A deep learning approach for PM2.5 exceedances forecasting in an urban area

Fine airborne particles with aerodynamic diameter lower than 2.5 μm (PM_2.5) pose a pivotal atmospheric threat. High PM_2.5 values are highly associated with numerous adverse health effects affecting the respiratory and cardiovascular systems. Apart from densely populated urban environments, where anthropogenic emissions are more intense and population density is constantly increasing, suburban and rural are also prone to experience severe PM_2.5 related pollution events. Eventually, PM_2.5 levels across many environments may report elevated concentrations that may lead to exceedances of the Word Health Organization (WHO) regulated thresholds severely affecting humans’ life, well-being, and ecosystems. A precise forecasting technique of the PM_2.5 concentrations could be essential to tackle air pollution control and warn vulnerable citizens.

In this study, a novel deep learning approach (long short-term memory, LSTM) was developed, to forecast the intraday air pollution exceedances across urban and suburban environments in northern Greece (Municipality of Thermi). A three-year dataset was used for the training (two years, 2021-2022) and testing (one year, 2023) of the proposed LSTM model. PM_2.5 measurements were provided by a dense low-cost PM monitoring network with approximately 28 sensors deployed across the greater measuring area.  In the scope of this study ground based PM_2.5 observations from 3 regions, that share a rather similar meteorological profile, the corresponding meteorological variables, acquired from the Copernicus Atmosphere Monitoring Service (CAMS) and operated by ECMWF, and time variables related to local emissions were utilized. These data were integrated into the LSTM-based methodology, to enhance the hourly intraday PM_2.5 concentrations forecasting capabilities. Additionally, the applicability of PM_2.5 forecast concentrations to capture the daily exceedances of air pollution was also assessed.

The proposed model's forecasting accuracy yielded promising outcomes, revealing correlation coefficients ranging between 0.68 and 0.93 among the observed PM_2.5 concentrations and LSTM forecasted data for various time horizons. Longer forecasting intervals though reported lower correlation coefficients values. Finally, regarding PM_2.5 threshold exceedances, the LSTM forecasting system correctly identified over 73% of such events in the examined area. These findings underline the model's capabilities in detecting potential breaches of WHO guideline limits and provide crucial local air quality management insights.