Hybrid models of machine-learning and mechanistic models for indoor particulate matter concentration prediction

While indoor PM2.5 concentrations are generally lower than those outdoors, daily exposure indoors can significantly exceed outdoor levels, given that people spend over 90% of their time inside. This study systematically compares traditional methods with machine learning(ML)-based approaches in the context of indoor particulate matter concentration prediction and control research, with the aim of combining their strengths for more effective outcomes. One-year indoor PM concentration data from a non-interference-operating target office were used to develop ML models. Input variables, including background concentrations (outdoor and hallway), office conditions, and weather parameters, were employed with the CNN-LSTM algorithm. The ML models, enhanced with ensemble techniques, achieved a prediction accuracy of R2 = 0.943 on the test set, not considered during model development. Using predictions under step-change events in outdoor PM2.5 levels, the ML model determined the infiltration factor, PM2.5 removal rates, and particle penetration rates, hence mechanistic model coefficients. Collaborative hybrid models, integrating ML and mechanistic approaches, yielded dynamic models (R2 = 0.928) applicable under general conditions, whereas conventional mechanistic models derived from controlled experiments were limited to controlled conditions.