Adaptive Charging Networks: A Framework for Smart Electric Vehicle Charging

Zachary J. Lee; George Lee; Ted Lee; Cheng Jin; Rand Lee; Zhi Low; Daniel Chang; Christine Ortega; Steven H. Low

Source-linked AI summary

Adaptive Charging Networks: A Framework for Smart Electric Vehicle Charging

Zachary J. Lee, George Lee, Ted Lee, Cheng Jin, Rand Lee, Zhi Low, Daniel Chang, Christine Ortega, Steven H. Low

arXiv:2012.02636v1Dec 4, 2020eess.SY

TL;DR

Large-scale EV charging requires systems that can manage growth, grid constraints, and operational costs under realistic hardware and behavior. The paper presents the ACN and its Adaptive Scheduling Algorithm, then evaluates them with real workloads and accurate infrastructure models. ASA improves energy delivery in congested settings and achieves substantially higher profits than uncontrolled or baseline charging in the reported cases.

Problem
Large-scale EV charging must accommodate rising demand while handling constrained, unbalanced infrastructure, non-ideal battery behavior, and quantized control signals.
Method
The paper develops the ACN and an Adaptive Scheduling Algorithm based on convex optimization and model predictive control for real-time managed charging.
Results
ASA consistently outperforms baseline algorithms in highly congested settings and reported profit cases reach $2,835, or 98.1% of offline optimal, in an idealized setting.
Takeaways & Limitations
The ACN provides a deployed framework for scaling EV charging while adapting scheduling objectives to constrained infrastructure and operator profit.
Takeaways & Limitations
Discrete EVSE pilot constraints are difficult to incorporate directly into optimization-based algorithms and require heuristics.

Abstract

from arXiv · show

We describe the architecture and algorithms of the Adaptive Charging Network (ACN), which was first deployed on the Caltech campus in early 2016 and is currently operating at over 100 other sites in the United States. The architecture enables real-time monitoring and control and supports electric vehicle (EV) charging at scale. The ACN adopts a flexible Adaptive Scheduling Algorithm based on convex optimization and model predictive control and allows for significant over-subscription of electrical infrastructure. We describe some of the practical challenges in real-world charging systems, including unbalanced three-phase infrastructure, non-ideal battery charging behavior, and quantized control signals. We demonstrate how the Adaptive Scheduling Algorithm handles these challenges, and compare its performance against baseline algorithms from the deadline scheduling literature using real workloads recorded from the Caltech ACN and accurate system models. We find that in these realistic settings, our scheduling algorithm can improve operator profit by 3.4 times over uncontrolled charging and consistently outperforms baseline algorithms when delivering energy in highly congested systems.

I. INTRODUCTION

The paper presents the Adaptive Charging Network (ACN), a large-scale smart charging system designed to support real-time control while addressing infrastructure and operational constraints. It develops practical scheduling algorithms and evaluates them using real ACN data and realistic scenarios.

Motivation: Smart charging is needed to accommodate projected EV growth while limiting grid stress, charging costs, and infrastructure expansion requirements.Large charging facilities are especially important for workplaces, apartments, shopping centers, airports, and fleets.
ACN overview: The ACN enables real-time monitoring and control of large-scale charging facilities and has been deployed at over 100 sites across the United States.The first ACN was built on the Caltech campus in 2016 and has since supported diverse sites including laboratories, universities, businesses, apartments, hotels, and public parking facilities.
Scheduling approach: The Adaptive Scheduling Algorithm framework combines convex optimization and model predictive control for online smart charging under practical system constraints.The framework targets energy delivery in congested infrastructure and operating-cost objectives using simulations based on real ACN workloads.
Research infrastructure: The ACN research ecosystem includes real charging data, an open-source simulator, and a framework for safely field-testing algorithms.These components are organized as ACN-Data, ACN-Sim, and ACN-Live.
Contribution: The work extends prior ACN research by providing a fuller account of the architecture and algorithms and evaluating them in realistic scenarios.The paper emphasizes insights gained from building and operating real-world charging systems.

III. ADAPTIVE CHARGING NETWORK ARCHITECTURE

The ACN is a cyber-physical charging system that integrates information collection, sensing, computation, actuation, electrical infrastructure, and driver input. Its control layer uses event-triggered updates to adjust charging schedules from live system data and driver requests.

