Ensembled models to improve residential indoor PM2.5 estimation for further Personal Exposure prediction

The present study is aimed to predict the PM2.5 concentration with a spatial and temporal resolution based on the limitation of the monitoring station, IDW and machine learning were applied to estimate daily residential indoor PM2.5 in Tainan. 62 predictor variables from daily indoor and outdoor PM2.5 observations, questionnaires, and Taiwan Environmental Protection Administration (Taiwan EPA) database were obtained from 2018 to 2020. The correlation analysis and boosted tree (BS) were used to identify the important variables. Then, 10-fold cross-validation was used to validate these models. Among these models, the multilayer perceptron (MLP) model has outperformed when R² exceeded 0.999 compared to other models. In addition, the higher fit in the predictive train-test was identified in the ensembled models and MLP model. The results demonstrated the ensembled approaches and deep learning show promise for modeling indoor air pollution and determined I/O ratio is an important factor. These models may apply to identify the indoor exposure assessment for epidemiology studies and serve as a model framework for other countries.