Researchers at Virginia Polytechnic Institute and State University are pioneering a new approach to energy management in smart buildings by integrating artificial intelligence, specifically utilizing Large Language Models (LLMs). This innovative framework for a Building Energy Management System (BEMS) aims to enhance energy efficiency by responding intelligently to user requests and interpreting building data in real time. Tianzhi He and Farrokh Jazizadeh, along with their team, have developed a prototype that moves beyond mere automation, creating a closed-loop system that learns from energy data to improve building operations.
The research leverages advanced LLMs, including ChatGPT and GPT-4, to manage key building systems such as HVAC, lighting, and appliances. This allows occupants to interact with their environments using natural language, tailoring automated routines to individual preferences. The findings indicate a shift from traditional energy management practices, with the team employing metadata schemas like Brick and Bot to provide structured data about building components and relationships. Vector databases, such as Milvus and FAISS, are utilized for efficient data storage and retrieval, while Retrieval-Augmented Generation (RAG) enhances the accuracy of LLM reasoning by merging it with external knowledge sources.
In addition to real-time management, the research delves into energy forecasting and optimization, enabling time series analysis and consumption prediction through LLMs. However, privacy and security concerns remain paramount as these models access sensitive building data. The exploration of multi-agent systems, wherein several LLM-powered agents collaborate on complex tasks, is also underway. To assess the viability of this technology, the researchers developed benchmarks like ElecBench, which evaluate LLM performance in building applications.
The framework comprises three modules: perception, central control, and action, forming a cohesive feedback loop that captures, analyzes, and responds to energy data and user requests. This structure differentiates it from traditional dashboard interfaces by embedding autonomous data analytics capabilities within LLMs. The prototype was rigorously tested against 120 user queries across four real-world residential energy datasets, focusing on metrics such as latency, functionality, capability, accuracy, and cost-effectiveness.
Results indicate high performance, with the system achieving 86% accuracy in device control, 97% in memory-related tasks, and 77% in energy analysis. Scheduling and automation tasks yielded 74% accuracy, demonstrating the system’s proactive energy management capabilities. However, challenges remain in complex cost estimation tasks, which resulted in a lower accuracy of 49%. The research team has identified this as a critical area for further development.
Statistical analysis through ANOVA tests confirmed the framework’s adaptability across diverse residential energy profiles, showcasing its potential for generalization in varying settings. The authors emphasized that while the system excels in user interaction and energy management, there is a trade-off between response accuracy and computational efficiency, warranting further exploration. Future work is expected to focus on refining cost prediction precision and optimizing system performance, ultimately enhancing the efficacy of LLM-driven BEMS.
This breakthrough presents a significant advancement in the integration of AI technology in energy management, bridging the gap between human-centered design and operational efficiency in smart buildings. As the landscape of energy management continues to evolve, the potential for LLMs to revolutionize the way buildings operate and interact with occupants is becoming increasingly apparent.
👉 More information
🗞Context-aware LLM-based AI Agents for Human-centered Energy Management Systems in Smart Buildings
🧠 ArXiv: https://arxiv.org/abs/2512.25055
Canva Hires Kshitiz Garg as Audio and Video Lead to Enhance Generative AI Research
Disney+ Launches Vertical Video Feature for AI Native Generation, Integrates OpenAI Collaboration
Zhipu AI Debuts in Hong Kong, Surging 12.4% to $7.4B Market Cap as First LLM Public Company
Nano Banana Prompts Cuts AI Image Generation Costs by 75% with Efficient Techniques
CraftStory Launches Image-to-Video Model for 5-Minute AI-Generated Videos
