One year of 1 h resolution ozone data allows us to produce detailed spectra from the two groups of models and the measured concentrations. In Fig. 1, the individual power spectra of ozone (plotted against the period in days for easier interpretation) from global and regional models are compared with the spectrum of the measured ozone. The time series of the rural monitoring stations have been averaged prior to producing the spectra. In almost all subsequent results, the measured time series should be interpreted as ensemble averages of all available rural monitoring stations with 1h temporal resolution. The analysis was not performed with spatially aggregated time series only in Figs. 7, 9 and 11, while a subset of the annual hourly time series was used in Fig. 8 (June–August).

Since ozone is a scalar quantity, its spectrum grows monotonically in log–log scale as expected (e.g. Galmarini and Thunis, 2000), showing a distinct peak around a period of 24 h, corresponding to the daily boundary layer evolution and photochemical production of ozone. This peak is captured well by the two groups of models. The global set tends to slightly underestimate the energy associated with this period, with only a single model that overestimates it. The regional-scale models are evenly distributed around the spectrum of the measured time series. The two groups behave remarkably similarly at scales smaller than the daily peak. The majority of the models overestimate the energy but capture the slope of the measured spectrum. As expected, the spectra of the global models are more scattered but yet very well behaved. A weak second peak is visible between 30 and 50 days, which could be easily attributed to the synoptic variability. Solazzo and Galmarini (2015) demonstrated that it could indeed be connected to meteorology and/or removal by dry deposition. Moving up the period scale, after the daily peak, all regional-scale model spectra are below the observed spectra, a behaviour that continues apart from a few exceptions up until the 60–70 day-period range. Out of seven global models however, only three under- or overestimate the energy in this period range, while the rest matches the observed spectrum. At 70–80 days a new peak appears in the observed time series, corresponding to the seasonal variability. Only one global model captures the observed time series; three models seem to anticipate it at smaller periods, and even in the regional-scale group there are a variety of behaviours including a monotonic increase of the energy throughout this period range. Beyond the 100-day period the ozone energy spectrum grows monotonically. The global model group matches the power line in trend and value very closely, whereas the regional scale group shows a more erratic behaviour.

This first test is important to assess the fundamental differences between the two sets of models with respect to the characteristics of the signal, the periodicities present in the latter and the ability to reproduce the power or the variance of the measured signal at the various frequencies (periods). In addition, it can give us an idea of the level of complementarity that exists between the two groups of models in the representation of the measured power spectrum. As clearly evident from Fig. 1, both groups of models show an internal coherence in the representation of the power spectra. A remarkable result is the capacity of global models to represent the high-frequency part of the ozone spectrum with an accuracy that is comparable with regional models. We can expect a complementarity in the behaviour of the two groups in the large-scale energy range, which should be regulating the long-range transport and background values. The global models have a better representation of that portion of the spectrum than the regional one.

A characterization of model performance for the individual members of the two groups beyond the information provided in Solazzo et al. (2017) and Solazzo Galmarini (2015) is also appropriate at this stage.

The Taylor diagrams presented in Fig. 2 provide an overview of the individual model performance across the year of reference. All model results underwent un-biasing (subtract the annual mean bias from the predicted hourly values, which produces a shift of the annual time series up or down by mean bias). We notice that the global models show a more scattered behaviour compared to the regional-scale models, with performance distributed across a wider range of standard deviation values. Among the global-scale models we find a clear outlier (model 5), whereas the rest tend to group in a rather narrow range of standard deviation values and correlations. Among the regional-scale models we can also identify an outlier, specifically model 9. The average root mean square error (RMSE) values over all stations ranges from 22.4 to 25.9 µg m-3 for the global models and 21 to 24.7 µg m-3 for the regional models and are thus comparable. Global models overestimate the observed standard deviation, while regional-scale models with the exception of model 9 are evenly distributed across the observed values. The correlation coefficient is comparable for the two groups of models.

Figure 3 presents two classical skill scores for categorical events also applied by Kioutsioukis et al. (2016), namely the probability of detection (POD) and false-alarm rate (FAR). The former represents the proportion of occurrences (e.g. events exceeding a threshold value) that were correctly identified, whereas FAR is the proportion of non-occurrences that were incorrectly identified. In other words they measure true and false positives. In this case the scores are calculated on the basis of the individual model performances at each station. POD and FAR plots are presented as probabilities above breakdowns for different threshold values, where the abundance of the observed data per concentration range is also given as a histogram. A binned analysis of the RMSE (Fig. S2) demonstrates that global models achieve lower RMSE at concentrations above 100 µg m-3; the opposite is true for concentrations below this threshold. This partially explains the facts of Fig. 3.

At the same time the global models also have a higher percentage of false positives as can be gleaned from the FAR index. This analysis is important to establish the capacity of the models to simulate extreme values.

