controller is approximated as a first-order transfer
“multiple objective” is not used in the sense of optimization, but function, with a gain KL and a time constant TL . This transfer
for engineering practices. It means that several goals are consid- function, together with intervehicle dynamics, is viewed as the
ered in the ACC design, such as good fuel economy, necessary controlled plant for the upper level controller. Its state-space
tracking capability, satisfactory driver feelings, and rear-end model is formulated to be [5], [27]
safety. However, to reduce acceleration alone is not sufficient ẋ = A(vf ) · x + B · u + G · ξ
in the design of ACC algorithms. The smoothing acceleration x = [⎡Δd Δv af ]T , u = af des , ξ = a⎤
p
could improve fuel economy; however, it also weakens the
0 1 −τh − r(2vf − vMean )
tracking capability and enlarges the range error. Moreover,
A = ⎣0 0 −1 ⎦
drivers are essential to evaluate the efficacy any ACC system;
⎡0 0 ⎤ −1/T
⎡ L⎤
ACC will be completely useless if drivers choose not to use
0 0
(e.g., due to undesirable ride comfort) no matter how it works in
B = ⎣ 0 ⎦, G = ⎣1⎦ (25)
fuel economy and tracking capability. Therefore, it is critical to
KL /TL 0
systematically consider multiple objectives, i.e., fuel economy,
tracking capability, and driver feelings, instead of focusing on a where [Δd, Δv, af ]T is the system state (distance error, relative
single one. The hierarchical control structure is shown in Fig. 3. speed, and longitudinal acceleration); af des is the control input
The upper level controller, synthesized by the MPC theory, aims (the desired acceleration); ap is the measurable disturbance
to reduce fuel consumption while comprehensively balancing (the acceleration of the preceding vehicle); vf is the vehicle
other two objectives, i.e., tracking capability and driver feel- speed; and τh , r, and vMean are the parameters of the quadratic
ings. Other demands, e.g., ride comfort, rear-end safety, and headway policy, representing the headway time, quadratic co-
physical limits of the vehicle, are formulated as constraints. The efficient, and average speed, respectively. Notice that (25) is
selection of MPC mainly attributes to its special capability in quasi-linear. The sampling rate if 10 Hz is used to linearize (25)
achieving optimal control while directly handling various input (i.e., sampling time Ts = 100 ms).
1204 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 16, NO. 3, JUNE 2015

B. Formulation of MPC Optimization Problem C. Constraint Softening to Address Computing Infeasibility


The upper level controller is formulated into a MPC problem A key issue with above-designed MPC is the potential “com-
by synthesizing multiple performance indexes such as fuel con- puting infeasibility” in numerical optimization because some of
sumption, tracking capability, and driver feeling, as well as in- the strict constraints on inputs and states can never be satisfied
equality constraints to reflect ride comfort, rear-end safety, and (they are referred to as hard constraints).
physical limitation [5]. The formulation of cost function and The bounds of hard constraints are not allowed to be violated,
constraints needs to achieve the following goals: 1) zero steady- but the tracking errors in ACC sometimes unavoidably reach
state tracking errors; 2) in accelerating cases, the intervehicle their bounds in highly dynamic traffic scenarios. To avoid the
states should be controlled within the permissible tracking saturation of tracking errors, a natural decision is to strengthen
range to avoid frequent cut-ins from adjacent lanes; 3) in brak- the acceleration or deceleration. This is often impossible in real-
ing cases, a rear-end collision must be avoided; 4) to depress ity because the control input is often located in its upper bound
acceleration to lower fuel consumption; and 5) The desired in an accelerating scenario, leaving no space to be strengthened.
responses or dynamics should satisfy driver desired distance The same is true for a braking scenario. Consequently, the states
feature, driver longitudinal ride comfort, and driver desired keep increasing or decreasing along original directions because
dynamic characteristics. The cost function (based on quadratic of vehicle body inertia. Once any bound saturation occurs, the
norm) in the predictive horizon k · Ts < t < (k + P ) · Ts is optimization problem becomes infeasible.
The constraint softening method in [6] is used to relax some
 )·T s
(k+P
hard constraints. Add a slack variable to (26), yielding a new
J= L(Δd, Δv, af , af des , jf des )dt cost function
k·T s
L = wyΔd Δd 2 + wyΔv Δv 2 Ψ = J + ρ · ε2 (ε ≥ 0) (30)
+ wu af des 2 + wdu jf des 2
where ρ is the weighting coefficient. The constraints (27) and
+ wya af R − af 2 (26)
(28) are transformed into soft constraints as

