Energy management of a smart home by using reinforcement learning methods

Home energy management system users schedule their real-time energy consumption thanks to the developments in communication technology and smart metering infrastructures. In this thesis, Online Home Energy Management System (ON-HEM), a data-driven strategy that uses model-free reinforcement learning algorithms (Q-Learning, Deep Q-Learning and Proximal Policy Optimization) to control the optimum energy consumption of a smart home system, is proposed. The proposed system includes power resources (grid, photovoltaic), communication networks and decentralized agents divided into four groups: deferrable, non-deferrable, power level controllable home appliances and electric vehicle. ON-HEM aims to reduce electrical costs and instantaneous high peak demands by considering the cost of user dissatisfaction with real-life data. Simulation studies were conducted with the proposed ON-HEM under different pricing approaches (Real-Time Pricing and Time-of-Use Pricing) with Q-Learning, Deep Q-Learning and Proximal Policy Optimization algorithms using PyCharm Professional Edition software. The findings showed both the superiority of policy-based Proximal Policy Optimization over value-based Deep Q-Learning and Q-Learning and the effectiveness of the proposed ON-HEM in reducing instantaneous high peak demands, electrical costs and user dissatisfaction costs. In terms of efficiency and reliability of the proposed system, ON-HEM was questioned whether the findings obtained in simulation studies with real-life data were meaningful using IBM SPSS Statistics software.