Data driven investigation of faults in HVAC systems with model, cluster and compare (MCC)

The complexity of modern HVAC systems leads to device mis-configuration in about 40% of buildings, wasting upto 40% of the energy consumed. Fault detection methods generate excessive alarms leading to operator alert fatigue, faults left unfixed and energy wastage. Sophisticated fault detection techniques developed in the literature are seldom used in practice. We investigate this gap by applying various fault detection techniques on real data from a 145,000 sqft, five floor building. We first find that none of these algorithms are designed to capture control loop configuration faults. We develop a novel algorithm, Model, Cluster and Compare (MCC) that is able to detect anomalies by automatically modeling and clustering similar entities in an HVAC system, in an unsupervised manner, and comparing them. We implemented MCC to detect faults in Variable Air Volume boxes in our building, and demonstrate that it successfully detects non-obvious configuration faults. We propose a two stage approach, where we design intelligent rules (iRules) based on anomaly exemplars from a mix of data driven algorithms. iRules are successful in capturing a large fraction of faults in our building, with only one false alarm and 78 anomalies detected out of 237 zones. Thus, comparative data mining is useful in filtering the large amount of data generated in modern buildings, but that human in the loop systems are better still.