An integrated data-driven/data assimilation approach for the forecast of PM10 levels in Northern Italy

Abstract

The EU Air Quality Directive 2008/50/EC recommends member states to ensure that timely information about actual and forecasted levels of pollutant concentrations are provided to the public. In order to follow these guidelines, prevent critical episodes and inform the public, environmental authorities need fast and reliable real time alarm systems. In this work, a performance comparison of different data driven model families has been performed using information provided by more than 100 monitoring stations in Northern Italy. The different models include linear (auto-regressive), non-linear (neural network), time variant (lazy learning) methods and their ensemble. Moreover, their inability to perform forecast where no monitoring stations are available is known as one of the main limitations related to this kind of models. To address this issue, an optimal interpolation technique has been introduced to integrate daily PM10 forecasted concentrations with the results of a deterministic chemical transport model, extending the forecast from the monitoring network sites to the whole area of interest. The validation shows very good performances for all stations, with high agreement in both mean value and 95th percentile computed over the whole year, a correlation coefficient usually higher than 0.7 and small values of root mean square error.