1 of 80

Electric Vehicle Fleet and Charging Infrastructure Planning��Presented by: Sushil Varma�Ph.D. Student, Georgia Tech���Motivated by my internship at Tesla�INFORMS TSL Best Student Paper Award – Finalist

2 of 80

Joint work with…

Francisco Castro

Assistant Professor

Anderson School of Management, UCLA

Siva Theja Maguluri

Associate Professor

ISyE, Georgia Tech

2

3 of 80

Motivation to Study Electric Vehicles

Technological Advancements

Government Backing

Climate Change Awareness

3

4 of 80

We study a Transportation System with EVs

4

5 of 80

We study a Transportation System with EVs

4

6 of 80

We study a Transportation System with EVs

4

7 of 80

We study a Transportation System with EVs

4

8 of 80

We study a Transportation System with EVs

4

System Operator’s Decisions:

  • Which EV to dispatch (based on location and state of charger)?

9 of 80

We study a Transportation System with EVs

4

System Operator’s Decisions:

  • Which EV to dispatch (based on location and state of charger)?

10 of 80

We study a Transportation System with EVs

4

System Operator’s Decisions:

  • Which EV to dispatch (based on location and state of charger)?

11 of 80

We study a Transportation System with EVs

4

System Operator’s Decisions:

  • Which EV to dispatch (based on location and state of charger)?

12 of 80

We study a Transportation System with EVs

4

System Operator’s Decisions:

  • Which EV to dispatch (based on location and state of charger)?
  • When to send to charge? For how long?

13 of 80

We study a Transportation System with EVs

4

System Operator’s Decisions:

  • Which EV to dispatch (based on location and state of charger)?
  • When to send to charge? For how long?

14 of 80

System Operator’s Decisions:

  • Which EV to dispatch (based on location and state of charger)?
  • When to send to charge? For how long?

We study a Transportation System with EVs

 

4

15 of 80

Infrastructure Planning Question

 

 

5

Target Service Level = Fraction of Customers Served

16 of 80

Preview of the talk

Infrastructure Planning Prescription

Trade-off between fleet size, number of chargers, and battery pack size

Near-Optimal Dispatching

Power-of-d Vehicles Dispatch Policy is near-optimal

Overview of the Model followed by the following four results

Simulations

Verifies the theoretical results and provides further insights

Time-Varying Arrivals

Phase Transition from EV-like to non-EV-like behavior

6

17 of 80

Analyzing Spatial Model

Idle/Charging

Busy

Spatial Abstraction

[Besbes-Lobel-Castro-22],

This work

Discretization Error in Spatial Quantities

Challenging to generalize to EV setting

Optimizes complexity v/s exactness trade-off

[Varma-Bumpensanti-Maguluri-Wang-21], [Varma-Castro-Maguluri-22], [Varma-Maguluri-22]

Discretization

Each region is a queue

Exact Analysis

(Non EV)

[Kanoria-2021], [Chen-Kanoria-Kumar-Zhang-23]

8

18 of 80

Model

 

7

19 of 80

Model

 

Abstracting out the spatial component

 

 

Idle/Charging

Busy

The state space involves SoC and state of all EVs (idle/charging/driving)

Dynamics is given by a system of ODEs

7

20 of 80

Preview of the talk

Infrastructure Planning Prescription

Trade-off between fleet size, number of chargers, and battery pack size

Near-Optimal Dispatching

Power-of-d Vehicles Dispatch Policy is near-optimal

Overview of the Model followed by the following four results

Simulations

Verifies the theoretical results and provides further insights

Time-Varying Arrivals

Phase Transition from EV-like to non-EV-like behavior

6

21 of 80

Trade-off between Fleet size and # of chargers

Infrastructure Planning Prescription

 

 

 

 

 

 

 

More Vehicles

More Chargers

 

 

No policy can achieve target service level

 

Admitted

Workload

Compensation for Downtimes

# of charging EVs to compensate driving

9

22 of 80

Infrastructure Planning Prescription

 

 

 

 

 

 

 

More Vehicles

More Chargers

 

 

No policy can achieve target service level

 

Admitted

Workload

Compensation for Downtimes

# of charging EVs to compensate driving

 

Power-of-d achieves target service

9

23 of 80

Global Stability Justification

 

