Dynamic Power Management with Hybrid Power Sources

1
Dynamic Power Management with Hybrid Power Sources
44th DAC, San Diego, CA June 7th, 2007

• Jianli Zhuo1, Chaitali Chakrabarti1,
• Kyungsoo Lee2, and Naehyuck Chang2
• 1EE, Arizona State University, U.S.
• 2CSE, Seoul National University, Korea

2
Fuel Cell in Portable Applications
Functionality proven but not optimized!
Ballard power system (www.ichet.org)
Fujitsu (pr.fujitsu.com)
Toshiba (www.engadget.com)
Toshiba, KDDI, and Hitachi (www.ubergizmo.com)
3
Outline

• Introduction
• Basics of a fuel cell (FC)
• FC characteristics
• FC-based hybrid power source
• Previous work
• Optimal FC output setting for a known DPM profile
• Fuel-efficient DPM algorithm
• Experimental results
• Conclusion and future work

4
Fuel Cell (FC)

• Fuel cell
• An electrochemical energy conversion device w/o
  combustion
• Uses external supply of fuel and oxygen
• Electrodes are catalytic and relatively stable
• Advantages of Fuel cell
• High energy density ? longer lifetime for same
  weight/size
• Instant recharge
• No performance degradation during discharge
• Clean, zero emission

5
Basic Operation of FC

• PEMFC operation
• Oxidation at anode
• Reduction at cathode

FC stack operation
6
FC Characteristics
Measured VI characteristic of a BCS 20W, 20 stack
FC

Current (A)
Polarization I-V-P curves for room temperature
fuel cell

• As the load current increases, the output voltage
  decreases
• As the load current increases, the output power
  first goes up and then down
• Load following range fuel flow control (Inlet
  hydrogen pressure)

7
Advantage of Hybrid Power Source

• Fuel cell/battery hybrid power source
• Fuel cell provides high energy density
• Battery provides high power density (peak power
  value) andfast load matching

voltage
current
Iave
Advantage of using a hybrid power source over a
fuel cell only source
8
FC- Battery Hybrid (Our Prototype)

• Fuel cell system
• Fuel processor generates hydrogen
• Temperature and cathode air flow control
• Power conditioning and charge management system

Demonstrated in Univ. Booth, DAC 06
9
Outline

• Introduction
• Previous work
• Others work on fuel cell
• Our previous work
• Optimal FC output setting for a known DPM profile
• Fuel-efficient DPM algorithm
• Experimental results
• Conclusion and future work

10
Others Work on FC System

• Membrane and fuel cell stack
• Hydrogen generation
• Fuel cell hybrid vehicles
• Hybrid automobiles are different from
  human-portable embedded systems
• Orders of magnitude slower in system dynamics,
  and thus larger time constants
• Need faithful tracking according to the user
  demands w/o explicit slack times
• Focus on bidirectional load (a load and a
  generator)
• Fuel cell usage in human-portable systems
• Expensive solutions for functional demonstration
• Functionality has been proven, but not optimized

11
Our Work on FC System

• Prototype of the FC-Battery hybrid system
• The prototype has been exhibited in Univ Booth,
  DAC06
• DVS algorithms for embedded systems powered by
  FC-B hybrid source
• When FC works at fixed output level (DAC06) the
  voltage scaling level of the DVS system is
  determined by the power model of the embedded
  system and the power state of the hybrid power
  source
• When FC works at multiple output levels
  (ISLPED06) by jointly applying DVS to the
  embedded system and FC control to the power
  source, the fuel consumption can be reduced
  further
• In both cases, we assumed constant FC efficiency
• New contributions of this work
• A more efficient FC system configuration with
  different efficiency curve
• The optimal FC control policy when the FC system
  efficiency isnot a constant
• Development of FC-aware DPM algorithm

12
Outline

• Introduction
• Previous work
• Optimal FC output setting for a known DPM profile
• Overview of DPM
• DPM-enabled embedded system powered by FC hybrid
  source
• Definitions
• FC system efficiency
• Motivational example
• Optimal FC output setting
• Fuel-efficient DPM algorithm
• Experimental results
• Conclusion and future work

13
Dynamic Power Management

• Principle of DPM
• Put the system into low power state when the idle
  time is long enough
• Break-even time

• DPM Techniques
• Prediction of the future idle periods
• Linear function, regression function, adaptive
  learning tree, etc.
• Stochastic control based on Markov chain model
• Aggregation of idle times to get longer idle
  duration
• Battery-aware DPM
• Battery scheduling
• Load profile shaping

14
System Overview

• Fuel cell system
• FC stack, DC-DC converter, BOP (fan controller,
  etc.)
• Charge storage
• Charging when IFgtIld, discharging when IF lt Ild
• Embedded system with DPM

15
Definitions

• Embedded system
• Power states
• Trans. overhead
• Task slots
• FC system
• FC stack output
• FC system output
• Charge storage
• Capacity
• State of charge
• Optimization goal
• Maximize lifetime ? minimize fuel consumption
• Fuel consumption charge consumption due to Ifc

16
FC System Efficiency (definitions)

• FC system efficiency
• FC system output power divided by the Gibbs free
  energy per unit time
• is proportional to the fuel flow
  rate, which is proportional to the FC stack
  current, i.e., , so
• FC stack efficiency

