Unlocking the potential of smart EV charging: A user-oriented control system based on Deep Reinforcement Learning
Christoforos Menos-Aikateriniadis, Stavros Sykiotis, Pavlos S. Georgilakis
Abstract
During the last years, Electric Vehicles (EV) have disrupted the transport market and are considered a key enabler towards economy-wide decarbonization. However, a high penetration of distributed energy resources on a residential level, such as EV, are expected to increase domestic demand and put a further stress on the operation of power distribution networks. This paper aims at suggesting real-time EV charging strategies based on Deep Reinforcement Learning (DRL). The end-users can select among different policy profiles, namely “cost savings” or “user-oriented”, depending on their personal optimization goals. Historical daily EV charging data from residential EV owners in Austin, Texas, USA have been analyzed to create user-specific tendencies based on the day of the week (weekday/weekend). The extracted end-user charging tendencies are combined with electricity consumption cost, solar PV self-consumption and EV battery constraints, including the number of daily charging activations. Experimental results on 6 different households show that the proposed Deep Q-Network (DQN) EV charging policies manage to reduce daily electricity costs by up to 49.83% when following a “user-oriented” energy policy while being able to charge up to 86.22% of the time in time periods that are consistent with users’ historical charging patterns. • Introduces two smart EV charging policies based on end-users optimization goals. • Categorizes end-users based on their “Overnight” or “Diurnal” charging tendency. • Achieves up to 49.83% cost savings and 86.22% compliance with historical EV patterns.