Publications

[Abstract available upon request.]

[Abstract available upon request.]

We propose an information-theoretic and Bayesian framework for principled selection among competing surrogate model architectures for building energy simulation. The method combines mutual information criteria — including AIC, BIC, and DIC variants adapted for stochastic surrogate models — with full Bayesian model evidence computed via thermodynamic integration. Applied across 20 real-world commercial and residential building datasets, the framework consistently identifies parsimonious models that generalize better than those selected by cross-validation alone, and provides interpretable posterior model probabilities that communicate modeling uncertainty to practitioners making calibration decisions.

We introduce Physics-Guided Bayesian Neural Networks (PG-BNNs), which embed first-principles constraints — including thermal dynamics equations and energy conservation laws — directly into the prior distributions of Bayesian neural networks through physics-informed regularization and structured prior construction. Evaluated on building thermal response prediction tasks, PG-BNNs produce significantly better-calibrated uncertainty estimates than standard BNNs, maintain physical plausibility under distribution shift, and improve predictive accuracy in data-scarce regimes by 15–27% over purely data-driven baselines. The approach enables reliable confidence bounds for model predictive control and fault diagnostics in building energy systems.

A comprehensive systematic review of 150+ papers (2015–2024) on uncertainty quantification (UQ) methods applied to machine learning-based building energy models. We survey Bayesian approaches (BNNs, GPs, variational inference), frequentist ensembles, conformal prediction, Monte Carlo dropout, and hybrid physics-data methods. The review establishes a taxonomy of uncertainty sources (aleatory vs. epistemic, input uncertainty, model structural uncertainty, and parameter uncertainty), synthesizes evaluation standards across studies, and identifies critical gaps: lack of calibration reporting, insufficient treatment of distribution shift, and rare application to control-oriented models. We propose a standardized UQ evaluation framework and identify high-priority research directions including integration with digital twins and equitable deployment under data scarcity.

[Abstract placeholder]

[Abstract placeholder]

[Abstract placeholder]

[Abstract placeholder]

[Abstract placeholder]

[Abstract placeholder]

[Abstract placeholder]

[Abstract placeholder]

[Abstract placeholder]

[Abstract placeholder]

[Abstract placeholder]

[Abstract placeholder]