Forecasting particulate matter concentration in Shanghai using a small-scale long-term dataset

Abstract This study applies machine learning techniques, specifically stacked generalization, to develop a 1-day-ahead air pollution forecasting model for Shanghai, one of the largest metropolitan areas in the world, taking advantage of the essential information on air quality that can be inferred from a small but long-term dataset of meteorological and pollutant concentration variables ( $$\text{PM}_{10}$$ PM 10 and $$\text{PM}_{2.5}$$ PM 2.5 ), consisting of only 4240 samples. We conducted a comprehensive analysis of daily-averaged meteorological observation data, including temperature (T), relative humidity (RH), wind speed (WS) and direction (WD) and precipitation (P) from 74 automatic weather stations over a 12-year period, from 2011 to 2021, and satellite-retrieved aerosol optical depth (AOD) at 550 nm and planetary boundary layer height (PBLH) for the same time interval. In addition, principal component analysis (PCA) was used to identify the most prevalent synoptic weather patterns in the Shanghai region to assess the origin of pollution advection sources. Thanks to the long-term trends information used for model training, combined with machine learning stacking generalization techniques, the developed model improved the prediction results of alternative methods for pollution forecasting with limited observations, obtaining RMSE and $$R^2$$ R 2 values of 11.93 $$\upmu \text{g}\, \text{m}^{-3}$$ μ g m - 3 and 0.72, respectively. Moreover, it was able to forecast most of the pollution peaks, such as those in January 2019 and November 2020, showing itself as a useful tool for policy making and alerting for health risks. The results of this study highlight the need for cohesive strategies that tackle both air quality and climate change to promote sustainable urban growth and environmental robustness.