Using the methodology proposed by Galmarini et al. (2013), in Fig. 4 we present the decomposition of the model errors according to specific timescales. In this figure, the individual model errors are shown as decomposed in the diurnal (< 6 h), inter-diurnal (6 h–1 day), synoptic (1–10 days) and long-term (> 10 days) timescales and the residual. The decomposition is performed using a Kolmogorov–Zurbenko filter (Rao et al., 1997) applied to the mean squared error (MSE) calculated from each model and the observed ozone time series. Such analysis can be very revealing as it identifies the scale and therefore the processes that are mainly responsible for the deviation of the model results from the measurements as well as possible persistence of errors at specific scales.

The figure reveals that most of the error is contained in the long-term and diurnal timescales. For regional-scale models, this is in agreement with the findings of Galmarini et al. (2013), Solazzo and Galmarini (2015) and Solazzo et al. (2017). The same behaviour is also found in the group of global models. What is remarkable is the similarity of the error values at the diurnal timescale across the two groups. This suggests that the difference in spatial resolution between the two sets of models does not seem to influence the error at the scale at which atmospheric boundary layer dynamics and daily emissions of ozone precursors are the dominant processes. Apart from a few exceptions (models 13 and 17 in the regional-scale group and models 1 and 5 in the global-scale group), all other models have very comparable errors at that scale. A comparable error across the two groups is found at the synoptic scale, although this is less surprising because this scale is explicitly resolved by the models in both groups and strongly depends on the quality of the meteorological driver used. Since both global and regional models employ assimilation of meteorological observations, they are able to represent the synoptic scale comparably and are less dependent on parameterizations employed. The long-term components have the largest error and also show the most variability across models. Remarkably, the regional-scale models seem to show smaller long-term error values than the global models, although the former show highly variable model-to-model errors. The strong dependence of the long-term error on boundary conditions (specifically lateral boundary conditions for regional-scale models and long-range transport in the case of a global model, upper-stratospheric air intrusions and surface emission of ozone precursors and direct ozone deposition) appears to influence the global-scale group concentrations more than the regional scale, though one should consider that almost all regional-scale models used boundary conditions from the same global model, which nevertheless does not have the smallest long-term error component.

A useful pre-characterization of an ensemble can be obtained by the construction of the Talagrand diagram (Talagrand et al., 1999). It is achieved by binning the range from the minimum to the maximum modelled concentrations with as many bins as the number of ensemble members plus 1. The bins are then filled with observed values accordingly. For example, if an observed value is lower than the lowest model value, it is assigned to the first bin; if it falls between the lowest and second-lowest model value, it is assigned to the second bin; and so on. If it exceeds the highest model value, it is assigned to the last bin. Figure 5 shows the Talagrand diagrams for the global and regional and the regional+global set of models. The figures reveal the tendency of the global model ensemble to be overdispersed as indicated by the accumulation of most of the observed data at the centre of the histogram and relatively few observations falling into the more extreme modelled bins. The regional-scale model ensemble shows a flat diagram, which is an indication of good group performance. A flat Talagrand diagram is an indication of the fact that the group members equally cover (by proportion) all the observed range of values, and the group variability does not show an excess or deficiency in the number of predictions in a specific range of observed values.

The first result obtained for a combined set of model results is shown in the third panel of Fig. 5, which presents the Talagrand diagram for the combination of the two groups of models. Note that the number of bins (x axis) has increased, corresponding to the new total number of models considered plus 1 (i.e. 7 global models plus 13 regional models plus 1). The diagram for the combined group of models qualitatively constitutes an improvement compared to those of the individual group ensembles. The combination of the bell-shaped diagram of the global set with the relatively flat shape of the regional set produces a new distribution within the range of modelled values of the observation, showing a flat region between bins 5 and 18 and an under-prediction region between bins 1 and 5 and bins 19 and 21, which now account for lower and higher values respectively compared to the same bins of the global and regional sets.

Prior to analysing the performance of the hybrid multi-model ensemble (mme_GR), let us concentrate on the individual ensembles (mme_R and mme_G) of the two groups for the sake of having an extra term of comparison beyond the measured concentrations against which to compare mme_GR. In this study, we would also like to build upon the research performed in other multi-model ensembles over the years; rather than calculating only the classical model average or median ensemble (mme), we shall also calculate three ensembles based on the findings of Potempski and Galmarini (2009), Riccio et al. (2012), Solazzo et al. (2012a, b, 2013), Galmarini et al. (2013), and Kioutsioukis and Galmarini (2014). We shall therefore refer to the ensemble made by the optimal subset of models that produce the minimum RMSE as mmeS (Solazzo et al., 2012a, b); the ensemble produced by filtering measurements and all model results using the Kolmogorov–Zurbenko decomposition presented earlier and recombining the four components that best compare with the observed components into a new model set as kzFO (Galmarini et al., 2013); and the optimally weighted combination as mmeW (Potempski and Galmarini, 2009; Kioutsioukis and Galmarini, 2014; Kioutsioukis et al., 2016).

