Reinforcement learning-based control for collective heating systems

Abstract:As the world deals with the impacts of excessive energy use on climate change and sustainable development, the demand for energy-efficient solutions is rising. This thesis investigates Reinforcement Learning (RL)-based control strategies for collective heating systems, focusing on reducing energy use for Space Heating (SH) and Domestic Hot Water (DHW) while maintaining occupant thermal comfort. The goal is to develop adaptive, energy-efficient control approaches for residential collective heating systems. Heating accounts for 78% of final energy use in European households. In Belgium, achieving the ambitious target of a 46% reduction in GHG emissions by 2030 compared to 2005 levels requires significant improvements in heating system efficiency. Deploying low-temperature emitters in new and renovated dwellings and implementing collective heating systems offer promising pathways for decarbonizing thermal energy supply. However, the increasing complexity of these systems poses challenges for optimal control, which is crucial for maximizing energy efficiency. Advanced control techniques such as MPC and RL have gained prominence in building energy systems. While MPC relies on solving complex optimization problems at each control step, which makes it computationally intensive for large-scale systems, RL offers more adaptability and computational efficiency after training, making it a suitable choice for collective heating systems. Despite its potential, RL faces challenges related to scalability, adaptability, and privacy. This thesis addresses these challenges through the development of energy-efficient, adaptive control strategies for collective heating systems, with a focus on scalability and privacy-aware implementation. The research is presented through five papers, organized into three chapters: Chapter 3 explores RL-based control performance in real-world settings. In the absence of a collective heating system testbed, a single-dwelling case study using a heat pump as the heat production unit is utilized to assess RL’s ability to improve energy efficiency and occupant comfort under dynamic conditions. Chapter 4 focuses on centralized control using single-agent RL. As the scope ex pands from single dwellings to multiple dwellings in collective heating systems, RL agents must manage increased complexity, balancing energy and cost reduction with thermal comfort across multiple dwellings. The effectiveness of centralized RL in handling SH and/or DHW control is a key area of investigation. Chapter 5 addresses the challenge of scalable, distributed control in collective heat ing systems. The goal is to maintain the autonomy of individual dwellings while ensuring overall energy efficiency. This involves navigating issues of privacy-aware control, coordination, and conflicting objectives among distributed RL agents.

Publication year:2025

Keywords:Engineering sciences. Technology

Accessibility:Open