Hybrid AI model for fault detection and energy consumption analysis of air handling unit systems with supervised and unsupervised learning

Abstract

Previous studies have primarily conducted fault detection based on datasets built in ideal laboratory environments. However, these datasets have limitations when applied to real-world air handling unit (AHU) operational data, which often contains missing values and noise and lacks clearly labeled faults. Additionally, there is still a lack of research quantitatively analyzing the impact of faults on building energy consumption. Accordingly, this study utilized real-world building operation data to predict faults in AHU systems and examined their relationship with energy consumption. We developed an optimized fault detection model for data with labeled faults using supervised learning methods, specifically long short-term memory (LSTM) and bidirectional LSTM (Bi-LSTM). For unlabeled fault data, we applied an unsupervised learning model, LSTM Autoencoder, to derive anomaly scores, which were validated using heat transfer efficiency, proportional-integral-derivative (PID) control performance, and CO₂ concentration. Furthermore, we employed Granger causality analysis and the Vector Autoregression (VAR) model to assess the significance of fault occurrences and their impact on energy consumption. The results showed that among the supervised learning models, Bi-LSTM outperformed LSTM in fault detection accuracy. Meanwhile, the unsupervised model effectively predicted performance degradation and energy consumption impacts by leveraging anomaly scores. This study features the potential for real-time optimization of building energy efficiency through the proposed approach, which can be applied to real-time AHU fault detection and predictive maintenance, helping to minimize energy consumption and maximize long-term operational efficiency in buildings.