Functions: Predictive Controllers

Functions: Predictive Controllers

All the predictive controllers in this module rely on a state estimator to compute the predictions. The default LinMPC estimator is a SteadyKalmanFilter, and NonLinMPC with nonlinear models, an UnscentedKalmanFilter. For simpler and more classical designs, an InternalModel structure is also available, that assumes by default that the current model mismatch estimation is constant in the future (same approach as dynamic matrix control, DMC).

Info

The nomenclature uses boldfaces for vectors or matrices, capital boldface letters for vectors representing time series (and also for matrices), and hats for the predictions (and also for the observer estimations).

To be precise, at the $k$th control period, the vectors that encompass the future measured disturbances $\mathbf{d̂}$, model outputs $\mathbf{ŷ}$ and setpoints $\mathbf{r̂_y}$ over the prediction horizon $H_p$ are defined as:

\[ \mathbf{D̂} = \begin{bmatrix} \mathbf{d̂}(k+1) \\ \mathbf{d̂}(k+2) \\ \vdots \\ \mathbf{d̂}(k+H_p) \end{bmatrix} \: , \quad \mathbf{Ŷ} = \begin{bmatrix} \mathbf{ŷ}(k+1) \\ \mathbf{ŷ}(k+2) \\ \vdots \\ \mathbf{ŷ}(k+H_p) \end{bmatrix} \quad \text{and} \quad \mathbf{R̂_y} = \begin{bmatrix} \mathbf{r̂_y}(k+1) \\ \mathbf{r̂_y}(k+2) \\ \vdots \\ \mathbf{r̂_y}(k+H_p) \end{bmatrix}\]

in which $\mathbf{D̂}$, $\mathbf{Ŷ}$ and $\mathbf{R̂_y}$ are vectors of nd*Hp, ny*Hp and ny*Hp elements, respectively. The vectors for the manipulated input $\mathbf{u}$ are shifted by one time step:

\[ \mathbf{U} = \begin{bmatrix} \mathbf{u}(k+0) \\ \mathbf{u}(k+1) \\ \vdots \\ \mathbf{u}(k+H_p-1) \end{bmatrix} \quad \text{and} \quad \mathbf{R̂_u} = \begin{bmatrix} \mathbf{r̂_u}(k+0) \\ \mathbf{r̂_u}(k+1) \\ \vdots \\ \mathbf{r̂_u}(k+H_p-1) \end{bmatrix}\]

in which $\mathbf{U}$ and $\mathbf{R̂_u}$ are both vectors of nu*Hp elements. Defining the manipulated input increment as $\mathbf{Δu}(k+j) = \mathbf{u}(k+j) - \mathbf{u}(k+j-1)$, the control horizon $H_c$ enforces that $\mathbf{Δu}(k+j) = \mathbf{0}$ for $j ≥ H_c$. For this reason, the vector that collects them is truncated up to $k+H_c-1$:

\[ \mathbf{ΔU} = \begin{bmatrix} \mathbf{Δu}(k+0) \\ \mathbf{Δu}(k+1) \\ \vdots \\ \mathbf{Δu}(k+H_c-1) \end{bmatrix}\]

in which $\mathbf{ΔU}$ is a vector of nu*Hc elements.

PredictiveController

ModelPredictiveControl.PredictiveController — Type

Abstract supertype of all predictive controllers.

(mpc::PredictiveController)(ry, d=[]; kwargs...) -> u

Functor allowing callable PredictiveController object as an alias for moveinput!.

Examples

julia> mpc = LinMPC(LinModel(tf(5, [2, 1]), 3), Nwt=[0], Hp=1000, Hc=1, direct=false);

julia> u = mpc([5]); round.(u, digits=3)
1-element Vector{Float64}:
 1.0

source

LinMPC

ModelPredictiveControl.LinMPC — Type

LinMPC(model::LinModel; <keyword arguments>)

Construct a linear predictive controller based on LinModel model.

