Agentic AI Detects Energy Anomalies, Recommends Solutions.

Managing energy consumption in office buildings often involves sifting through noisy alerts from appliance-level monitoring, which facility managers find difficult to act upon. This research introduces a comprehensive, end-to-end agentic AI pipeline designed to streamline this process. The system integrates deep time-series forecasting, specifically a hybrid Singular Spectrum Analysis (SSA) and Long Short-Term Memory (LSTM) model, with a per-appliance LSTM Variational Autoencoder (VAE) for anomaly detection. This setup flags abnormal daily consumption episodes across seven tracked office appliances. The core innovation lies in its three-stage LangChain pipeline, which leverages large language models (LLMs) for intelligent reasoning. A Context Agent dynamically retrieves relevant information, including model reliability, baseline data, and expert knowledge, and conditionally adds more context based on the anomaly's characteristics. A Diagnosis Agent then converts this evidence into a structured JSON diagnosis, which a Report Agent renders into a human-readable narrative with actionable maintenance recommendations. The system also features a reflective memory layer that incorporates operator feedback, continuously improving its recommendations. Benchmarking across 16 scenarios, including various spikes and shutdowns, showed the best LLM backend achieving 90.4/100, with even a local 7B-parameter model successfully passing all scenarios, demonstrating robust performance and efficiency in context retrieval.