PREV'AIR pollution

Introduction >> Verification


Throughout each forecast period ("summer" period or "winter" period), standard statistical tests are conducted with the aim of comparing expected pollutant concentrations with the available measurements and thus to assess the performance of the air quality forecast models used in the PREV'AIR system.

The results of these statistical tests are for the use of specialists. They respond to our aim of achieving transparency regarding the limits of the PREV'AIR system and the performance of the tools used.


The available measurements are compared to concentrations simulated by the models, spatially interpolated to the measurement points for the various compounds modelled.

All the measurements available for a given forecast period (i.e. from 1 April to 30 September for the "summer" period; from 1 October to 31 March for the "winter" period) are used. A distinction is made between rural stations, on the one hand, and stations in suburban areas, on the other hand.

The agreement between the amounts simulated by the models and the observations for each forecast period (from the previous day to the day after the following day) are assessed using cumulative statistical indicators, updated on a daily basis. These statistical indicators are as follows:

  • Normalised bias: The normalised bias - expressed as a percentage - provides information on the tendency of the model to over- (positive bias) or under- (negative bias) estimate observations.
  • Square root of normalised mean square error (NMSE): The NMSE - expressed as a percentage - provides information on the relative difference between the observations and the amounts simulated by the model: the lower the NMSE, the closer the simulated amounts are to the observations.
  • Correlation: The higher the correlation - a number between -1 and 1, with no dimension - the better the model reproduces the temporal variations of the observations.
  • E-X%: This indicator - expressed as a percentage - refers to the percentage of days for which the simulated and observed amounts differ by less than X%.

Furthermore, we also calculate the average observed and simulated amounts and the number of observations used to calculate the statistical indicators.

WARNING!!! The calculated statistical indicators use observation data provided in real time, which are not subject to a complete checking procedure. Since these data may be adjusted later, the calculated indicators may change accordingly. An assessment of the forecasts will be carried out subsequently with the help of approved observations.