The first arc of stations was characterised by arrival times between 15 and 18 hours after the release start; the stations of the second arc had times of arrival equal to or beyond 30 hours after the release start. The transit duration of the tracer cloud was significantly longer at the stations of the second arc compared to those of the first arc. The concentration peak did not decrease passing from the first to the second arc stations.

Only one model did not predict any tracer concentrations in all the 11 stations, but at each station there were models that did not predict any tracer arrival. The stations where most of the models predicted a tracer episode were B5, NL1, D44 (central and South in the first arc) and DK2 and D42 (central-North in the second arc). The station where most of the models failed was H2, placed at the extreme South of the second arc.

For each of the selected stations the following quantities were computed for each model and compared with the corresponding measured quantities:

• time of arrival

• cloud duration

• time of occurrence of maximum

• maximum concentration

• cumulative concentration after 24 and 48 hours

Time of arrival, temporal duration and peak concentration were well predicted by almost all the models in the two stations NL1 (centre of the first arc) and D42 (centre of the second arc). The best four models predicted the time of arrival within 9 hours of that measured for all the 11 stations. Predicted duration of three best-performing models differs from the measured ones between only –6 to +9 hours for almost all the stations.

The following statistical indices were also calculated using the 20 observation-prediction pairs for each of the 11 selected stations:

• FMT, quantifying the overlapping of the predicted and observed time series. The more a predicted time series coincides with the corresponding measured one, the higher the FMT. Five models had very low FMTs at all the 11 stations (all FMTs<15% with many zeros), five models had high FMTs at all 11 stations (almost all the FMTs>15%, with no zeros and more than one FMT>50%). The remaining models performed well either in the Northern part or in the Southern part of the arcs (some with very high values of FMT) but had zero FMT values in the remaining stations.

• NMSE, giving the absolute deviation of predictions with respect to measurements (a good performance gives a low NMSE). Nine well-performing models (which include the five models that performed well for FMT) had the majority of the eleven NMSEs less than 10. In general the lower values are at the central stations of the two arcs, while models had a generally bad performance either at the Southern or at the Northern stations of the two arcs.

• Bias, representing the overprediction (if positive) or the underprediction (if negative) of predicted concentrations. In general there was an under-estimation in the stations at the North and, especially, South of the arcs. For instance, at both the southern stations of the two arcs only one model overestimated the concentration. Three models under-estimated at all eleven stations. The performance of the other models was characterised by a high over-estimation in a few stations (generally belonging to the first arc) but low under-estimations in the majority of the stations.

• Pearson Correlation Coefficient, indicating the extent of agreement in the predicted and observed temporal trends. It ranges from -1 to +1, with best performances characterised by positive and close-to-one values. Nine models had positive close-to-one PCCs in almost all 11 stations selected for the time analysis.

The global analysis considered the observed-predicted concentration pairs at all times in all the stations.

The results of the global analysis were described in terms of FA2, FA5, FOEX, bias, NMSE and Pearson correlation coefficient.

Bias and FOEX are both used to evaluate the overprediction or underpredictions of a model. The difference between these two indices is that FOEX counts the number of over or underpredictions, while bias quantifies the amount of over or underprediction. Most of the models had negative FOEX, indicating that most of the values are underpredicted by the models. At the same time, almost all the models had positive bias, which means that summing all the overpredictions and subtracting all the underpredictions the result is positive, this means that a very few high values are overpredicted by a large amount.

The FA2 and FA5 take into account the percentage of predictions within a factor 2 or 5 respectively of the measurements. Only four models had a FA5 greater than 40%.

Only five models had a value lower than 10 for the global NMSE, which represents the absolute deviation of predictions from measurements. Two of these models also had the FA5 greater than 40% confirming that the agreement is on the majority of points. Five models had NMSEs greater than 100, with two of them showing NMSEs greater than 1000.

Three models had the global PCC greater than or equal to 0.5.

The predictions of models that used analysed meteorology improved in the FA5 (six models had FA5 greater than 40%) but worsened in terms of NMSE; in fact nine models had NMSE greater than 100, including four models with a NMSE greater than 1000. The other indices for predictions with analysed meteorology were almost unvaried with respect to predictions with forecast meteorology.