Abstract
Buildings account for a significant portion of global energy consumption, a matter of increasing concern as the construction sector continues to expand worldwide. This growth in building construction exacerbates the challenges related to energy conservation and emissions reduction, especially given the simultaneous demand for improved indoor comfort reflecting rising living standards. In this context, the early stages of building design present a crucial opportunity for enhancing energy efficiency, central to which is the design of energy-efficient buildings. Such designs not only reduce operational costs but also contribute to lowering greenhouse gas emissions, with the early design decisions playing a pivotal role in shaping a building's energy consumption profile.
The building design process is inherently dynamic and complex, influenced by a multitude of factors including cultural context, climatic conditions, and personal design preferences, involving numerous stakeholders. The state-of-the-art method for estimating the energy and environmental performance of various design options is building performance simulation. However, the labor-intensive setup and computational costs of simulation models often relegate their use to the later stages of the design process, rather than as an active component throughout.
The advent of machine learning algorithms offers a new approach to tackling analytical problems in building design, leading to the development of surrogate models. These models aim to emulate the outputs of high-fidelity building simulation models based on given inputs, offering a significant advantage in computational efficiency. This allows for the exploration of vast regions of the design space and provides instantaneous feedback on the sustainability of a building during design sessions.
This doctoral research focuses on using deep learning (DL) to support energy-related design decisions in the early stages by providing information on the dynamic (hourly) cooling loads of buildings. While energy sustainability in buildings involves various aspects including heating loads, lighting, acoustics, and natural ventilation, this research zeroes in on cooling loads due to their complexity, sensitivity to design and operational parameters, and significant impact on energy efficiency and peak electrical demands. Cooling loads, characterized by their dynamic nature and sensitivity to factors like insulation and building orientation, often outweigh heating demands in modern buildings, particularly in warmer climates.
Addressing the challenge of representing design flexibly at early stages, this research employs a deterministic approach, simulating thousands of models and comparing multiple deep learning architectures to develop a robust surrogate model closely emulating the output of a full Building Energy Model (BEM). The focus on a deterministic rather than probabilistic approach allows for precise predictions of cooling loads for specific inputs, essential for evaluating the impact of individual design choices on the overall energy performance.
The research further explores integrating these DL models into existing design environments, using the Rhinoceros software with its parametric modeling tool Grasshopper as a test bed. This integration facilitates the reuse of design information and reduces modeling efforts, merging physics-based methodologies with advanced data-driven techniques focused on early design stages. Through three main experiments involving increasingly complex data and model architectures, the research aims to develop versatile, theoretically sound surrogate models for early-stage building design.
This work bridges the gap in modeling approaches, laying the foundation for innovative solutions that balance energy efficiency with practical design considerations. It contributes significantly to the global effort to decarbonize the building sector, aligning architectural innovation with sustainability goals. Despite the existence of a prediction gap between DL predictions and dynamic physics-based simulations, the research demonstrates through a test case that DL models are suitable for quickly performing comparative assessments of designs, providing a foundation for informed decision-making in performance-oriented building design.