Implementation of an adaptive control strategy to regulate heating systems in residential building

In recent years, control systems able to predict the continuous adjustments of dynamic factors, which allow the adaptability in the building energy management, have become necessary due to the increasing complexity of HVAC systems, and the rapid change in occupants' behaviour. Classic control systems, including On/Off or PID, can not perform these tasks because they do not provide any prediction capabilities. Moreover, model-based predictive control strategies, such as Model Predictive Control (MPC), are complex to apply because they both need a model for the optimisation, which is difficult to achieve and have a high computational cost. For these reasons, recent researches are focusing on model-free control strategies, and in particular on the application of Reinforcement Learning (RL). Since RL does not require a prior known model, the agent learns the best action through trial-and-error interactions within the environment, following an action-reward process. In this dissertation, a control algorithm based on Soft Actor-Critic (SAC) is implemented to control a radiant floor heating system in an existing residential building. The initial phase consists of the construction of geometrical and energy models. It is necessary to implement the control agent, which is then tested in the training stage to estimate potential energy savings and temperature violations' reduction. As a consequence, through a sensitivity analysis, conducted on the hyperparameters to determine the best configuration, an energy saving of 5% and a significant decline in the sum of temperature violations are obtained. After the training phase, the agent is tested in the deployment phase by analysing four different scenarios to examine its adaptivity in different conditions. In conclusion, the agent obtains a reduction of temperature violations in all scenarios, and, at the same time, the energy-saving obtained ranges between 4% and 6%.