The controller minimizes the following objective function at each discrete time $k$:

\[\begin{aligned} \min_{\mathbf{ΔU}, ϵ} \mathbf{(R̂_y - Ŷ)}' \mathbf{M}_{H_p} \mathbf{(R̂_y - Ŷ)} + \mathbf{(ΔU)}' \mathbf{N}_{H_c} \mathbf{(ΔU)} \\ + \mathbf{(R̂_u - U)}' \mathbf{L}_{H_p} \mathbf{(R̂_u - U)} + C ϵ^2 \end{aligned}\]

subject to setconstraint! bounds, and in which the weight matrices are repeated $H_p$ or $H_c$ times by default:

\[\begin{aligned} \mathbf{M}_{H_p} &= \text{diag}\mathbf{(M,M,...,M)} \\ \mathbf{N}_{H_c} &= \text{diag}\mathbf{(N,N,...,N)} \\ \mathbf{L}_{H_p} &= \text{diag}\mathbf{(L,L,...,L)} \end{aligned}\]

Time-varying and non-diagonal weights are also supported. Modify the last block in $\mathbf{M}_{H_p}$ to specify a terminal weight. The $\mathbf{ΔU}$ includes the input increments $\mathbf{Δu}(k+j) = \mathbf{u}(k+j) - \mathbf{u}(k+j-1)$ from $j=0$ to $H_c-1$, the $\mathbf{Ŷ}$ vector, the output predictions $\mathbf{ŷ}(k+j)$ from $j=1$ to $H_p$, and the $\mathbf{U}$ vector, the manipulated inputs $\mathbf{u}(k+j)$ from $j=0$ to $H_p-1$. The slack variable $ϵ$ relaxes the constraints, as described in setconstraint! documentation. See Extended Help for a detailed nomenclature.

This method uses the default state estimator, a SteadyKalmanFilter with default arguments.

Arguments

model::LinModel : model used for controller predictions and state estimations.
Hp=10+nk: prediction horizon $H_p$, nk is the number of delays in model.
Hc=2 : control horizon $H_c$.
Mwt=fill(1.0,model.ny) : main diagonal of $\mathbf{M}$ weight matrix (vector).
Nwt=fill(0.1,model.nu) : main diagonal of $\mathbf{N}$ weight matrix (vector).
Lwt=fill(0.0,model.nu) : main diagonal of $\mathbf{L}$ weight matrix (vector).
M_Hp=diagm(repeat(Mwt,Hp)) : positive semidefinite symmetric matrix $\mathbf{M}_{H_p}$.
N_Hc=diagm(repeat(Nwt,Hc)) : positive semidefinite symmetric matrix $\mathbf{N}_{H_c}$.
L_Hp=diagm(repeat(Lwt,Hp)) : positive semidefinite symmetric matrix $\mathbf{L}_{H_p}$.
Cwt=1e5 : slack variable weight $C$ (scalar), use Cwt=Inf for hard constraints only.
optim=JuMP.Model(OSQP.MathOptInterfaceOSQP.Optimizer) : quadratic optimizer used in the predictive controller, provided as a JuMP.Model (default to OSQP optimizer).
additional keyword arguments are passed to SteadyKalmanFilter constructor.

Examples

julia> model = LinModel([tf(3, [30, 1]); tf(-2, [5, 1])], 4);

julia> mpc = LinMPC(model, Mwt=[0, 1], Nwt=[0.5], Hp=30, Hc=1)
LinMPC controller with a sample time Ts = 4.0 s, OSQP optimizer, SteadyKalmanFilter estimator and:
 30 prediction steps Hp
  1 control steps Hc
  1 slack variable ϵ (control constraints)
  1 manipulated inputs u (0 integrating states)
  4 estimated states x̂
  2 measured outputs ym (2 integrating states)
  0 unmeasured outputs yu
  0 measured disturbances d

Extended Help

Extended Help

Manipulated inputs setpoints $\mathbf{r_u}$ are not common but they can be interesting for over-actuated systems, when nu > ny (e.g. prioritize solutions with lower economical costs). The default Lwt value implies that this feature is disabled by default.