Figure 6 shows the effect of the various ensemble treatments for the two groups of models separately and presented as a Taylor diagram. The correlation has increased and narrowed between 0.90 and 0.95 for both groups. As expected, the best ensemble treatment of the two individual groups is mmeW, which in the case of the global models is comparable to mmeS and in the case of the regional-scale models is farther apart from mmeS. The fact that the optimal partition of the error in terms of accuracy and diversity in an equally weighted sub-ensemble (mmeS) and the analytical optimization of the error in a weighted full-ensemble (mmeW) are comparable for the global models implies that this group better replicates the behaviour of an independent and identically distributed (i.i.d., represented by the square in all panels) ensemble around the true state set (on average). The range of improvement of the RMSE is comparable for the two groups of models.

Of the entire set of ensemble treatments proposed, mmeS is the only one that works with an identified subset of elements. The elements chosen in this context are those that minimize a specific metric (e.g. RMSE). The combination of all possible permutations of a pre-defined subset and for all possible subsets allows us to identify the subgroup of models that performs best (Solazzo et al., 2012a, b). This group is the one that best reduces the redundancies and optimizes the complementarity of the model results (Kioutsioukis and Galmarini, 2014). Other methods have been devised to determine the optimal number of models (Bretherton et al., 1999; Riccio et al., 2012) that are equally effective as the one used here, though they do not allow identifying the members of the subset. Beyond the use of the mmeS for the current analysis, given the diversity in the number of models comprising the two ensembles we have calculated the effective numbers of models (Bretherton et al., 1999) for the regional and global sets in the attempt to verify whether the effective numbers were close for the two sets. Figure 7 shows the Neff obtained for the global set and the regional set. At over two-thirds of the stations, the mmeS used three to four global models and three to five regional models. In other words, roughly half of the global models (3–4 out of 7) produce the best result when constructing the mmeS globally, while in the case of the regional-scale models less than half (3–5 out of 13) of all models are required. Figure 7 also provides the frequency of contribution of the individual models to the mmeS, thus confirming the dominance of three global and four regional models determined with the Neff analysis. What is presented in Fig. 7 is the analysis for the aggregated set of model results at all available monitoring points. We also would like to determine the spatial variability of this result, i.e. to answer the question of whether Neff is uniform throughout the domain or whether there are sub-regions that require more or fewer models to construct mmeS.

In order to have a more objective assessment of the result presented in Fig. 7, we introduce a metric which samples only the diversity of the model results (see Sect. 4.3). Following Pennel and Reichler (2011) and Solazzo et al. (2013) we introduce the metric dm, defined for M models at location i as

dm,i=em,i∗-Rm,mmemmei∗, where mmei=1M∑m=1Mem,i,em,i=modm,i-obsiσobs and the ∗ version of em,i and mmei is obtained by normalizing them with σe and σmmei respectively. Rm,mme is the correlation between the individual and average model results. Therefore only the uncorrelated portion of the individual result is retained in dm as a measure of the diversity, whereas the correlated portion is filtered out. Applying this metric, the model results have been decomposed by means of the Kolmogorov–Zurbenko filter described earlier, and Neff has been calculated across the domain for the most relevant components – long term (> 4 days, LT), synoptic (< 4 days, SY) and diurnal (< 1 day, DU) – according to the definitions presented by Solazzo et al. (2017) and references therein. Figure S3 presents the results for the two groups of models. For the long-term component, Neff results shown in Fig. 7 are largely confirmed with an overall spatial homogeneity of Neff. The global model set appears to require a larger number of models than the average in critical areas like northern Italy, where the resolution would be insufficient to capture the inhomogeneity of the concentration field due to the complex terrain in that region (similarly in the western part of the domain). At the synoptic scale, the regional-scale models require slightly more models on average than the numbers presented in Fig. 7, and in some portions of the domain almost all available models are required. The number of required models increases even further at the diurnal scale. In the case of the global set, the average Neff is the same across these two scales, and more models are required in the Po valley (Italy) at the synoptic scale and western Poland at the diurnal scale.

Given the fact that there is redundancy in the two groups of models and a disparity exists in the overall and effective number of models in the two groups, a strategy has to be devised so that no pre-determined weight is assigned to one of the two groups, thus masking the potential outcome of this study or creating false results. This goal is accomplished by applying the following strategy.

