Next Generation Adaptive and Intelligent Algorithms for the Control of Complex and Dynamic Systems

boorishadamantAI and Robotics

Oct 29, 2013 (3 years and 1 month ago)

66 views

Next Generation Adaptive
and Intelligent Algorithms
for the Control of Complex
and Dynamic Systems


Dr. Sukumar Kamalasadan

Department of Engineering and Computer Technology

University of West Florida

Pensacola, FL
-
32514

Sukumar Kamalasadan Ph.D.

2

10/29/2013

Presentation Outline


Overview



Part I: Theoretical Design and Algorithms



Part II: Current Research Projects


Speed Control of Synchronous Generator.


Multi
-
Machine Power System Control and Angular Stability.



Part III: Other Research Projects and Directions


Smart
-
Grid Applications


Wide Area Monitoring and Control based on scalable intelligent
supervisory loop concept.


Distributed Power Generation Control and Grid Interface.



Summary

Overview


Main focus


Modeling and control of dynamic systems


Mathematical modeling


Using Computational Intelligence


Simulation using computer algorithms


Designing and developing novel control, optimization
and identification techniques


Real
-
time implementation of scalable algorithms


Integrating research elements to teaching


Dissemination and Outreach


This talk is about one particular dynamic system

Sukumar Kamalasadan Ph.D.

4

10/29/2013

Overview:

Importance of Modern Power System Control


Fast acting MIMO devices such as generators, Distributed
Generation (DG) and their integration, tight and congested
transmission systems, deregulated power system …


Shows multiple behavior such as: discrete changes (transformer
taps), deterministic operations (voltage and speed control),
stochastic behavior (load forecasting), optimal needs (power
transactions with constraints).


Existence of multiple controllers that increases the system
complexity and controller interactions.


Advances in high speed digital processor and computer
architecture enhance the feasibility of modern control design
techniques:


Operates in real
-
time


Provide some elements of learning and adaptation

Sukumar Kamalasadan Ph.D.

5

10/29/2013

Overview: Existing control topologies for
Generator/Power Systems


Linear controllers such as conventional Automatic Voltage Regulators (AVR)
(voltage control) and Governor (speed control).


Conventional Power System Stabilizers (CPSS) used for damping of
generator oscillations, used in industry (P. Kundur, O.P. Malik et. al.).


Model based controllers for generators (adaptive controllers) has been
proposed and used (adaptable and simple in architecture) (K.S. Narendra,
Ghandakly et. al.)

Provide linear adaptation but no learning and memory.


Nonlinear controllers and adaptive nonlinear controller (Feedback
Linearization, backstepping)


Useful but often cannot cover entire domain.


Neural network based designs (Venayagamoorthy, Harley, Lee)

Provide
learning and adaptation especially with time delayed system

Not always
needed.


Proposed Solution:

Provide hybrid control architecture that is system
-
centric
in nature.


Sukumar Kamalasadan Ph.D.

6

10/29/2013

Overview: Intelligent Power System Control
and Analysis


Why Hybrid Intelligent Control Architecture?


Operates in a decentralized way while exhibiting
desirable system
-
wide characteristics (Complex tasks
can be made simpler).


Produces effective local decisions that contribute
towards a coherent and effective overall system
(Emerging behavior).


Ability to interact and coordinate with existing design
and are adaptable (organizational behavior).


Capable of providing efficient and effective signals
based on system needs (case based approach).


Provide adaptation, learning and model
-
less control.


Sukumar Kamalasadan Ph.D.

7

10/29/2013

Overview: Current Research Efforts:

Focus Areas


Hybrid intelligent control
-

Theoretical formulations, design
and development such as,


Issues related to stability, adaptation and global contributions in
changing plant conditions.


Reliability, robustness and adaptability.


System modeling, algorithmic development, implementation.


New and Suitable computational intelligence techniques:


Methods in online and offline learning.


Issues such as tuning, autonomous action.


Power System Control and Stability


Generator control.


Wide Area Controllers (WAC).