The objective function follows this nomenclature:

VARIABLE	DESCRIPTION	SIZE
$H_p$	prediction horizon (integer)	`()`
$H_c$	control horizon (integer)	`()`
$\mathbf{ΔU}$	manipulated input increments over $H_c$	`(nu*Hc,)`
$\mathbf{D̂}$	predicted measured disturbances over $H_p$	`(nd*Hp,)`
$\mathbf{Ŷ}$	predicted outputs over $H_p$	`(ny*Hp,)`
$\mathbf{U}$	manipulated inputs over $H_p$	`(nu*Hp,)`
$\mathbf{R̂_y}$	predicted output setpoints over $H_p$	`(ny*Hp,)`
$\mathbf{R̂_u}$	predicted manipulated input setpoints over $H_p$	`(nu*Hp,)`
$\mathbf{M}_{H_p}$	output setpoint tracking weights over $H_p$	`(nyHp, nyHp)`
$\mathbf{N}_{H_c}$	manipulated input increment weights over $H_c$	`(nuHc, nuHc)`
$\mathbf{L}_{H_p}$	manipulated input setpoint tracking weights over $H_p$	`(nuHp, nuHp)`
$C$	slack variable weight	`()`
$ϵ$	slack variable for constraint softening	`()`

source

LinMPC(estim::StateEstimator; <keyword arguments>)

Use custom state estimator estim to construct LinMPC.

estim.model must be a LinModel. Else, a NonLinMPC is required.

Examples

julia> estim = KalmanFilter(LinModel([tf(3, [30, 1]); tf(-2, [5, 1])], 4), i_ym=[2]);

julia> mpc = LinMPC(estim, Mwt=[0, 1], Nwt=[0.5], Hp=30, Hc=1)
LinMPC controller with a sample time Ts = 4.0 s, OSQP optimizer, KalmanFilter estimator and:
 30 prediction steps Hp
  1 control steps Hc
  1 slack variable ϵ (control constraints)
  1 manipulated inputs u (0 integrating states)
  3 estimated states x̂
  1 measured outputs ym (1 integrating states)
  1 unmeasured outputs yu
  0 measured disturbances d

source

ExplicitMPC

ModelPredictiveControl.ExplicitMPC — Type

ExplicitMPC(model::LinModel; <keyword arguments>)

Construct an explicit linear predictive controller based on LinModel model.

The controller minimizes the following objective function at each discrete time $k$:

\[\begin{aligned} \min_{\mathbf{ΔU}} \mathbf{(R̂_y - Ŷ)}' \mathbf{M}_{H_p} \mathbf{(R̂_y - Ŷ)} + \mathbf{(ΔU)}' \mathbf{N}_{H_c} \mathbf{(ΔU)} \\ + \mathbf{(R̂_u - U)}' \mathbf{L}_{H_p} \mathbf{(R̂_u - U)} \end{aligned}\]

See LinMPC for the variable definitions. This controller does not support constraints but the computational costs are extremely low (array division), therefore suitable for applications that require small sample times. The keyword arguments are identical to LinMPC, except for Cwt and optim which are not supported.

This method uses the default state estimator, a SteadyKalmanFilter with default arguments.

