Large-Scale Adaptive Electric Vehicle Charging Zachary J. Lee , - PowerPoint PPT Presentation

Large-Scale Adaptive Electric Vehicle Charging Zachary J. Lee , Daniel Chang, Cheng Jin, George S. Lee, Rand Lee, Ted Lee, Steven H. Low

I don’t have to convince you EV are coming… https://i.ytimg.com/vi/tj6B489H_zg/maxresdefault.jpg

We assume EV charging will look like this… http://o.aolcdn.com/hss/storage/midas/f31dd15c97d6237dd816c5d186980528/200403157/DP6V4737.jpg

But the future of EV charging in cities looks like this…

Capital Costs Prohibitively Expensive

Transformer Line Currents The Need for Uncontrolled Charging Adaptive Charging Adaptive 400 Current (A) Charging 200 0 06:00 10:00 14:00 18:00 22:00 06:00 10:00 14:00 18:00 22:00

Physical Charging Testbed

What good is a real testbed? • Working with real systems allow us to understand their limitations. • Without a proper understanding of these limitations our algorithms may look great on paper but be practically useless.

Utility Company Caltech Substation t 0 The Adaptive Charging Garage Loads 480 V Main Switch (Lighting, Fans, 800 A Panel Network Elevators, etc.) 50 kW 3­phase Transformer t 1 400 VDC 150 kVA, 480V/208V 208 V EV Switch Panel • 54 controllable level-2 EVSEs 420 A • 50 kW DC Fast Charger. ... x19 • Oversubscription of transformers, cables and breakers. • Demonstration environment for demand response, pricing schemes, and renewables integration.

54+ charging stations

150 kW of Capacity

585 MWh of energy delivered

1.8 million mile equivalent

610 $% avoided tons of CO #

What can we do with this system?

Data Collection

11,000 Charging Sessions since April 2018

Average Number of Sessions Average Length of Sessions Charging Session Average Total Energy Delivered Statistics Average Energy Delivered per Session Maximum Concurrent Sessions

Arrival Statistics

Session per Day

Simultaneous Sessions

Online Scheduling

Scheduling Problem Maximizing profit. Charging quickly. Maximizing renewable energy use. Following demand response signals. SCH max U k ( r ) No discharging. Maximum charging rate. r Relaxation of allowable rate set. 0 ≤ r i ( t ) ≤ ¯ r i ( t ) t < d i , i r i ( t ) = 0 No charging after departure. t ≥ d i , i d i − 1 Total energy delivered must be less than X r i ( t ) δ ≤ e i energy requested. t = a i f j ( r 1 ( t ) , ..., r N ( t )) ≤ R j ( t ) Infrastructure constraints. t ∈ T , j

Unbalanced Three-Phase Constraints I a0 ≤ R 0,a = 640 A I a1 ≤ R 1,a = 180 A I a1 ≤ R 2,a = 180 A I a3 ≤ R 3,a = 420 A a Garage DC Fast t 1 ­ primary t 1 ­ secondary EVSEs Loads Charger aux I a I a DC I ab evse I ca evse P I ca P I ab aux aux I c I c DC I b I b DC P I bc I bc evse b c | I 3 | I evse − I evse a | | R 3 ,a ≤ = ab ca

Unbalanced Three-Phase Constraints • We assume that we know/can measure the voltage phase angles at the EVSEs. • Since EVSEs can be modeled as constant current loads with unity power factor, we thus know the phase angles of their currents. • Since the magnitude of the current phasor is the only variable, these constraints are second-order cone constraints and the optimization problem is tractable. | 2 | I 3 ,a | 2 | I evse − I evse = ab ca | cos φ ca ) 2 + ( | I evse ( | I evse | cos φ ab − | I evse | sin φ ab − | I evse | sin φ ca ) 2 = ab ca ab ca R 2 ≤ 3 ,a

Unbalanced Three-Phase Constraints • We assume that we know/can measure the voltage phase angles at the EVSEs. • Since EVSEs can be modeled as constant current loads with unity power factor, we thus know the phase angles of their currents. • Since the magnitude of the current phasor is the only variable, these constraints are second-order cone constraints and the optimization problem is tractable. | I evse − I evse | I 3 ,a | | = ab ca | I evse | + | I evse | ≤ ab ca R 3 ,a ≤

125 100 Price ( $ ) Profit ($) 75 Phase Aware 50 Constraints A ffi ne Constraints 25 SOC Constraints Unconstained 0 0 . 0 0 . 2 0 . 4 0 . 6 0 . 8 1 . 0 Infrastructure Capacity (% Nominal) X U ( r ) := ( p ( t ) − c ( t )) r i ( t ) t ∈ T i ∈ V

Imperfect Actuation • Control is done via a pilot signal. Charging Rate (A) 30 • Pilot signal is only an upper bound on charging current. 20 • Battery management system is free to charge at any rate lower than the pilot. 10 Pilot Actual 0 16:30 16:45 17:15 17:30 17:45 17:00 18:00 Time

Model Predictive Control We use model predictive • collect active compute new optimal yes recompute? control to account for charging sessions schedule using SCH deviations. no Schedule is recomputed • periodically or when changes update energy update pilot signals occur in the system. remaining and from most remaining duration recent schedule

Actual Charging Behavior Charging Rate (A) 30 20 10 Pilot Simple Actual 0 16:30 16:45 17:15 17:30 17:45 17:00 18:00 Battery Time Two-Stage Battery Model Model 30 Current (A) 20 10 0 0 . 00 0 . 25 0 . 50 0 . 75 1 . 00 State of Charge

× 10 7 Robustness to Non-Ideal Charging Behavior 1 . 48 1 . 47 1 . 46 U(r) 1 . 45 Robustness Ideal Noiseless 1 . 44 σ 2 = 1 A σ 2 = 2 A 1 . 43 σ 2 = 3 A 1 . 42 10 20 30 40 50 60 Maximum Recompute Period (minutes) X X U ( r ) := ( T − t ) r i ( t ) t ∈ T i ∈ V

Results

Profit Maximization

Conclusions • We should consider the unique challenges of large- scale charging infrastructure. • Adaptive scheduling can significantly reduce the capital and operating costs of large-scale charging systems. • Experience with real systems can inform how we design practical algorithms. • Real time data from our testbed can be found at caltech.powerflex.com.

Future Work • Demonstrating how large-scale EV charging can be used to flatten the “duck curve” • Demonstrating the viability of large-scale EV charging in demand response markets • Analyzing user behavior to design predictive scheduling algorithms

Releasing Dataset and Simulator email zlee@caltech.edu to be notified of the release

Questions