Real-time Multi-model Anomaly Detection for Interactive Smart Building Management

Abstract

The increasing amount of operational data recorded by building automation systems (BAS) has driven growing interest in data-driven models for fault detection, diagnosis, and predictive maintenance. However, most studies rely on public datasets such as ASHRAE RP-1312, which contain engineered faults, raising doubts regarding model robustness when handling noisy faults found in real world data. In this study, the dataset is recorded from two chillers and three variablespeed pumps between September 2025 to June 2026 in the FUSE building at Mason Square in Arlington, VA. Correlation analysis together with other data wrangling methods were used to aggregate raw data into hourly observations and bring down the independent features and response variables from 140 and 25 to 80 and 6, respectively. Time series analysis techniques were used to extract dominant lags and design engineered features for model development. We trained singleoutput Random Forest, Extra Trees, XGBoost, LightGBM, long short-term memory (LSTM), and support vector regression (SVR) models for 3 energy consumption and 3 temperature-related variables, implementing hyperparameter optimization and walk-forward time-series cross-validation to preserve temporal order and prevent data leakage. The best models were then validated using a separate dataset with known faults before being integrated into a structured, intelligent, and proactive dashboard framework for real-time anomaly detection for both energy efficiency and thermodynamic performance monitoring.