CEE PhD Defense: Eleanor Ho, "A Multi-Scale Investigation of Building Energy Use: From Attribution to Prediction, for Design and Operational Decision-Making"

Abstract:

Buildings account for roughly one third of global energy use, and both their design and operation shape how much they consume. This dissertation investigates building energy use across spatial scales, asking how building and urban form attributes affect energy use, and at what level of modeling detail we can support design and operational decisions. At the urban scale, an empirical study of hourly smart meter data from 170,224 utility customers shows that population density has significant, non-linear, and size-dependent effects on building energy use intensity, with reductions of up to 32-74% per density level across building types, consistent across monthly, daily, and hourly scales. At the fleet scale, the analysis moves from a static snapshot to time-varying attribution, asking which external drivers a building's energy use is coupled to and how that dependency structure evolves over time, by applying time-varying graphical lasso to metered electricity and open data for a fleet of buildings. Results across ten temporal lenses show that the dominant driver depends on the lens itself: solar and traffic dominate at daily and sub-daily resolution, and temperature emerges at hourly resolution when days of the same month are merged. At the building scale, a single multi-zone campus building and a controlled temperature setback experiment serve as one testbed for asking whether a physics-based model can predict a building's response to operational changes, and how much physics that takes. An EnergyPlus model calibrated against the measured setback response first shows that it can, reproducing the building's response to setpoint changes and evaluating operational strategies before implementation. The dissertation then asks how much of that physics is enough: a hybrid model combining an uncalibrated EnergyPlus simulation with downstream machine learning predicts cooling loads under unseen temperature setbacks within ASHRAE Guideline 14 tolerances, while pure machine learning baselines fall short without weeks of measured data from the new regime. This enables portfolio-scale deployment without demanding costly per-building calibration or field experiments. Together, these studies move from quantifying magnitude, to attributing drivers, to prediction, in support of design and operational decision-making.