Performance-based drift detection for active machine learning model adaption: A comparative analysis across 35 HVAC devices
K Derzsi, Florian Stinner, L Patrun, Jonas Klingebiel, Dirk Müller · Journal of Physics Conference Series · 2025
Abstract Building operations contribute significantly to global CO2 emissions, making optimal energy system control crucial for climate change mitigation. While machine learning models can predict building system behavior to enable advanced control strategies, model performance is prone to deterioration due to concept drift in real-world data streams. This study investigates active model adaptation with concept drift detection across 35 HVAC devices in 14 German non-residential buildings on two years of monitoring data. We compared static models against various combinations of machine learning algorithms and performancebased drift detection methods. Results revealed model adaptation effectiveness correlates with baseline performance characteristics. Improvements were observed for devices exhibiting wider model performance distributions, while negative effects were discovered for devices with narrower ranges. Notably, drift detection timing proved more critical than retraining frequency. Active model adaptation achieved improvements of 6.44 to 35.58 % across device types, with significant variations based on machine learning algorithm and concept drift detection method combinations.