Modelling indoor household PM2.5 using positive matrix factorization and machine learning algorithms

Abstract

People tend to spend most of their time indoors, yet the concentration and composition of indoor fine particulate matter (PM2.5) remain poorly understood in the Maltese Islands, with existing receptor modelling studies focusing solely on ambient air. The first objective of this study was to carry out long-term indoor air sampling of PM2.5 followed by chemical characterisation, in order to identify and quantify the main natural and anthropogenic sources of indoor PM2.5 at an urban background site in Malta using Positive Matrix Factorisation (PMF). The second objective was to explore the use of Machine Learning (ML) algorithms to model and predict indoor PM2.5 concentrations in several households in Malta and Gozo. PMF was used to identify and quantify the major sources of indoor PM2.5. Quartz and PTFE filters were collected and analysed gravimetrically and chemically using ICP-MS, IC, and an OC-EC aerosol analyser to determine concentrations of PM2.5, 18 elements, 5 ions, organic carbon (OC), and elemental carbon (EC). Eight contributing factors were identified, seven outdoor sources and one indoor source, contributing 68% and 26% to indoor PM2.5, respectively. Cooking and e-cigarette use were the main contributors to the indoor factor. Uniquely for Malta, a fireworks factor was isolated indoors, responsible for most of the measured Sb and Ba, raising concerns due to the toxicity of these elements. An RF-SHAP model was integrated with the indoor PMF model to investigate the influence of key drivers on indoor PM2.5 concentrations. An outdoor PMF analysis was also conducted, and a corresponding RF-SHAP model (CV RMSE: 2.79 µg m−3; R²: 0.80) was used to refine the outdoor source contributions. Transboundary contributions (Saharan dust and ammonium sulfate) were higher outdoors (58%) than indoors (33%) due to reduced infiltration when windows are closed. Local anthropogenic sources (Industrial, Fireworks, Traffic, Shipping) contributed more to outdoor PM2.5 (33%) than indoor (25%), with increased indoor infiltration during warmer months coinciding with peak fireworks activity. For the second objective, continuous PM measurements were taken using aerosol spectrometers at seven non-smoking residences. RF and XGBoost models were developed to predict indoor PM2.5 at six-hourly intervals. At sites with low indoor PM generation, predictions were mainly influenced by outdoor PM1 levels. At a site with high indoor emissions, indoor relative humidity was a key predictor, especially during cooking. The RF model performed best overall (RMSE: 30.65 µg m−3; IOA: 0.66).