where Ts = 100 ms is the sampling time, k · Ts is the ⎪ a
af
+ ε · vmin
af
≤ af des ≤ af max + ε · vmax

⎨ f min
current time, P · Ts is the prediction length, and af R is the af af
af min + ε · vmin ≤ af ≤ af max + ε · vmax
reference acceleration from a microscopic car-following model (31)

⎪ Δdmin + ε · vminΔd
≤ Δd ≤ Δdmax + ε · vmax Δd
for average human drivers (with coefficients formulated as ⎩
Δvmin + ε · vmin ≤ Δv ≤ Δvmax + ε · vmax
Δv Δv
functions of vehicle speed vf ). The constraints in the predictive
#
horizon k · Ts < t < (k + P ) · Ts are where vmin #
and vmax are the relaxation coefficients with the
a ≤a ≤a subscript number representing parameters such as Δd and Δv.
f min f des f max
jf min ≤ jf des ≤ jf max (27) A new optimization problem is formulated to be
 af min ≤ af ≤ af max min Ψ(t), t ∈ [k · T s, (k + P ) · T s]
Δdmin ≤ Δd ≤ Δdmax u(t)
(28)
Δvmin ≤ Δv ≤ Δvmax subject to
d ≥ dsafe , dsafe = max{TTC · Δv, ds0 }. (29)
(a) car-following model : (25)
In aforementioned equations, af min , af max , jf min , and (b) soft constraints : (31)
jf max are bounds of longitudinal acceleration and jerk; dsafe (c) hard constraints : (29). (32)
is the distance safety threshold; Δdmin , Δdmax , Δvmin , and
Δvmax are the bounds of tracking errors; and TTC is shorted In Problem (32), the slack variable ε will become positive if
from “time-to-collision”, which is used to restrain the mini- the hard constraints (27) and (28) cannot be satisfied. In this
mum distance to maintain a safe intervehicle clearance. The case, hard constraints (27) and (28) will be replaced with the
constraints on acceleration and jerk are used to meet the driver soft constraints (31), in which the bounds are extended to guar-
ride comfort demand. In naturalistic driver experiments, vehicle antee feasible solutions. This implies the violation of original
cut-ins and driver interventions are less likely to occur if the hard constraints. However, the violation is not completely free
tracking errors are bounded, thereby smoothing the tracking and the violation degree is limited by penalizing ε in the cost
process and helping to reduce the fuel consumption [5]. The function, so that a balance between constraint violation and
turning of weighting coefficients wyΔd , wyΔv , wu , wya , and control optimality is expected. When all hard constraints hold,
wdu in (26) affects the performance of MPC controllers, in the slack variable ε becomes zero and Problem (32) can be
particular on how to balance different objectives. As shown in reduced to original problem without any softening constraints.
Table II, wya is intentionally selected to be large to increase In addition, it must be pointed out that the hard constraint
the penalty on vehicle acceleration, and wyΔd and wyΔv are for safety [i.e., (29)] is not softened here because any rear-
selected to be small to relax the limitation on tracking errors. end collision should be avoided in the car-following process.
Nevertheless, the relaxation on tracking errors is not unlimited, Due to the demand on fuel economy, vehicle acceleration is
but confined by (28) to satisfy driver expected range. This both constrained and penalized in the performance index. This
choice is useful to depress the acceleration level, thus helping design might cause potential risk of rear-end collision in highly
improving fuel consumption of ACC vehicle. dynamic traffic conditions. Note that (29) requires that the
LI et al.: COMPUTATION OF A MODEL PREDICTIVE CONTROLLER AND ITS APPLICATION TO FUEL ECONOMY–ORIENTED ACC 1205

