A review of the state-of-the-art data-driven modeling of building HVAC systems

An accurate model for a building's HVAC system plays an essential role in providing thermal comfort for the building while maintaining efficiency. Recent advancements in artificial intelligence, along with the development of tools in various programming languages and platforms, enable researchers to quickly implement these technologies to tackle challenges across different engineering fields. For building energy system modeling, machine learning (ML) and deep learning (DL) models are gaining attention due to their high accuracy and less modeling effort than physics-based white-box models. However, the data-driven model is less robust than the physics-based model due to the lack of physical interpretation. Therefore, a comprehensive review to summarize the latest progress of data-driven approaches in building energy systems is needed. This study addresses the limitation of the early review by 1) performing a comprehensive review of recent advancements in data-driven approaches for building HVAC systems modeling and highlighting key findings; 2) summarizing the attributes of various data types and sources used in data-driven models across different tasks; 3) identifying essential features for different HVAC systems and analyzing the strengths and weaknesses of the feature selection methods; 4) reviewing the state-of-the-art techniques like hybrid and physics-informed neural network (PINN) model which improve the explainability and prediction with limited data; and 5) evaluating recent hyperparameter tuning methods for data-driven models and offer practical recommendations. Based on the review on the papers published in the last 15 years and related to the topics mentioned above, potential future research directions in this field are proposed.