Physical deployment: The Caltech deployment expanded from 54 custom-built level-2 EVSEs to 126 commercial level-2 EVSEs, one 50 kW DCFC, and four 25 kW DCFC across three garages.The system had delivered over 1,103 MWh by July 7, 2020.
III. ADAPTIVE CHARGING NETWORK ARCHITECTURE: The ACN combines five interacting subsystems: information, sensors, actuation, physical infrastructure, and drivers.The information system computes control actions, sensors observe the electrical network, actuators regulate vehicle charging, and drivers provide data and availability decisions.
Control algorithms: The control layer recalculates charging schedules when vehicles or driver requests change, when utility demand-response signals arrive, or when a time threshold is exceeded.The system uses event-based scheduling updates with periodic recomputation, such as every five minutes.
Actuation: The J1772 pilot signal communicates an upper current bound to the vehicle BMS, which may draw less because of vehicle limits or battery state.Charging control therefore sets a maximum allowed current rather than guaranteeing the vehicle’s actual charging rate.
Sensing and data: Sensors measure power, current, and voltage and feed system-state information and energy-usage feedback to the control algorithm.The data layer also stores configuration, driver, session, and time-series measurements for real-time visualization.

D. Physical system

The Caltech ACN’s physical system combines three-phase service, line-to-line EVSE connections, oversubscribed feeder capacity, and stochastic driver demand. These properties create unbalanced-network and modeling challenges that practical scheduling algorithms must address.

D. Physical system: The California Parking Garage contains 78 EVSEs and a 50 kW DC fast charger distributed across two 150 kVA transformer-fed switch panels.The DC fast charger is a balanced three-phase load and is not directly controlled in the described system.
D. Physical system: The ACN uses two 150 kVA delta-wye transformers and line-to-line 208 V EVSE connections, so stochastic demand can produce unequal phase loading.Balanced operation cannot be assumed, making transformer protection challenging.
D. Physical system: The system demonstrates substantial oversubscription: each 80 A pod feeder serves eight EVSEs rated up to 32 A, a 3.2 times ratio.Smart charging allows EVSE capacity to exceed existing feeder capacity when charging is coordinated.
E. Drivers: Real deployments also face unpredictable driver arrivals, departures, and inaccurate mobile-app inputs.The authors explore machine-learning prediction and pricing schemes to improve driver information and system control.
A. Infrastructure modeling: Unbalanced three-phase infrastructure is a practical gap for prior charging algorithms because many assume single-phase or balanced three-phase systems, whereas ACN EVSEs are line-to-line connected.The paper identifies this topology as a reason existing methods may not transfer directly to large U.S. charging systems.
A. Infrastructure modeling: ACN modeling represents infrastructure limits through resource capacities, uncontrollable loads, and a matrix mapping individual EVSE currents to aggregate network currents.The resulting current constraints can be expressed as convex second-order cone constraints or more conservative affine constraints.

B. Battery Management System behavior

Practical EV charging must account for deviations between pilot signals and actual charging, discrete EVSE controls, and constraints that prevent unnecessary charging interruptions.

B. Battery Management System behavior: Actual EV charging current can deviate significantly from the EVSE pilot signal, requiring accurate battery models or robust scheduling.The deviation arises from each EV’s onboard charger and battery management system.
B. Battery Management System behavior: EVSEs support discrete pilot signals, including limited sets such as {0, 8, 16, 24, 32} A for some 32 A chargers.Other EVSE types provide finer 1 A control.
B. Battery Management System behavior: Allowable charging rates depend on both the EVSE and the modeled behavior of the EV’s battery management system.The allowable-rate set can vary with the EV and time.
B. Battery Management System behavior: Charging is constrained to remain non-zero until the EV’s demand is met, reducing contactor wear and improving user experience.This rule prevents repeated charging interruptions while the vehicle still has unmet demand.
B. Battery Management System behavior: Discrete charging constraints complicate optimization and therefore require heuristics to recover implementable schedules.The constraint is tied to each EV’s requested energy and the charging horizon.

V. ONLINE SCHEDULING FRAMEWORK

The online scheduling framework repeatedly optimizes charging for active EVs over a finite horizon, applies the current control action, and updates the system state using measured delivery.

