Context-Dependent Network Agents

EPRI/ARO CINS Initiative CDNA Consortium CMU, RPI, TAMU, Wisconsin, UIUC

The CDNA Consortium

Carnegie Mellon University Texas A&M University

Prof. Pradeep Khosla Prof. Bruce Krogh Dr. Eswaran Subrahmanian Prof. Sarosh Talukdar
Rensselaer Polytechnic Institute

Prof. Garng Huang Prof. Mladen Kezunovic

University of Illinois at UrbanaChampaign

Prof. Lui Sha

University of Minnesota

Prof. Joe Chow Prof. Bruce Wollenberg

CDNA Objective
s

Improve
x

x
s

agility and robustness (survivability) of large-scale dynamic networks that face new and unanticipated operating conditions. U.S. Power Grid Local networks

Target Networks:
x x

CDNA Approach
s

Improve
x

decision-making competence of components distributed throughout the network, particularly existing and future control devices, such as relays, voltage regulators and FACTS.

Why CDNA?
s

centralized real-time control is

x

infeasible in many situations because of the distribution of information and growing number of independent decision makers on the grid intractable - robust control algorithms simply dont scale, the problems are NP hard undesirable - we contend that centralized solutions are less robust against major network upsets and less adaptive to new situations

Why CDNA? (contd.)

s

control devices are already pre-programmed for anticipated situations BUT one-size fits all strategies are conservative in most cases, and wrong in some (the most critical!) situations necessary communication and computation technology for CDNA exists today

Key Research Issues

s

modeling
x x x x

operating modes contingencies impact of restructured power systems device capabilities/influence

Key Research Issues - 2

s

state estimation
x x x

using local information network state estimation real-time constraints adaptive mode switching coverage

hybrid control
x x

Key Research Issues - 3

s

learning
x x

distributed learning state-space decomposition collaboration strategies moving off-line techniques for asynchronous algorithms online

coordination
x x

Decentralized Large Area Power System Control

Bruce Wollenberg University of Minnesota

Objectives
s

Research goal is to show how all standard functions built on a power flow calculation can be accomplished without a large area (centralized) model and computer system Each region of the power system retains its own control system, models it own power network and communicates with immediate neighbors Functions that now require central computing
x x x

Security Analysis Optimal Power Flow Available Transfer Capability

LARGE AREA CON TRO L SYSTEM

R E G IO N A CON TR OL SYSTEM

Typical Power Pool or ISO

Trends:
C

R E G IO N C CON TROL SYSTEM B R E G IO N B CONTROL SYSTEM

- Getting larger - Standard data formats - Less functionality in regional systems

D R E G IO N D CONTR OL SYSTEM

Examples: - California ISO - Midwest ISO

R E G IO N E CON TR OL SYSTEM

R E G IO N A CONTROL SYSTEM

Networked Control Systems

R E G IO N C CONTROL SYSTEM B R E G IO N B CONTROL SYSTEM D C

- Region can be any size - Can extend to any number of regions

R E G IO N D CONTRO L SYSTEM

- Aggregate has same functionality as large area control system - Can new functionality be added that would not be available in a central system?

R E G IO N E CONTROL SYSTEM

Collaborative Nets
Eduardo Camponogara and Sarosh Talukdar Institute for Complex Engineered Systems Carnegie Mellon University

Controlling Large Networks

Operating goals fall into categories:
s s s s

Costs & profits Safety Regulations Equipment Limits

Limitations:
s

s s

No organization can cope with all operating goals Need of diverse skills Multitudes of agents

Control Solution:
s

Delegate goals to separate organizations

Organization: Agent:

A network of agents and communication links. Any entity that makes and implements decisions such as relays, control devices, and humans.

Multiple Organizations in the Power Grid

Protection Systems Agents Relays Generator Control Governors, exciters optimization soft. Reduce cost s.t. constraints Seconds Security Systems Simulation & learn. tools, humans Prevent cascading failures Hours, days

Goals

Keep equipment under limits 0.01 to 0.1secs

Reaction Time

Low Large Fast

Agent Skills Number of Agents Agent Speed

High Small Slow

Organizations Do Not Collaborate

Protection Systems Generator Control Security Systems

Current Scenario:

s s s

Agents in separate organizations do not talk Agents might work at cross-purpose Organizations might interfere with one another

How do we make individual agents more effective? How do we prevent interference between organizations?

