PID Controller Design for Nonlinear Systems Using
Discrete-Time Local Model Networks
4. Workshop f¨ur Modellbasierte Kalibriermethoden
Nikolaus Euler-Rolle, Christoph Hametner, Stefan Jakubek
Christian Mayr(AVL List GmbH)
Feedback Control of Nonlinear Systems
Motivation
Implementation ofTwo-Degrees-of-Freedom controlusing local model networks Feedforward part improves the dynamic performance
- Reference tracking - Deadtime - Input saturation
Controller design on (semi)-physical process modelsinstead of testbed runs Opportunity of inexpensivefeasibility studies and rapid prototyping
PID Plant
w
w
*
u
*
u
y
Feedback Control of Nonlinear Systems
Motivation
Implementation ofTwo-Degrees-of-Freedom controlusing local model networks Feedforward part improves the dynamic performance
- Reference tracking - Deadtime - Input saturation
Controller design on (semi)-physical process modelsinstead of testbed runs Opportunity of inexpensivefeasibility studies and rapid prototyping
Approach
Globally nonlinear process model (based on input/output measurements) Design of nonlinear PID controllers with guaranteed global stability Fully automated generation of a dynamic feedforward control
Controller Design Workflow
Testbed
Maps Signals DoE [n, q, u]LMN
SS-Model
Local PIDs
DoE Optimisation y Simulation Parameter Controller Maps Performance Stability Identif icationController Design Workflow
Testbed
Maps Signals DoE [n, q, u]LMN
SS-Model
Local PIDs
DoE Optimisation y Simulation Parameter Controller Maps Performance Stability Identif ication Dynamic FF-ControlLocal Model Network
Overview 1000 1500 2000 2500 5 10 15 20 25 30 In je ct io n M a ss , m g / st ro ke Engine Speed, rpm local globalLocal Model Network
Globally nonlinear dynamical system represented by local linear models Found by system identification Local stability proof & controller design using linear methods
⇒ Global approach necessary (due to transition, model interpolation...)
o for nonlinear systems
Typical PID Controller Structure
Example: Engine Control Unit
-min max P-Part I-Part DT1-Part anti windup Map Map Feedforward- Feedback-Control n n n q q q u w y e ufb uffFeedback Controlled Local Model Network
Concept
One local controller (LC) per local model (LM)
Scheduling of parameters according to the validity functions of local models (Parallel
DistributedCompensator)
KP ID(Φ) =PΦiK (i) P ID
Formal split into inputs used for controlu
and disturbancesz !"#$%"&&'% (')*+# ,-!. ,- / ,!!. ,! / + 01'%2$*#+! 1"*#$! ('1'#('#$
3
,-!/ , -,!!/ ,! -,-!- ,!!-Nonlinear process is approximated by a local model network
Trade-Off: model fit↔simple controller design
Closed-Loop State-Space Representation
Including Error Signal Adaptation
Input Scheduler q−1I v(k) B(Φ) B(Φ) E(Φ) A(Φ) KP ID(Φ, e) we(Φ, e) cT x(k+ 1) x(k) ˆ y(k) w(k) ˆ z(k) z(k) f(Φ) -System P re -F il te r
Figure:Local model network with PID controller in state-space representation
State Equation
x(k+ 1) = [A(Φ)−B(Φ)KP ID(Φ, e)]x(k) +B(Φ)G(Φ, e)w(k) +E(Φ)ˆz(k)
+f(Φ) +B(Φ)we(Φ, e)
ˆ
Overview of the Design Procedure
Controller Design
Basic calibration (linear design methods per local model) Generation of a suitable performance sequence (DoE)
- Operating range (e.g.: 1000–4000 rpm, 0–70 mg/stroke) - Holding time
- Gradients (e.g.: engine speed) - Filtering
Nonlinear, multi-objective optimisation of controller parameters considering
- Performance - Stability
Multi-objective optimisation of the parameters of the error signal adaptation (optional)
Multi-Objective Genetic Algorithm
Objective Function min fm(xopt) subject to gj(xopt)≥0 hk(xopt) = 0 x(ilb)≤xi≤x(iub) fSStability(by Lyapunov’s direct method) fP Performance(by a closed-loop simulation)0 fP fS Paretofrontier
GA Population
1 n · · · · · · Individuals G en o m e G en o m e F it n es s F it n es s S ta b il it y S ta b il it y P er fo rm a n ce P er fo rm a n ceFitness Function: Stability
Lyapunov’s Direct Method for Discrete-Time Systems
Stability of Dynamic Systems
A positive definite, scalar Lyapunov function
V(k) =V(x(k))with state vectorx(k)
proves globalasymptoticstability if:
o V(x(k) =0) = 0 o V(k)>0 forx(k)6= 0 o V(k)→ ∞ askx(k)k → ∞
o V(k+ 1)< V(k) ∀k∈N+
or globalexponentialstability if:
o V(k+ 1)≤α2V(k) ∀k∈N+ with decay rate0< α <1
Results in Linear Matrix Inequalities (LMIs), which are solved by optimisation