TABLE I
T HREE MPC A LGORITHMS I NVOLVED IN R EAL -T IME S IMULATION

TABLE II
K EY PARAMETERS FOR THE MPC-BASED U PPER L EVEL C ONTROLLER

TABLE III
K EY PARAMETERS FOR THE MB AND CSC S TRATEGIES

minimum intervehicle distance must be greater than the safety


threshold dsafe during the entire prediction horizon. When a
rear-end collision risk emerges, the intervehicle distance rapidly
reduces because of the high-speed difference between the front
and following vehicles. In order to satisfy (29) in the prediction
horizon, the optimization of Problem (32) will automatically
relax soft constraints (31), including the lower bound of acceler-
ation, by increasing the slack variable. Consequently, the MPC
will strengthen the deceleration by applying more braking force
to avoid potential rear-end collision. This balancing mecha-
nism helps to reduce the risk of rear-end collisions caused
by restraining the vehicle acceleration for high fuel economy
considerations.

V. R EAL -T IME S IMULATION OF FAST MPC-BASED ACC


Three MPC algorithms are investigated through real-time
simulation for performance evaluation. As shown in Table I,
algorithm A (the baseline) has no scale reduction strategy;
algorithm B includes only the MB strategy; and algorithm Fig. 4. Simulation results of three MPC algorithms. (a) Acceleration.
(b) Relative speed. (c) Distance error. (d) Slack variable.
C includes both the MB and CSC strategies. The simulation
platform is constructed based on an industrial computer with
A. Case 1: Constant Accelerating Scenario
2.4-GHz CPU and 512-MB memory, whereas the software is
designed based on the MATLAB/xPC toolbox. The MPC algo- The leading vehicle runs first at 15 m/s and starts accelerating
rithms are implemented as S-functions with codes programmed 5 s at the deceleration of −2 m/s2 until to the speed 1 m/s. The
in C/C++ for fast computation. All the numerical optimization ACC car is controlled to achieve fuel economy-oriented ACC.
problems in the MPC are solved by using the Dantzig–Wolfe Simulation results of the control input (af des ), relative speed
optimization method [18]. Some key parameters for the MPC- (Δv), distance error (Δd), and slack variable (ε) are given in
based upper level controller are given in Tables II and III. The Fig. 4. The x-marked line, solid line, and dashed line represent
scale reduction ratios for MB and CSC are around 50/12 (P/Q) algorithm A, B, and C, respectively. The dotted line represents
and 50/26 (P/Ω), respectively. original upper/lower bounds (i.e., hard constraints). The three
1206 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 16, NO. 3, JUNE 2015

Fig. 5. Computational efficiency of three MPC algorithms. (a) Computational


time at each time step. (b) Peak and mean computing time.

