The overall purpose of the AQUISS (Air Quality Information Services) application is to make improved Air Quality Forecasts available at the level of cities and urban agglomerations.

There are two comprehensive sources of Air Quality (AQ) Data in Europe. First, the Member States operate large and complex monitoring networks that measure AQ. Since decades, concentrations of pollutants such as SO2, NO2, CO, PM10, PM2.5 and Ozone have been automatically monitored by governmental agencies throughout the EU. These data are accumulated in national AQ databases and through reporting obligations of the Member States are further integrated at European level at European Environment Agency (EEA) in the so-called AirBase 1. This data are high quality point data, i.e. it shows data at the location of the monitoring stations only. EEA offers near real-time access to the data free of charge without any limitation, mot excluding commercial use 2.

Second, EU-wide AQ modelling services have been made available by the research community, initially stemming from the MACC project (now part of the COPERNICUS service CAMS 3). Due to the complexity of the dispersion modelling process and the required computational power these data are provided as wide-scale grid data (grid sizes vary between 50 * 50 km and 15*15 km). CAMS is also free of charge without limitations, including commercial use.

Although both data sets are freely available through the legal framework of INSPIRE 4, COPERNICUS and EEA data policy 5, they do not deliver what is needed in cities and urban agglomerations, as important elements of the value chain are missing and key market requirements (particularly spatial resolution) are not met. AQUISS aims at overcoming this situation.

Forecast methodology

At the core of the product is a neural network (NN) which forecasts the three trace gases O3, PM10 and NO2 at the sites of several AQ monitoring stations inside and around an urban agglomeration, based on the CAMS 4-day forecast service for these gases and the weather forecast stemming from the Global Forecast System (GFS) 10 . This NN (to be exact: one NN per trace gas per AQ monitoring station) is trained once with

  • historical CAMS data,
  • historical weather data and
  • historical AQ station monitoring data,

and will typically be updated once per year using newly available data.

Training of neuronal network (NN)


For 2016, for instance, one NN will be trained for each pollutant with data from (typically) 5 years (2011 to 2016) and will then be used for the forecasting period of the year 2016. A calendar serves as additional input to account for anthropogenic effects such as variations due to weekends, holiday, bridge days and special days like local festivities.

Once trained, these NN’s compute the next-day forecast at the AQ stations for each trace gas once per day, based on the CAMS AQ forecast and the GFS weather forecast for the next day.

The result of this step are AQ forecast time series at the location of each station in the city for each of the three pollutants O3, PM10 and NO2 for the following day. These forecasts are point data and do only represent the forecasts at the sites where stations are located.


1 www.eea.europa.eu/themes/air/air-quality/map/airbase
2 www.eea.europa.eu/legal/eea-data-policy#toc-6
3 Copernicus Atmospheric Monitoring Service, see atmosphere.copernicus.eu
4 inspire.ec.europa.eu
5 www.eea.europa.eu/legal/eea-data-policy