Reinforcement Learning for Building Energy Management: Sensitivity to Data and Training Strategies

This thesis investigates the application of Reinforcement Learning techniques to Building Energy Management Systems (BEMS), with a focus on the sensitivity of learning performance to training strategies and data diversity. After reviewing the current state of the art, the study analyses how different training configurations affect the agent’s ability to learn and generalise control policies. Two training approaches are compared. In the first, the environment resets at the end of each episode, with the simulation restarting from the same initial date and weather conditions. In the second approach, the reset is ignored, enabling the agent to experience temporal continuity and different outdoor and operating conditions across episodes. The Soft Actor-Critic (SAC) algorithm is employed to control the HVAC system of a multi-zone office building simulated in EnergyPlus and wrapped within a Gym API, offering a physics-based and customisable environment. The agent acts on the temperature set points during the winter season to minimise energy use while maintaining a comfortable indoor temperature. The results show that the diversity of the training data has a stronger influence on the policy performance than the mere number of training iterations, with the temporal sliding-window approach leading to more robust and generalisable control strategies. Compared to the baseline controller, improvements of 6% in temperature control and 3% in energy savings were achieved over the entire testing period. The improvement is particularly significant in the warmest month, highlighting the potential of RL-based control strategies for adaptive and energy-efficient building operations.