24 of 80

Takeaways

 

 

 

 

More Vehicles

More Chargers

 

 

Power-of-d achieves target service

 

No policy can achieve target service level

 

 

10

 

25 of 80

Takeaways

 

 

 

 

More Vehicles

More Chargers

 

 

Power-of-d achieves target service

 

No policy can achieve target service level

 

 

 

10

 

26 of 80

Takeaways

 

 

 

 

More Vehicles

More Chargers

 

 

Power-of-d achieves target service

 

No policy can achieve target service level

 

 

 

10

 

27 of 80

Takeaways

 

 

 

 

More Vehicles

More Chargers

 

 

Power-of-d achieves target service

 

No policy can achieve target service level

 

 

 

10

 

28 of 80

Takeaways

 

 

 

 

More Vehicles

More Chargers

 

 

Power-of-d achieves target service

 

No policy can achieve target service level

 

 

 

 

10

 

29 of 80

Takeaways

 

 

 

 

More Vehicles

More Chargers

 

 

Power-of-d achieves target service

 

No policy can achieve target service level

 

 

Pareto Frontier: Characterizes the trade-off between fleet size and charger density

 

 

10

 

30 of 80

Takeaways

Partially charged EVs help!

Downtimes due to charging

Capacity Planning

Nominal Capacity

Buffer

Reason

M/M/n

[Halfin-Whitt]

Stochasticity

Spatial

[Castro-Besbes-Lobel]

Spatial effects

Spatial + EVs

[This Work]

Spatial + SoC

Power-of-d achieves target service

 

 

 

 

More Vehicles

More Chargers

 

 

No policy can achieve target service level

 

 

EV v/s non-EV System

Staffing for

non-EV system

 

10

 

 

31 of 80

Takeaways

Partially charged EVs help!

Downtimes due to charging

Capacity Planning

Nominal Capacity

Buffer

Reason

M/M/n

[Halfin-Whitt]

Stochasticity

Spatial

[Castro-Besbes-Lobel]

Spatial effects

Spatial + EVs

[This Work]

Spatial + SoC

No policy can achieve target service level

 

 

EV v/s non-EV System

10

 

Power-of-d achieves target service

 

32 of 80

Preview of the talk

Infrastructure Planning Prescription

Trade-off between fleet size, number of chargers, and battery pack size

Near-Optimal Dispatching

Power-of-d Vehicles Dispatch Policy is near-optimal

Overview of the Model followed by the following four results

Simulations

Verifies the theoretical results and provides further insights

Time-Varying Arrivals

Phase Transition from EV-like to non-EV-like behavior

6

33 of 80

Which EV to dispatch?

15%

90%

85%

20%

50%

25%

45%

 

11

Challenge: Need to pick an EV not too far with not too low SoC

Our Idea: Develop a dispatching policy by making connections to load balancing in queueing

34 of 80

Load Balancing

Policy

Delay

Overhead

Equivalent Policy

SoC Balancing

Pickup Time

Join the Shortest Queue (JSQ)

Good

High

Highest SoC Dispatch

Good

High

Random Routing

Poor

Low

Closest Dispatch

Poor

Low

Good

Moderate

Power of d Dispatch

Good

Moderate

 

Challenge: Need to pick a less loaded queue with a low overhead

Simplicity & connections to load balancing makes it amenable to analysis

 

35 of 80

Frequent Charging Sessions?

 

36 of 80

Preview of the talk

Infrastructure Planning Prescription

Trade-off between fleet size, number of chargers, and battery pack size

Near-Optimal Dispatching

Power-of-d Vehicles Dispatch Policy is near-optimal

Overview of the Model followed by the following four results

Simulations

Verifies the theoretical results and provides further insights

Time-Varying Arrivals

Phase Transition from EV-like to non-EV-like behavior

6

37 of 80

Simulation Setup

 

12

Converges to a steady state

Workload Served

Stable Aggregate SoC

Very less idle EVs

38 of 80

Validating the Scaling Results

 

 

 

 

 

 

 

 

% error

 

1

0.9

0.8

0.7

4.1%

0.2%

3.2%

3.6%

 

0.508

0.547

0.599

0.639

 

0.529

0.548

0.580

0.617

Provides Empirical validation of the Theory

 

13

More EVs and Less Chargers

39 of 80

How to compute 90% fleet size?

