1. The document describes an optimal control problem for maximizing the energy extracted from an oscillating system comprising a spring, dampers, and a control input.
2. The system is modeled using state equations and the optimal control problem is formulated to maximize the extracted energy over time, subject to constraints on the control input.
3. Pontryagin's maximum principle is applied to derive necessary conditions for optimality, resulting in a set of differential equations that can be solved numerically to obtain the optimal control input over time.
1. The document describes an optimal control problem for maximizing the energy extracted from an oscillating system comprising a spring, dampers, and a control input.
2. The system is modeled using state equations and the optimal control problem is formulated to maximize the extracted energy over time, subject to constraints on the control input.
3. Pontryagin's maximum principle is applied to derive necessary conditions for optimality, resulting in a set of differential equations that can be solved numerically to obtain the optimal control input over time.
1. The document describes an optimal control problem for maximizing the energy extracted from an oscillating system comprising a spring, dampers, and a control input.
2. The system is modeled using state equations and the optimal control problem is formulated to maximize the extracted energy over time, subject to constraints on the control input.
3. Pontryagin's maximum principle is applied to derive necessary conditions for optimality, resulting in a set of differential equations that can be solved numerically to obtain the optimal control input over time.
1. The document describes an optimal control problem for maximizing the energy extracted from an oscillating system comprising a spring, dampers, and a control input.
2. The system is modeled using state equations and the optimal control problem is formulated to maximize the extracted energy over time, subject to constraints on the control input.
3. Pontryagin's maximum principle is applied to derive necessary conditions for optimality, resulting in a set of differential equations that can be solved numerically to obtain the optimal control input over time.
Let x(t ) the state of the system with input u(t ) that satisfies the state equation [1] = F (x, u) , x
(1)
x (0) = x0
(2)
uU
(3)
t [0, T ]
(4)
with
with T fixed. The set U shows the constraints applied to u.
The goal is to determine u, defined in [0, T ] which maximizes the cost functional J defined as J (u) = (xT ) +
L (u, x) dt
(5)
where xT = x (T ) .
(6)
(xT ) is the cost associated with the terminal state xT . The Lagrangian L (u, x) is the function to be maximized in [0, T ].
1.2
Pontriagyn maximum principle
Along optimal path for (x, u, (t )) it is verified the following necessary conditions for maximizing J = F (x, u) , x with x (0) = x0 uU
t [0, T ]
(7)
= T F (x, u) + L (x, u)
(8)
and
subjected to the terminal condition
(T ) = (T ) .
(9)
The vector (t ) is designated by co-state and Eq. (8) by adjoint equation.
For each t , the Hamiltonian H defined by H (x, u, ) = T F (x, u) + L (x, u)
(10)
is maximum for the optimal input u. If the maximum is for an u in the interior of U then H = 0. ui
(11)
In the case that the optimum is in the boundary of U then Eq. (11) does not apply.
Oscillator system Let as consider a simple oscillator system comprising a spring, k, a energy extraction damper,c,
and control damper, G, with damping as function of, u, see Fig. 1. The equation that describes the motion of this system is m z + c z + g u z + k z = f cos(t )
(12)
where the dot denotes time derivative, m is the block mass, f and are the modulus are angular frequency of the external force. The initial conditions are z(0) = z0 and z(0) = v0 . The extracted energy, e, can be computed using the instantaneous power of the damper c which is given by e = c z2 .
(13)
The initial condition is e(0) = 0.
The system of equations (12) and (13) can be written as a first-order system in vectorial form as = F (x, u) , x
(14)
T T x = v z e = x1 x2 x3 ,
(15)
defining
Figure 1: Oscillating system with spring, k, energy extraction damper,c, and control damper, g . and
F1 ( f cos(t ) c x1 g u x1 k x2 ) /m
, F (x, u) = x1
F2 = 2 F3 c x1
(16)
with initial conditions
T x0 = v0 z0 0 .
(17)
We want to determine u(t ) for 0 t T such that
J (u) = e(T ),
(18)
is maximum subjected to the constraint 0 u 1.
