Prototype of a commissioning-friendly fault detection tool for residential heat pumps using machine learning algorithms

Abstract

Heat pumps are central to residential decarbonization, yet undetected operational faults can reduce system efficiency by up to 25%. Existing fault detection methods rely on extensive labeled data, detailed physical models, or specialist configuration — none of which are available immediately after commissioning. This article presents a lightweight, commissioning-friendly fault detection framework based on short-horizon time-series forecasting using a regularized ARX model. Trained on as few as three fault-free reference days, the framework evaluates daily deviations between predicted and measured flow temperature using three complementary residual metrics combined through a logical OR rule. Evaluated on 12 real-world residential heat pump datasets and one synthetic dataset across air-to-water, hybrid, and ground-source configurations, the framework achieved zero missed faults in eight of twelve systems with a mean false positive rate of 3.9 days per system.