Control of other electric machines.


Control of energy sources, integration of Distributed Generation
(DG) with mega grid.

Sukumar Kamalasadan Ph.D.

8

10/29/2013

Part I: Design Concept: Hybrid Architecture
for Coordinated Control


Three Design structure with System Supervision


Systems that shows parametric uncertainty;


A conventional adaptive module (such as Model Reference
Adaptive Control) to adaptively monitor system output and develop
control action.


Systems that shows modal changes;


Intelligent module to recover these changes and develop a desired
reference model trajectory. Important in the presence of multiple
modes of operations.


Systems that shows functional changes and/or influenced
by external disturbances;


An intelligent module to approximate the changing nonlinear
function such as offline/online trained neural network.



Sukumar Kamalasadan Ph.D.

9

10/29/2013

Part I: Design Concepts
:

Intelligent Adaptive Control : Supervisory Loop
Approach

Adaptive Controller

(Controller 1)


System

Under Consideration

Plant

Output

Error

Input

Signal

Reference

Output

Adaptive

Control Law

Reference Model/
Parameter Estimation






Sukumar Kamalasadan Ph.D.

10

10/29/2013

Part I: Design Concept: Hybrid Intelligent
Control Architecture

Sukumar Kamalasadan Ph.D.

11

10/29/2013

Part I: Design Concept: System
-
Centric
Controllers: Design Scenarios

Fuzzy Reference Model
Generator (FRMG)

Monitor

Adaptive
Controller

System under
consideration

Σ

Monitor

RBFNN

Controller

Creative

Controller


Figure 1: Scenario 1: Proposed Framework

Fuzzy Reference
Model Generator

Monitor

Adaptive
Controller


Multi
-
machine
System

Σ


Figure 2: Scenario2: Proposed Framework

Fuzzy Reference
Model Generator

Monitor

Adaptive
Controller

Multi
-
machine
System

Σ

Monitor

RBFNN

Controller


Figure 3: Scenario 3: Proposed Framework


Hypothesis for System
-
Centric Controllers



Changes in Modes of Operation: Fuzzy Reference Model
Generator (FRMG).


Nonlinear Behavior (ability to cope up with system
nonlinearity) but the target of operation known: RBFNN
Controller (with supervisory learning).


Nonlinear Behavior and target unknown: Reinforcement
learning.



Challenges


Controller’s Integrity, Design and Development Issues


Implementation Issues, continuous
-
discrete interplay


Sukumar Kamalasadan Ph.D.

12

10/29/2013

Part I: Design Concept: Controller 1

Is a stable matrix of order (n
-
1) X(n
-
1)
such that

The

Model

Reference

Adaptive

Controller

can

be

formulated

as

Where

theta

is


and

omega

is


and


Start

Calculate error from

outputs

Adaptive Mechanism

Calculate control value

Continue

Calculate theta

Calculate Omega

where,

e represents the error,

represents the fuzzy contribution

represents the adaptive factor

1) Model based design, 2)
Adaptation capability, however no
memory, no learning 3) Able to expand to the next level for
plant drastic changes


Adaptation

Regressor

Sukumar Kamalasadan Ph.D.

13

10/29/2013

Part I: Switching Mechanism


Design
Concept

Fuzzy

system

can

be

represented

as

A

reference

model

in

a

state

space

form

will

be


Modal

transitions

can

be

included

as

In

general

it

can

be

written

as

Start

System Auxiliary States

Fuzzy Logic Scheme

Defuzzification

Reference Model

Fuzzification

Rule Base

1) Multiple Model switching, 2)
Stable 3) Able to work
coherently with model based adaptive controller 3) Need offline
design and knowledge base development


Further details: Kamalasadan et al. (2004), (2005), (2006), (2007)

Sukumar Kamalasadan Ph.D.

14

10/29/2013

Part I: Design Concept:

Growing Dynamic RBFNN Controller

New
Node

Existing Node
Movement

y(t)

Center
Movement

Number of
nodes
required for
a Static
Network

Active Nodes

Static
Network

Nodal Region

Train offline
-

Adaptive Online

μ
1

μ
2

μ
n

.

