Transcript Next Generation Systems & Smarter Planet

Authors Jagabondhu Hazra, Kaushik Das, Deva Seetharam, Amith Singhee
IBM Research
Presented by Anand Seetharam
Stream-computing Based
Synchrophasor Applications for
Power Grid
© 2011 IBM Corporation
Introduction
 Need for real-time situational awareness of the grid for stable, economic operation
– Increasing volatility from renewables
– Increasing energy usage
 Improved Sensing Technology
– Conventional sensors (e.g. Remote Terminal Units) in SCADA systems provide one
measurement every 4-10 seconds
– Phasor Measurement Units (PMUs) could provide upto 120 phasor and frequency
measurements per second
– PMUs provide more precise measurements with time stamp having microsecond
accuracy
– Phasor measurements are time synchronized across national-scale grid via GPS clock:
“synchrophasors”
 Potential for unprecedented real-time visibility into the grid state across a wide area
(regional/national)
2
© 2011 IBM Corporation
Phasor Measurement System Example (NASPInet)
 PMUs collect real-time data and through a communications system deliver the data from
many PMUs to a local data concentrator, Phasor Data Concentraretor (PDC).
 Concentrated data are relayed on a wide-band, high-speed communications channel to a
higher capability data concentrator sometimes called Super Phasor Data Concentrator
(SPDC)
 SPDC feeds the consolidated data from all the PDCs into analytical applications such as a
wide-area visualization, state estimator, stability assessment, alarming, etc.
Super PDC
(SPDC)
Master DB
Phasor Data
Concentrator
(PDC)
Phasor Data
Concentrator
(PDC)
GPS
satellite
Local DB
Local DB
PMU
PMU
PMU
PMU
PMU
PMU
3
PMU
PMU
PMU
PMU
© 2011 IBM Corporation
Real-Time Synchrophasor Applications
 Application examples
– Dynamic state estimation
– Voltage stability monitoring
– Oscillation monitoring
– Real-time grid stability control
 Requirements from application framework
– Low latency data processing: 100 ms – 1 s
– High data rates (throughput): 1000’s PMU x 120 /s
– Synchronization of data streams: Network jitter, different reporting rates
– Integration of analysis engines: State estimation, voltage stability, oscillation monitor
– Reconfigurable: Changes in grid
– Highly available: Configuration change, software upgrades
– Expandable: new data sources, new analytics
4
© 2011 IBM Corporation
Stream Computing: A New Paradigm
Streaming Algorithms used to analyze massive amount of real time data
'Useful information' extracted in 'low memory' in 'low time' complexity
Computations can be done in parallel to improve performance
Applications - database, networking and machine learning
Methods – sampling, sketches and clustering
5
© 2011 IBM Corporation
Stream Computing: A New Paradigm
Traditional Computing
Historical fact finding with data-atrest
 Batch paradigm, pull model
 Query-driven: submits queries to static data
 Relies on Databases, Data Warehouses
Queries
Queries
6
Data
Data
Results
Results
Stream Computing
Real time analysis of data-in-motion
Streaming data
 Stream of structured or unstructured data-in-motion
Stream Computing
 Analytic operations on streaming data in real-time
Data
Queries
Results
© 2011 IBM Corporation
Programming Streams
 Application specified as a data flow graph
– Data streams (tuples of data)
– Operators (operations on these tuples of data)
• Operators are triggered by arrival of tuple on input port
– Subscription model
 IBM InfoSphere Streams derived from System S
– Stream Processing Core: execution engine
SPADE: programming language and compiler
Computing Node
7
Processing
Element (PE)
PE Container
(PEC)
© 2011 IBM Corporation
Synchrophasors and Stream Computing
Synchrophasor systems can take advantage of stream computing because
– High volume of data: too much to store and mine
– Data streaming by, faster than a database can handle
– Complex analytics: correlation from multiple sources and/or signals
– Time Sensitive: responses required in under a couple of hundred milliseconds especially
for the control applications.
8
© 2011 IBM Corporation
Proof of Concept Application - Real time voltage stability monitoring
9
© 2011 IBM Corporation
Voltage Stability
Voltage stability is the ability of a power system to maintain steady acceptable voltages at all
buses in the system under normal operating conditions and after being subjected to a
disturbance.
Causes of voltage instability
 Disturbance
– During fault, angular difference between generators increases quickly which causes
depressed voltages
 Motor stall
– When terminal voltage of a motor goes below 80% of nominal, motor torque falls below
load torque and the motor slows to a standstill where it draws a large reactive current
further depressing voltage and force nearby motors to stall.
 Reactive power deficiency
– Reactive power available to a portion of the grid falls below that required by customers,
transmission lines, and transformers in that portion of the grid.
1
© 2011 IBM Corporation
Voltage Instability – An important industry problem
Power Blackouts caused by voltage instability
– 1996 US west blackouts
– 1997 Brazil blackout
“Voltage collapse is still the biggest single threat to
– 2003 US/Canada blackout
the transmission system. It’s what keeps me
– 2003 Italy blackout
awake at night.”
– 2003 South Sweden/Denmark
-Phil Harris, PJM President and CEO, March 2004
– 2005 Moscow blackout
– 2007 Colombia blackout
Lessons learn from history
 In general control devices are tuned under normal loading condition and hence are effective
under normal condition
 Most of the control devices do not perform satisfactorily during abnormal condition
 Need for intelligent online monitoring and decision making tools
Real time voltage stability monitoring and control using Synchrophasors with high
end communication & middleware architectures could be effective in ensuring the
voltage stability of the grid.
1
© 2011 IBM Corporation
Voltage stability index
 Voltage magnitudes, in general, do not give a good indication of proximity to voltage collapse
 Voltage stability index gives better idea about how far the current operating condition is from
voltage collapse
VR / ES
Voltage stability index is given by:
Operating point
1.0
0.8
Normal range
VM arg in
Stable
0.6
0.4
0.2
0.0
0.0
Unstable
Where,
NB Number of buses in the system
Pi Active power injection at bus i
Vi Voltage magnitude
Voltage phase angle at bus I
B
Admittance matrix
PR / PRMAX
At Vcritical, value of stability index is 0.5
Critical voltage
PM arg in
0.2
0.4
0.6
0.8
1.0
Fig. Character of a PV curve
1
© 2011 IBM Corporation
Test Grid
2
7
8
9
Generators : 3
3
G
G
zone1
zone3
6
5
: 3
Tr. Lines
: 6
Transformers : 3
4
1
Loads
zone2
PMUs
: 9
PDCs
: 3
SPDCs
: 1
G
Fig. 9 bus system
1
PMUs
PDC1
PDC2
PDC3
2, 5, 7
1, 4, 6
3, 8, 9
© 2011 IBM Corporation
SPADE application graph
PDC1
PDC2
PDC3
PDCs
1
SPDC
Application
© 2011 IBM Corporation
Real Time Stability Monitoring
Gradual overloading by 20%
Load buses 5, 6, 8
2
7
8
9
3
G
G
5
6
4
1
G
Fault
initiated
In the presence of a fault
1
© 2011 IBM Corporation
Aggregation Experiments
 Communication network may degrade: application needs to gracefully adapt by reducing
traffic
 Data prioritization
– Filter out phasors with V > Vth
 Data dropping
– Filter out phasors with insignificant change since last reading (phasor(t) ~= phasor(t-1))
– |Vt – Vt-1| <= Vth AND |δt – δt-1| <= δth
 Data clustering
– “Compress” N phasors into k < N cluster centers
 Partial computations
– VSI calculation for a bus needs phasors for the bus and its neighbors
– Distribute VSI calculation among different nodes instead of transporting all phasors to
one node
1
© 2011 IBM Corporation
Aggregation Experiments: Accuracy
 Aggregation methods tested on the IEEE
14 bus grid
 Accuracy ~1-10%
 Data prioritization has worst accuracy
– Too much information lost
 Partial computation has best accuracy
– All phasors are used
 Data dropping shows state-dependent
behavior
– Filtering depends on state evolution
1
© 2011 IBM Corporation
Aggregation Experiments: Reduction in Traffic
1
© 2011 IBM Corporation
Discussions

