𝞥-GPT: Building Physics-Informed LLMs for Sustainable Design and Operation
Research Team
Our Motivation:
The building sector contributes 37% of global carbon emissions, making it a key target for climate change mitigation. Emission reductions in existing buildings can be achieved through optimal control of systems such as Heating, Ventilation, and Air Conditioning (HVAC), which alone accounts for 59% of energy use in residential buildings and 36% in commercial buildings.
However, there is a significant bottleneck in the wide-scale implementation of advanced technologies like Model Predictive Control (MPC) and Reinforcement Learning (RL) in building system control. Despite their potential benefits, these technologies face challenges related to scalability and interpretability, which hinder broader adoption and impact. This limitation restricts stakeholders from making informed decisions that could optimize building operations effectively.
This research aims to develop a scalable framework that leverages Large Language Models (LLMs) enhanced with Retrieval-Augmented Generation (RAG) to incorporate domain-specific knowledge, creating domain-guided digital feedback loops to support energy efficiency improvements in facility operations. By addressing the scalability and interpretability challenges of existing approaches, LLMs offer the potential to make advanced building control more accessible and actionable for facility managers and stakeholders.
Research Contribution
We propose the development of 𝞥-GPT, a novel framework that integrates domain-specific knowledge from building physics and building system fundamentals into LLMs. This innovative approach facilitates interactive analysis and optimization to enhance energy efficiency, specifically focusing on building operations. By establishing digital feedback loops between data, simulation, LLM reasoning, and human interaction, our framework aims to lower the barrier to advanced building analytics and control, eliminating the need for complex digital twin development.
For this year's scope, our objective is to demonstrate the effectiveness of integrating domain-guided LLMs in decision-making processes for HVAC operations. In the long term, we plan to extend this framework to other essential building processes such as energy retrofit planning, construction management, worker safety, and manufacturing.
Problem
Practical Problem
The scalability of advanced HVAC control strategies, such as MPC and RL, faces significant practical challenges despite their proven energy-saving potential. Real-world adoption is limited by building-specific requirements, which demand substantial effort for tailoring models and extensive training time. Additionally, these models need ongoing recalibration to accommodate changing conditions, such as operational schedules and occupancy. The decision-making processes of these strategies are often too complex, making it difficult for operators to interpret, trust, or override control actions. These factors hinder widespread adoption, leaving most buildings reliant on simpler rule-based systems and creating a gap between theoretical benefits and the practical deployment of advanced controls.
Solution
The solution, 𝞥-GPT, consists of four integrated components: a Modeler for creating and calibrating building energy models, a Retriever for identifying historical data and relevant knowledge from the knowledge base in response to current operational conditions or users' questions, a Generator for providing actionable recommendations with rationale, and a Judge for evaluating outputs against building physics principles and operational constraints.
For this year's scope, 𝞥-GPT focuses specifically on HVAC control optimization. It interacts with stakeholders, building systems, and domain knowledge sources to generate and validate control strategies, such as setpoint adjustments and system scheduling. The framework is designed to be extensible to broader building performance applications, including energy retrofits and construction planning, in future phases.
Added Value For The Industry
𝞥-GPT eliminates the primary barriers to advanced HVAC control adoption: the need for extensive modeling and training efforts and the black-box decision-making that facility managers cannot trust or understand. This framework enables the immediate deployment of sophisticated control strategies through natural language interactions with domain-guided LLMs.
By making proven control technologies accessible to the vast majority of buildings that still use basic rule-based systems, this approach directly addresses operational energy costs. 𝞥-GPT generates explainable recommendations alongside control decisions, such as setpoint adjustments, bridging the critical gap between the potential of advanced controls and real-world implementation constraints.
Cooperation Partners
Timeline
| Date | Activity | Outcome |
|---|---|---|
| Q4 2025 | Testbed building selection and HVAC operational data collection | |
| Q1 2026 | Initial HVAC control testing | |
| Q2 2026 | Full framework evaluation | |
| Q3 2026 | System refinement and validation with facility managers |
