Surrogate models to compute optimal air quality planning policies at a regional scale

Secondary pollutants (such as PM10) derives from complex non-linear reactions involving precursor emissions, namely VOC, NOx, NH3, primary PM and SO2. Due to difficulty to cope with this complexity, Decision Support Systems (DSSs) are essential tools to help Environmental Authorities to plan air quality policies that fulfill EU Directive 2008/50 requirements in a cost-efficient way. To implement these DSSs the common approach is to describe the air quality indices using linear models, derived through model reduction techniques starting from deterministic Chemical Transport Model simulations. This linear approach limits the applicability of these surrogate models, and while these may work properly at coarse spatial resolutions (continental/national), where average values over large areas are of interest, they often prove inadequate at sub national scales, where the impact of non linearities on air quality are usually higher. The objective of this work is to identify air quality models able to properly describe the relation between emissions and air quality indices, at a sub national scale. In this context, artificial neural networks, identified processing long-term simulation output of a 3D deterministic multi-phase modelling system, are used to describe the non-linear relations between the control variables (precursor emissions reduction) and a pollution index. These models can then be used with a reasonable computing effort to solve a multi-objective (air quality and emission reduction costs) optimization problem, that requires thousands of model runs and thus would be unfeasible using the original process-based model. A case study of Northern Italy is presented.