[Paper] Context-aware LLM-based AI Agents for Human-centered Energy Management Systems in Smart Buildings

Source: arXiv - 2512.25055v1

Overview

A new research effort proposes turning large language models (LLMs) into the “brain” of building‑energy‑management systems (BEMS). By wiring an LLM into a three‑layer architecture—sensing, central control, and actuation—the authors demonstrate how natural‑language queries can drive real‑time, context‑aware decisions about heating, cooling, lighting, and appliance scheduling in smart homes. The prototype is evaluated on realistic residential datasets, showing that LLM‑powered agents can already answer many energy‑management questions with high accuracy while still leaving room for improvement on cost‑prediction tasks.

Key Contributions

• Context‑aware BEMS framework that couples LLM reasoning with sensor streams, a central decision engine, and actuator/user‑interface modules.
• Closed‑loop feedback design enabling the agent to continuously refine its actions based on real‑time energy data and user feedback.
• Prototype implementation tested on four publicly available residential energy datasets and 120 natural‑language queries.
• Comprehensive benchmark suite covering latency, functional breadth, accuracy, and cost‑effectiveness, plus statistical validation (ANOVA) of generalizability.
• Empirical performance numbers: 86 % accuracy for device‑control commands, 97 % for memory‑related queries, 74 % for scheduling/automation, 77 % for energy‑analysis, and 49 % for cost‑estimation.
• Guidelines for trade‑offs between computational load (LLM inference cost) and response quality, laying groundwork for standardized evaluation of LLM‑based BEMS agents.

Methodology

1. Perception Layer – Sensors (smart meters, temperature/humidity probes, occupancy detectors) stream raw measurements into a time‑series store.
2. Central Control (LLM “brain”) – A pre‑trained LLM (e.g., GPT‑4‑style) is fine‑tuned with domain‑specific prompts and a lightweight retrieval system that pulls the latest sensor snapshots. The model performs three core tasks:
   • Interpretation: Translate a user’s natural‑language request (“How much did the HVAC cost last month?”) into a structured query.
   • Reasoning: Run on‑device analytics (e.g., regression, rule‑based heuristics) or call external APIs, then synthesize a textual answer.
   • Planning: Generate actionable commands for actuators (e.g., “Turn off living‑room lights at 10 pm”).
3. Action Layer – A middleware translates the LLM’s output into MQTT/REST calls that drive smart plugs, thermostats, or send feedback to the user interface.
4. Evaluation – The authors scripted 120 diverse queries covering control, memory, scheduling, analysis, and cost estimation. Each query was run on the prototype, and metrics such as response latency, functional correctness, and answer accuracy were recorded. ANOVA tests confirmed that performance differences across datasets were not statistically significant, indicating good generalizability.

Results & Findings

Latency stayed under 1.2 s for most queries, making the interaction feel conversational. The cost‑effectiveness metric (energy consumed by the compute hardware vs. savings suggested by the agent) was positive for control and analysis tasks but marginal for cost‑estimation due to the higher inference cost of the LLM.

Practical Implications

• Developer‑ready API pattern – The three‑module design maps cleanly onto microservice architectures (sensor ingestion → LLM inference service → actuator gateway), enabling rapid prototyping in existing smart‑building platforms.
• Natural‑language interface – Facility managers and occupants can ask “Why is my electricity bill high this month?” and receive actionable diagnostics without learning a new UI.
• Plug‑and‑play integration – Because the LLM only needs structured sensor snapshots, legacy BEMS hardware can be retrofitted with a thin software layer rather than a full hardware overhaul.
• Energy‑saving automation – Automated scheduling (e.g., pre‑cooling rooms based on occupancy forecasts) can be deployed with minimal rule‑authoring, reducing the engineering effort for custom control logic.
• Cost‑prediction research direction – The identified weakness in financial forecasting points to a niche for hybrid models that combine LLM reasoning with domain‑specific econometric engines.

For developers, the paper suggests a practical roadmap: start with a lightweight retrieval‑augmented LLM, expose a RESTful “ask‑BEMS” endpoint, and iteratively add actuator adapters. Open‑source toolkits like LangChain or LlamaIndex can handle the prompt‑engineering and retrieval layers, while MQTT brokers manage the actuation side.

Limitations & Future Work

• Inference overhead – Running a large LLM on‑premises can be costly; edge‑optimized models or model‑distillation techniques are needed for large‑scale deployments.
• Cost‑estimation accuracy – The current prototype lacks integration with utility tariff APIs and detailed appliance‑level metering, limiting financial predictions.
• Security & privacy – Directly exposing sensor data to an LLM raises concerns about data leakage; robust sandboxing and encryption are required.
• Scalability to commercial buildings – The study focused on residential datasets; larger, multi‑zone commercial environments will introduce additional complexity (HVAC zoning, demand‑response programs).
• User study – Real‑world usability testing with occupants and facility managers was not performed; future work should assess acceptance, trust, and behavioral impact.

Overall, the research opens a promising path toward conversational, context‑aware energy management, while highlighting the engineering challenges that must be tackled before widespread industry adoption.

Authors

• Tianzhi He
• Farrokh Jazizadeh

Paper Information

• arXiv ID: 2512.25055v1
• Categories: cs.AI, cs.HC
• Published: December 31, 2025
• PDF: Download PDF