Concept of a Wide Area Defense Concept of a Wide Area Defense System for the Power GridSystem for the Power Grid

Chen-Ching LiuUniversity College Dublin

National University of Ireland, Dublin

Seminar at NTUA and IEEE PES Greece, June 20091

Catastrophic Power Outages

2

Western Electricity Coordinating Council (WECC) system - Aug 10th, 1996 Blackout

PDCI Remedial Action Schemes (RAS) began to actuate. Shunt and series capacitors were inserted.

15:47:40-15:48:57 p.m. Generators at the McNary power house supplying 494 MVAR trip. The system begins to experience “mild oscillations”.

15:48:51 p.m. Oscillations on the POI reached 1000MW and 60-kV peak-to-peak.

Lines throughout the system begin to experience overloads as well as low voltage conditions. Additional lines trip due to sagging.

The WECC broke into 4 asynchronous islands with heavy loss of load.

7.5 million people lost power.

Mild .224 Hz oscillations were seen throughout the system and began to appear on of the PDCI.

15:48 p.m. Keeler-Allston 500-kV line contacts a tree due to inadequate right-of-way maintenance. Additionally the Pearl-Keeler line is forced out of service due to the Keeler 500/230-Kv transformer being OOS.

Initiating events System becomes unstable Blackout

Shunt capacitor banks were switched in to raise the voltage but the oscillations were not being damped.

AZ

CA CO

ID

MT

NENV

NM

ND

OR

SD

UT

WA

WY

With the loss of these 2 lines, 5 lines are now out of service, removing hundreds of MVAR.

3

Eastern Interconnection –August 14th, 2003 Blackout

2:14 p.m. FE’s control room lost alarm functions followed by a number of the EMS consoles.

1:31 p.m. Eastlake 5 generation unit trips and shuts down.

2:02 p.m. Stuart-Atlanta 345-kV line tips off due to contact with a tree.

1:07 p.m. FE turns off their state estimator for troubleshooting.

Initiating events System becomes unstable Blackout

AL

AR

CT

DE

FL

GA

IL IN

IA

KS

KY

LA

ME

MD

MA

MI

MN

MS

MO

NE

NH

NJ

NY

NC

ND

OH

OK

PA

RI

SC

SD

TN

VT

VAWV

WI

265 power plants tripped off line and 50 million people are without power.

Low voltage/ high load conditions and system disturbances propagate through the system tripping transmission lines and generators.

4:08:59 p.m. Galion-Ohio and Central-Muskinghun 345-kV lines trip on Zone 3 causing major power swings through New York and Ontario and into Michigan.

4:05:57 p.m. The loss of 138-kV lines overloads the Sammis-Star line.

2:54 p.m. The primary and secondary alarms servers failed.

3:05:41-3:57:35 p.m. 3 345-kV lines trip due to contact with trees. This overloads the underlying 138-kV system and depressed voltages.

3:39:17-4:08:59 p.m. 16 138-kV lines trip due to overloading.

4:13 p.m. most of the North East and parts of Canada blacked out. There are only a few islands which remain operating.

4

Hydro-Québec Blackout-April 18th, 1988

8:08 p.m. In response to the loss of the Church Fall generation station, a signal to initiate load shedding is sent to the central control center

8:08 p.m. 3200 MW of generation from the Church Falls generation complex is isolated due to isolation of transmission lines

8:08 p.m. Wet snow and freezing rain cause flash over on all three phases at the Arnaud substation

Initiating events System become unstable Blackout

18,500 MW of load was lost because 3,200 MW of automatic load shedding failed to occur

With the loss of the James bay transmission lines the Le Grand network was separated from the rest of the system. Shortly afterward three transformers at Le Grand 4 failed

1.7 seconds after the failure of the load shedding system the first of the James bay transmission lines tripped, followed by the other two

The load shedding signal was not received at the central control center due to a faulty contact in the communications system

15 seconds later the Manicouagan-Montréal transmission system collapsed. This lead to the loss of all DC interconnections as well as 8 generators at the Beauharnois generating station which had been isolated to serve the Ney York Power Authority (NPA)

5

Blackout Propagation(without defense systems)

Complete System Collapse

Triggering Event

6

Occurrences and Extent of Blackouts in North America

Number of customers affected100 102 104 106 108

Num

ber

of B

lack

outs

10

100

1000

7

Strategic Power Infrastructure Defense (SPID)

