The Stability of Nonlinear MPC
非线性MPC的稳定性
1. Systems
x ˙ = f ( x ( t ) , u ( t ) ) , x ( 0 ) = x 0 , u ( t ) ∈ U \dot{\mathbf{x}}=\bm{f}\big(\mathbf{x}(t),\mathbf{u}(t)\big),\quad \mathbf{x}(0)=\rm{x}_0,\\ \quad\mathbf{u}(t)\in U x˙=f(x(t),u(t)),x(0)=x0,u(t)∈U
2 Assumptions
- f \bm{f} f is twice continuously differentiable and f ( 0 , 0 ) = 0 \bm{f}(\bm{0},\bm{0})=\bm{0} f(0,0)=0, i.e. x = 0 \mathbf{x}=\bm{0} x=0 is an equilibrium under zero input.
- U U U is compact, convex and 0 \bm{0} 0 is contained in U U U.
- The jacobian linearization of system ( 1 ) (1) (1) is stabilizable.
3. Standard MPC formulation
min
u
(
⋅
)
J
(
x
(
t
)
,
u
(
⋅
)
)
\min _{u(\cdot)} J(\mathbf{x}(t), \mathbf{u}(\cdot))
u(⋅)minJ(x(t),u(⋅))
with
J ( x ( t ) , u ( ⋅ ) ) = ∫ t t + T p ( ∥ x ( τ ) ( x ( t ) , t ) ∥ Q 2 + ∥ u ( τ ) ∥ R 2 ) d τ + ∥ x ( t + T p ; x ( t ) , t ) ∥ P 2 (2) J(\mathbf{x}(t), \mathbf{u}(\cdot))=\int_{t}^{t+T_{p}}\left(\|\mathbf{x}(\tau)(\mathbf{x}(t), t)\|_{Q}^{2}+\|\mathbf{u}(\tau)\|_{R}^{2}\right) \mathrm{d} \tau \\ \quad\quad\quad\quad+\left\|\mathbf{x}\left(t+T_{\mathrm{p}} ; \mathbf{x}(t), t\right)\right\|_{P}^{2}\tag{2} J(x(t),u(⋅))=∫tt+Tp(∥x(τ)(x(t),t)∥Q2+∥u(τ)∥R2)dτ+∥x(t+Tp;x(t),t)∥P2(2)
subject to
x
˙
=
f
(
x
,
u
)
,
x
(
t
;
x
(
t
)
,
t
)
=
x
(
t
)
u
(
τ
)
∈
U
,
τ
∈
[
t
,
t
+
T
p
]
x
(
t
+
T
p
;
x
(
t
)
,
t
)
∈
Ω
(3)
\dot{\mathbf{x}}=\mathbf{f}(\mathbf{x}, \mathbf{u}), \quad \mathrm{x}(t ; \mathbf{x}(t), t)=\mathbf{x}(t)\\ \mathbf{u}(\tau) \in U, \quad \tau \in\left[t, t+T_{\mathrm{p}}\right]\\ \mathbf{x}\left(t+T_{\mathrm{p}} ; \mathbf{x}(t), t\right) \in \Omega \tag{3}
x˙=f(x,u),x(t;x(t),t)=x(t)u(τ)∈U,τ∈[t,t+Tp]x(t+Tp;x(t),t)∈Ω(3)
4. Main ideas
- Since the finite horizon optimization property of MPC can not guarantee the stability, while the infinite may do more, such as the LQR.
- By introducing terminal state cost and terminal inequality constraint, the terminal states can be forced to be in the terminal region, which is selected as the attraction region around origin and is invariant under local linear feedback controllers.
- When trajectory under optimization goes into the terminal region at t + T p t+T_p t+Tp, the states will stay in the region when imposing designed linear state feedback.
- It can be proved that the infinite integral cost ∫ s ∞ ( ∥ x ( s ) ( x ( t ) , t ) ∥ Q 2 + ∥ u ( s ) ∥ R 2 ) d s \int_{s}^{\infty}\left(\|\mathbf{x}(s)(\mathbf{x}(t), t)\|_{Q}^{2}+\|\mathbf{u}(s)\|_{R}^{2}\right) \mathrm{d} s ∫s∞(∥x(s)(x(t),t)∥Q2+∥u(s)∥R2)ds starting from invariant set Ω \Omega Ω is always less equal than the terminal cost in ( 2 ) (2) (2).