TABLE IV
C OMPARISON OF C OMPUTATIONAL E FFICIENCY

algorithms yield extremely close outputs, indicating quite simi-


lar performance on the optimality of the solutions. Note that the
xPC kernel also consumes some CPU time to support the code
execution. It is compensated for by running a prosimulation
without any MPC algorithm.
Fig. 5 compares the computational efficiency of these three
algorithms: 1) computational time et each time step; and 2) peak
and mean computing time. Table IV summarizes overall com-
putational performance. The critical factor of the MPC-based
ACC is the peak CPU time because the MPC algorithm must be
successfully executed during a fixed time step (Ts = 100 ms).
With a 50/12 (≈4.2) scale reduction ratio in MB, the peak CPU
time is reduced to 53% of that of the baseline algorithm. The Fig. 6. Computing errors between three MPC algorithms. (a) Error of control
proposed CSC strategy further cuts the peak CPU time by 65% input. (b) Error of relative speed. (c) Error of distance error. (d) Error of slack
variable.
with only a 50/26 (≈1.9) scale reduction ratio (relative to B).
Hence, the CSC method is more effective in improving the
comutational speed, thereby enabling real-time implementation the nominal values in Fig. 4, these absolute errors are quite
of MPC. small. Notice that the CSC strategy somewhat affects the pre-
The solution generated by algorithm A is considered to be cision of slack variable ε (with a relative error around 1%),
optimal since no scale reduction is used. Compared with algo- but with very limited effect on optimal solutions. Therefore,
rithm A (baseline), the relative input/state errors of algorithm B it can be concluded that the proposed fast MPC algorithm
and C are given in Fig. 6. For algorithm B and C, the absolute (by integrating the MB and CSC strategies) can improve the
errors of af des , Δv, and Δd are less than 0.5 × 10−2 m/s2 , computational speed significantly with negligible effect on the
0.5 × 10−2 m/s and 1 × 10−2 m, respectively. Compared with optimality of MPC solutions.
LI et al.: COMPUTATION OF A MODEL PREDICTIVE CONTROLLER AND ITS APPLICATION TO FUEL ECONOMY–ORIENTED ACC 1207

Fig. 7. Speed profiles from naturalistic traffic flow. Fig. 9. Distribution of Δv and its computation errors. (a) Relative speed.
(b) Computation errors.

Fig. 8. Computational efficiency of three MPC algorithms. (a) Average. Fig. 10. Distribution of Δd and its computation errors. (a) Distance error.
(b) Maximum. (b) Computation errors.

B. Case 2: Naturalistic Traffic Flow Scenario TABLE V


C OMPARISON OF L OSS OF O PTIMALITY
The effectiveness of the proposed fast MPC algorithm is also
tested in the naturalistic traffic flow scenarios. The purpose is
to evaluate the effect of the MB and CSC strategies on the
control optimality in a comprehensive manner. The control
optimality is measured by the errors of inputs or states in
algorithm B (MB) and C (MB+CSC) compared with algorithm
A (baseline). Fig. 7 presents the speed profiles of the leading
vehicle in naturalistic traffic flow tests. The profiles are acquired
from driver test data in and around Beijing, including both city
road (the fourth ring road) and highway (the JingJun highway)
tests. The speed profiles of four drivers are used with time
length around 55, 27, 50, and 57 min, respectively (totally in less than 0.03% relative speed error and less than 0.04%
around 3 h). distance error. Algorithm C achieves less than 0.08% relative
Fig. 8 presents the computing times of three MPC algo- speed error and less than 0.048% distance error. Therefore, the
rithms. The MB strategy reduces the average computing time MB and CSC strategies have negligible effect (with less than
by 45.1% and the additional use of CSC further reduces the 0.1% in errors) on the control optimality in the representative
average time by 34.6% (totally by 79.7%). For the maximum naturalistic traffic flow tests. It is summarized that the proposed
computation time (i.e., the peak time), the MB strategy reduces fast MPC could maintain the desired control optimality in
the value by 43.5%, and the additional use of CSC further different driving scenarios.
reduces the value by 46.8% (totally by 90.3%). Hence, the
proposed CSC strategy is more effective in reducing the peak
VI. F IELD E XPERIMENTS AND D ISCUSSIONS
time, which is also more critical for real-time implementations.
This accords with our analysis on the pros and cons of CSC The fast MPC algorithm proposed in Section III is further
and MB. evaluated through field tests. As shown in Fig. 11, the ex-
Fig. 9 shows the distribution of Δv in algorithm A and perimental platform is built based on a passenger car with
the computation errors in algorithm B/C; Fig. 10 shows the identical parameters as before. The millimeter wavelength radar
distribution of Δd in algorithm A and the computation errors in is equipped to detect frontal moving or static targets, and two
algorithm B/C. Table V summarizes their standard deviations actuators, including an electronic throttle and electronic vac-
(both states and their computation errors) as the comparison uum booster, are used to track desired engine torque commands
index to evaluate the loss of optimality. Algorithm B results and braking force.
1208 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 16, NO. 3, JUNE 2015