Design self-healing strategies and adaptive reconfiguration schemes

To achieve autonomous, adaptive, and preventive remedial control actions

To provide adaptive/intelligent protection

To minimize the impact of power system vulnerability

8

SPID System

FailureAnalysis

Self-HealingStrategies

VulnerabilityAssessment

Informationand

SensingReal-Time

Securit

y Robustne

ss

Dependability

Power Infrastructure

•Satellite, Internet•Communication system monitoring and control

Hidden failure monitoring

Adaptive: load shedding, generation rejection, islanding, protection

Fast and on-line power & comm.

system assessment

9

Multi-Agent System for SPID

REACTIVE LAYER

COORDINATION LAYER

DELIBERATIVE LAYER

Knowledge/Decision exchange

Protection Agents

GenerationAgents

Fault Isolation Agents

FrequencyStabilityAgents

ModelUpdate Agents

CommandInterpretation

Agents

Planning Agent

Restoration Agents

HiddenFailure

Monitoring Agents Reconfiguration Agents

VulnerabilityAssessment

Agents

Power System

Controls

Inhibition Signal

Controls

Plans/Decisions

EventIdentification

Agents

Triggering Events

Event/AlarmFiltering Agents

Events/Alarms

Inputs

Update Model CheckConsistency

Comm.Agent

10

Multi-Agent System for SPID

• Subsumption Architecture (Brooks) for Coordination

• Agents in the higher layer can block the control actions of agents in lower layers

Load SheddingAgent

Global View/Goal(s)

R

Under Frequency RelayLocal View/Goal(s) Tripping Signal

Inhibition Signal

11

Cascaded Events

Some Basic Patterns• Line tripping due to overloading• Generator tripping due to over-excitation• Line tripping due to loss of synchronism• Generator tripping due to abnormal voltage and frequency system condition• Under-frequency/voltage load shedding

Identifying the basic patterns of cascaded events and explore how these patterns can be

combined into sequences

12

Cascaded Zone 3 Operations

13

Zone 3 Relay Operations Contributed to Causes of Blackouts.

Heavy Loaded Line

Low Voltage High Current

Lower Impedance Seen by Relay

Loss of Transmission Lines

Other Heavy Loaded Lines

Zone 3 Relay

Operation(s)

Catastrophic Outage

Prediction of Zone 3 Relay Tripping Based on On-Line Steady State Security Assessment

14

Case Relay Status Contingency Description

1 N/A Secure 3 phase fault at bus 1

2 Zone3 Insecure 3 phase fault at bus 2

.. …

N Secure 3 phase fault at bus N

Case Relay Status Contingency Description

1 N/A Secure 3 phase fault at bus 1

2 Zone3 Insecure 3 phase fault at bus 2

.. …

N Secure 3 phase fault at bus N

Case Relay Status Contingency Description

1 N/A Secure 3 phase fault at bus 1

2 Zone3 Insecure 3 phase fault at bus 2

.. …

N Secure 3 phase fault at bus N

Case Relay Status Contingency Description

1 N/A Secure 3 phase fault at bus 1

2 Zone 3 Insecure 3 phase fault at bus 2

.. …

N N/A Secure 3 phase fault at bus N

…

Contingency Evaluation Performed On Line

Every Several Minutes

Contingency Evaluation

Post-Contingency Power Flow

Post-Contingency Apparent Impedance

Corrected Post-ContingencyApparent Impedance

FISFuzzy Inference System (FIS) Developed Using Off-Line Time-Domain Simulations

Impedances Obtained by Power Flow and Time Domain Simulation

15

Post-Contingency Impedance Obtained by Power Flow Does Not Coincide with Impedance Obtained by Time-Domain Dynamic Simulations

Impedance Obtained by Steady State Power Flow

Zone 3 Circle

Correction Term

Impedance Locus

Zon

e 3

Rea

ch

R

X

0

Load Angle

Automatic Development of Fuzzy Rule Base

16

Wang & Mendel’s algorithm is a “learning” algorithm:1) One can combine measured information and human linguistic information into a common framework2) Simple and straightforward one-pass build up procedure3) There is flexibility in choosing the membership function

Pre-determine number of membership functions N

Give input and outputdata sets

In this example, N is 7

(Inp1, Inp2, Out) = (10, 1, -2)(Inp1, Inp2, Out) = (8, 3, -1)(Inp1, Inp2, Out) = (5, 6, -4)(Inp1, Inp2, Out) = (2, 8, -5)…

