Deep Reinforcement Learning based Energy Management System for Controlling Storage Systems in a Cluster of Buildings: Cypriot use case

In this paper, a deep reinforcement learning (DRL) energy management system is developed to control various energy storage systems such as a cooling storage system, domestic hot water (DHW) storage tank, and also electrical storage systems (batteries) to increase the share of solar energy consisting of solar photovoltaic (PV) systems and solar collectors. The main aim is to study a cluster of three buildings in Cyprus, to enhance the energy flexibility and efficiency in both demand and supply sides. To this end, the collected measured data is preprocessed and utilized in CityLearn, a data-driven simulation environment, to train a DRL controller for managing distributed energy systems. A Long Short-Term Memory (LSTM) model is applied to evaluate the indoor dynamic temperature in each building, and then the developed DRL algorithm coordinates energy systems across the buildings. The results are compared with an optimized Rule-Based Controller (RBC) to demonstrate reductions according to proposed key performance indicators (KPIs): Average Daily Peak, Thermal Resilience Index, and Zero Net Energy. The results show that the proposed approach in this cluster of buildings significantly reduces the demand by up to 29%, especially during peak hours by consuming renewable energy sources (RES), and improves the overall performance of the building energy system.