A support vector machine model to forecast ground-level PM 2.5 in a highly populated city with a complex terrain

Physical models are essential to describe the behavior of pollutants especially in high latitudes, and they have been regarded as immensely precise. In the tropics, however, these models have lower accuracy due to the absence of a simple theoretical framework to describe tropical dynamics. Hence, the development of predictive nonlinear models with machine learning has increased, as they are able to quantify the different dynamic processes regarding air quality and to obtain accurate predictions in less computational time than their physical counterpart. This study constructs and evaluates a support vector machine (SVM) to forecast ground-level PM2.5 in a populated city with complex topography. The simulations were built for days with red Air Quality Index (AQI), to assess whether the model could represent the behavior of days with high values and data with fast and substantial changes in the PM2.5 tendency. The SVM is trained with an air quality monitoring network using the radial basis function kernel. A spatial interpolation is also conducted to determine and describe the behavior of the AQI in the city of Bogota. This work uses statistical scores (root mean square error (9.302 μg/m3), mean BIAS (1.405 μg/m3), index of agreement (0.732), and correlation coefficient (0.654)) to validate the capability of an SVM model of simulating, with high precision, the concentrations of PM2.5 in a city with complex terrain in the short term and also to demonstrate the potential of the SVM to be used as a forecast model in other tropical cities.