.

.

.

.

.

.

.

.

Bias

Bias

y
1

y
p



μ

σ

Sample Basis Function

X
1

X
2

X
n

α
11

α
p1

Input Layer

Hidden Layer

Output Layer

μ
=Center positions

h=hidden neurons

σ
=Gaussian functions

α
=Weights

ε
= Distance

e=y
i
-
f(x
i
)

Sukumar Kamalasadan Ph.D.

15

10/29/2013

Part I: Controller 2


Design Concept

The

neuro
-
controller


can

be

written

as

Adding

hidden

units
:


Tuning

laws

are

Where

P

is

positive

definite

matrix

and

B

is

the

gradient

Growth

parameter

criterion

Add

new

unit

with


α
(h+1)
=e
i
,
μ
(h+1)=X
i,

σ
(h+1)=k||X
i
-
μ
||

1) Function approximation based design, 2) Learn offline,
Adaptation online, associative memory 3) nonlinear and
supervisory learning 4) Unique algorithm that can grow and
prune and provide sequential learning 5) Able to expand to the
next level for optimal control/reinforcement learning

Start

Get System States

RBFNN Structure

Calculate distance and Output

Update Weight and

Generate Control Value

Generate Nodes

Calculate Centers and radii

Grow or prune?

Growing and

Pruning Stage

No

Yes

Sukumar Kamalasadan Ph.D.

16

10/29/2013

Part I: Creative Controller

DHP

based

controller

Action

update

Critic

Error

Sukumar Kamalasadan Ph.D.

17

10/29/2013

Part I: Under nonlinear Optimal Condition??

Plant

Critic Network

Action Network

Supervisory Learning

(Earlier Designs)

Scheduler Block


(a=ka
E
+(1
-
k)a
S
)

V
ref

+

+

X(t)

A(t)

Exploration

Shaping (Prediction)

J(t)

1.0

TDL

Transport lag

X(t)

X(t)

X(t
-
Δ
t
)

F
-
1
(X,X
d
)

TDL

Nonlinear dynamic programming for Reinforce Learning (RL)

RBFNN based supervisory learning (SL)

Coherency


Supervised Actor
-
Critic Reinforcement Learning Evolved
from (Rosentein, Barto et al, 2004)

Dynamic Programming, Given U (utility function), solve the Bellman

Equation to get J; use J to calculate optimal actions

Sukumar Kamalasadan Ph.D.

18

10/29/2013

Part I: Overall algorithmic functional flowchart

Current Status


Performed theoretical analysis including
stability while switching for MRAC with
FRMG block.


Developed algorithms for adaptive
controller and design basis for FRMG.


Performed theoretical analysis including
stability for MRAC
-
FRMG block with
Supervisory Learning (SL), RBFNN
controller.


Developed algorithms for a novel RBFNN
controller.


Developed the novel supervisory loop
based algorithms.

Current theoretical Work


Analyze the strategy for creative controller
using dynamic programming in presence of
optimal conditions


Adaptive Mechanism


Controller Output

Limit Reached?

Return

Disturbances/

Uncertainties/

Constraints

Start

Multi
-
machine Power
System

Is error >
Threshold


Adaptive

Controller

Is Output Desirable?

Model

Fuzzy Model

Generator

RBFNN

Controller

Input or Output
Constraints?

Creative Control

Return

J function

minimization


No

No

No

No

Yes

Yes

Yes

Yes

Sukumar Kamalasadan Ph.D.

19

10/29/2013

Part II: Control of Synchronous Generator

Single Machine Infinite Bus System (SMIB)

P
ref

Governor

Z=Re+jXe

G

T

Exciter

CPSS

MRAC

FRMG

AVR

V
t

V
ref

-

+

U
t

Σ

Δω

+

+

+

U
ps
s

U
ad

Generator Model

Representation of System Dynamics




Sukumar Kamalasadan Ph.D.

