Forecast-Aided Q-Learning for Decentralized Game-Theoretic EV Charging and Discharging at Solar-Powered Supply Equipment in Smart Cities: Learning When to Plug In

Decentralized electric vehicle supply equipment (D-EVSE) powered by photovoltaic (PV) plants and local storage can serve EV charging and vehicle-to-grid discharging in smart cities, but existing decentralized game-theoretic (D-GT) schemes are myopic in time: EVs plug in as soon as they need energy, so the morning demand peak drains the storage before the PV output ramps up, while midday PV energy is curtailed. This paper adds a learning layer that determines when a flexible EV plugs in, while the D-GT best response still determines where. A long short-term memory (LSTM) network trained on six years of measured PV data (NREL PVDAQ) forecasts each station’s PV output one hour ahead; for plug-in demand, the mean daily profile of real charging sessions (Caltech ACN-Data) is used because it is more accurate than an LSTM on validation data. Each station broadcasts the resulting forecast energy balance, and a tabular Q-learning agent, trained on these forecasts rather than on realized values, maps it and the station’s state to a plug-in-now or wait-one-hour decision. On a six-station Ottawa network over 30 unseen days with 3000–8000 EVs per day, the resulting scheme (FQL-DGT) increases the average satisfaction at 8000 EVs from 86.5% to 91.8%, reduces unserved requests from 2.2% to 0.6% and the mean queueing time from 9.8 to 5.3 min, and decreases grid imports by 22% (5000 EVs) and 9% (8000 EVs). At the 60-min horizon used by the decision, the PV forecast has a normalized RMSE of 13.57%, against 20.92% for persistence. Ablations show that learning the congestion dynamics provides most of the gain and forecasting a smaller additional part at high load; a linear autoregressive PV forecast, although less accurate than the LSTM, serves the decision layer equally well. The mean ex post regret of the switching station remains at or below 0.024, and a decision takes approximately 0.11 ms.

Authors

Institutions

Publication Details

Journal
Electronics
Published
2026-11-12
DOI
https://doi.org/10.3390/electronics15225266
Primary Topic
Electric Vehicles and Infrastructure
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Forecast-Aided Q-Learning for Decentralized Game-Theoretic EV Charging and Discharging at Solar-Powered Supply Equipment in Smart Cities: Learning When to Plug In

Turki G. Alghamdi
Electronics
Electric Vehicles and Infrastructure
article

Forecast-Aided Q-Learning for Decentralized Game-Theoretic EV Charging and Discharging at Solar-Powered Supply Equipment in Smart Cities: Learning When to Plug In

Turki G. Alghamdi
article en

Abstract

Decentralized electric vehicle supply equipment (D-EVSE) powered by photovoltaic (PV) plants and local storage can serve EV charging and vehicle-to-grid discharging in smart cities, but existing decentralized game-theoretic (D-GT) schemes are myopic in time: EVs plug in as soon as they need energy, so the morning demand peak drains the storage before the PV output ramps up, while midday PV energy is curtailed. This paper adds a learning layer that determines when a flexible EV plugs in, while the D-GT best response still determines where. A long short-term memory (LSTM) network trained on six years of measured PV data (NREL PVDAQ) forecasts each station’s PV output one hour ahead; for plug-in demand, the mean daily profile of real charging sessions (Caltech ACN-Data) is used because it is more accurate than an LSTM on validation data. Each station broadcasts the resulting forecast energy balance, and a tabular Q-learning agent, trained on these forecasts rather than on realized values, maps it and the station’s state to a plug-in-now or wait-one-hour decision. On a six-station Ottawa network over 30 unseen days with 3000–8000 EVs per day, the resulting scheme (FQL-DGT) increases the average satisfaction at 8000 EVs from 86.5% to 91.8%, reduces unserved requests from 2.2% to 0.6% and the mean queueing time from 9.8 to 5.3 min, and decreases grid imports by 22% (5000 EVs) and 9% (8000 EVs). At the 60-min horizon used by the decision, the PV forecast has a normalized RMSE of 13.57%, against 20.92% for persistence. Ablations show that learning the congestion dynamics provides most of the gain and forecasting a smaller additional part at high load; a linear autoregressive PV forecast, although less accurate than the LSTM, serves the decision layer equally well. The mean ex post regret of the switching station remains at or below 0.024, and a decision takes approximately 0.11 ms.

ElectronicsVol. 15(22)
Jouf University (SA)
Affordable and clean energy, Industry, innovation and infrastructure, Sustainable cities and communities, Climate action
Electric Vehicles and Infrastructure
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon