Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
Autonomous Energy Systems: Transportation
Brennan Borlaug, Kalpesh Chaudhari, Rob Fitzgerald, Venu Garikapati, Yanbo Ge, Matt Moniot, Clement Rames, Nick Reinicke, Jinghui Wang, Eric Wood
NREL | 2
• Overview of two research areas related to transportation and grid impacts– Learned ride-hailing fleet control, load management– Consensus charging overview
Presentation Overview
NREL | 3
• Overview of two research areas related to transportation and grid impacts– Learned ride-hailing fleet control, load management– Consensus charging overview
• Integration into broader AES simulation framework
Presentation Overview
Ride Hailing Modeling, Managed Loads
NREL | 5
Model Overview, Inputs and Outputs
• Battery size• Fleet size• Occupancy, etc.
• Station power levels• Locations, plugs, etc.
• O-D locations• Pickup times
• Passenger pooling willingness
• $/kWh by TOD
NREL | 6
Model Overview, Inputs and Outputs
• Battery size• Fleet size• Occupancy, etc.
• Station power levels• Locations, plugs, etc.
• O-D locations• Pickup times
• Passenger pooling willingness
Broad Outputs:• Vehicle utilization• Fleet performance• Station economics• … and many more!• $/kWh by TOD
Highly Integrated Vehicle Ecosystem
NREL | 7
Trip Demand:Ride Austin TNC Data
Animation Web link
Fleet Visualization (Austin, TX)
https://i.imgur.com/4UpXQjH.gifv
HIVE 0.4.0+ Model Structure
NREL | 9
Data-Driven Fleet Control
• Version: 0.1.0, heuristic based decision making• Version: 0.4.0: Refactored HIVE in an “RL gym” to enable model training• Exploring opportunities for improved performance beyond heuristics
NREL | 10
Data-Driven Fleet Control
• Version: 0.1.0, heuristic based decision making• Version: 0.4.0: Refactored HIVE in an “RL gym” to enable model training• Exploring opportunities for improved performance beyond heuristics
Agent“Fleet Manager”Environment
State• Current & upcoming
requests• Current & upcoming
prices to charge• Fleet average SOC• Fleet composition by state
NREL | 11
Data-Driven Fleet Control
• Version: 0.1.0, heuristic based decision making• Version: 0.4.0: Refactored HIVE in an “RL gym” to enable model training• Exploring opportunities for improved performance beyond heuristics
Agent“Fleet Manager”Environment Action
• % Fleet Charging• % L2 vs DCFC
State• Current & upcoming
requests• Current & upcoming
prices to charge• Fleet average SOC• Fleet composition by state
NREL | 12
Data-Driven Fleet Control
• Version: 0.1.0, heuristic based decision making• Version: 0.4.0: Refactored HIVE in an “RL gym” to enable model training• Exploring opportunities for improved performance beyond heuristics
Agent“Fleet Manager”Environment Action
• % Fleet Charging• % L2 vs DCFC
State• Current & upcoming
requests• Current & upcoming
prices to charge• Fleet average SOC• Fleet composition by state
Reward• Marginal Income• Request revenue – charging costs
NREL | 13
Data-Driven Fleet Control
• Version: 0.1.0, heuristic based decision making• Version: 0.4.0: Refactored HIVE in an “RL gym” to enable model training• Exploring opportunities for improved performance beyond heuristics
Agent“Fleet Manager”Environment
Reward• Marginal Income• Request revenue – charging costs
Action• % Fleet Charging• % L2 vs DCFC
State• Current & upcoming
requests• Current & upcoming
prices to charge• Fleet average SOC• Fleet composition by state
NREL | 14
Sample Scenario, Downtown Chicago
• Requests: Two days of request data - one for training, one for test• Fleet: 350 fleet vehicles with 50 kwh battery• Infrastructure 4 fast charging stations, 1 base with slow charging• Charging Costs: variable by time of day, based on data from ComEd• Forecasting: perfect for upcoming prices and requests
NREL | 15
Results, RL-Trained Fleet Manager
N training steps
Rew
ard
($10
00)
NREL | 16
Results, RL-Trained Fleet Manager, cont.
NREL | 17
Future Work, HIVE
• Extended study over 10+ days of request data• Expose fleet manager to more complicated rate structures
(demand charges)• Simulate more constrained scenarios – limited infrastructure,
larger geographic area• Expand scope of state/action space to include control of more
than just charging– Fleet rebalancing
Consensus Charge Control
NREL | 19
Background – Typical Control Hierarchies
Centralized data acquisition
from vehicles
Optimized charge profiles passed to
EVs from centralized node
A) No Control B) Centralized Control
All vehicles permitted to charge as demanded
Node
AES Simulation Framework
Node
AES Simulation Framework
Node
AES Simulation Framework
NREL | 20
Proposed: Consensus-Based Control
C) Consensus-based Distributed ControlRather than communicating with a centralized node, vehicles communicate amongst each other to develop a charging profile through consensus
Step 1: Communication among EVs Step 2: Aggregate optimized charge profiles from individual EVs
EV
EV
EV
Pev,i
Pev,i Pev,i
EV
EVEV
EVSE
Pev,i Pev,i
Pev,i
Node objective
Pev,i - Charging profile for each EV= min. Pev,iWhere,
NREL | 21
Consensus Control – Deeper Dive
Plim = Grid supply/capacity constraint
Pcs,iPcs,i
Pcs,i
EVSE
EVSEEVSEPaggPcs,i Pcs,i
Pcs,iEVSE
EVSE
EVSE
Step 3: Communication among Charging Stations
Step 4: Aggregate optimized load profilefrom individual charging stations
Step 5: If Pagg
NREL | 22
Consensus Control: Parameters Assumed
Focus: Demand charge mitigation, by flattening the charging profileParameters:
– EVi is the vector specified by (ai, di, ei, Pmax,i)– ai is the arrival time of EVi.– di is the departure time of EVi.– ei is the charging energy demand of EVi. ei (t) =0 if t < ai or t≥di– Pmax,i is the peak charging rate of EVi.– ri(t) is the instantaneous charging rate of EVi. ri (t) =0 if t
NREL | 23
Consensus Control: Performance Evaluation, cont.
