Professional Documents
Culture Documents
Sciencedirect: © 2017, Ifac (International Federation of Automatic Control) Hosting by Elsevier Ltd. All Rights Reserved
Sciencedirect: © 2017, Ifac (International Federation of Automatic Control) Hosting by Elsevier Ltd. All Rights Reserved
navigability/sustainability in river systems based on MPC. time) to simulate the hydrodynamic systems. Regarding
Sutrisno et al. (2012) propose a scheme to determine the the CB system illustrated in Fig. 1, the one-dimension
weighting values that are proportional to the control load three-velocities (D1Q3) LB method is used to model the
for each subsystem. This weighting scheme is then applied irrigation canal in one dimension l (lattice spacing ∆l)
to control the irrigation canal using Feasible-Cooperation and with three velocities [v1 v2 v3 ] = [0 v (−v)] where
MPC and Nash-bargaining MPC. Igreja et al. (2011) pro- v = ∆l/∆t (∆t is time step). Modeling one-dimensional
pose a distributed MPC algorithm for a water delivery fluid flows with D1Q3 LB method in the presence of an
canal using linearized and discretized Saint-Venant model. external force F , the following relations result:
Negenborn et al. (2009) focus on distributed MPC strategy 1 ∆t
in which a discrete-time linear integral delay model is used fi (l + v∆t, t + ∆t) = fi (l, t) + (fie − fi ) + ωi 2 vi F
τ cs
for prediction. Their distributed control algorithm is based (2)
on augmented Lagrangian duality method. A methodol-
ogy for the optimal management of a combined irrigation where i ∈ {1, 2, 3}, τ is constant relaxation time, ωi and cs
and water supply system based on MPC is proposed in are the parameters that are determined by the geometry
Puig et al. (2012). Van Overloop et al. (2015) present a of the lattice and chosen to obtain the isotropy for the
decentralized MPC scheme for large-scale systems whose model. The fie are the equilibrium distribution functions
components can exchange information through a network. calculated by Pham et al. (2010). The macroscopic vari-
ables h and u are determined by:
3
2.1 Prediction model
h= f i = f1 + f 2 + f 3
i=1
The water height and flow dynamics in such canal are usu- 3
(3)
ally modeled by the one-dimension Shallow Water equa-
tions, also known as Saint-Venant equations Malaterre q = hu = vi fi = v(f2 − f3 )
i=1
et al. (2007). The Shallow Water equations for fluid dy-
namics in one dimension x are formulated as follows: Furthermore, we consider the coupling of two reaches
∂t (h) + ∂x (hu) = 0 modeled by D1Q3 LB method with a gate in a submerged
(1) regime. The flow rate through gate Qg depends on the
1
∂t (hu) + ∂x ( gh2 + hu2 ) = F difference between the water heights of upstream h and
2 downstream h around the gate. That is:
where h is the water height, u is (depth-average) horizontal
velocity of the flow, and g is the acceleration of gravity. Qg = Bg αθ 2g(h − h ) (4)
The flow rate Q at a particular location in a rectangular where Bg is the gate width, α is the gate coefficient and θ
canal of width B is deduced by: Q = Bhu. The force term is the gate opening. In addition, the outflow Q of upstream
F = gh(I − J) is calculated from the bed slope I (where reach is equal to the inflow Q of downstream reach, that
I = ∂x hb with hb is the bed profile) and the friction J at is: Q = Q = Qg . The manipulated variable θ is deduced
the bottom of the reach. J is deduced by the Manning- from following gate equations:
Strickler equations (see Pham et al. (2010)). vB(f2 − f3 ) = vB (f2 − f3 ) = Bg αθ 2g(h − h ) (5)
For further use, we consider the linearization of the LB
dynamics (2) around an equilibrium height, h = he , and
an equilibrium velocity, u = ue (we deduce the equilibrium
flow rate: Qe = Bhe ue ). A small derivations i , i =
{1, 2, 3} around the equilibrium points are expressed by:
i = fi − fie (he , ue ). We consider here the downstream
configuration of our case study (as shown in Fig. 2). In
this configuration, we assume that some measurements are
available such as the water height hus at upstream, the
water height hds at downstream and the flow rate Q at the
downstream of each reach. The control action is applied on
the gate to adjust gate opening θ (i.e., manipulated input).
The controlled variable is assumed as the downstream flow
rate Q of each reach. We also consider a lateral discharge
Qp at point lp of each reach as the perturbation. As a
result, the general linearized discrete-time model for reach
i of length L (corresponding to N discretized points lj ,
j = 1, . . . , N ) used in MPC scheme is presented as follows:
Fig. 1. The “Canal de la Bourne” (CB) irrigation canal x i (k + 1) =A i x i (k) + B i u i (k)
modeled by Lattice Boltzmann method.