V. ONLINE SCHEDULING FRAMEWORK: The framework combines convex optimization and model predictive control with practical constraints for online EV charging.It is designed to address unbalanced infrastructure, modeling inaccuracies, and discrete EVSE controls.
V. ONLINE SCHEDULING FRAMEWORK: At each time step, ASA identifies active EVs with unmet energy demand and remaining connection time.The active set excludes vehicles whose demand is met or whose departure is due.
V. ONLINE SCHEDULING FRAMEWORK: ASA recomputes an optimal schedule when an event occurs or the elapsed time since the last computation exceeds a threshold.Events include changes in vehicle presence or charging requests.
V. ONLINE SCHEDULING FRAMEWORK: The optimizer returns a finite-horizon schedule, whose current-period rate is applied as the pilot signal before the state is updated.Future schedule entries are reindexed at each computation, while measured energy delivery updates remaining demand.

B. Utility Functions Uk

The utility function combines weighted operator objectives, allowing the scheduler to trade off charging speed, profit, load behavior, fairness, and completion under changing conditions.

B. Utility Functions Uk: ASA represents multiple operator objectives as a weighted sum of utility functions with time-dependent priorities.The weights can change between optimization computations.
B. Utility Functions Uk: The charging-speed objective rewards delivering energy earlier in the scheduling horizon.Its reward is strictly decreasing in time.
B. Utility Functions Uk: Profit maximization uses charging revenue, time-varying energy cost, and predicted net load from other loads and on-site generation.Future load and generation are unknown and require prediction.
B. Utility Functions Uk: The fairness regularizer promotes equal sharing among EVs and can reduce line losses when other objectives are tied.It favors lower, more evenly distributed charging rates when equivalent solutions exist.
B. Utility Functions Uk: A non-completion penalty prioritizes energy delivery when demands cannot all be met, while remaining inactive when every demand is fully satisfied.For p > 1, EVs with higher remaining demand receive priority under infeasibility.

C. Feasible set Rk

The feasible set bounds charging rates, departure timing, and total delivered energy while relaxing discrete controls to preserve convexity and guaranteeing that optimization remains feasible.

C. Feasible set Rk: Charging rates are non-negative and bounded by each EV’s battery-management limits and the EVSE’s maximum pilot signal.Vehicle-to-grid operation is excluded.
C. Feasible set Rk: Departure constraints prevent an EV from charging after its session ends, while energy constraints cap delivery at its requested amount.The energy constraint is an inequality rather than an equality.
C. Feasible set Rk: Relaxing discrete charging rates keeps the scheduling problem convex, after which heuristics recover a feasible discrete solution.The continuous relaxation is necessary for tractable optimization with practical EVSE limits.
C. Feasible set Rk: Using an inequality for delivered energy ensures OPT always has a feasible solution, with the objective encouraging full delivery whenever possible.The zero charging schedule remains feasible even when all requests cannot be satisfied.

D. Quantization of pilot signal

The scheduler relaxes discrete EVSE pilot-rate constraints to preserve convexity, then recovers feasible discrete charging rates through rounding and capacity reclamation.

D. Quantization of pilot signal: Discrete pilot signals are omitted from the convex scheduling constraints and replaced with a relaxed nonnegative-rate constraint.The relaxation is necessary because discrete pilot constraints are generally intractable for large problems.
D. Quantization of pilot signal: The first optimized charging rate is rounded down to the nearest value allowed by each EV’s discrete rate set.
D. Quantization of pilot signal: Unused capacity after rounding is reclaimed by prioritizing EVs with the largest gap between optimized and rounded rates.Rates are incremented only when the resulting charging vector remains feasible.
D. Quantization of pilot signal: The minimum-rate constraint assumes that infrastructure can feasibly provide a minimum charging rate to every EV.

E. Battery tail capacity reclamation

The rampdown procedure reclaims capacity when battery-management systems draw less current than requested near full state of charge, while allowing temporary throttling to recover.

E. Battery tail capacity reclamation: Rampdown reclaims wasted capacity caused by battery-management systems limiting current as batteries approach full state of charge.
E. Battery tail capacity reclamation: When measured current falls sufficiently below the pilot signal, the next-period pilot upper bound is reduced to measured current plus a small margin.The margin σ is typically around 1 A.
E. Battery tail capacity reclamation: When the discrepancy becomes small again, the pilot upper bound is increased by σ up to the vehicle or EVSE limit.This lets the algorithm respond if the battery-management system’s throttling was temporary.
E. Battery tail capacity reclamation: The current implementation uses one pilot upper bound for all time periods within a scheduler subproblem.More advanced schemes could make this bound depend on time or the charging decisions.

