Ensemble Forecasts of Air Quality in Eastern China -- Part 2. Evaluation of the MarcoPolo-Panda Prediction System, Version 1.

An operational multi-model forecasting system for air quality has been developed to provide air quality services for urban areas of China. The initial forecasting system included seven state-of-the-art computational models developed and executed in Europe and China (CHIMERE, IFS, EMEP MSC-W, WRF-Chem-MPIM, WRF-Chem-SMS, LOTOS-EUROS and SILAMtest). Several other models joined the prediction system recently, but are not considered in the present analysis. In addition to the individual models, a simple multi-model ensemble was constructed by deriving statistical quantities such as the median and the mean of the predicted concentrations. The prediction system provides daily forecasts and observational data of surface ozone, nitrogen dioxides and particulate matter for the 37 largest urban agglomerations in China (population higher than 3 million in 2010). These individual forecasts as well as the multi-model ensemble predictions for the next 72 hours are displayed as hourly outputs on a publicly accessible web site (www.marcopolo-panda.eu). In this paper, the performance of the predictions system (individual models and the multi-model ensemble) for the first operational year (April 2016 until June 2017) has been analysed through statistical indicators using the surface observational data reported at Chinese national monitoring stations. This evaluation aims to investigate a) the seasonal behavior, b) the geographical distribution and c) diurnal variations of the ensemble and model skills. Statistical indicators show that the ensemble product usually provides the best performance compared to the individual model forecasts. The ensemble product is robust even if occasionally some individual model results are missing. Overall and in spite of some discrepancies, the air quality forecasting system is well suited for the prediction of air pollution events and has the ability to provide alert warning (binary prediction) of air pollution events if bias corrections are applied to improve the ozone predictions. [ABSTRACT FROM AUTHOR]