Improving Overall Performance of Nets

The suggested answer is based on: 1.) The use of a common framework to specify agent tasks. 2.) The implementation of a sparse, collaborative net that can cut across the hierarchic organizations. 3.) The design of collaboration protocols to promote effective exchange of information. Protection Systems Generator Control Security Systems

C-Net

C-Net

What Is A Collaborative Net?

A flat organization of dissimilar agents that can integrate hierarchic organization. Properties:
s

Agents are autonomous within the C-Net. They have initiative, make and implement decisions. Agents collaborate with their neighbors. The collaboration protocol determines: x what information is exchanged, x in which way, and x how agents make use of it.

Advantages:
s s s

Disadvantages:
s

Quick Fault Tolerant Open

No structural coordination. if necessary, it can emerge from the collaboration protocol. Unfamiliar.

The Rolling Horizon Formulation

The dynamic control problem A framework to solve dynamic control problems as a series of static optimization problems. Minimize f(x,dx/dt,u,t) Subject to h(x,dx/dt,u,t)=0 g(x,dx/dt,u,t)<=0 The static opt. problem (P) Minimize f(X,U) Subject to H(X,U) = 0 G(X,U) <= 0 The prediction model is embedded in the H(X,U). G(X,U) approximates the operating constraints in g(x,dx/dt,u,t).

The steps of the rolling horizon formulation: 1.) Choose a horizon [t0,..,tN], I.E. a set of time
points where t0 is the current time. 2.) Let x(tn) be the state predicted at time tn. x(t0) is the current state. 3.) Let u(tn) be the planned actions at time tn. 4.) Let X=[x(t0),,x(tN)] and U=[u(t0),,u(tN)] 5.) Choose a model to predict x(tn+1) from x(tn) and u(tn). Possibly, a discrete approximation of the dynamic equations (e.g., Eulers step).

The Rolling Horizon Algorithm

s s

A model is used to predict the future state of the physical network over a set of discrete points in time (horizon). An optimization procedure computes the control actions, over the horizon, that minimize error. Design Issues:
s

Steps of the Algorithm: 1.) The current time it t0. 2.) Sense the current state x(t0) 3.) Instantiate the static optimization problem (P). 4.) Solve (P) to obtain the control actions U=[u(t0),,u(tN)]. 5.) Implement the control action u(t0). 6.) Pause and let the physical network progress in time. The horizon rolls forward. 7.) Repeat from step 1.

The horizon has to be long enough to avoid present actions with poor long-term effects. Accuracy of the prediction model.

The Rolling Horizon

Plan ahead

Control

implemented control model predicted control

t0 now

t1

t2

t3

t4

Time

The Rolling Horizon

Update plans frequently

Control

plans at t0

plans at t1

t0

t1 now

t2

t3

t4

Time

A Framework for Specifying Agent Tasks

Break up the static optimization problem, (P), into a set of M small, localized subproblems, {(Pm)}. Assemble M agents into a C-Net, so that each agent matches one subproblem. Agent m and its subproblem (Pm) Proximate variables (xm,um): It senses the values of a subset xm of x. It sets the values of a subset um of u. Neighborhood variables (ym) Variables sensed or set by neighbors. Remote variables (zm): All the other variables.
Ag1 C-Net (P) (P1) (P2) (P3) (P4)

It has partial perception of, and limited authority over, the physical network.

Matching Agents to Subproblems

The rolling horizon formulation of (Pm) Minimize fm(Xm,Um,Ym,Zm) Subject to Hm(Xm,Um,Ym,Zm) = 0 Gm(Xm,Um,Ym,Zm) <= 0

Exact: If its not sensitive to remote vars. The matching between agent-m and its subproblem (Pm) fm = fm(Xm,Um,Ym) Hm = Hm(Xm,Um,Ym) Gm = Gm(Xm,Um,Ym) Near: If it is weakly sensitive to remote vars.

Collaboration Protocols
A protocol prescribes: a) the data exchanged by agents, b) in which way, and c) how agents use the data to solve their problems. Voting In setting the values of its controls, each agent takes the votes of its neighbors into account. Two protocols Proximate Exchange Each agent broadcasts its plans to nearby agents which, in turn, take these plans into account. Semi-synchronous, semi-parallel (mutual help). Synchronization between neighbors. Parallel work if agents are non-neighbors. Asynchronous, parallel.

s no sr e V i

Equivalence and Convergence

Equivalence: Two Questions: When are the solutions to the network of subproblems, {(Pm)}, solutions to (P)? Convergence: When does the effort of the collaborative agents converge to a solution of {(Pm)}? Sufficient conditions for equivalence and convergence: 1.) Coverage: 2.) Density: 3.) Convexity: 4.) Feasibility: 5.) Int-Pt-Mtd: The C-Net must provide complete coverage of the network. The matching of agents to subproblems must be exact. (P) must be convex. (P) must be strictly feasible. The agents must use an interior-point-method.