In the present case (xT ) = e(T ) and L (x, u) = 0, see Eq. (5). Using this, the adjoint equation is given by = T F (x, u) . The gradient of F is given by
F =
F1 x1 F2 x 1 F3 x1
F1 x2 F2 x2 F3 x2
F1 x3 F2
x3 F3 x3
(19)
(c + g u) /m k/m 0
0 ,
2c x1
(20)
resulting in
1 1 (c + g u) /m + 2 + 23 c x1
= . 1 k/m 2
3 0 The final condition of is (T ) = (T ), giving
1 (T )
x1
(T ) =
x 2
3 (T ) x 3
x = xT
= 0 . 1
(21)
(22)
For each t , the Hamiltonian H
( f cos(t ) c x1 g u x1 k x2 ) /m
x1
H (x, u, ) = T F = 1 2 3
c x12
(23)
= 1 ( f cos(t ) c x1 g u x1 k x2 ) /m + 2 x1 + 3 c x12 is maximum for the optimal input u. The conditions to maximize H are u(t ) =
1,
if (1 (t )x1 (t )) 0
0,
otherwise
(24)
The usual solution of this problem is:
1. Set u(t ) = 0. 2. Compute (14) using the (17). 3. Compute (21) backward (from T to 0) to impose the final condition (22). 4. Determine u(t ) using (24). 5. If not converged go to step 2.
Demonstration of the Pontriagyn maximum principle
Let us define the functional J as J =J
T [ x F (x, u)] dt .
(25)
The vector is introduced so J is optimal and satisfies the system of ordinary differential equations (1) for all t [0, T ]. In other words, T
T [ x F (x, u)] dt ,
(26)
is a J constraint that is zero for the optimal solution1 . The vector is called co-state. Introducing Eq. (5) in Eq. (28) we get J = (xT ) +
L (u, x) dt
= (xT ) +
T [ x F (x, u)] dt
0 T
L (u, x) + F (x, u) dt T
0 1
Similar to a Lagrange multiplier.
(27) T
dt . x
Defining the Hamiltonian function
H (x, u, ) = T F (x, u) + L (x, u) ,
(28)
we get J = (xT ) +
H (u, x, ) dt
dt . T x
(29)
Let us assume that uopt is the function which maximizes then functional J then a small perturbation u results in a decrease of J . The system will follow another path x + x and the variation of the objective function is negative
J = J (x + x, v) J x, uopt < 0,
(30)
where v = uopt + u. Using Eq. (29) we get
J = (xT + xT ) (xT ) +
[H (v, x + x, ) H (u, x, )] dt
T x dt .
(31)
Integrating by parts the last term of Eq. (31)
Z
x dt = (T ) x (T ) (0) x (0)
T x dt
(32)
Since the optimal control does not change x (0) we have x (0). As a result, J = (xT + xT ) (xT ) T (T ) x (T ) ZT Z + [H (v, x + x, ) H (u, x, )] dt + 0
T x dt .
(33)
Performing a first order Taylor series expansion of the terms on x we get
(xT + xT ) (xT ) + (xT ) xT ,
(34)
H (v, x + x, ) H (v, x, ) + H (v, x, ) x
(35)
and
Replacing both expansions in Eq. (33)
Z
J = (xT ) (T ) x (T ) +
i T
H (u, x, ) + x dt
(36)
[H (v, x, ) H (u, x, )] dt .
We can choose T = H (u, x, ) ,
(37)
satisfying T (T ) = (xT ) ,
(38)
to assure that the first term of the right-hand-side of Eq. (36) is zero. This results in J =
[H (v, x, ) H (u, x, )] dt .
(39)
The states x and co-states are computed for uopt and independently of v. If uopt is the optimum then
H uopt , x, H (v, x, )
(40)
v U . This needs to be proved because we compute uopt by maximizing H (u, x, ).
Suppose that there is an time instant, t1 where a function w satisfies
H (w (t1 ) , x (t1 ) , (t1 )) > H uopt (t1 ) , x (t1 ) , (t1 )
(41)
Since H is a continuous function, there is an interval [t1 , t1 + ] where this inequality holds. Let w = uopt except in this interval. Using this choice, the variation is J =
t1 +
H (w (t ) , x (t ) , (t )) H uopt (t ) , x (t ) , (t ) dt > 0
t1
However, this contradicts the hypothesis that uopt is the optimal control.
References [1] J. L. M. Lemos, Lectures on Optimal Control, Instituto Superior Tcnico, 2012.