40 of 80

Validating the Scaling Results

 

Series

A

B

C

D

4.1%

0.2%

3.2%

3.6%

Less

chargers

More

EVs

 

 

 

41 of 80

Pickup and Drive to Charger Times

Pickup err.

A

B

C

D

To Charger err.

4%

7%

11%

15%

3%

1%

2%

6%

Pickup time is insensitive to fleet size and # of chargers

Drive to charger time increases as the charger density decreases

Verifies the spatial abstraction

 

 

42 of 80

 

Po2

CD

 

 

14

43 of 80

Power of d v/s Closest Available Dispatch

  • The workload served under CAD is much smaller than Po2 and even CD
  • Wild Goose Chase results in large pickup times for CAD
  • Large pickup times induce low SoC, implying inefficient system operation
  • Po2 and CD actively drop customers to maintain stable SoC

15

44 of 80

Further comparisons with natural policies

 

CD: Closest Dispatch

CAD: Closest Available Dispatch

Po2: Power of 2

45 of 80

Preview of the talk

Infrastructure Planning Prescription

Trade-off between fleet size, number of chargers, and battery pack size

Near-Optimal Dispatching

Power-of-d Vehicles Dispatch Policy is near-optimal

Overview of the Model followed by the following four results

Simulations

Verifies the theoretical results and provides further insights

Time-Varying Arrivals

Phase Transition from EV-like to non-EV-like behavior

6

46 of 80

Time-Varying Arrival Rate

 

 

 

 

 

 

 

 

Incoming Workload

EV Driving

EV Charging

 

 

 

 

16

Small Peak Amplitude

47 of 80

Infrastructure Planning Prescription

 

 

 

 

 

 

 

More Vehicles

More Chargers

 

 

No policy can achieve target service level

 

Admitted

Workload

Compensation for Downtimes

# of charging EVs to compensate driving

 

Power-of-d achieves target service

9

* Global stability of the underlying ODE is conjectured

48 of 80

Universal Lower Bound

Tight Upper Bound

Formulate a system of ODEs - tracks the evolution of the state of system

Generic ODEs satisfied by any policy

Detailed ODEs for Power-of-d Dispatch

Useful bounds on the Fixed Point

Fixed point translates into bounds on fleet size and number of chargers

Existence, Uniqueness, and Characterization of Fixed Point

 

 

Effective Arrival Rate

 

 

Aggregate Charge Rate

Aggregate Discharge Rate

Aggregate SoC

Rate of driving to charger

Rate of finishing charge session

# of EVs charging

18

49 of 80

Universal Lower Bound

Tight Upper Bound

Formulate a system of ODEs - tracks the evolution of the state of system

Generic ODEs satisfied by any policy

Detailed ODEs for Power-of-d Dispatch

Useful bounds on the Fixed Point

Fixed point translates into bounds on fleet size and number of chargers

Existence, Uniqueness, and Characterization of Fixed Point

 

 

 

Effective arrival rate minus the total service rate

# Cars arriving minus leaving the chargers

 

Aggregate charge rate minus the discharge rate

 

50 of 80

Proof Idea

Step 1: Abstract out the spatial component

19

51 of 80

Proof Idea

Step 1: Abstract out the spatial component

19

52 of 80

Idle/Charging

Busy

Proof Idea

Step 1: Abstract out the spatial component

All charging/idling EVs with the same SoC are homogenous

 

Step 2A: Define state space

19

53 of 80

Proof Idea

Idle/Charging

Busy

Step 1: Abstract out the spatial component

Step 2A: Define state space

All charging/idling EVs with the same SoC are homogenous

 

19

54 of 80

Proof Idea

Step 1: Abstract out the spatial component

Step 2A: Define state space

All charging/idling EVs with the same SoC are homogenous

 

Idle/Charging

Busy

19

55 of 80

Proof Idea

Idle/Charging

Busy

Step 1: Abstract out the spatial component

Step 2A: Define state space

All charging/idling EVs with the same SoC are homogenous

 

19

56 of 80

Proof Idea

Idle/Charging

Busy

All charging/idling EVs with the same SoC are homogenous

 

