Application of Discrete Wavelet Transform and Tree-Based Ensemble Machine Learning for Modeling of Particulate Matter Concentrations

The study of air pollution is an extremely important and urgent problem to be solved on a global and local scale. In this field, huge arrays of measurement data are accumulating, for the analysis of which various approaches based on mathematical, statistical, and machine learning (ML) methods are developed. In this paper, we investigate the application of different discrete wavelet transforms (DWT) families, coupled with state-of-the-art ML algorithms to predict concentrations of particulate matter PM10. Average daily data for this pollutant and several meteorological time series for a period of 630 days were used. A hybrid type models with wavelet decomposition of the initial time series and the application of predictive ensembles (Arcing, Arc-x4) were obtained. All models are cross-validated. The models are applied for short-term pollution forecasts.