Energy Management System Using Deep Reinforcement learning

Abstract

The main objective of this study is to implement an Energy Management System (EMS) using Deep Q-Learning, a well-established deep reinforcement learning algorithm that combines Q-learning with deep neural networks. The proposed hybrid renewable energy system consists of a suitable combination of wind turbines, photovoltaic panels, and battery storage to supply a residential load connected to the grid. The EMS establishes a strategy for battery charging and discharging to ensure efficient operation. To achieve optimal performance and address the intermittency of renewable energy generation, the use of an EMS is essential. Deep Q-Learning is employed to minimize the total cost of energy drawn from the grid, particularly during peak hours. Two scenarios are simulated over a 72-hour period, with and without grid injection, under a bi-hourly pricing where energy costs vary across two time blocks each day. The net cost reaches $2.63 when grid injection is not permitted, while it decreases to -$1.82 when energy injection into the grid is allowed.