FIScreated

automatically

17

Case A

86

157

181

182

113

32

31

79

74

66 65

184

75 73

7669

68

70

7782

9591

92

93

94

96

97

98

83

168

169

170

171

114

172

173

111

120

121

122

123

124

125

115

100

101

105106

117

116

132

133

134

104

118

108

107

110

89

90

88

87180 81 99 84

156

161162

85

36 186187

183

185

188

189

ß 21

6

11

18

190

5

17

8 9

10

72914

192193194

139

13

21

20

2627

148147

195

191

163

158 159167 155 165

164

166

16045 44

64

63

37

178176

179177

142

46 4839

230-287kV

109

112

11-22kV

Substation

12

138

28

135

126

127

128

129

130

131

175

174

67

72197

196

198

200

78

199

80

71

Relay Location

Fault Location

Impedance on R-X (Case A)

18

-0.01

0

0.01

0.02

0.03

0.04

0.05

0.06

0.07

0 0.02 0.04 0.06 0.08 0.1

R [PU]

X [

PU

]

Impedance LocusZone 3 CircleZ Obtained by Power Flow CalculationCorrected Z Obtained by FISPost-Contingency Z Obtained by Time-Domain Simulation

Pre-fault

Line Tripping by Zone 3 relay

Case A

Z obtained by power flow solution is outside Zone 3 circle.

Load Shedding Studies have shown that the August 10th 1996 blackout could have been

prevented if just 0.4% of the total system load had been dropped for 30 minutes.

According to the Final NERC Report on August 14, 2003, Blackout, at least 1,500 to 2,500 MW of load in Cleveland-Akron area had to be shed, prior to the loss of the 345-kV Sammis-Star line, to prevent the blackout.

19

Automatic Load Shedding

Under Voltage

Under Frequency

Rate of Frequency Decrease

Remedial Action Scheme

20

Adaptive Self-healing:Load Shedding Agent

• A control action might fail

• Unsupervised adaptive learning method should be deployed

• Reinforcement Learning– Autonomous learning method based on interactions with the

agent’s environment

– If an action is followed by a satisfactory state, the tendency to produce the action is strengthened

21

Load Shedding Options

0 50 100 150 200 25058.6

58.8

59

59.2

59.4

59.6

59.8

60

frequency with 20 % load shedding

10% load shedding freq

uenc

y

Time (multiples of 0.02 sec)

22

Adaptive Self-healing:Load Shedding Agent

• 179 bus system resembling WSCC system• ETMSP simulation• Remote load shedding scheme based on

frequency decline + frequency decline rate• Temporal Difference (TD) method is used for

adaptation: Need to find the learning factor for convergence

23

Adaptive Load Shedding Agent

State 1 State 2 State 3

Freq := 59.5

Dec.rate > threshold value

Freq := 58.8

Dec.rate > threshold value

Freq := 58.6

Dec.rate > threshold value

179 bus system24

Adaptive Load Shedding Agent

0 20 40 60 80 100 120 140 160 1800

0.5

1

1.5

2

2.5

a=0.55

a=0.75

Number of trials

Nor

mal

ized

fre

quen

cy

Expected normalized system frequency that makes the system stable

“The load shedding agent is able to find the proper control action in an adaptive manner based on responses from the power system”

“The load shedding agent is able to find the proper control action in an adaptive manner based on responses from the power system”

25

Flexible Grid Configuration to Enhance Robustness

Flexible Grid Configuration can play a significant role in defending against catastrophic events.

Power infrastructure must be more intelligent and flexible.

To allow coordinated operation and control measures to absorb the shock and minimize the potential damages caused by radical events .

26

Cascading Events

• A Cascading Event Refers to a Series of Tripping Initiated by One or Several Component Failures in a Power System– Here the initial component(s) failure is

designated as “shock” to the power infrastructure

27

Simulated Cascading Events (179 Bus System)

• Compute Power Flows after Tripping

– Six lines are found on limit violation – Trip these lines

• Identify New Network Configuration and Solve Power Flows Again

– Fifteen lines are found with limit violations – Trip these lines

• Continue This Simulation Procedure

– Finally system collapses: most transmission lines are tripped and most loads are lost

28

2-Area Partitioning Algorithm (from VLSI)

• Spectral 2-way Ratio-Cut Partitioning – Theorem Given an edge-weighted graph G = (V,