Examples

julia> model = LinModel([tf(3, [30, 1]); tf(-2, [5, 1])], 4);

julia> mpc = ExplicitMPC(model, Mwt=[0, 1], Nwt=[0.5], Hp=30, Hc=1)
ExplicitMPC controller with a sample time Ts = 4.0 s, SteadyKalmanFilter estimator and:
 30 prediction steps Hp
  1 control steps Hc
  1 manipulated inputs u (0 integrating states)
  4 estimated states x̂
  2 measured outputs ym (2 integrating states)
  0 unmeasured outputs yu
  0 measured disturbances d

source

ExplicitMPC(estim::StateEstimator; <keyword arguments>)

Use custom state estimator estim to construct ExplicitMPC.

estim.model must be a LinModel. Else, a NonLinMPC is required.

source

NonLinMPC

ModelPredictiveControl.NonLinMPC — Type

NonLinMPC(model::SimModel; <keyword arguments>)

Construct a nonlinear predictive controller based on SimModel model.

Both NonLinModel and LinModel are supported (see Extended Help). The controller minimizes the following objective function at each discrete time $k$:

\[\begin{aligned} \min_{\mathbf{ΔU}, ϵ}\ & \mathbf{(R̂_y - Ŷ)}' \mathbf{M}_{H_p} \mathbf{(R̂_y - Ŷ)} + \mathbf{(ΔU)}' \mathbf{N}_{H_c} \mathbf{(ΔU)} \\& + \mathbf{(R̂_u - U)}' \mathbf{L}_{H_p} \mathbf{(R̂_u - U)} + C ϵ^2 + E J_E(\mathbf{U_e}, \mathbf{Ŷ_e}, \mathbf{D̂_e}, \mathbf{p}) \end{aligned}\]

subject to setconstraint! bounds, and the custom inequality constraints:

\[\mathbf{g_c}(\mathbf{U_e}, \mathbf{Ŷ_e}, \mathbf{D̂_e}, \mathbf{p}, ϵ) ≤ \mathbf{0}\]

The economic function $J_E$ can penalizes solutions with high economic costs. Setting all the weights to 0 except $E$ creates a pure economic model predictive controller (EMPC). As a matter of fact, $J_E$ can be any nonlinear function to customize the objective, even if there is no economic interpretation to it. The arguments of $J_E$ and $\mathbf{g_c}$ include the manipulated inputs, predicted outputs and measured disturbances, extended from $k$ to $k + H_p$ (inclusively, see Extended Help for more details):

\[ \mathbf{U_e} = \begin{bmatrix} \mathbf{U} \\ \mathbf{u}(k+H_p-1) \end{bmatrix} , \quad \mathbf{Ŷ_e} = \begin{bmatrix} \mathbf{ŷ}(k) \\ \mathbf{Ŷ} \end{bmatrix} , \quad \mathbf{D̂_e} = \begin{bmatrix} \mathbf{d}(k) \\ \mathbf{D̂} \end{bmatrix}\]

The argument $\mathbf{p}$ is a custom parameter object of any type, but use a mutable one if you want to modify it later e.g.: a vector. See LinMPC Extended Help for the definition of the other variables.

Tip

Replace any of the arguments of $J_E$ and $\mathbf{g_c}$ functions with _ if not needed (see e.g. the default value of JE below).

This method uses the default state estimator:

if model is a LinModel, a SteadyKalmanFilter with default arguments;
else, an UnscentedKalmanFilter with default arguments.

Warning

See Extended Help if you get an error like: MethodError: no method matching Float64(::ForwardDiff.Dual).

