NASPI Working Group Meeting
March 23-24, 2015
Capturing Real-Time Power System
Dynamics: Opportunities and
Challenges
Zhenyu (Henry) Huang, PNNL, zhenyu.huang@pnnl.gov
Ning Zhou, Binghamton University, ningzhou@binghamton.edu
Ruisheng Diao, PNNL, ruisheng.diao@pnnl.gov
Shaobu Wang, PNNL, shaobu.wang@pnnl.gov
Steve Elbert, PNNL, steve.elbert@pnnl.gov
Da Meng, PNNL, da.meng@pnnl.gov
Shuai Lu, PNNL, shuai.lu@pnnl.gov
Acknowledgement: the work is supported by the US Department of Energy Office of Advanced Scientific Computing
Research and Office of Electricity Delivery and Energy Reliability.
1
Frequency trending away from nominal
requires tools capturing real-time dynamics
Dynamic state estimation vs. traditional static state estimation
Dynamic security assessment
Dynamic model validation
(all identified as killer apps in NASPI phasor roadmap,
https://www.naspi.org/Badger/content/File/FileService.aspx?fileID=539)
Frequency
60 Hz
0.019
0.018
0.017
0.016
0.015
0.01
0.014
0.013
0.012
0.011
0.01
Emergency situation
0.015
2007
2,008
2,009
2,010
Year
2,011
Normal operation
2,012
0.005
2
Mean of Absolute Frequency Deviation, Hz
Root Mean Square Error of Frequency Signal, Hz
0.02
Dynamic paradigm for grid operation and
planning
National Driver: clean and efficient power grid as well as being
affordable, reliable, and secure  dynamic and fast
Technical Approach: combine model prediction and measurement
observations to determine where the grid is, where the grid is going,
and where the grid could be (what-ifs).
Dynamic States
Require good dynamic models
Better Reliability
Data
collection
cycle
Clean Energy
Integration
Dynamic security assessment
Look-ahead dynamic simulation
Dynamic state estimation
1/30 sec
2/30 sec
3/30 sec
1 min
Better Asset
Utilization
Time
3
Phasor measurement offers great
opportunities for capturing dynamics
Time-synchronized, high-speed measurement at 30 samples per
second, able to capture the majority of grid dynamics
NASPI: www.naspi.org
4
General formulation of dynamic state
estimation
Two-step process
Prediction through model simulation
Correction using real-time phasor measurement
Mathematical algorithms: Extended Kalman Filter, Unscented Kalman
Filter, Ensemble Kalman Filter, Particle Filter, …
Q
x(k-1)
“Prediction”
R
x’(k)
Dynamic Simulation
z(k)
“Correction”
x(k)
Measurement Eq’s
dx/dt = f(x,α) +ν
z = h(x,α) +ε
Prediction
Cycle
milliseconds
Correction Cycle
~1/30 second
5
Performance evaluation – estimation
accuracy (Ensemble Kalman Filter)
Excellent tracking with realistic evaluation conditions
States tracking (Basic EnKF)
x 10
1.5
Speed Devi (pu)
Relative Rotor Angle (rad)
3% measurement noise; 40 ms measurement cycle (phasor measurement)
5 ms interpolation cycle; modeling errors considered; unknown inputs;
unknown initial states
1
0.5
0
5
10
15
20
-3
15
10
5
0
0
5
Blue/green tracking red is the goal
Ed' (pu)
Eq' (pu)
100 MC Ests
0.65
0.95
0.9
0.85
0
True
Mean
10 Mean+/-3*Std
15
20
0.6
0.55
0.5
5
10
Time (sec)
15
20
0
5
10
Time (sec)
15
20
6
Performance evaluation – computational
speed (Ensemble Kalman Filter)
Current codes scale to ~1,000 cores
Current computational performance meets the real-time requirement
for regional systems
Challenge: real-time performance for interconnection-scale systems.
Computational time per estimation step
10000
2017?
1000
TeraFLOPS
100
10
June 2014
Oct 2012
June 2014
1
Oct 2012
0.1
0.01
0.001
0.0001
30.000
30 ms
3.000
0.300
Seconds
0.030
0.003
7
Performance evaluation – impact of nonGaussian noises (Ensemble Kalman Filter)
Non-Gaussian noise discovered in phasor measurements.
Such non-Gaussian noises could result in large estimation errors.
Challenge: new methods to accommodate non-Gaussian noises.
Histogram (pValue=0.00%)
-0.1
120
-0.12
100
-0.14
80
0.2
pValue=0.00%
0.1
-0.16
-0.18
Error of rotor angle in rad.
