Smart Crystal Ball on a Budget: Reinforcement Learning and Prediction for Budget-Friendly Comfort

Current smart thermostats adjust temperature settings to ensure thermal comfort for users. However, it is beneficial to enhance such thermostats to be adaptive to dynamic energy pricing and budget constraints without compromising user comfort. We present a financially autonomous thermostat achieved through reinforcement learning techniques. Our main contribution is the Crystal-gazer agent, a reinforcement learning-based agent that integrates weather predictions to deliver more advanced and adaptive solutions for budget-efficient and user-centric smart thermostats. The results achieved through simulations using real-world data demonstrate that when the budget is insufficient to achieve maximum satisfaction, the Crystal-gazer is highly effective in maximizing budget usage and achieves near-optimal temperatures approximately $\mathbf{9 7. 6 6 \%}$ of the time.