Conference Agenda

Modeling energy behavior across cities and regions requires detailed representations of building stock characteristics to enable accurate analysis of energy use.

Current approaches such as ResStock rely on pre-compiled datasets of building stock to model building characteristics distributions.

While effective, this method is not sustainable as the underlying data quickly become outdated, requiring extensive new data collection every few years to maintain accuracy. Moreover, such frameworks are typically limited to specific regions, such as the United States, restricting their applicability to global contexts.

With the rapid advancement of Generative AI and Large Language Models (LLMs), there is a growing opportunity to transform how building stock data are generated and updated.

Traditional static databases can no longer meet the scalability, adaptability, and contextual diversity required for modern energy modeling.

This paper proposes a Retrieval-Augmented Generation (RAG) framework that dynamically collects, validates, and synthesizes building-related data from trusted online sources.

The proposed model leverages LLMs with domain-specific retrieval mechanisms to identify, extract, and augment key energy-relevant building characteristics (e.g., construction type, HVAC systems, envelope performance, and occupancy profiles) across different regions.

The resulting system can automatically reproduce datasets for any location in the world, ensuring that urban energy models remain continuously updated, geographically extensible, and data-driven.

The rapid expansion of data centers, driven by artificial intelligence (AI) and cloud computing, has made them one of the fastest-growing contributors to global electricity demand. In the United States, data centers are projected to account for up to 9% of total electricity generation by 2030, underscoring the urgent need for scalable modeling tools to quantify their energy and environmental impacts. The current study introduces a national-scale framework for generating and simulating prototype energy models of U.S. data centers using the Automatic Building Energy Modeling (AutoBEM) platform and the Model America Version 2 (MAv2) dataset. The framework automatically integrates spatial information from the IM3 Data Center Atlas, links each facility to the corresponding MAv2 footprint, and produces OpenStudio/EnergyPlus models that include local TMY3 weather data, internal load profiles for high and low information technology equipment (ITE) densities, and representative HVAC configurations such as CRAC and CRAH systems. A demonstration using the 1001 Texas Data Center validated the framework’s performance, with simulated end-use shares aligning closely with published benchmarks for IT equipment (40–50%) and cooling systems (30–40%). Regional simulations further showed agreement with Electric Power Research Institute (EPRI) inventory estimates within plausible uncertainty ranges: Chicago (278.9 MW vs. 342.0 MW) and New York Tri-State (245.8 MW vs. 178.0 MW). This consistency confirms AutoBEM’s ability to capture regional-scale energy use patterns and represents a foundation for developing a digital twin of the national data center fleet to support forecasting, scenario analysis, and climate impact assessment.

The growing power density in AI- and cloud-driven data centers heightens cooling challenges, highlighting the need for reliable system models. Prototype building models offer standardized and reproducible frameworks, yet existing data center prototypes mainly emphasize long-term energy assessment, lacking the model fidelity and standardized interfaces required for developing and testing advanced control strategies.

To enable control-oriented studies, we developed the first high-fidelity prototype building model for an air-cooled data center based on co-simulation between Modelica and EnergyPlus. Compared with existing models, the proposed prototype offers these key advantages: (1) It provides a more faithful representation of the HVAC system serving the data center, with features such as underfloor air distribution (UFAD), and detailed patterns of IT equipment operation. (2) It establishes a standardized co-simulation interface for integrating and testing control strategies. These improvements enable more accurate dynamic simulations and facilitate the development and evaluation of advanced control algorithms under realistic operating conditions.

The results of the enhanced prototype demonstrate physically consistent interactions between dynamic IT heat generation and HVAC thermal responses under different control conditions. With the unified control interface, the model provides a practical yet detailed test case for evaluating data center performance.

Computer room air conditioning (CRAC) systems are commonly used in small to medium-sized data centers due to their straightforward installation and control. A

key operational challenge throughout a CRAC system’s life cycle is the degradation of its coefficient of performance (COP). Conventional data center energy models often assume new or homogeneous equipment operating conditions. However, in data centers with multi-unit CRAC configurations, factors such as differing equipment ages, varied maintenance schedules, and individual unit faults lead to heterogeneous performance degradation. In practice, this heterogeneous nature leads to a significant yet often overlooked opportunity for energy optimization. This paper presents a case study of a four-unit CRAC system exhibiting degradation to

demonstrate an optimal staging control strategy that explicitly accounts for heterogeneous COP conditions. A dynamic testbed model was developed in Modelica, incorporating a rule-based state graph control that prioritizes the use of the most efficient CRAC units. The proposed COP degradation-aware control was benchmarked against a traditional uniform part load ratio (PLR) staging approach, showing a 4.74% reduction in total cooling energy consumption. These results highlight the importance of considering real-world degradation effects when designing advanced staging controls to enhance the energy efficiency of aging data center facilities.

System‑level simulation helps data center designers de‑risk thermal and control system decisions that are costly or impractical to trial‑and‑error in live facilities. This lightning talk presents a minimal reference topology—IT rack → CDU → chilled‑water loop → chiller/condenser → air‑handling equipment—built using the Modelica Buildings Library and executed in Modelon Impact. Using fast what‑if studies, we compare air‑ vs. liquid‑cooled configurations, assess cooling system control strategies, and examine setpoint adjustments to reveal energy and resiliency trade‑offs across climate conditions and load profiles. We also show how to incorporate vendor models or higher‑fidelity data as requirements grow. Attendees leave with a repeatable workflow and three practical what‑if examples.