Fig. 11. Configuration of experimental platform. Fig. 13. Powertrain state of ACC vehicle, including acceleration pedal angle,
engine speed in revolution per minute, gear range, and engine fuel rate.

Fig. 14. Computing time of fast MPC controller in each step.

straints. To satisfy driver feeling, the ACC limits af des and


af within an acceptable range (around af max ) while enlarging
the slack variable to avoid infeasible solutions. The computing
Fig. 12. Intervehicle and vehicle state, including acceleration, relative speed, time, which critically relies on active constraints in the pre-
distance error, and slack variable. diction horizon, does not exceed 40 ms in the whole process.
Therefore, the proposed fast MPC algorithm could satisfy the
The xPC controller is identical in setup to the one used in computational speed requirement. It should be pointed out that
real-time simulation. In field tests, aside from executing the algorithm A and algorithm B might not satisfy the 100-ms
MPC algorithm (with “MB+CSC” strategy), it also fulfills other time requirement for single-step calculation, which is not ex-
functions including radar signals processing, target vehicle perimentally involved in field tests due to safety concern.
selection, lower level controller, inverse vehicle model calcula-
tion, throttle/brake switching logic, and controller area network
VII. C ONCLUSION
(CAN) communication. They are built by MATLAB/Simulink
as separated functional modules. This paper has presented an effective scale reduction frame-
A typical accelerating scenario was experimentally studied in work for computationally intensive MPC controllers to enable
order to systematically test predefined multiple objectives, e.g., their real-time implementation in engineering practice, and also
fuel economy, tracking capability, and driver feeling, which evaluates its effectiveness based on a fuel economy-oriented ve-
also outputted the overall computing time of xPC controller at hicular ACC system. The “MB” strategy, which aims to reduce
each time step. Fig. 12 shows the real acceleration (af ), the the number of manipulated variables; and the proposed “CSC”
relative speed (Δv), the distance error (Δd), and the slack strategy, which aims to squeeze the constraint set by relaxing
variable (ε). Fig. 13 shows the accelerator position, engine some of the inequality constraints in the prediction horizon,
speed, gear position, and fuel injection rate. The computing are used to improve the computational efficiency. Satisfying
time as each time step is given in Fig. 14. the tradeoff between control optimality and the computational
In this scenario, the preceding vehicle accelerates at 0.6 m/s2 , speed can be achieved by properly designing the blocking and
which exceeds the boundary af max (0.5 m/s2 ) in the con- compression matrices.
LI et al.: COMPUTATION OF A MODEL PREDICTIVE CONTROLLER AND ITS APPLICATION TO FUEL ECONOMY–ORIENTED ACC 1209