17
FC System Efficiency

• FC stack efficiency follows the same trend as
  stack voltage Vfc
• FC system efficiency Determined by stack
  efficiency, DC-DC converter efficiency, and the
  loss due to BOP (the controller current)
• PWM DC-DC constant-speed fan ? constant
  efficiency
• PWM-PEM DC-DC variable-speed fan ? linear
  efficiency

18
Motivational Example

• Assumption
• The load current profile has been generated by a
  DPM policy
• Goal
• For the given load current profile, determine the
  FC output setting such that the fuel consumption
  is minimized
• Power configurations
• FC system load following range is 0.3 A1.2 A,
• Charge storage element capacity is 200 A-s,
  initial state is 0.
• Load profile

19
Example
(a) Conv-DPM no FC control
39 A-s
(b) ASAP-DPM FC output follows the load
faithfully
16 A-s
(c) FC-DPM the most efficient FC output setting
13.45 A-s
20
Optimal FC control (1)

• No DPM state transition overhead
• Objective function is

• Assumptions
• FC has a very large load following range
• The charge storage has unlimited capacity
• We add the constraint that the charging and
  discharging amounts are equal in each task cycle

• Solution by Lagrange method

21
Optimal FC control (2)

• We assumed unlimited load following range
• If the load following range is limited, then the
  FC output is bound by the range.
• We assumed unlimited charge capacity
• If the charge capacity is limited, then we have
  an additional constraint
• We remove the constraint Cend Cini
• They may not be equal because of the load
  following range constraint, charge capacity
  constraint, and the task execution time
  variations
• In this case, the constraints are changed to

22
Optimal FC control (3)

• Consider state transition overhead
• Assumptions
• During state transition, FC output is same as
  that in active period
• Transition between RUN and STANDBY exists for
  every slot, so we only need to take care of the
  overhead of STANDBY??SLEEP
• We greedily assume the next idle slot will be in
  SLEEP mode, and we take into account the power
  down overhead in advance.
• Change of the functions
• Boolean variable 1 -- SLEEP, 0
  STANDBY (in idle period)
• The objective function and the constraint are now
• We can use a method similar to that in the
  previous slides to solve the above optimization
  problem.

23
Outline

• Introduction
• Previous work
• Optimal FC output setting for a known DPM profile
• Fuel-efficient DPM algorithm
• Prediction based DPM for the embedded system
• Optimal FC control policy for the power source
• Experimental results
• Conclusion and future work

24
Prediction-based DPM

• Traditional DPM aims at load energy minimization,
  which helps in fuel consumption minimization
• Lower load energy(constant voltage)? lower
  IF,iTiIF,aTa ?lower FC output current ? lower
  FC stack current ? lower fuel consumption
• The embedded system can use many existing DPM
  algorithms with energy-minimization objective
• We borrow a simple prediction-based DPM algorithm
  
• Proposed by C.H. Hwang and A. Wu in ICCAD97
• The length of the idle period if predicted as a
  linear combination of the predicted length and
  the actual length of the previous idle period.

25
FC system output control

• On the power source side, in order to determine
  the optimal FC output, we need the information of
  the active period as well
• The length of the active period is derived using
  a similar prediction function as that used for
  the idle period
• The current (power) of the active period
• We can assume that Ild,a is the same for all
  active period
• We can assume that Ild,a is the average of the
  past history
• We can also use some pre-known task parameters
• Then we can use the function derived in optimal
  FC control logic to determine the desired FC
  output level, and the corresponding fuel flow
  rate.

26
Flow chart of FC-DPM
27
Outline

• Introduction
• Previous work
• Optimal FC output setting for a known DPM profile
• Fuel-efficient DPM algorithm
• Experimental results
• Real trace based MPEG encoding/writing
• Random task trace
• Conclusion and future work

28
Experimental setting

• Power source
• Load following range is 0.31.2 A,
• Use a super capacitor as the charge storage,
  capacity is 1 F
• Embedded system (DVD camcorder)
• 4X speed DVD writer, 16MB buffer size, 5.28MB/sec
  writing speed
• Load timing profile Active period is 3.03 sec,
  Idle period 8, 20 sec.
• Power states characterization

29
Experimental results - 1
Current profile segments (when the active period
is fixed)
FC-DPM saves 24.4 fuel compared to ASAP-DPM The
lifetime extension is about 32
30
Experimental results - 2
Current profile segments (synthetic task trace)

• Synthetic random tasks based on the previous
  camcorder trace
• Ta 2, 4 sec
• Ti 5, 25 sec
• Ild,a 1,1.33 A

31
Outline

• Introduction
• Previous work
• Optimal FC output setting for a known DPM profile
• Fuel-efficient DPM algorithm
• Experimental results
• Conclusion and future work

32
Conclusion and future work

• Conclusion
• Measured and characterized the FC system
  efficiencies.
• Proposed an optimization framework which
  explicitly takes into account the characteristics
  of the hybrid power source and the FC system
  efficiency factor.
• Developed an FC-aware DPM algorithm, which can
  saveup to 24.4 fuel consumption for the
  experiment settingunder consideration.
• Next step
• Combination of DPM and DVS (will be presented in
  ISLPED07)
• Consideration of the power loss in the charge
  storage element (non-ideal efficiency)
• Formal control method to implement the fuel flow
  rate control
• Acknowledgement
• NSF grant (CSR-EHS 05059540)
• LG Yonam Research Foundation
• ICT at Seoul National University

33
Questions and Answers

• ? Thank you!

47.2 Dynamic Power Management with Hybrid Power
Sources