Deep Reinforcement Learning based Building Energy Management System

Abstract

With growing popularity of rooftop photovoltaic panels and increasing electricity costs smart buildings seek for the efficient ways of their electricity management. Due to intermittent nature of solar power and time varying electricity prices energy storage becomes profitable. Smart scheduling of energy transfer among components of energy management system (EMS) is the key to reduction of building expenditures on electricity. However, modern EMS environment has dynamic nature which changes not only throughout the day, but also throughout weeks, months and seasons, thus, management of this building energy system becomes a complicated task. Deep Reinforcement Learning provides a great tool for solving complex control problems like this. In this work, firstly, control task is modelled for further optimization problem formulation. Secondly, Deep Reinforcement Learning (DRL) based algorithm is proposed as a solution to this scheduling problem for Building Energy Management System. Proposed approach is compared with other methods with varied amount of available input data. DRL based algorithm was proved to be the best at minimizing customers’ expenditures on electricity when real-life dataset with large variance of input data depending on season is used.