20

10/29/2013

Part II: Design and Implementation:

Modeling of Power System Components (SMIB)

Conventional Power System Stabilizer (CPSS) Model

K
stab

+/
-
0.8 p.u.

Exciter Model

IEEE Type I exciter

T
1
=0.2 T
2
=0.2 T
w
=10s K
stab
=8

Sukumar Kamalasadan Ph.D.

21

10/29/2013

Part II: Design and Implementation:

Fuzzy Reference Model Generator Design

Knowledge Base Design


The membership function of the load torque is
defined over a domain interval of [0, 1.2].


The membership function of the electric power
is defined over a domain interval of [0, 1.5].


The membership function of is defined over
a domain interval of [0, 1.5].


Each membership function is covered by five
fuzzy sets.


The fuzzy rules are derived by studying and
simulating the response of the process.


25 fuzzy rules are used to perform the fuzzy
switching to evaluate the value of

Sukumar Kamalasadan Ph.D.

22

10/29/2013

Part II: Design and Implementation:

RBFNN Design and off
-
line learning

Training of RBFNN network


At first a Pseudo Random power deviation
Δ
P
ref
and
exciter input deviation while CPSS in place (
Δ
V
field
)
is generated using Matlab
®

environment. The input
are saved
.


These signals are then fed to the generator model.
The resulting output speed deviation in
δ

and the
output terminal voltage deviation (
Δ
V
t
) are saved.


These values are time delayed by one, two and three
time periods. These time delayed signals are the
inputs to the RBFNN network.


Initially 10 hidden neurons are used and 2000 such
samples are included.


RBFNN then estimate speed deviations and terminal
voltage deviation for the subsequent period
(projection).


The output is then compared with the generator
output. The difference is the error signal.


The error signals are used to calculate change in
weight, width and RBFNN centers.


At this point the nodes growth and
pruning is not performed.


These steps are repeated until the
error is minimized to a threshold value


Once the error reaches the threshold
value the network is used for online
post
-
control phase learning.

RBFNN

Δ
P

Δ
V
f

TDL

Δ
Vt

δ

Sukumar Kamalasadan Ph.D.

23

10/29/2013

Part II: Design and Implementation:

Control and Model Development

Algorithmic Implementation


Step 1
: Plant output is used to calculate the regression
vector.


Step 2
: The output of the plant being fed to the error
block and the error between the plant output and the
reference model output is used to update the adaptive
mechanism. Adaptive vector theta is calculated
.


Step 3
: The FRMG monitoring changes in Pe and Te and
calculating values for omega at each time stamp. Based
on the error dynamics and the monitor block this is fed to
the reference model to update the model parameter
.


Step 4:

Input is being fed to RBFNN and the network
output is calculated
.


Step 5:

Gradient is fed back to RBFNN and W is updated.


Step 6:

Based on this error, centers and width are
updated
.


Step 7:

MRAC control signal is calculated based on the
delayed input from adaptive mechanism and applied to
the plant along with RBFNN signal
.


Step 8:

Reference model is updated
based on the fuzzy tuning and
requirements of the plant investigating
the monitor module
.


Step

9:

Error calculations are
performed

RBFNN

Sukumar Kamalasadan Ph.D.

24

10/29/2013

Part II: Design and Implementation

Case 1: Simulation Results

System Operating Conditions

Time
(sec)

Disturbance

0.1

Three Phase Fault

10

25% Mechanical Power
Increase

Case 1


Power =0.83 pu.


Power Factor= 0.85
lag.


Terminal
Voltage=1.062 pu.


State Initial Conditions



As the power system stress is not known
a multiple disturbance profile is used. It
can cause small signal or transient
instability.


The purpose is to assess the stability and
the deviation of all parameters.


Main parameters under observation are
angle, speed, voltage and power.


Small signal stability can cause local
mode oscillations and this test can show
-

case oscillatory or non
-
oscillatory
instability.


