Exploration of Reinforcement Learning Techniques as a Control Strategy for Heating Systems

Abstract

This thesis explores machine learning techniques to enhance the control of heating systems and compares them with traditional control methods which are often unsuitable for dynamic and nonlinear systems. The focus of study is Reinforcement Learning due to its adaptive nature. By thoroughly exploring this approach, an algorithm is selected for implementation based on its documented performance. In later sections, a simplified plate-based heat exchanger is modeled mathematically and then realized in a simulation software, where the algorithms are tested and compared with existing methods, focusing on key metrics such as response time, stability, and energy efficiency. After testing in simulation, the approach is applied to real hardware. The demonstration exhibits potential to address the shortcomings of current methods and improve the overall output. However, there are some challenges identified in the current implementation, including computational cost during model training, safety constraints when applied to real-time setup and convergence difficulties due to simplification in reward function. Future work includes integrating identified strategies with traditional solutions and exploring advanced algorithms for improved results. In conclusion, this research marks a step toward sustainable control strategies, providing valuable insights to meet present and future demands.