Urban population exposure to tropospheric ozone: A multi-country forecasting of SOMO35 using artificial neural networks

Abstract

Urban population exposure to tropospheric ozone is a serious health concern in Europe countries. Although there are insufficient evidence to derive a level below which ozone has no effect on mortality WHO (World Health Organization) uses SOMO35 (sum of means over 35 ppb) in their health impact assessments. Is this paper, the artificial neural network (ANN) approach was used to forecast SOMO35 at the national level for a set of 24 European countries, mostly EU members. Available ozone precursors' emissions, population and climate data for the period 2003-2013 were used as inputs. Trend analysis had been performed using the linear regression of SOMO35 over time, and it has demonstrated that majority of the studied countries have a decreasing trend of SOMO35 values. The created models have made majority of predictions (≈60%) with satisfactory accuracy (relative error <20%) on testing, while the best performing model had R² = 0.87 and overall relative error of 33.6%. The domain of applicability of the created models was analyzed using slope/mean ratio derivate from the trend analysis, which was successful in distinguishing countries with high from countries with low prediction errors. The overall relative error was reduced to <14%, after the pool of countries was reduced based on the abovementioned criterion.

Keywords: ANN; Europe; Forecasting; GRNN; SOMO35.