Model reduction
Model reduction refers to a reduction of complexity, sometimes by approximation, of a model by a simpler model. Many control-design workflows benefit from model reduction, some examples are
- A plant model is derived by linearization of a detailed model implemented in ModelingToolkit. Such a model may contain 1000s of states, but may be accurately described in a particular operating point by a low-order linear model.
- When only subsets of all inputs and outputs of a large MIMO model are required, some states of the original model may no longer be required.
- Some control-design techniques, such as $\mathcal{H}_\infty$ design or
glover_mcfarlane
may result in high-order controllers. Such a controller may sometimes be reduced while maintaining most of the stability margin. - Some modes in a large model may be unobservable or uncontrollable.
Reduced-order models require less effort to simulate and may improve the numerical performance of some algorithms. Model reduction of linear time-invariant models is a well-developed field, and this page lists some of the available functionality. For reduction of nonlinear models, consider a linearization-based approach, or build a nonlinear surrogate model.
Model reduction using balanced truncation is available through the functions
- Model reduction GUI, available as
JuliaSimControl.app_modelreduction
. minreal
obtain a minimal realization containing only controllable and observable modes.sminreal
obtain a structurally minimal realization of a state-space model.baltrunc2
for standard model reduction.frequency_weighted_reduction
for reduction with frequency focus.baltrunc_coprime
for normalized-coprime factor reduction.baltrunc_unstab
for unstable models.controller_reduction
reduce the order of controllers with guaranteed stability margins. See example.
Reduction of very large models
Balanced truncation requires the solution to a Lyapunov equation which may be prohibitively expensive for large systems. For systems of order above about 500, a method based on frequency-domain fitting may be substantially faster if the desired model order is less than about 100. The following example illustrates the procedure.
using JuliaSimControl, ControlSystemIdentification, Plots
ny,nu,nx = 5,5,1000 # number of outputs, inputs and states
Ts = 1 # Sample time
G = ssrand(ny,nu,nx; Ts, proper=true); # Generate a random system
N = 200 # Number of frequency points
w = range(0, stop=pi/Ts-1/N, length=N) # Frequency vector
frd = FRD(w, G); # Build a frequency-response data object
nxr = 60 # Reduced model order
@time Gh, x0 = subspaceid(frd, G.Ts, nxr; r=nxr+1, zeroD=true); # Fit frequency response
sigmaplot([G, Gh], w[2:end], lab=["Full order" "Reduced order"])
The frequency-fitting method does not have support for exact DC matching like the balanced-truncation method does, but there exists an option for frequency-based weighting which can achieve similar results. See subspaceid
for additional details.