E), the second smallest eigenvalue λ2 of the graph’s Laplacian matrix Q yields a lower bound on the cost c of the optimal ratio cut partition, with c = e(U,W)/(|U|·|W|) ≥ (λ2/n)

– Cut-Size: e(U,W) ≥ (λ2/n) (|U|·|W|)

29

Area-Partitioning Algorithm

5

2 1

6 4

3 1

0.1 0.1 0.1

2

1 1

6-Bus System

30

Area-Partitioning Algorithm

Partition {U, W} Cut Set Size e(U,W)

e(U,W) / (|U|·|W|)

λ2 / n

{(4), (5 6 3 2 1)} 1.1 1.1 / (1 5) = 0.22

 

{(4 5), (6 3 2 1)} 1.2 1.2 / (2 4) = 0.15

 

{(4 5 6), (3 2 1)} 0.3 0.3 / (3 3) = 0.0333

0.1966 / 6 = 0.0328

{(4 5 6 3), (2 1)} 2.1 2.1 / (4 2) = 0.2625

 

{(4 5 6 3 2), (1)} 1.1 1.1 / (5 1) = 0.22

 

Partition Results of 6-Bus System

31

Flexible Grid Configuration to Absorb the Shock

• Solve Power Flows of Area One – All 35684.71 MW loads are supplied, no line

flow constraints violations

• Solve Power Flows of Area Two – Seven lines on limit violation:

• (Bus158-Bus164), (163-8), (64-163), double lines (16-19), and double lines (150-154)

32

Flexible Grid Configuration to Absorb the Shock

• Use “Power Redispatching & Load Shedding” in Area Two– Totally, 188 + 64.4 + 60 = 312.4 MW load are shed

Bus #

Original Load (MW)

Load Shed (MW)

Load Supplied (MW)

8 239 188 51

16 793.4 64.4 729

154 1066 60 1006

33

Split System into Two Areas

33 32

31 3 0

35

80

78

74

79 66

75

77

76 72

82

81

86 83

84

85

114

115

118 119

103

107

108

110 102

104

109

230 kV 345 kV 500 kV

34

65

71

69

70

87 88

99 36

73

89 90

124 125

169 171

170 172 173

111

120

121 122

123

91 - 94 95 - 98

13 2 133

135 134 104

174, 176, 178

113 100

101

105 106

117 116

68

67

112

180

156 157 161 162 168

167 165

158

159 155 44

45 160

166

163

5 11

6

8

9

18 17

4

3

7

14

12 13

138 139

147

15

19

16

142

37 64 63

153 145 151 152

136 49 48

146 154

149

143

43

175, 177, 179

48 38

57 58

54 51

52 53 42

55 41 62 56

40

39

150

137

61

148 22

23 25

24

20 21

29 28

2

10

164

50

42

47

46

59 60(3)

27 26

141 140

144

48

48

48

34

Flexible Grid Configuration to Absorb the Shock

• Shed Load vs. System Total Load– K=1

– K=2

– K=3

%51.0%100MW 60785.41

MW 312.4

%982.0%100MW 60785.41

MW 597

%755.1%100MW 60785.41

MW 1067

35

Flexible Grid Configuration to Avoid Cascaded Failures

• Step 1 : Compute power flows after initial tripping event(s).

• Step 2 : Convert power network to an edge-weighted graph G, weight of each edge is absolute value of real power flow.

• Step 3: Multilevel graph partitioning with minimum edge-cut.

Network is separated into k areas to minimize generation / load imbalance.

Initial Graph G0

G1

G2

Gk

G2

G1

G0

Projected Partition

Refined Partition

Coarsening Phase

Initial Partitioning Phase

Uncoarsening and

Refinement Phase

•Graph COARSENED to a smaller number of vertices, • Bisection PARTITIONING of much smaller graph,• UNCOARSENING back toward original graph.• Coarsest graph small, Coarsening can be parallelized, Partitioning efficiency high.

36

Emergency Control with Multilevel Graph Partitioning

CommunicationNetwork

Power Transmission Network

Central Control Unit

CB1

CB2

SCU1

SCU2SCU n

Relay1

Relay2

SCU – Substation Control UnitCB – Circuit Breaker

• CCUs acquires system data, generates system separation strategy.• SCUs receive system separation commands from CCU and send breaker opening commands to specific auxiliary relays.

An adaptive relaying architecture for controlled islanding

• Partitioning a 22,000 bus and 32,749 branch system into 2, 3, 4 islands with 0.07s, 0.081s and 0.09s on 2 GHz Pentium CPU and 1GB RAM.

