Adaptive Control Strategies for enhancing energy efficiency and comfort in buildings

In recent years machine learning has expanded into new fields, including the building energy area. The concept of energy flexibility in buildings has become overwhelmingly in the interests of several scholars, making applications increase over time. Nowadays, control systems under analysis make use of algorithms belonging to a particular machine learning branch, called Deep Reinforcement Learning (DRL). DRL has become increasingly popular because it can control systems in which the dynamic process is very complex, as is mainly the case with HVAC systems. The current state of the art has very few, if any, real applications, but a lot of studies on the subject that are thought to become applicable in the shortest possible time. In this thesis work, it is implemented a DRL control method for a radiant heating system, installed on a real building for office use. In the initial phase, it was necessary to calibrate the energy model, based on the real and available temperature profiles, made available by the real Energy Management System. Through a trial-and-error approach, a calibrated model was obtained, as the metrics provided by ASHRAE were respected. After that, the calibrated energy model was used for the implementation of a DRL control agent, using a Soft Actor-critic (SAC) algorithm, in order to evaluate possible energy savings but above all the presence of the desired occupants comfort conditions during occupancy, compared to the baseline already present, based on the climate curve. In the initial training phase, a sensitivity analysis was carried out on the hyperparameters in order to choose the best configuration from those explored. The best agent's configuration allows to obtain energy savings of 5 % and at the same time to improve the internal comfort of the occupants, evaluated through the reduction of a sum of temperature violations compared to a fixed comfort range. The agent was then used for a static deployment phase on the current radiant system. Evaluating five different deployment scenarios, the excellent flexibility of DRL control logic concerning changes in the initial boundary conditions was proved. As a result, the comfort degree for each scenario has improved significantly, and in some cases even managed to make temperature violations null, opposite to the huge baseline values. At the same time, the energy savings obtained varied between 6 % and 8 %, except for the scenario in which the internal temperature setpoint was increased, where the energy saving achieved is in the order of 20 %.