Figures shows that oscillations are
greater in the presence of PSS alone and
the Adaptive with FRMG could damp
these oscillations effectively.

Sukumar Kamalasadan Ph.D.

25

10/29/2013

Part II: Design and Implementation

Case 1: Simulation Results

Sukumar Kamalasadan Ph.D.

26

10/29/2013

Part II: Design and Implementation

Case 1: Simulation Results

Voltage in p.u.

Sukumar Kamalasadan Ph.D.

27

10/29/2013

Part II: Design and Implementation

Case 1: Simulation Results

Sukumar Kamalasadan Ph.D.

28

10/29/2013

Part II: Design and Implementation

Case 1: Simulation Results

Sukumar Kamalasadan Ph.D.

29

10/29/2013

Part II: Design and Implementation

Case 2: Simulation Results

Case 2

In this experiment two intelligent loops viz, FMRG augments the MRAC and

RBFNN based neuro
-
controller is being used for Multiple Input Multiple Output

control of the system. The system is running under the following specifications:

Power =0.28 pu.

Power Factor= 0.24 lag.

Terminal Voltage=1.062 pu.




Conclusions:


Different operating points behaved differently. In the first case, RBFNN did not
provide much control contribution. With a change in operating point, the
contribution was noticeable. This confirms the need for such control.

In all these case system supervision concept performed better than individual
control.



Sukumar Kamalasadan Ph.D.

30

10/29/2013

Part II: Design and Implementation

Case 2: Simulation Results

Sukumar Kamalasadan Ph.D.

31

10/29/2013

Part II: Design and Implementation

Case 2: Simulation Results

Sukumar Kamalasadan Ph.D.

32

10/29/2013

Part II: Design and Implementation

Case 2: Simulation Results

Sukumar Kamalasadan Ph.D.

33

10/29/2013

Part II: Control of Two Machine Infinite Bus
System

Sukumar Kamalasadan Ph.D.

34

10/29/2013

Part II: Control of Two Machine Infinite Bus
System

Case 3

Both machines P=0.8 and Q=0.4 p.u.

100ms short circuit in bus 2
-
3

Machine Parameters

Machine Operating points

Sukumar Kamalasadan Ph.D.

35

10/29/2013

Part II: Control of Two Machine Infinite Bus
System

100ms short circuit in bus 2
-
3

36

Part II: Control of Two area Power System

3/8/2009


University of West Florida,
Copyright © 2009

Power System

Plant

Critic Network

Action Network

Supervisory Learning

(Earlier Designs)

Scheduler Block


(a=ka
E
+(1
-
k)a
S
)

V
ref

+

+

X(t)

A(t)

Exploration

Shaping (Prediction)

J(t)

1.0

TDL

Transport lag

X(t)

X(t)

X(t
-
Δ
t
)

F
-
1
(X,X
d
)

TDL

37

3/8/2009


University of West Florida, Copyright © 2009

Part II: Control of Two area Power System

Sukumar Kamalasadan Ph.D.

38

10/29/2013

Part III: Other Research Projects and
Directions (Five year plan)


Smart Grid Applications


Real
-
time test bed for power system modeling and control.


Various projects.



Wide Area Monitoring and Control based on scalable
intelligent supervisory loop concept.


Theory, development and simulation studies.



Distributed Power Generation and Grid Interface.


Integrating Fuel
-
Cell and Micro
-
Turbine Models.


Control system development and assessment.

39

Part III: Smart Grid Applications

3/8/2009


University of West Florida, Copyright © 2009

40

3/8/2009


University of West Florida, Copyright © 2009

Smart Controllers

Part III: Smart Grid Applications

41

3/8/2009


University of West Florida, Copyright © 2009

Part III: Smart Grid Applications

42

3/8/2009


University of West Florida, Copyright © 2009

Part III: Smart Grid Applications

43

Part III: Wide Area Monitoring and Control

3/8/2009


University of West Florida, Copyright © 2009

Sukumar Kamalasadan Ph.D.