+ Bp i p i (k) + Bd i d i (k)
By coupling with the equations (1), the LB method y i (k) =Cy i x i (k) + Dy i u i (k) + Dyd i d i (k) (6)
is demonstrated in several works Pham et al. (2010) z i (k) =Cz i x i (k) + Dz i u i (k) + Dzd i d i (k)
that it is an efficient and powerful numerical tool (in
terms of accuracy, numerical stability and computational q i (k) =Cq i x i (k) + Dq i u i (k) + Dqd i d i (k)
6759
Proceedings of the 20th IFAC World Congress
6566
Toulouse, France, July 9-14, 2017 Le-Duy-Lai Nguyen et al. / IFAC PapersOnLine 50-1 (2017) 6564–6569
where we define discrete-time local states of each reach as: 2.3 Cost function
x i (k) = [1i (l1 , k) 3i (l1 , k) . . . 2i (lj , k) 1i (lj , k) 3i (lj , k)
. . . 2i (lN , k) 1i (lN , k)]T ∈ R(N −2)×(N −2) , inputs as: The regulatory control of irrigation canals expects an auto-
u i (k) = 2i (l1 , k) ∈ R, perturbation as: p i (k) ∈ R, matic adjustment of controlled structures with reliability
and durability. For instance, a gate will move up and down
interaction variables between subsystems as: d i (k) = permanently to regulate flow rate at the end of a reach
3i (lN , k) ∈ R, controlled variable as: z i (k) = Q̃ i (lN , k) ∈ close to reference value (set-point). In other words, a local
R, (where Q̃ i (k) = Q i (k) − Qe ), measured output as: controller seeks to achieve the objectives as follows (see
y i (k) = [h̃ i (l1 , k) h̃ i (lN , k)], (where h̃ i (k) = h i (k) − Negenborn et al. (2009); Puig et al. (2012)):
he ), estimated flow rate through controlled gate as: • Minimize the deviation of controlled variable from
q i (k) = Q̃ i (l1 , k) ∈ R, and corresponding matrices as: the reference value (perform the regularization) (first
A i ∈ R(N −2)×(N −2) , B i ∈ R(N −2)×1 , Bpi ∈ R(N −2)×1 , term of (9))
Bdi ∈ R(N −2)×1 , Cyi ∈ R2×(N −2) , Dyi ∈ R2×1 , Dyd
i
∈ • Minimize the change in deviation of controlled vari-
i
able (e.g., in order to encourage smooth change of
R2×1 , Czi ∈ R1×(N −2) , Dzi ∈ R1×1 , Dzd ∈ R1×1 , Cqi ∈ flow rate) (second term of (9))
i
R1×(N −2) , Dqi ∈ R1×1 , Dqd ∈ R1×1 . • Minimize the change in sequence of computed actions
(avoid the oscillation) (third term of (9))
Therefore, we present the cost function for each controller:
2.2 Structural and operational constraints
(Np −1)
i
J (k) = (||z i (k + n|k) − r i (k)||2Z1
The operational management of irrigation canals is subject
n=0 (9)
to constraints on the canal structure (e.g., limits of gate
opening, reliability and robustness of equipment, limits of + ||∆z i (k + n|k)||2Z2
reservoir capacity) as well as on canal operations (e.g., + ||∆u i (k + n|k)||2U1 )
overflow risks, bank stability, minimum off-take levels).
The control solutions for the irrigation canal illustrated where z i is the controlled variable over the prediction
in Fig. 2 and modeled by the discrete-time systems in (6) horizon Np , r i (k) is the reference value, u is system input
must satisfy constraints on water height, gate opening, and over prediction horizon Np , and Z1 , Z2 , U1 are weighting
on variation of gate opening as follows: matrices of appropriate dimensions.
hmin ≤ h i ≤ hmax
2.4 Decentralized control algorithm
θmin ≤ θ i ≤ θmax (7)
∆θmin ≤ ∆θ i ≤ ∆θmax The details on the decentralized construction can be
synthesized in Algorithm 1.
where the water height is given by: h i (k) = h̃ i (k) + he
with h̃ i (k) = 1i (k) + 2i (k) + 3i (k). The gate opening is Algorithm 1 Decentralized MPC scheme
deduced from gate equations (5), that is: 1: Inputs: Initial state x i (1), initial input u i (1), pre-
i i
vB(2 (k) − 3 (k) + f2e − f3e ) dicted perturbation p i (k) and reference trajectory
θ i (k) = (8)
i-1 i-1 r i (k)
α 2g(hds (k) − hds (k)) 2: for k = 1 : kmax do
i-1
3: measure the flow rate Q i at downstream of each
i
where hds (k), hds (k) are given by water level sensors or reach
computed from measured outputs y i-1 (k) and y i (k). 4: estimate the interaction variable d i (k) from gate
equations (5)
5: obtain Uki ∗ (x i (k)) by solving optimization prob-
lems (9) for prediction horizon N p under the con-
straints (7)
6: apply the first element uki ∗ of Uki ∗ to the subsystem
7: determine the state x i (k + 1) and outputs
z i (k), y i (k) at time instant k from (6)
8: go to the next step (k + 1)
9: end for
6760
Proceedings of the 20th IFAC World Congress
Toulouse, France, July 9-14, 2017 Le-Duy-Lai Nguyen et al. / IFAC PapersOnLine 50-1 (2017) 6564–6569 6567
account for the interactions between subsystems, informa- matrices of appropriate dimensions. Therefore, the reced-
tion exchange among controllers is required (see more de- ing horizon control algorithm can be implemented in a dis-
tails in Nguyen et al. (2016a)). Moreover, the cooperation tributed (and cooperative) control fashion as summarized
involved in this strategy requires that all local objectives of in Algorithm 2.
