A data-centric approach to terminal unit’s fault categorization and optimal positioning in building HVAC systems using ensemble learning

Abstract This paper focuses on the fault detection and diagnosis of terminal units (TUs) in a building located in London, utilizing real operational historical data to assess their performance and optimal placement across multiple floors. While precise locations of the TUs are unavailable, our method analyzes their operational behaviour for one month, applying popular machine learning models to detect and analyze faults effectively. By examining each TU individually and in the aggregate, we identify behavioural patterns that inform decisions regarding their positioning within the building. The dataset comprises over 2 million data points collected from 730 TUs, enabling a comprehensive analysis of their functionality and the impact of suboptimal thermostat placements. Our study employs three machine learning models-traditional multi-class Support Vector Machines and two ensemble methods: Random Forest (RF), and Adaptive Boosting (AdaBoost)-to classify TU behaviors into normal operation, heating faults, and cooling faults. Results indicate that RF outperforms the other models with an accuracy of 99.89%, while AdaBoost achieves an accuracy of 85% and SVM shows 47% accuracy. The findings underscore the potential of a data-driven approach to inform retrofitting decisions and enhance the reliability of HVAC systems. This research contributes valuable knowledge toward optimizing TU placement, ultimately leading to improved energy efficiency and indoor environmental quality.