Arguments

model::SimModel : model used for controller predictions and state estimations.
Hp=nothing: prediction horizon $H_p$, must be specified for NonLinModel.
Hc=2 : control horizon $H_c$.
Mwt=fill(1.0,model.ny) : main diagonal of $\mathbf{M}$ weight matrix (vector).
Nwt=fill(0.1,model.nu) : main diagonal of $\mathbf{N}$ weight matrix (vector).
Lwt=fill(0.0,model.nu) : main diagonal of $\mathbf{L}$ weight matrix (vector).
M_Hp=diagm(repeat(Mwt,Hp)) : positive semidefinite symmetric matrix $\mathbf{M}_{H_p}$.
N_Hc=diagm(repeat(Nwt,Hc)) : positive semidefinite symmetric matrix $\mathbf{N}_{H_c}$.
L_Hp=diagm(repeat(Lwt,Hp)) : positive semidefinite symmetric matrix $\mathbf{L}_{H_p}$.
Cwt=1e5 : slack variable weight $C$ (scalar), use Cwt=Inf for hard constraints only.
Ewt=0.0 : economic costs weight $E$ (scalar).
JE=(_,_,_,_)->0.0 : economic or custom cost function $J_E(\mathbf{U_e}, \mathbf{Ŷ_e}, \mathbf{D̂_e}, \mathbf{p})$.
gc=(_,_,_,_,_,_)->nothing or gc! : custom inequality constraint function $\mathbf{g_c}(\mathbf{U_e}, \mathbf{Ŷ_e}, \mathbf{D̂_e}, \mathbf{p}, ϵ)$, mutating or not (details in Extended Help).
nc=0 : number of custom inequality constraints.
p=model.p : $J_E$ and $\mathbf{g_c}$ functions parameter $\mathbf{p}$ (any type).
optim=JuMP.Model(Ipopt.Optimizer) : nonlinear optimizer used in the predictive controller, provided as a JuMP.Model (default to Ipopt optimizer).
additional keyword arguments are passed to UnscentedKalmanFilter constructor (or SteadyKalmanFilter, for LinModel).

Examples

julia> model = NonLinModel((x,u,_,_)->0.5x+u, (x,_,_)->2x, 10.0, 1, 1, 1, solver=nothing);

julia> mpc = NonLinMPC(model, Hp=20, Hc=1, Cwt=1e6)
NonLinMPC controller with a sample time Ts = 10.0 s, Ipopt optimizer, UnscentedKalmanFilter estimator and:
 20 prediction steps Hp
  1 control steps Hc
  1 slack variable ϵ (control constraints)
  1 manipulated inputs u (0 integrating states)
  2 estimated states x̂
  1 measured outputs ym (1 integrating states)
  0 unmeasured outputs yu
  0 measured disturbances d

Extended Help

Extended Help

NonLinMPC controllers based on LinModel compute the predictions with matrix algebra instead of a for loop. This feature can accelerate the optimization, especially for the constraint handling, and is not available in any other package, to my knowledge.

The economic cost $J_E$ and custom constraint $\mathbf{g_c}$ functions receive the extended vectors $\mathbf{U_e}$ (nu*Hp+nu elements), $\mathbf{Ŷ_e}$ (ny+ny*Hp elements) and $\mathbf{D̂_e}$ (nd+nd*Hp elements) as arguments. They all include the values from $k$ to $k + H_p$ (inclusively). The custom constraint also receives the slack $ϵ$ (scalar), which is always zero if Cwt=Inf.

More precisely, the last two time steps in $\mathbf{U_e}$ are forced to be equal, i.e. $\mathbf{u}(k+H_p) = \mathbf{u}(k+H_p-1)$, since $H_c ≤ H_p$ implies that $\mathbf{Δu}(k+H_p) = \mathbf{0}$. The vectors $\mathbf{ŷ}(k)$ and $\mathbf{d}(k)$ are the current state estimator output and measured disturbance, respectively, and $\mathbf{Ŷ}$ and $\mathbf{D̂}$, their respective predictions from $k+1$ to $k+H_p$. If LHS represents the result of the left-hand side in the inequality $\mathbf{g_c}(\mathbf{U_e}, \mathbf{Ŷ_e}, \mathbf{D̂_e}, \mathbf{p}, ϵ) ≤ \mathbf{0}$, the function gc can be implemented in two possible ways:

Non-mutating function (out-of-place): define it as gc(Ue, Ŷe, D̂e, p, ϵ) -> LHS. This syntax is simple and intuitive but it allocates more memory.
Mutating function (in-place): define it as gc!(LHS, Ue, Ŷe, D̂e, p, ϵ) -> nothing. This syntax reduces the allocations and potentially the computational burden as well.

