Evaluating the Scalability of a PM_2.5 Indoor Concentration Prediction Model Using Transfer Learning

School of Architecture and Building Science, Chung-Ang University, Seoul 06974, South Korea

Abstract

Indoor air quality (IAQ) directly impacts occupant health and productivity, with particulate matter (PM2.5) identified as a major contributor to respiratory and cardiovascular diseases. This study aims to improve the efficiency of IAQ management by analyzing the performance variability and adaptability of an indoor PM2.5 concentration prediction model, originally trained at an established test site (Site A), when applied to a different environment (Site B). Additionally, a transfer-learning based self-learning framework is proposed to minimize user intervention. Using data from Site A collected between January and April 2025, including indoor and outdoor PM2.5 concentrations, indoor humidity, air handling unit (AHU) airflow rate, damper opening ratio, occupant count, and air purifier operating status, a deep neural network (DNN) model was developed to predict indoor PM2.5 concentrations 10 minutes in advance. The model was deployed for seven days across four AHU zones at Site B, utilizing Transfer-learning and self-learning techniques. The initial Site A model achieved a mean absolute error (MAE) of 0.5 μg/m³, a coefficient of variation of the root mean square error (CvRMSE) of 10.21%, and a coefficient of determination (R²) of 0.98. At Site B, prediction accuracy temporarily declined due to variations in outdoor damper operation but improved as the model adapted through continuous retraining. Results demonstrated high accuracy in certain zones, with performance differences influenced by occupant density and activity levels. This research establishes a foundation for a real-time, adaptive PM2.5 prediction system applicable to diverse indoor environments. Future work will assess scalability across multiple sites to enhance model generalization and support the development of fully automated IAQ management systems.