Step 2B: Define transitions under Power-of-d

Step 1: Abstract out the spatial component

Step 2A: Define state space

 

 

19

57 of 80

Idle/Charging

Busy

Proof Idea

All charging/idling EVs with the same SoC are homogenous

 

Step 2B: Define transitions under Power-of-d

Step 1: Abstract out the spatial component

Step 2A: Define state space

 

 

19

58 of 80

Idle/Charging

Busy

Proof Idea

All charging/idling EVs with the same SoC are homogenous

 

Step 2B: Define transitions under Power-of-d

Step 1: Abstract out the spatial component

Step 2A: Define state space

 

 

19

59 of 80

Proof Idea

Idle/Charging

Busy

Step 1: Abstract out the spatial component

Step 2A: Define state space

All charging/idling EVs with the same SoC are homogenous

 

Step 2B: Define transitions under Power-of-d

 

 

 

 

 

 

 

 

19

60 of 80

 

61 of 80

Takeaways

 

20

62 of 80

Overview of my Work

Stochastic Matching Network: Matching queue and its applications in matching markets

Fundamental Results in Stochastic Networks

Applications to Matching Markets

Stochastic Processing Network: The equivalent classical queueing model

Single Server Queue

[VM, Performance’22]

State-Dependent Arrivals

EV-Based Ride-Hailing [VCM, MS (Under Review)]

Charging and Spatial Matching

Heavy-traffic Theory of Matching Queues

[VM, Performance’22]

Dynamic Pricing

 

Online Marketplaces [VBMW, OR’21] [VCM, SIGMETRICS’21]

Dynamic Pricing and Matching

Parallel Server Queues

[VM, SIGMETRICS (Under Review)]

Production System Flexibility

21

63 of 80

Questions?

CREDITS: This presentation template was created by Slidesgo, and includes icons by Flaticon, and infographics & images by Freepik

22

64 of 80

Bonus Slides

65 of 80

Global Stability Justification

 

66 of 80

 

67 of 80

Frequent Charging Sessions?

 

68 of 80

How significant is the reduction in the second order term?

 

69 of 80

Trade-off: Fleet size, pack size and # of chargers

70 of 80

CAD is operating inefficiently as it tries to serve all customers

This results in high pickup times

Which reduces the SoC to the minimum

With no partially charged EV advantage, it reinforces the pickup times to be high

Po2 and CD preemptively drops customers to maintain a stable non-zero SoC

Further comparisons with natural policies

71 of 80

Further comparisons with natural policies

 

CD: Closest Dispatch

CAD: Closest Available Dispatch

Po2: Power of 2

72 of 80

Idle/Charging

Busy

Model

 

 

Effective arrival rate minus the total service rate

# Cars arriving minus leaving the chargers

 

Aggregate charge rate minus the discharge rate

 

For lower bound, coarse tracking of states is sufficient

73 of 80

Idle/Charging

Busy

Model

 

 

Effective arrival rate minus the total service rate

# Cars arriving minus leaving the chargers

 

Aggregate charge rate minus the discharge rate

 

For lower bound, coarse tracking of states is sufficient

74 of 80

Idle/Charging

Busy

Model

 

 

Effective arrival rate minus the total service rate

# Cars arriving minus leaving the chargers

 

Aggregate charge rate minus the discharge rate

 

For lower bound, coarse tracking of states is sufficient

75 of 80

Universal Lower Bound

 

Average time EV spends driving to serve a customer:

 

Trip Time

Pickup Time

Drive to

Charger Time

 

76 of 80

Universal Lower Bound

 

Average time EV spends driving to serve a customer:

 

Trip Time

Pickup Time

Drive to

Charger Time

 

77 of 80

Universal Lower Bound

 

 

Average time EV spends driving to serve a customer:

 

Trip Time

Pickup Time

78 of 80

Universal Lower Bound

 

 

Average time EV spends driving to serve a customer:

 

Trip Time

Pickup Time

Use partially charged EVs for pickup

79 of 80

Universal Lower Bound

 

 

Average time EV spends driving to serve a customer:

 

Trip Time

Use partially charged EVs for pickup

80 of 80

Universal Lower Bound

 

 

Average time EV spends driving to serve a customer:

 

Trip Time

Use partially charged EVs for pickup

Fleet size requirement