44

10/29/2013

Part III: Wide Area Monitoring and Control

Wide Area Controller (WAC)

Bus 1

Bus 9

Bus 2

Bus 7

Bus 10

Bus 6

Bus 12

Bus 8

Bus 3

Bus 11

Bus 4

Bus 5

Infinite Bus

Gen 2

Gen 4

Gen 3

STATCOM

P
45

P
25

P
78

P
16

P
46

V
ref

V
ref

V
ref

ω
2

ω
3

ω
4

Sukumar Kamalasadan Ph.D.

45

10/29/2013

Part III: Wide Area Monitoring and Control:
Scalability: Supervisory Loop Approach

Energy

Management

Center


ANN Agent


Intelligent

Control


Intelligent

Control

Intelligent

Control


Intelligent

Control

Intelligent

Control

Intelligent

Control

Intelligent

Control

Intelligent

Control



PMU



Agent



PMU



PMU



PMU



PMU


PMU

PMU


PMU


Voltage Stability

Assessment Tool

Area 1

Area 2

Area 3

Area 4

Wide Area Controller

Sukumar Kamalasadan Ph.D.

46

10/29/2013

Part III: Distributed Power Generation and
Grid Interface: Concept


Objectives



Intelligent control of distributed Generation


Control (measurement) strategies of voltage and speed of the DG
system based on intelligent controllers (agents)


Integration of renewable energy based power generation to the grid


Development of test bed and hardware in the loop experiments based
on simulations


Practical Implementation and Integration of the proven research
activities to power distribution grid and testing


General Conceptual Implementation of DG Grid Interface


Control station:


Supervisory controller for DG system including protection


Coordination with nearest substation


Database for power flow, generation and load dynamics


Intelligent agents interaction

47

Part III: Distributed Power Generation
and Grid Interface

3/8/2009


University of West Florida, Copyright © 2009

Micro
-
grid and Interface

48

3/8/2009


University of West Florida, Copyright © 2009

Fuel Cell Model

Part III: Distributed Power Generation
and Grid Interface

49

3/8/2009


University of West Florida, Copyright © 2009

PV System Model

Part III: Distributed Power Generation
and Grid Interface

50

3/8/2009


University of West Florida, Copyright © 2009

Islanding Mode

Connected to the Grid

Part III: Distributed Power Generation
and Grid Interface

Sukumar Kamalasadan Ph.D.

51

10/29/2013

Current Research Support and Future
Considerations


Current Support


National Science Foundation CAREER Grant


(2008
-
2012)


Internal Grant from the University of West Florida
(UWF)


(2008
-
2009)


Under Consideration


Office of Naval Research (ONR)


NSF Power Control and Adaptive Network (PCAN)


NSF Course, Curriculum and Lab Improvement (CCLI)

Sukumar Kamalasadan Ph.D.

52

10/29/2013

Research Collaborations


Areas



Mathematical Modeling of
physical systems such as power
systems, energy systems,
avionics and robotics.


Developing computer algorithms
in the form of control,
optimization, identification of
systems through mathematical
models


Developing computational
intelligence based (neural
network, fuzzy systems,
biologically inspired
computational intelligent
techniques) algorithms that can
augment traditional controllers.


Applying control, optimization
and identification algorithms for
dynamic systems models.


Real
-
time implementations



People


Graduate students who
are interested in these
area


Research faculty who
are interested in
collaborations.

Sukumar Kamalasadan Ph.D.

53

10/29/2013

Summary


Intelligent Adaptive Controllers based on the supervisory
loop concept can be expanded to agent based control and
monitoring.


This approach is found to be scalable and useful for power
system control, identification and optimization.


Intelligent tool in the form of agents can be developed and
feasible for dynamic voltage stability assessment and
improvements.


These approaches are expandable to modular
technologies, DG control and grid interface, distribution
system and in reconfigurable and survivable modes.


For modern power system, these techniques would have
significant impact especially in the areas of power system
control, stability, reliability and security.