6.) Serial Work: The agents run the semi-synchronous, semi-parallel protocol.

Relaxing Sufficient Conditions in Practice

We believe that the following conditions can be relaxed in practice: 1.) Density: 2.) Convexity: 3.) Serial work: Near matching of agents to problems are likely to be adequate. It is impractical in real-world networks. Serial work within a neighborhood is too slow.

A prototypical network: A forest of pendulums. - One agent at each pend. - Agents control two forces: Horizontal & Orthogonal. - Agents collaborate with nearest neighbors.

The Dynamic Control Problem

Problem: Drive pendulums to the pre-disturbance mode, that is, minimize cumulative error (from desired trajectory) and total control-input cost. Minimize Subject to C1 Three Control Solutions:
t = t =

f ( x, u ) =

h ( x, x , u ) = 0

t =0

x ~ dt + b x
2

t =0

u 2 dt

A centralized, nonlinear optimization package that solve the stat. opt. prob. (P). A centralized, feedback linearization controller. A collaborative net, with one agent at each pendulum, that solves {(Pm)}.

C2

C-Net

C-Net and C1: Experimental Set-up

Goal: Set-up: Evaluate the loss in quality of the Collaborative Net solution. C-Nets and C1s restore synchronous mode of pendulums. At each sample time t, 1.) solve the network of subproblems, {(Pm)}, with the C-Net, 2.) record the obj-function evaluation of the C-Net, F(C-Net), 3.) solve the static optimization problem, (P), with C1, and 4.) record the obj-function evaluation of C1, F(C1). Output Data: Scenarios: A list of obj-function-evaluation pairs [F(C-Net),F(C1)]. Place pendulums in a line to form forests of 2 to 9 pendulums. 2-Pendulum Forest Add 1 Pend. 3-Pendulum Forest

C-Net and C1: Results

C-Net Excess: The difference in quality between the C-Net and C1 solutions. F(C-Net) is the obj-function evaluation attained by the C-Net. F(C1) is the obj-function evaluation attained by controller C1. C-Net excess = [F(C-Net) F(C1)] / F(C1) C-Net Penalty: The mean value of the C-Net excess.

C-Net penalty is low

C-Net Penalty (%)

Number of Pendulums

C-Net and C2: Experimental Set-up

Goal: Set-up: Output Data: Scenario: Evaluate the performance of the C-Net and the feedback linearization controller, C2, a traditional control technique. C-Net and C2 restore synchronous mode of pendulums. The cumulative error and input-cost, f(x,u), for the C-Net & C2. A forest with 9 pendulums placed in grid.

C-Net and C2: Results

Objective: Minimize

f ( x, u ) =

t =

t =0

2 x ~ dt + b x

t =

t =0

u 2 dt

Control-Input Cost (b) 10e-4 10e-3 10e-2 10e-1

Objective Function Evaluation: f(x,u) C2 (feedback lin) 9.56 10.49 17.05 82.64 C-Net 11.89 12.32 16.00 32.07

The lower the f(x,u), the better the solution C-Net performance improves

C-Net and C2: Trajectory of Pendulums

Pendulums under control of C2 (feedback linearization) C2 immediately drives pendulums to the desired trajectory.

Pendulums under control of the C-Net The C-Net waits until it becomes cheaper to drive pendulums.

Conclusion
The experiments show that C-Nets are promising. Current research effort:
s

Development of collaboration protocols that allow agents to work asynchronously and in parallel, at their own speed. - Use of safety margins to guarantee feasibility, and foster effective work between slow and fast agents. A taxonomy of collaboration protocols.

What else have we done?

s

Employed C-Nets to recover synchronous operation of generators in power networks IEEE-14, -30, -57. Preliminary work on the decomposition of (P) into {(Pm)}: - Models and algorithms to specify neighborhood perception.

Hybrid Control Strategies

controllers u1

C1 C1 Cn

u1

PLANT

un

performance monitors

M1 M1 Mn Decision Module