Commissioning oriented fault detection in heat pump operation using machine learning algorithms

Abstract

Heat pumps now play a significant role in Europe’s residential heating sector. As deployment continues to expand, there is an increasing need for scalable and commissioning-friendly fault detection (FD) solutions that operate with minimal instrumentation and limited historical data. Existing machine-learning-based FD approaches rely heavily on long process histories or simulation data and require extensive expert configuration, making them unsuitable for early-stage deployment. This paper proposes a lightweight, data-driven FD framework based on short-horizon time-series forecasting using autoregressive–exogenous (ARX) models. The method requires only three consecutive no-fault days for training and incorporates a daily re-initialization loop to mitigate cumulative prediction drift. Deviations between predicted and measured flow-temperature profiles—selected as the primary diagnostic variable due to its universal availability and strong sensitivity to common fault modes—are evaluated using a composite set of complementary residual metrics, including root mean squared error, maximum absolute error, and correlation-based pattern matching.

The framework was tested on seven real-world datasets, with detailed results presented for two representative heat-pump systems: a ground-source heat pump (GSHP) and an air-to-water system. Across both cases, the ARX model achieved strong forecasting accuracy and reliably identified deviations indicative of abnormal behavior, while requiring minimal configuration effort. These results demonstrate the potential of lightweight forecasting-based FD methods as practical commissioning tools and motivate further work toward hybrid approaches that integrate rule-based and data-driven diagnostics for broader real-world deployment.