Scientific Machine Learning for building energy consumption

Due to its impact on global energy consumption, it is fundamental to take action in order to reduce the building sector’s energy use, which means achieving energy-efficient structures. Reaching this goal involves implementing advanced energy management and control strategies, which adeptly handle internal loads uncertainties and external disturbances. Consequently, it is crucial to develop a proper building model. White-box models strongly rely on the system’s physics, leading to an interpretable and generalizable solution. They require detailed knowledge and expertise and significant computational resources during execution. Black-box models, on the other hand, only count on historical data, therefore they are limited to their training dataset, lacking in interpretability. To address these considerations, interpretable machine learning, also known as Scientific Machine Learning (SciML), emerged. Its goal is to incorporate physical information about the system, ensuring interpretability without compromising the accuracy and computational efficiency seen in traditional machine learning models, such as neural networks, for which the quality and uniformity of the training and testing dataset are crucial. The aim of this thesis is to develop a Physics Informed Neural Network (PiNN) model with the objective of predicting the temperature within the environment. This model will incorporate the ordinary differential equation directly into its structure, thereby integrating the physical knowledge of the domain within the model itself. This work demonstrates the PiNN model's ability to rely less on the training dataset, showing greater generalizability compared to a traditional neural network model. It manages to generate predictions consistent with real-world data even with a leaner training set, making it versatile and promising. This capability opens up potential applications in various fields where accurate temperature prediction is essential, paving the way for further developments and applications in the field of thermal and environmental modeling.