Smart Buildings Achieve 86% Energy Savings with AI Agents and LLM Technology

Smart buildings promise significant energy savings, but realising this potential requires systems that understand both building data and occupant needs, a challenge that researchers are now addressing with artificial intelligence. Tianzhi He and Farrokh Jazizadeh, from Virginia Polytechnic Institute and State University, along with their colleagues, present a new framework for a Building Energy Management System (BEMS) powered by Large Language Models (LLMs). This innovative system moves beyond simple automation by interpreting natural language requests and responding intelligently to user needs, creating a closed-loop system that learns from energy data and adjusts building operations accordingly. The team’s prototype demonstrates promising results in areas like device control and scheduling, achieving high accuracy in many tasks, and establishes a crucial benchmark for evaluating the potential of LLMs to create truly human-centered and efficient energy management in smart buildings.

Investigations focus on utilizing LLMs, such as ChatGPT and GPT-4, to directly control building systems including HVAC, lighting, and appliances, enabling interaction through natural language interfaces and the creation of automated routines based on user preferences. Many studies address connecting LLMs to building data sources, employing metadata schemas like Brick and Bot to provide structured knowledge about building components and relationships. Vector databases, including Milvus and FAISS, are used to store and retrieve relevant data for LLM reasoning, and Retrieval-Augmented Generation (RAG) combines LLMs with external knowledge to improve accuracy and context.

Research also extends to energy forecasting, optimization, and data mining, leveraging LLMs for time series analysis and predicting energy consumption. Investigations address privacy and security concerns associated with LLMs accessing sensitive building data, and explore multi-agent systems where multiple LLM-powered agents collaborate on complex tasks. Scientists are developing benchmarks, such as ElecBench, to assess the performance of LLMs in building-related applications. The team developed a three-module framework, perception, central control, and action, creating a closed feedback loop that captures, analyzes, and responds to energy data and user requests. This system surpasses traditional dashboard interfaces by employing the autonomous data analytics capabilities of LLMs to provide insights into energy consumption, predict costs, and schedule devices intelligently. Rigorous evaluation involved experiments using 120 user queries applied to four real-world residential energy datasets.

The study assessed the system’s responsiveness and reliability using metrics including latency, functionality, capability, accuracy, and cost-effectiveness, revealing promising performance across several key areas. The system achieved 86% accuracy in device control, 97% accuracy in memory-related tasks, 77% accuracy in energy analysis, and 74% accuracy in scheduling and automation. While complex cost estimation tasks presented challenges, yielding 49% accuracy, the team identified this as a focal point for future development. ANOVA tests demonstrated the framework’s generalizability across diverse residential settings and data variations, confirming its adaptability to different household energy profiles. The research team constructed a framework comprising perception, central control, and action modules, creating a closed-loop system capable of capturing, analyzing, and responding to energy data and user requests, ultimately managing connected appliances with greater intelligence. Experiments involved evaluating the prototype system with 120 user queries, utilizing four distinct real-world residential energy datasets to assess performance across multiple metrics, including latency, functionality, capability, accuracy, and cost-effectiveness. The results demonstrate promising performance in several key areas, with the system achieving 86% accuracy in device control tasks, effectively managing appliances based on user input.

Furthermore, the team measured 97% accuracy in memory-related tasks, indicating a strong ability to retain and utilize relevant information, and 77% accuracy in energy analysis, successfully interpreting consumption patterns. Scheduling and automation tasks were completed with 74% accuracy, showcasing the system’s capacity for proactive energy management. While the system excelled in these areas, tests revealed an accuracy of 49% in more complex cost estimation tasks, identifying a clear area for future refinement and optimization. ANOVA tests confirmed the generalizability of the framework, demonstrating its adaptability across diverse residential energy profiles, and the breakthrough delivers a benchmarking study that formalizes the assessment of LLM-based BEMS AI agents. The system integrates perception, central control, and action modules, enabling it to interpret energy data and respond to user requests via natural language. By employing the analytical capabilities of these models, the system offers insights into energy use, predicts costs, and schedules devices, addressing shortcomings in conventional energy management approaches. Evaluations using a diverse set of user queries and real-world energy datasets demonstrate promising results, with high accuracy in device control and memory-related tasks, achieving 86% and 97% respectively.

The system also performs well in scheduling and energy analysis, reaching 74% and 77% accuracy, although cost estimation currently presents a greater challenge with 49% accuracy. Statistical analysis confirms the framework’s ability to generalize across different building types, indicating consistent performance in varied residential settings. The authors acknowledge that complex cost estimations require further refinement and highlight a trade-off between response accuracy and computational efficiency as an area for ongoing investigation. Future work will likely focus on improving the precision of cost predictions and optimizing the system’s performance to balance accuracy with resource demands, ultimately enhancing the effectiveness of language-model driven building energy management.