Speaker
Description
In this talk, we will demonstrate the performance of a neural network approach in predicting the daily average concentration of NO2, PM10 and PM2.5 in different countries. The neural network is trained with archived observations where meteorological parameters and chemical transport predictions serve as inputs. The archived and forecasted chemical transport predictions are obtained through the Copernicus Atmospheric Monitoring System (CAMS) while the high resolution meteo data is delivered by the local meteo authorities. To minimize the number of input features we first identified the most relevant meteorological parameters that drive the pollution concentration and only then perform the training.
This approach can overcome the common underestimation associated with chemical transport models and is computationally more efficient. However, for each country a tailored training phase is required where we select the most relevant input parameters. The forecasted pollution concentration at the different stations is then used to build regional air quality forecast maps by means of geostatistical kriging interpolation with the RIO model. Finally, we will show applications of operational deployments in Croatia, Delhi (India) and several Chinese regions and cities.