0.00
200.00
400.00
600.00
800.00
1000.00
1200.00
100 120 140 160 180 200 220 240 260 280 300
Peak Power vs. No. of Vehicles
Real-time No Control
Real-time Central
Real-time Hierarchical w consensus
0.00
200.00
400.00
600.00
800.00
1000.00
1200.00
100 120 140 160 180 200 220 240 260 280 300
Peak Power vs. No. of Vehicles
24h Horizon No Control
24h Horizon Central
24h Horizon Hierarchical w consensusPe
ak P
ower
No. of Vehicles No. of Vehicles
Peak
Pow
er
NREL | 24
Consensus Control: Performance Evaluation, cont.
0.00
10.00
20.00
30.00
40.00
50.00
60.00
100 120 140 160 180 200 220 240 260 280 300
Comparison for 24h Horizon
24h Horizon % reduction Central
24h Horizon % reductionHierarchical w consensus
0.00
10.00
20.00
30.00
40.00
50.00
60.00
100 120 140 160 180 200 220 240 260 280 300
Comparison for Real-time
Real-time % reduction Central
Real-time % reductionHierarchical w consensus
No. of Vehicles
% P
eak
Redu
ctio
n
No. of Vehicles
% P
eak
Redu
ctio
n
NREL | 25
Consensus Control: Performance Evaluation, cont.
050
100150200250300350400450500550600650700750800850900950
1000
100 130 150 200 220 240 260 290
Solution time (24h Horizon)
24h Horizon Central
24h Horizon Hierarchical w consensus
Tim
e (s
econ
ds)
No. of EVs
020406080
100120140160180200220240260280300320340360
100 130 150 200 220 240 260 290
Solution time (Real-time)
Real-time Central
Real-time Hierarchical w consensusTi
me
(sec
onds
)No. of EVs
Transportation Topics, AES Simulation Framework
NREL | 27
Load Shifting Strategy, Comparison
HIVE: “Plug” control• HIVE has no formal charge control, but can load
shift by controlling when vehicles are sent to charge• Fleetwide charging peaks are controlled through
strategic dispatch & charge instructions• Main incentive is to recharge vehicles quickly so
additional passengers may be served
Time of Day
Char
ge L
oad
No grid feedback, static decision-making
NREL | 28
Load Shifting Strategy, Comparison
HIVE: “Plug” control• HIVE has no formal charge control, but can load
shift by controlling when vehicles are sent to charge• Fleetwide charging peaks are controlled through
strategic dispatch & charge instructions• Main incentive is to recharge vehicles quickly so
additional passengers may be served
Consensus: “Charge” control• Consensus control can affect the charging rate during an
event, but has no control over plug-in / plug-out times• Charging peaks are controlled within the confines of a
dwell. Greater dwell time correlated with greater flexibility• Likely to be best integrated at fleet depots where vehicles
have long overnight stays
Time of Day
Char
ge L
oad
Time of Day
Char
ge L
oad
No grid feedback, static decision-making Direct integration with AES framework
NREL | 29
Proposed Electric Vehicle-Grid Integration Framework
• Grid constraints• Electricity price• Additional load data
• Optimized charge schedule• Optimized charging cost• Cost incentive for wait times
• Generation• Storage
values
• Optimized charge schedule for • Demand charge mitigation• Operating cost reduction• Charging stations integrated with building &
renewables• Quantify communication requirements• Enable peer-to-peer transactions for price negotiation
HIVE Electric Vehicle-Grid IntegrationUsing Distributed Control
AES Simulation Framework
(Distribution Grid)• Arrival and departure time• Start and end SOC• Energy requirement• Vehicle parameters• Connected charger details• Desired charging cost• Maximum allowable charging
delay
BuildingsRenewables &
Storage
• Controllable loads• Uncontrollable loads • Optimized
load profile
Electric Vehicles Loads DER Grid
This work was funded by the US Department o Energy Vehicle Technologies Office.
NREL | 31
Consensus Control: Performance Evaluation
Title SlidePresentation Overview 1Presentation Overview 2Ride Hailing Modeling, Managed LoadsModel Overview, Inputs and Outputs 1Model Overview, Inputs and Outputs 2Fleet Visualization (Austin, TX)HIVE 0.4.0+ Model StructureData-Driven Fleet Control 1Data-Driven Fleet Control 2Data-Driven Fleet Control 3Data-Driven Fleet Control 4Data-Driven Fleet Control 5Sample Scenario, Downtown ChicagoResults, RL-Trained Fleet ManagerResults, RL-Trained Fleet Manager, cont.Future Work, HIVEConsensus Charge ControlBackground – Typical Control HierarchiesProposed: Consensus-Based ControlConsensus Control – Deeper DiveConsensus Control: Parameters AssumedConsensus Control: Performance Evaluation, cont. 1Consensus Control: Performance Evaluation, cont. 2Consensus Control: Performance Evaluation, cont. 3Transportation Topics, AES Simulation FrameworkLoad Shifting Strategy, Comparison 1Load Shifting Strategy, Comparison 2Proposed Electric Vehicle-Grid Integration FrameworkThanks! Questions?Consensus Control: Performance Evaluation