Household Energy Cost Optimization Using Deep Reinforcement Learning

Abstract: This thesis aims to address the rising energy costs by using IoT technology and reinforcement learning. We use historical sensor data to fit a deep reinforcement learning model that is capable of optimizing the control of a heating system in a way that minimizes energy costs, while maintaining a comfortable indoor temperature. This model-free approach uses neural networks to simulate the thermodynamic behavior of an existing building, making it more cost-effective than using building simulation software. Using the final Deep Q-Network model, a cost reduction of up to 25% was achieved.