Applying reinforcement learning to maximise photovoltaic self-consumption for electric vehicle charging

Abstract

Recently, and in line with the ongoing trend of electrification, many office and industrial buildings are investing in rooftop photovoltaic (PV) installations and electric vehicle (EV) charging stations. This combination and the ability to shift EV charging in time provides opportunities to apply the flexibility of EV charging for local peak shaving and load balancing. This study applies a reinforcement learning algorithm, fitted Q-iteration (FQI), to coordinate the charging of an EV fleet in an office building with the objective of maximising self-consumption of the local generation by shifting the morning peak of EV charging to the afternoon when there is more PV generation. The performance of the proposed algorithm is evaluated using real-world data collected at the EnergyVille research institute in Belgium. The data analysis shows that each EV arriving in this building can offer on average 4 h of flexibility – the number of hours that the EV's charging can be delayed – a day. Simulation results show that FQI achieves a 32.1% increase in self-consumption compared to the business-as-usual model and an 11% decrease compared to a theoretical optimal control. Simulation results also show the scalability of the proposed method with the number of available charging stations.