R EFERENCES [26] N. Zeilinger, M. Morari, and C. Jones, “Soft constrained model predictive
control with robust stability guarantees,” IEEE Trans. Autom. Control,
[1] F. Borrelli, A. Bemporad, M. Fodor, and D. Hrovat, “An MPC/hybrid
vol. 59, no. 5, pp. 1190–1202, May 2014.
system approach to traction control,” IEEE Trans. Control Syst. Technol.,
[27] S. E. Li, K. Li, and J. Wang, “Economy-oriented vehicle adaptive cruise
vol. 14, no. 3, pp. 541–552, May 2006.
control with function of multiple objectives coordination,” Veh. Syst. Dyn.,
[2] B. Bageshwar, W. Garrard, and R. Rajamani, “Model predictive control of
vol. 51, no. 1, pp. 1–17, Jan. 2013.
transitional maneuvers for adaptive cruise control vehicles,” IEEE Trans.
Veh. Technol., vol. 53, no. 5, pp. 365–374, Sep. 2004.
[3] D. Corona, M. Lazar, B. Schutter, and M. Heemels, “A hybrid MPC
approach to the design of a smart adaptive cruise controller,” in Proc.
IEEE Int. Conf. Control Appl., Munich, Germany, 2006, pp. 231–235. Shengbo Eben Li received the M.S. and Ph.D. de-
[4] P. Falcone, F. Borrelli, H. E. Tsengz, J. Asgari, and D. Hrovat, “A hi- grees from Tsinghua University, Beijing, China, in
erarchical model predictive control framework for autonomous ground 2006 and 2009, respectively.
vehicles,” in Proc. Amer. Control Conf., Seattle, WA, USA, 2008, He is an Assistant Professor with the Department
pp. 3719–3724. of Automotive Engineering, Tsinghua University.
[5] S. Li, K. Li, R. Rajamani, and J. Wang, “Model predictive multi-objective His active research interests include nonlinear opti-
vehicular adaptive cruise control,” IEEE Trans. Control Syst. Technol., mal control, autonomous vehicle control, and lithium
vol. 19, no. 3, pp. 556–566, May 2011. ion battery management. He is the author of more
[6] J. Maciejowski, Predictive Control With Constraints. London, U.K.: than 60 journal/conference papers and the coinventor
Pearson Education, 2002. of more than ten patents.
[7] J. Sun, R. Ghaemi, and I. Kolmanovsky, “Developments in receding hori- Dr. Li received the Award for Science and Tech-
zon optimization-based controls: Towards real-time implementation for nology of China ITS Association (2012), Award for Technological Invention
nonlinear systems with fast dynamics,” in Advances in Control Theory in Ministry of Education (2012), National Award for Technological Invention
and Applications, G. Tao and J. Sun, Eds. Berlin, Germany: Springer- in China (2013), and Honored Funding for Beijing Excellent Youth Researcher
Verlag, 2008. (2013).
[8] Y. Wang and S. Boyd, “Fast model predictive control using online opti-
mization,” IEEE Trans. Control Syst. Technol., vol. 18, no. 2, pp. 267–278,
Mar. 2010.
[9] A. Bemporad, M. Morari, V. Dua, and E. N. Pistikopoulos, “The explicit
linear quadratic regulator for constrained systems,” Automatica, vol. 38, Zhenzhong Jia received the B.E. degree in mea-
no. 1, pp. 3–20, Jan. 2002. surement, control technology, and instruments and
[10] F. Borrelli, A. Bemporad, and M. Morari, “A geometric algorithm for the M.E. degree in mechanical engineering from
multi-parametric linear programming,” J. Optim. Theory Appl., vol. 118, Tsinghua University, Beijing, China, in 2005 and
no. 3, pp. 515–540, Sep. 2003. 2007, respectively, and the M.S. degree in mechani-
[11] L. Imsland, N. Bar, and B. Foss, “More efficient predictive control,” cal engineering, the M.S. degree in mathematics, and
Automatica, vol. 41, no. 8, pp. 1395–1403, Aug. 2005. the Ph.D. degree in naval architecture and marine
[12] A. Alessio and A. Bemporad, “A survey on explicit model predictive engineering, from the University of Michigan, Ann
control,” in Proc. Int. Workshop Assessment Future Directions Nonlinear Arbor, MI, USA, in 2009 and 2014, respectively.
Model Predictive Control, Pavia, Italy, Sep. 2008, pp. 345–369. His research interests are in the field of robotics,
