Prediction of daily PM10 concentration using machine learning

Particulate matter with a diameter less than 10 μm (PM
₁₀
) is one of the main urban air pollutants that have significant impacts on human health and environment. It severely affects visibility as well as local and global radiative balances. Air quality measurements collected over ten years from the air quality station of Zrektouni located in Casablanca (Morocco) are analyzed through the use of three supervised machine learning models: Artificial Neural Networks with back propagation (ANN), Random Forest (RF) and Support Vector machines with Regression (SVR). More precisely, daily average of NO2, CO, SO2, O3, temperature, humidity, 850-hPa Geopotential Heights (G-850), precipitation (RR), pression (P), and wind speed (ff) are used as predictors for the PM
₁₀
daily average. In order to improve the learning model performances, a resampling strategy with fivefold cross validation and grid search are applied using different statistical indicators. In addition, the learning model results are compared to the Multiple Linear Regression (MLR) model, which provides comprehensive insights into the relationship between the predictors and PM
₁₀
. It turns out that the ANN gives the best predictive performance.