Near-term projection of anthropogenic emission trends using neural networks

Abstract

This study analyses for the first time the global time series 1970-2008 of the Emissions Database for Global Atmospheric Research (EDGAR) for 10 chemical species and more than 3000 subsectors with neural networks, looking for non-linear behaviours that several species bear in common. The application of the different neural network types, suggests that General Regression Neural Networks (GRNN) are more adequate to train a typical Gaussian trend with a very low error. As such, GRNN are very suitable to complete the database for some missing data points within the time-series of the database, but they are not so good for a projection outside the database time period. Instead Multi Layers Perceptron (MLP) is very adequate for projecting for a subsequent year to the database time period, even though MLP is characterised by a slightly higher absolute mean error than the GRNN. By means of the Principal Component Analysis (PCA) we identified which chemical substances are driven similarly by the activity data over the almost 40 years time period. We observed that in all the geographic aggregations the emission trends of CO2, SO2 and NOX can be grouped into one cluster, and the emission trends of CH4 and the particulates in another cluster. The best interval time for the prediction proved to be eleven years and projections seemed to be reliable for three consecutive years following the last year of the database time-series.