the controllers have to be coordinated towards optimizing
global objective (e.g., all controlled variables rapidly reach Algorithm 2 Distributed MPC scheme for N controllers
their set-points). 1: for i = 1, . . . , N do
2: Inputs: Initial state x i (1), initial input u i (1), pre-
3.1 Definition and management of shared information dicted perturbation p i (k) and reference trajectory
r i (k)
We consider now the downstream configuration of our 3: for k = 1 : kmax do
i+1
case study as shown in Fig. 2. When we couple two 4: receive the flow rate Qg through downstream
reaches i-1 and i with a gate controlled by controller
i+1
C i , the upstream point lus of the gate also is the gate from downstream controller C and the
downstream point of the reach i-1 , and the downstream reference value r of upstream controller C i-1
i-1
point lds of the gate is the upstream point of the 5: determine the interaction variable d i (k) from
reach i . The flow rate Qg through the gate is com- gate equations (5)
puted from the gate equations (5). In order to deter-
6: obtain Uki ∗ (x i (k)) by solving optimization
mine interaction variable d i (k) and to compute the gate problems (9) for prediction horizon N p under
opening θ i , we may identify from the linearized LB constraints (7)
model (6) and the gate equations (5) that the neces-
7: apply the first element uki ∗ of Uki ∗ to the sub-
sary information exchanged among the controllers con-
system
i-1 i-1 i+1
tains: {1 (lLi−1 , k), 2 (lLi−1 , k), 3 (lus , k)}. Appropri- 8: determine the state x i (k + 1) and outputs
ately, sharing information from controller C i to its neigh- z i (k), y i (k) at time instant k from (6)
9: send the estimated flow rate through gate
bors can be chosen as: [3i (lus , τ ) 1i (lLi , τ ) 2i (lLi , τ )]T .
In addition, the gate equations (5) permit the computation Q i (k) = q i (k)+Qe and reference value r i (k+1)
of these interaction variables, once the neighbor controllers 10: go to the next step (k + 1)
share the flow rate Qg through the gate at their position. 11: end for
Timing for sharing information depends on synchroniza- 12: end for
tion mechanism and also the coordination method for each
group of neighboring controllers (see more detail in Nguyen
et al. (2016b)). 4. SIMULATION RESULTS
6761
Proceedings of the 20th IFAC World Congress
6568
Toulouse, France, July 9-14, 2017 Le-Duy-Lai Nguyen et al. / IFAC PapersOnLine 50-1 (2017) 6564–6569
zr1
Table 2. Simulation parameters of irrigation 1.5
canals rr1
zr2
LB parameters
Spacing step (∆x) 1m 1 rr2
Time step (∆t) 0.05s zr3
Relaxation time (τi ) 0.8
rr3
Linearization around equilibrium points 0.5
Water height at equilibrium (he ) 0.16m Qp
Boundary conditions
Upstream height of first reach (Hus ) 0.20m 5 10 15 20 25 30 35 40 45
Time (s)
Downstream height of final reach (Hds ) 0.10m
Simulation parameters
Simulation time (T ) 45s Fig. 4. Distributed cooperative control with perturbation
Prediction horizon (Npred ) 0.5s Qp - Variation of downstream flow rates (zri = Qri −
Weighting matrices Z i = Qi = R i = I Qe ) with regards to equilibrium points Qe of three
Reference value (r)(m3 /s) rr1 = 2.6e − 3
reaches (ri = {r1, r2, r3}).
rr2 = 1.8e − 3
rr3 = 1.0e − 3
5 10 15 20 25 30 35 40 45
Time (s)
6762
Proceedings of the 20th IFAC World Congress
Toulouse, France, July 9-14, 2017 Le-Duy-Lai Nguyen et al. / IFAC PapersOnLine 50-1 (2017) 6564–6569 6569
−3 DCC: Outputs and reference values that: (1) it respects the constraints on the controlled and
x 10 manipulated variables; (2) it avoids excessive variations
3
on the manipulated variables allowing better use of ac-
tuators; and (3) it rejects external perturbations. Future
work will concentrate on more comparisons with other lin-
2.5
ear/nonlinear optimization-based control approaches over
different benchmarks available in the literature. Also, dif-
zri = Qri − Qe (m3/s)
6763