Stream computing is a compelling framework for data collected from PMU
- Highly parallelized and scalable
- Data flow abstraction
- Reconfigurability, expandability
Streaming Algorithms can be used and tailored for smart grid applications.
http://people.cs.umass.edu/~mcgregor/courses/CS711S12/index.html

Data from PMU considered as flows; flow algorithms and packet inspection algorithms can
be applied


Fault detection techniques used in networks can be used in smart grids as well.
1
© 2011 IBM Corporation
Performance Scalability in Streams
 Multi-level parallelism
– Across nodes
– Across PEs
– Across operator threads
 Compile-time data stream optimization
– Inter-node: network transport
– Inter-PEC (intra-node): shared memory between
processes
– Inter-PE (intra-PEC): pointer passing between
threads
– Inter-operator (intra-PE): direct function calls
 Compile-time operator-PE mapping
– Minimize inter-PE traffic
– Optimize processor utilization (not too low, not too
high)
– Statistics collection driven compilation
From [Amini et al, DM-SSP ’06]
700 PEs on 85 dual-core Xeon 3.06
2
© 2011 IBM Corporation
Synchrophasor Application Needs and Streams
 Low latency
– Data transport and PE management is highly optimized
 High data rates
– High parallelism and data transport optimization
 Synchronization of data streams
– Inbuilt operators like Barriers, Join in InfoSphere Streams. Also custom operators.
 Integration of analysis engines
– Edge adaptors (operators) for network, file and pipe connections
– Analytics can be implemented/interfaced via primitive operators in C++/Java
 Reconfigurable
– Operators can have state-based behavior and state can be modified dynamically
 Expandable
– Subscription model allows dynamic operator additions/upgrades and stream
addition/upgrades
– Fully dynamic application composition and re-composition possible
– New applications can dynamically subscribe to data from running applications
 Highly available
– Subscription model maintains PE independence: graceful fail-over
2
© 2011 IBM Corporation
Features of Streams
 Stream-centric design
– Process, analyze as soon as available: no intermediate archiving
 Operator / data-flow graph level declarative programming model
– High abstraction level keeps things simple
 Hides complexities of infrastructure
– Data streaming manipulations: e.g., language support for data types and building block
operations
– Application decomposition in a distributed computing environment: e.g., application
layout, resource optimization
– Computing infrastructure and data transport: e.g., shipping data streams between
operators, thread management
 C++, Java programming interface available
– Customized operators
2
© 2011 IBM Corporation