The keyword argument nc is the number of elements in LHS, and gc!, an alias for the gc argument (both gc and gc! accepts non-mutating and mutating functions).

The optimization relies on JuMP automatic differentiation (AD) to compute the objective and constraint derivatives. Optimizers generally benefit from exact derivatives like AD. However, the NonLinModel state-space functions must be compatible with this feature. See Automatic differentiation for common mistakes when writing these functions.

Note that if Cwt≠Inf, the attribute nlp_scaling_max_gradient of Ipopt is set to 10/Cwt (if not already set), to scale the small values of $ϵ$.

source

NonLinMPC(estim::StateEstimator; <keyword arguments>)

Use custom state estimator estim to construct NonLinMPC.

Examples

julia> model = NonLinModel((x,u,_,_)->0.5x+u, (x,_,_)->2x, 10.0, 1, 1, 1, solver=nothing);

julia> estim = UnscentedKalmanFilter(model, σQint_ym=[0.05]);

julia> mpc = NonLinMPC(estim, Hp=20, Hc=1, Cwt=1e6)
NonLinMPC controller with a sample time Ts = 10.0 s, Ipopt optimizer, UnscentedKalmanFilter estimator and:
 20 prediction steps Hp
  1 control steps Hc
  1 slack variable ϵ (control constraints)
  1 manipulated inputs u (0 integrating states)
  2 estimated states x̂
  1 measured outputs ym (1 integrating states)
  0 unmeasured outputs yu
  0 measured disturbances d

source

Move Manipulated Input u

ModelPredictiveControl.moveinput! — Function

moveinput!(mpc::PredictiveController, ry=mpc.estim.model.yop, d=[]; <keyword args>) -> u

Compute the optimal manipulated input value u for the current control period.

Solve the optimization problem of mpc PredictiveController and return the results $\mathbf{u}(k)$. Following the receding horizon principle, the algorithm discards the optimal future manipulated inputs $\mathbf{u}(k+1), \mathbf{u}(k+2), ...$ Note that the method mutates mpc internal data but it does not modifies mpc.estim states. Call preparestate!(mpc, ym, d) before moveinput!, and updatestate!(mpc, u, ym, d) after, to update mpc state estimates. Setpoint and measured disturbance previews can be implemented with the R̂y, R̂u and D̂ keyword arguments.

Calling a PredictiveController object calls this method.

See also LinMPC, ExplicitMPC, NonLinMPC.

Arguments

Info

Keyword arguments with emphasis are non-Unicode alternatives.

mpc::PredictiveController : solve optimization problem of mpc.
ry=mpc.estim.model.yop : current output setpoints $\mathbf{r_y}(k)$.
d=[] : current measured disturbances $\mathbf{d}(k)$.
D̂=repeat(d, mpc.Hp) or Dhat : predicted measured disturbances $\mathbf{D̂}$, constant in the future by default or $\mathbf{d̂}(k+j)=\mathbf{d}(k)$ for $j=1$ to $H_p$.
R̂y=repeat(ry, mpc.Hp) or Rhaty : predicted output setpoints $\mathbf{R̂_y}$, constant in the future by default or $\mathbf{r̂_y}(k+j)=\mathbf{r_y}(k)$ for $j=1$ to $H_p$.
R̂u=mpc.Uop or Rhatu : predicted manipulated input setpoints $\mathbf{R̂_u}$, constant in the future by default or $\mathbf{r̂_u}(k+j)=\mathbf{u_{op}}$ for $j=0$ to $H_p-1$.

Examples

julia> mpc = LinMPC(LinModel(tf(5, [2, 1]), 3), Nwt=[0], Hp=1000, Hc=1);

julia> preparestate!(mpc, [0]); ry = [5];

julia> u = moveinput!(mpc, ry); round.(u, digits=3)
1-element Vector{Float64}:
 1.0

source