Conference Agenda

Recent advances in artificial intelligence (AI) have driven broad application across industries, with machine learning systems now demonstrating superhuman performance in numerous tasks. However, this leap in performance has often come at the cost of increased model complexity, turning many systems into opaque “black boxes” and raising concerns about how decisions are made. This lack of transparency poses challenges for deploying AI in sensitive yet high-impact domains, such as energy benchmarking, where trust and accountability are crucial.

Energy benchmarking in commercial buildings is often the first step toward evaluating performance and determining whether further interventions are needed. Yet, progressing beyond benchmarking remains challenging. Identifying buildings suitable for upgrades or pinpointing opportunities for energy savings and operational cost reduction is often hindered by the complexity and cost of in-depth assessments. To overcome these barriers, this study integrates data-driven modeling, machine learning and Explainable AI (XAI) to enhance both energy benchmarking and retrofit evaluation processes. Building stock-scale counterfactuals showed that envelope upgrades with high-performance HVAC achieved ~9% average savings, while lighting modernization with advanced controls yielded ~11% median savings.

Rapid early-phase design requires building energy models (BEMs) that can be generated, simulated, and optimized as quickly as design ideas evolve. While recent AI-driven tools translate text or images into BEMs, they lack guidance for navigating design trade-offs or iterative optimization. This study presents a context-aware, AI-directed framework that interprets natural language goals, proposes design variants, instantiates EnergyPlus models via Honeybee, and executes batch simulations with ranking and rationale. The proposed framework uses code-compliant guardrails and override options, enabling flexible, rule-based optimization. By automating intent parsing and variant evaluation, the framework closes the loop between design input and performance-driven model iteration.

High-fidelity building energy simulation (BES) tools such as EnergyPlus are costly to scale for optimization, urban modeling, and Artificial Intelligence (AI) workflows due to non-differentiability and orchestration overheads. We present a portable high-throughput computing (HTC) framework that distributes and monitors EnergyPlus ensembles across heterogeneous backends (on-premises clusters, national facilities, and cloud services) with fault-tolerant scheduling and an AI-ready data layer. One backbone supports real-time heat-event re-simulation for neighborhood alerts and dataset creation for training differentiable surrogates and inverse models. Benchmarks show near-linear wall-time reductions to resource limits and reproducible outputs. Coupling portability, fault-tolerance, and standardized interfaces, the framework democratizes large-scale AI-ready EnergyPlus studies.

Benchmarking building energy performance serves as an important tool for comparing energy use, identifying retrofit opportunities, and supporting energy-efficiency policies in the commercial building sector. This paper investigates deep learning approaches for predicting the annual electricity consumption of small office and retail strip mall buildings to support large-scale benchmarking. The paper leverages a previously developed building energy modeling workflow that integrates OpenStudio Analysis and EnergyPlus to generate 384 synthetic datasets for surrogate model training and validation. The datasets encompass two building types, 6 prototype HVAC systems, and sixteen ASHRAE climate zones representing the continental United States. Each sample within the datasets is produced by systematically varying key building parameters such as equipment efficiency, insulation levels, and geometric characteristics to capture a wide range of realistic design and operational conditions.

Traditionally, benchmarking approaches rely on linear regression models, decision trees, or, shallow neural networks. However, these methods may be limited in capturing nonlinear interactions among diverse building features. To address this, different neural network architectures are evaluated and their prediction performance is assessed at three scales: (1) a global model, trained on all use-cases with one-hot encoded categorical inputs; (2) separate models for each building–system–climate combination; and (3) a hierarchical multi-task model that combines these two approaches. The networks are further refined using autoencoders, which reduce input dimensionality, and physics-informed loss functions, which promote physically consistent predictions. Results show that the tuned global model consistently achieves lower CVRMSE and higher R² than the separate models.