Sufficient but not necessary condition Common Quadratic Lyapunov Function
LMI Problem
P ≻0
Fitness Function: Performance
Requirements
Assessment of the closed-loop performance for a given set of parameters
Representative synthetic reference is generated by DoE
Desired trajectory is PT1-filtered
Fitness Function
Closed-loop simulation of the reference cycle for each genome
Sum of squared errors
fP =P∀k(ˆy(k)−ydmd(k))2 Time 0 5 10 15 0 0.5 1 1.5 2
Pareto-Optimal Solutions
0.995 1 1.005 1.01 1.015 1.02 1.025 1.03 1 1.5 2 2.5 3 3.5 4x 10 7 A B Performance Stability fPFeedforward Control
State of the Art: Static Model Inversion
Steady state input is found by static model inversion
u(Φ) = [cT(I−A(Φ))−1B(Φ)]−1(w(Φ)−cT(I−A(Φ))−1(E(Φ)ˆz(Φ) +f(Φ)))
Stored in a map
Dynamic Feedforward Control
PID Plant
w
w
*
u
*
u
y
Dynamic Feedforward Control
Generation using Local Model Networks
Benefits
Automatic generationof a dynamic feedforward control law for nonlinear dynamic systems
Exploits thegeneric model structureof local model networks
Model complexitymay be arbitrarily high
Applicableonlinefor any reference trajectory without pre-planning
Properties
Based on an open-loop state-space model
Realised by afeedback linearizinginput transformation Restricted to globally minimum-phase local model networks
Feedback Linearization
Undamped Nonlinear Oscillation
Consider an undamped oscillator with a nonlinear spring force characteristic
f(y), which is to be stabilized using constantcand inputu
¨
y+f(y) =cu
Figure:Air suspension
Exact Linearization
For this second order system, the state variables are chosen as
y=x1 ˙
y= ˙x1=x2 ¨
Feedback Linearization
Undamped Nonlinear Oscillation
Consider an undamped oscillator with a nonlinear spring force characteristic
f(y), which is to be stabilized using constantcand inputu
¨
y+f(y) =cu
Figure:Air suspension
Exact Linearization
For this second order system, the state variables are chosen as
y=x1 ˙ y= ˙x1=x2 ¨ y= ¨x1= ˙x2=cu−f(y)=v 1 s 1 s v y
Feedforward Control
Undamped Nonlinear Oscillation
Exact Linearization
For a two times differentiable desired trajectoryw, the nonlinear feedforward control inputu∗can be found from
v= ¨! w=cu∗−f(w) → u∗=w¨+f(w) c 1 s 1 s y u u∗ w ¨ w C u∗ =w¨+f(w) c
Demonstration Example
Automatic Feedforward Control Design
Wiener Model G(z) = P(z) U(z)= 0.6z−3 1−1.3z−1 + 0.8825z−2 −0.1325z−3 y(k) =f(p(k)) = arctan(p(k)) Figure:Wiener Model approximated by an LMN:
ˆ y ( k − 1 ) u(k−3) 6 5 4 3 2 1 −3 −2 −1 0 1 2 3 −1 −0.5 0 0.5 1
Feedforward Controlled Simulation
Wiener Model yW ie n e r u w , ˆy 40 60 80 100 120 140 160 180 200 220 40 60 80 100 120 140 160 180 200 220 −1 0 1 −3 0 3 −1 0 1Feedforward Controlled Simulation
Wiener Model yFFC w ˆy Samples 0 50 100 150 200 250 300 350 400 450 500 −1.5 −1 −0.5 0 0.5 1 1.5 PID Plant w w* u* u yTwo-Degrees-of-Freedom Control
Wiener Model y2DoF yFFC w Samples ˆy 0 50 100 150 200 250 300 350 400 450 500 −1.5 −1 −0.5 0 0.5 1 1.5 PID Plant w w* u* u yTwo-Degrees-of-Freedom Control
Wiener Model y2DoF yPID w Samples ˆy 0 50 100 150 200 250 300 350 400 450 500 −1.5 −1 −0.5 0 0.5 1 1.5 PID Plant w w* u* u yConclusion & Outlook
Two-Degrees-of-Freedom Control
Nonlinear PID controller design using local model networks Multi-objective optimisation of controller parameters considering
Stability Performance
Automatic feedforward control law generation for minimum-phase local model networks
Outlook
Application of a Lyapunov function to check internal stability Considering input constraints
Fitness Function: Stability
Common Quadratic Lyapunov Function for Closed-Loop Systems
Exponential stability with decay rateαof the closed-loop feedback system is shown, if symmetric matricesP andXijexist and the following conditions are fulfilled:
P ≻0 infn0< α <1 :ΛT ijPΛij+Xijα 2 Po ˜ X= X11 X12 · · · X1I X12 X22 · · · X2I . . . . .. . . . X1I X2I · · · XII ≻0 ∀i∈ I,∀i≤j≤I using Λij=Gij+Gji 2 , Gij=Ai−Bik T P ID,jC.
Fitness Function: Stability
Common Quadratic Lyapunov Function for Closed-Loop Systems
Exponential stability with decay rateαof the closed-loop feedback system is shown, if symmetric matricesP andXijexist and the following conditions are fulfilled:
P ≻0 infn0< α <1 :ΛT ijPΛij+Xijα2P o ˜ X= X11 X12 · · · X1I X12 X22 · · · X2I . . . . .. . . . X1I X2I · · · XII ≻0 ∀i∈ I,∀i≤j≤I using Λij=Gij+Gji 2 , Gij=Ai−Bik T P ID,jC.