VI. APPLICATIONS

The applications evaluate ACN scheduling with real operational data and realistic simulations spanning user arrivals, infrastructure constraints, control quantization, and battery behavior.

VI. APPLICATIONS: The evaluation targets two operational objectives: delivering energy quickly under constrained infrastructure and maximizing operating profit.
A. Data Collected: ACN data records arrivals, departures, driver estimates, delivered energy, control signals, and actual charging rates at 4-second resolution.The data can be used with ACN-Sim to evaluate scheduling algorithms.
A. Data Collected: Weekdays contain many flexible sessions with higher total energy delivery, whereas weekends contain fewer, shorter, relatively inflexible sessions.This workload variation limits simple solutions based on installing many level-1 chargers.
A. Data Collected: Weekday arrivals peak from 7:00–10:00 and departures from 16:00–19:00, while weekend arrivals and departures are more uniformly distributed.
B. Practical Scenarios: The simulations vary information availability, EVSE control granularity, battery behavior, and three-phase infrastructure using five practical scenarios.ACN-Sim uses 5-minute slots, 5-minute scheduler calls, and a maximum 12-hour optimization horizon.

C. Energy delivery with constrained infrastructure

ASA-QC is evaluated for energy delivery under transformer congestion against LLF, EDF, and round-robin baselines, with results showing strong performance in highly constrained settings and sensitivity to non-ideal EVSEs.

C. Energy delivery with constrained infrastructure: The experiment varies transformer capacity from 20 to 150 kW and measures the percentage of total energy demand met by ASA-QC and three baseline schedulers.The actual transformer capacity is 150 kW, whereas a conventional system of this size would require 362 kW.
C. Energy delivery with constrained infrastructure: ASA-QC performs near optimally within 0.4% in scenario II and outperforms EDF by as much as 14.1% at 30 kW transformer capacity.It performs better than the baselines in almost all cases, especially under high congestion.
C. Energy delivery with constrained infrastructure: Non-ideal EVSEs substantially reduce ASA-QC performance because rounding optimal pilots restricts the feasible set.At transformer capacities below 60 kW, non-ideal EVSEs instead improve LLF and EDF performance, possibly through minimum-current phase balancing.
C. Energy delivery with constrained infrastructure: Non-ideal batteries have relatively little effect on ASA-QC compared with the baselines, indicating algorithmic robustness.
C. Energy delivery with constrained infrastructure: ASA-QC optimizes infrastructure constraints to balance electrical phases and prioritize current and anticipated congestion, unlike baselines using narrower information.Round-robin uses presence only, EDF uses departure time, and LLF also uses energy demand.

U PM := uEC + uDC + 10−4uQC + 10−12uES

The ACN combines real-time managed charging with ASA, a flexible model-predictive-control framework designed for practical infrastructure constraints. Across realistic workloads and models, ASA approaches offline-optimal profit and substantially outperforms uncontrolled charging while supporting research and deployment at scale.

Results: ASA-PM and ASA-PM with a peak hint came within 3.6% and 1.9% of optimal profit, respectively, and exceeded all baseline algorithms.The comparison used a 150 kW transformer capacity and uncontrolled charging as a baseline.
Results: Peak hints increased the share of energy demands met from 95.6% to 97.8% by allowing higher capacity utilization earlier without increasing cost.The algorithm still did not meet all demands because it lacked knowledge of future arrivals and acted conservatively about peak increases.
Method: ASA addresses unbalanced three-phase infrastructure, quantized pilot signals, and non-ideal battery behavior through model predictive control and preprocessing and postprocessing heuristics.Its objectives can be configured for different operator goals, including energy delivery and profit maximization.
Results: $2,600, or 90% of offline-optimal profit, was achieved with non-ideal batteries and EVSEs; simulations also yielded 3.4 times the profit of uncontrolled charging.In an idealized setting, profit reached $2,835, or 98.1% of offline optimal.
Research infrastructure: The ACN Research Portal extends the system’s role through open data, an open-source simulator, and field testing of charging algorithms.These components are ACN-Data, ACN-Sim, and ACN-Live.