[13] A. Zheng, “A computationally efficient nonlinear MPC algorithm,” in controls, and hybrid power systems.
Proc. ACC, Albuquerque, NM, USA, 1997, pp. 1623–1627.
[14] R. Cagienard, P. Grieder, E. Kerrigan, and M. Morari, “Move blocking
strategies in receding horizon control,” J. Process Control, vol. 17, no. 6,
pp. 563–570, Jul. 2007.
[15] P. Tøndel and T. Johansen, “Complexity reduction in explicit model pre- Keqiang Li received the M.S. and Ph.D. degrees
dictive control,” in Proc. 15th IFAC World Congr., Barcelona, Spain, 2002. from Chongqing University, Chongqing, China, in
[16] J. Sun, I. Kolmanovsky, R. Ghaemi, and S. Chen, “A stable block model 1988 and 1995, respectively, and the B.Tech. degree
predictive control with variable implementation horizon,” Automatica, from Tsinghua University, Beijing, China, in 1985.
vol. 43, no. 11, pp. 1945–1953, Nov. 2007. He is Professor of automotive engineering with
[17] M. Kothare, V. Nevistic, and M. Morari, “Robust constrained Tsinghua University. His main areas of research in-
model predictive control for nonlinear systems—A comparative study,” terest include vehicle dynamics and control for driver
in Proc. IEEE Conf. Decision Control, New Orleans, LA, USA, 1995, assistance systems and hybrid electrical vehicle. He
pp. 2884–2885. has authored over 90 papers and holds 12 patents in
[18] R. Fletcher, Practical Methods of Optimization. Chichester, U.K.: Wiley, China and Japan.
1987. Dr. Li has been a Senior Member of the Society
[19] D. Mayne, J. Rawlings, C. Rao, and P. Scokaert, “Constrained model of Automotive Engineers of China, and on the editorial boards of International
predictive control: Stability and optimality,” Automatica, vol. 36, no. 6, Journal of ITS Research and International Journal of Vehicle Autonomous Sys-
pp. 789–814, Jun. 2000. tems. He has been a recipient of the “Changjiang Scholar Program Professor.”
[20] W. Kwon and A. Pearson, “On feedback stabilization of time-varying
discrete-linear system,” IEEE Trans. Autom. Control, vol. AC-23, no. 3,
pp. 479–481, Jun. 1978.
[21] J. Rawlings and K. Muske, “The stability of constrained receding horizon
control,” IEEE Trans. Autom. Control, vol. 38, no. 10, pp. 1512–1516, Bo Cheng received the B.S. and M.S. degrees from
Oct. 1993. Tsinghua University, Beijing, China, in 1985 and
[22] S. Keerthi and E. Gilbert, “Optimal infinite-horizon feedback laws for a 1988, respectively, and the Ph.D. degree from Tokyo
general class of constrained discrete-time systems: Stability and moving- University, Tokyo, Japan, in 1998.
horizon approximations,” J. Optim. Theory Appl., vol. 57, no. 2, pp. 265– He is a Professor with Tsinghua University and the
293, May 1988. Dean with the Suzhou Automotive Research Insti-
[23] D. Mayne and H. Michalska, “Receding horizon control of nonlinear tute, Tsinghua University. His active research inter-
systems,” IEEE Trans. Autom. Control, vol. 35, no. 7, pp. 814–824, ests include autonomous vehicles, driver assistance
Jul. 1990. systems, active safety, and vehicular ergonomics. He
[24] T. Yang and E. Polak, “Moving horizon control of nonlinear systems is the author of more than 70 journal/conference
with input saturation, disturbances and plant uncertainty,” Int. J. Control, papers and holds more than 20 patents. He also is
vol. 58, no. 4, pp. 875–903, Oct. 1993. Chairman of the Academic Board of SAE-Beijing, Committee Member of
[25] A. Zheng and M. Morari, “Stability of model predictive control with National 863 Plan, Member of Academic Committee of China Transportation
mixed constraints,” IEEE Trans. Autom. Control, vol. 40, no. 10, Technology Project, and Member of Academic Committee of Ministry of
pp. 1818–1823, Oct. 1995. Public Security Traffic Management.

Das könnte Ihnen auch gefallen