Frequency
kV
0
60
40
-0.1
-0.2
Error for Gaussian noise
Error for Non-Gaussian noise
-0.3
-0.2
20
-0.22
0
5
10
15
20
25
t/s
30
35
40
45
50
0
-0.22
-0.4
-0.5
-0.2
-0.18
-0.16
Measurements
-0.14
-0.12
-0.1
0
5
10
15
t/s
8
Filtering technology comparison for
dynamic state estimation
Extended
Unscented
Ensemble
Particle Filter
Kalman Filter Kalman Filter Kalman Filter
Accuracy
The 2nd best
with 0%
diverged
33% diverged
The best with 20% diverged
0% diverged
(PF 2000)
Efficacy of
interpolation
Number of samples
needed
Sensitivity to missing
data
High
High
Low
High
None
Small
Medium
Large
Low
Low
Low
Low
Sensitivity to outliers
Low
Low
Medium
High
Computation time
(non-parallel)
Shortest
Same order
as EKF
longer than
EKF
Same order
as EnKF
9
Extending the formulation for dynamic model
validation and calibration
Treating parameters as augmented states
dα/dt = g(x,α) = 0
Dynamic states and parameters are estimated simultaneously,
suitable for real-time applications
Q
x(k-1),α(k-1)
“Prediction”
Dynamic Simulation
R
x’(k) ,α’(k)
dx/dt = f(x,α) +ν
dα/dt = g(x,α)
z(k)
“Correction”
Measurement Eq’s
x(k),α(k)
z = h(x,α) +ε
Prediction
Cycle
milliseconds
Correction Cycle
~1/30 second
10
Calibration of Generator Dynamic Model
(Data from August 18 2002, 1500 MW Navajo units drop)
Before calibration: Significant mismatch between simulation and
measurement
-740
5
10
15
20
25
30
35
40
90
-745
70
-750
-755
50
P (MW)
Q (MW)
-760
-765
30
10
-770
-775
-10
5
10
15
20
25
-780
P3_ref
P3_sim
P4_ref
P4_sim
35
40
Q3_ref
Q3_sim
Q4_ref
Q4_sim
-30
-785
-790
30
-50
t (s)
11
t (s)
Calibration of Generator Dynamic Model
(Data from August 18 2002, 1500 MW Navajo units drop)
5.5
380
5
360
Exciter Gain
Inertia Constant
After calibration: model and measurement match.
4.98
4.5
Inertia
4
3.5
10
20
t in seconds
30
40
reactive power (MVar)
real power (MW)
-740
-750
-760
Measurement
Model
-770
-780
340
0
10
30
20
t in seconds
40
311
320
300
0
Exciter Gain
0
10
20
t in seconds
30
40
20
0
-20
Measurement
Model
-40
-60
0
10
20
t in seconds
30
40
real power (MW)
-650
-700
Measurement
Model
-750
-800
0
30
20
t in seconds
10
reactive power (MVar)
Multi-Event Model Verification: good match
across multiple events after calibration
40
240
220
200
Measurement
Model
180
0
10
20
30
t in seconds
40
Event 1: July 16 2002, 2588 MW RAS generation drop
40
-740
-750
Measurement
Model
-760
-770
reactive power (MVar)
real power (MW)
-730
20
0
-40
-60
0
5
10
t in seconds
15
Measurement
Model
-20
0
5
10
t in seconds
15
Event 2:Verification (July 28 2003, 1252 MW Palo Verde #3 trip and 1500 MW of other generation loss)
Summary
Capturing dynamics is necessary because the power grid is trending
to be more dynamic (frequency deviates from nominal more often with
larger amplitude).
A dynamic paradigm is proposed to capture emerging dynamics and
understand where the grid is, where the grid is going, and where the
grid could be (what-ifs).
Phasor measurements offer opportunities for implementation with
Kalman filters.
Great performance for dynamic state estimation and dynamic model
validation, ready for pilot testing and adoption.
Challenges remain in real-time computational performance and nonGaussian noise impact.
14
References
Ning Zhou, Da Meng, Zhenyu Huang, and Greg Welch, “Dynamic State
Estimation of a Synchronous Machine using PMU Data: A Comparative
Study,” IEEE Transactions on Smart Grid, vol. 6, no. 1, pp. 450-460, January
2015.
Zhenyu Huang, Ning Zhou, Ruisheng Diao, Shaobu Wang, Steve Elbert, Da
Meng, and Shuai Lu, “Capturing Real-Time Power System Dynamics:
Opportunities and Challenges”, 2015 IEEE Power and Energy Society
General Meeting, Denver, CO, USA, July 26 – 30, 2015.
Zhenyu Huang, Pengwei Du, Dmitry Kosterev, and Steve Yang, “Generator
Dynamic Model Parameter Calibration Using Phasor Measurements at the
Point of Connection”, IEEE Transactions on Power Systems, Volume: 28,
Issue: 2, Page(s): 1939 - 1949, 2013.
Zhenyu Huang, Kevin Schneider, and Jarek Nieplocha, “Feasibility studies of
applying Kalman filter techniques to power system dynamic state estimation,”
in Proc. 8th Int. Power Eng. Conf. (IPEC), Singapore, Dec. 2007, pp. 376–
382.
15
Questions?
16