Spatio-temporal learning in predicting ambient particulate matter concentration by multi-layer perceptron

Abstract

In this work, a novel spatio-temporal air quality prediction framework is proposed, and its development and efficacy as a predictive tool are described. The framework exploits data from the suite of models of the Copernicus Atmosphere Monitoring Service (CAMS), fed to an artificial neural network for the removal of bias. The method inherently considers spatial and temporal correlations, because it is applied simultaneously to all monitoring stations of a given region, using past observations and past and future forecasts. The methodology is tested on twelve months of CAMS forecasts of daily surface particulate matter (PM10) in 2017 and is verified against observations measured at 413 monitoring stations from the Italian air quality network.The raw data from the CAMS system, although they contain valuable information, show very poor performance, due to a large negative bias that does not allow the correct prediction of critical conditions. The model bias is found to have a strong seasonal dependency, with a large positive bias in winter and a small negative bias during summer months. The correction applied through the neural model allows to correct the original predictions and to practically eliminate the bias, increasing the performance forecast even to four days ahead. It is concluded that neural networks can be used to develop reliable air quality early-warning systems based on a network of automated monitoring stations and real-time ensemble predictions from deterministic models.