• Fast computational speed makes it possible to determine partitioning strategy and identify new network configuration in on-line environment.

37

Flexible Grid Configuration to Avoid Cascaded FailuresControlled islanding on a 200-bus system:199 buses, 31 generators, 248 branches.

Sequence of cascading events:•At t=0 s, three transmission lines out-of-service.•At t=60 s, line 72-197 tripped due to line fault.•At t=120 s, line 78-196 tripped due to line fault. Generator G70 at bus 70 overloaded.•At t=240 s, generator G70 tripped by over-excitation protection. At t= 260 second, system collapses.

157

113

35

34

33

32

31

79

30

66 65

184

75

70

7782

9591

92

93

94

96

97

98

168

169

170

171

114

172

173

111

120

121

122

123

124

125

115

119

100

101

105106

117

116

132

133

134

104

118

108

102

107

110

89

90

88

87180 81 99 84

156

161162

85

36 186187

183

188

189

ß 21

6

11

18

5

17

8 9

103

2 4

72914

24

16 15

25

19

23

22

193194

139

13

21

20

2627

148147

195

163

158 159167 155 165

164

166

16045 44

64

63

37

38

178176

179177

145

154

146

144

151

142

152143

137

150

149

40

46 4839

59

60

55

4258

43

50

57 54 51 53

52

41

56

140

500kV345-360kV230-287kV

109

112

11-22kV

103

12

138

28

49

47

135

126

127

128

129

130

131

141

175

174

62

67

72

198

200

199

80

71

74

196

19778

69 76

182

181

73

185

6886

83 190

191

192

136

153

61

North Island

South Island

Cut Set Load-Generation (MW)

Bus 174-175, 176-177, 178-179, 155-156, 155-

167, 188-189(1,2)

North: Gen=37862 , Load=37104South: Gen=24517, Load=23794

38

Flexible Grid Configuration to Avoid Cascaded Failures

0

0. 2

0. 4

0. 6

0. 8

1

1. 2

0 30 60 90 120 150 180 210 240 270

Time (sec)

Loa

d B

us V

olta

ge (

pu)

Load bus voltages without/with islanding strategy

0.6

0.7

0.8

0.9

1

1.1

1.2

0 40 80 120 160 200 240 280 320 360 400

Loa

d B

us V

olta

ge (

pu)

Time (sec)

Load bus at North IslandLoad bus at South Island

• System islanding initiated at 241s.• Islanding strategy results in balanced generation / load in both islands.• All loads in the system are served. Load shedding scheme not applied. • Islanding strategy successfully prevents the collapse of the system.

39

Conclusion

• Cascading failures remain a grand challenge

• New communication, information and computer technologies enable wide area protection and control

• Connectivity also brings cyber vulnerability

• “Smart” grid?

• “Self-healing” grid?

40

Further Information

• J. Li, C. C. Liu, “Power System Reconfiguration Based on Multilevel Graph Partitioning,” IEEE PES Power Tech, Bucharest, Romania, 2009.

• K. Yamashita, J. Li, C. C. Liu, P. Zhang, and M. Hafmann, “Learning to Recognize Vulnerable Patterns Due to Undesirable Zone-3 Relay Operations,” IEEJ Trans. Electrical and Electronic Engineering, May 2009, pp. 322-333.

• J. Li, K. Yamashita, C. C. Liu, P. Zhang, M. Hoffmann, “Identification of Cascaded Generator Over-Excitation Tripping Events,” PSCC, Glasgow, U.K., 2008.

• H. Li, G. Rosenwald, J. Jung, and C. C. Liu, “Strategic Power Infrastructure Defense,” Proceedings of the IEEE, May 2005, pp. 918-933.

• J. Jung, C. C. Liu, S. Tanimoto, and V. Vittal, “Adaptation in Load Shedding under Vulnerable Operating Conditions,” IEEE Trans. Power Systems, Nov. 2002, pp. 1199-1205.

• H. You, V. Vittal, J. Jung, C. C. Liu, M. Amin, and R. Adapa, “An Intelligent Adaptive Load Shedding Scheme,” Proc. 2002 PSCC, Seville, Spain, June 2002.

• C. C. Liu, J. Jung, G. Heydt, V. Vittal, and A. Phadke, “Strategic Power Infrastructure Defense (SPID) System: A Conceptual Design,” IEEE Control Systems Magazine, Aug. 2000, pp. 40-52.

41