- The optimal value function can thus be proved to be non-increasing. This further leads to the stability.
- Although the local state feedback is used to demonstrate the stability, MPC is implemented in online manner at every time step and the linear feedback control law is never used. The finite horizon control can thus be proved in a quasi-infinite way.
5. Method
-
Solve the linear state feedback control problem to find a locally stabilizing feedback gain K after jacobian linearization around origin.
x ˙ = A x + B u \dot{\mathbf{x}}=A\mathbf{x}+B\mathbf{u} x˙=Ax+Bu A = ∂ f ∂ x ( 0 , 0 ) , B = ∂ f ∂ u ( 0 , 0 ) A=\frac{\partial{\bm{f}}}{\partial{\mathbf{x}}}(\bm{0}, \bm{0}),\quad B=\frac{\partial{\bm{f}}}{\partial{\mathbf{u}}}(\bm{0}, \bm{0}) A=∂x∂f(0,0),B=∂u∂f(0,0) u = K x such that A k = A + B K is asymptotically stable \mathbf{u}=K\mathbf{x}\quad \text{such that} \quad A_k=A+BK \quad\text{is asymptotically stable} u=Kxsuch thatAk=A+BKis asymptotically stable -
Solve the Lyapunov equation with a suitable κ ∈ [ 0 , + ∞ ] \kappa\in[0,+\infty] κ∈[0,+∞]. κ \kappa κ should be selected (usually depends on the eigenvalues of A k A_k Ak) to make sure that the equation has unique positive definite and symmetric solutions P P P.
-
Find the largest possible set Ω β = { x ∣ x ⊤ P x ≤ β } \Omega_\beta=\{\mathbf{x}|\mathbf{x}^{\top}P\mathbf{x}\leq\beta\} Ωβ={x∣x⊤Px≤β}, such that K x ∈ U K\mathbf{x}\in U Kx∈U for all x ∈ Ω β \mathbf{x}\in\Omega_\beta x∈Ωβ. This ensures that the state feedback control respects the control constraints.
-
Find the largest possible set Ω α = { x ∣ x ⊤ P x ≤ α , α ≤ β } \Omega_\alpha=\{\mathbf{x}|\mathbf{x}^{\top}P\mathbf{x}\leq\alpha, \alpha\leq\beta\} Ωα={x∣x⊤Px≤α,α≤β}, such that the condition of non-increasing-value-function is satisfied in Ω α \Omega_\alpha Ωα.
6. Feasibility and stability
The feasibility at any t > 0 t>0 t>0 can be guaranteed if the feasibility is satisfied at t = 0 t=0 t=0. The stability is guaranteed by the feasibility and non-increasing property mentioned above.
7. Remarks
- If there is no linear feedback controller can locally asymptotically stabilize the nonlinear system, the attraction region
Ω
α
\Omega_\alpha
Ωα contracts to the origin. In accordance, the terminal inequality
(
3
)
(3)
(3) reduces to equality constraint
x ( t + T p ) = 0 . \mathbf{x}(t+T_p)=\bm{0}. x(t+Tp)=0. - The conditions above are only sufficient and not necessary. The fact
that the linearized system is not stabilizable does not imply that
there exists no linear feedback controller being able to stabilize
the nonlinear system locally.
Reference:
8. Path Following
- Transfer the original problem into the error dynamic problem with some conditions such that x ˙ e = g ( x e , u e ) \dot{\mathbf{x}}_e=\mathbf{g}(\mathbf{x}_e,\mathbf{u}_e) x˙e=g(xe,ue), g ( 0 , 0 ) = 0 \mathbf{g}(\bm{0},\bm{0})=0 g(0,0)=0.
- Solve the LMI problem based on PLDI to obtain the penalty matrix P P P and feedback gain K K K.
- Find the largest invariant terminal region.