We want to compare three equally populated ensembles of just global, just regional, and mixed global and regional models. We will therefore reduce the ensemble of regional-scale models and include extra models in the ensemble of global models beyond the effective number calculated in Figs. 7 and S3 so that the joint ensemble will not be too small. In order to accomplish this, we select the global models contributing most to the global ensemble beyond those identified by Neff. We begin by assuming that six models comprise a reasonably abundant ensemble (as also indicated by the effective number of regional-scale models), and as such the single-scale ensembles will be based on six members. Taking advantage of the various techniques developed to build an ensemble presented earlier, we define the following sets:

(mme_GR) hybrid ensemble of rank 6 (ensemble of six members) composed of the three best global models and the three best regional models

Among them, the mmeS_GR is the only ensemble product that allows unbalanced contributions from global and regional models.

The comparison of the ensemble performances will be restricted to the months of June–August, when the photochemical production of ozone is at its maximum and the number of exceedances is expected to peak throughout the continent. The results of the comparison of the mme, mmeS and mmeW for the regional (_R), global (_G) and hybrid cases (_GR) are shown in Fig. 8. The elements common to the three panels are as follows:

The hybrid ensemble of rank 6 composed of the three best global models and the three best regional models (mme_GR) when compared to mme_G (best six global models) and mme_R (best six regional models) does not show improved performance; rather its skill is inferior to both mme_G and mme_R.

The partition of global and regional models in mmeS (Fig. 9) shows that the contribution of regional models is more frequent. Specifically, at two-thirds of the stations, the optimum hybrid ensemble of rank 6 consists of one or two global models and five or four regional models respectively. At only 15 % of the stations, mmeS consists of an equal number of global and regional models. The maximum number of global models in the mmeS_GR ensemble is four, achieved at roughly 1 % of the stations. Conversely, at around 10 % of the stations the hybrid ensemble utilized only regional models. The second panel of Fig. 9 also gives the spatial distribution of the number of global models contributing to the hybrid ensemble, clearly indicating a preference for regional models in the northeastern part of the domain. This “spatial” preference is not observed in the JJA hourly time series or the annual daily maximum time series (Fig. S4), both being high-ozone datasets. This is in line with the relatively higher RMSE of the global models at low concentrations (Fig. S2).

In Fig. 10, POD and FAR show a net improvement over the mmeW_G results when the hybrid ensemble is considered, with a minimum in false positives and a maximum in true positives that closely match the mmeW_R results.

The real improvement of the hybrid ensemble with respect to the single-scale model ensembles becomes evident when analysing Fig. 11. The panels in the figure are the collective representation of three of the most important characteristics of an ensemble as proposed by Kioutsioukis and Galmarini (2014), i.e. diversity, accuracy and error. On the x and y axes respectively “diversity” and “accuracy” are presented. The former represents the average square deviation of the single models from the mean of the models, whereas the latter is the square of the average deviation of the individual model results from the observed value. As presented by Krogh and Vedelsby (1995), the difference of the diversity and accuracy defines the quadratic deviation of the ensemble average from the observed value. From the definition it follows that, in order for the ensemble result to be closer to the observed value, one has to find the right trade-off between accuracy and diversity (A–D). A mere increase in diversity does not guarantee a minimization of the ensemble error since it might produce a reduction in the accuracy. What one hopes to obtain is the right combination of models that provides the maximum accuracy and maximum diversity. In the plots of Fig. 11, the optimal condition is achieved when the model results are concentrated in the upper left quadrant of the plot toward the (x=100/(numberofmodels), y=1) point. In the plot, the accuracy parameter is presented as a deviation from the best model performance. The dots represent the estimate of the two parameters at every location where measurements are available. The colour scale is based on the RMSE. The two upper panels give the A–D mapping for the mme_R and mme_G ensembles; the lower two panels give the map for the hybrid ensembles, i.e. mme_GR and mmeS_GR. The difference in nature of the two ensembles is clear from the two panels. Ensemble mme_G is less diverse and more accurate than mme_R (x values: 69 in G and 66 in R; y: 0.75 in G and 0.66 in R). The combination of the two ensembles produces an improvement only in diversity (mme_GR). However, if the models are selected as in mmeS, both accuracy and diversity increase (mmeS_GR). The real advantage of the combination is visible in a slight increase of the diversity as compared to mme_GR and a marked improvement of the accuracy from 0.70 to 0.81. The error decreases from a median value of 17.9 to 15.6 and from an interquartile range of 5.1 to 3.8.

To answer the question of whether the multi-scale ensemble is more skillful, we consider the two optimal single-scale ensembles of rank 6, namely the global (mmeS-G) and the regional (mmeS-R), and the optimal multi-scale ensemble of rank 6 (mmeS-GR) that is constructed from elements of the optimal single-scale ensembles. The multi-scale ensemble achieves an improved diversity by at least 4 % compared to the single-scale ensembles, even reaching 10 % for the daily maximum time series (Table 3). It reflects the independent development of global and regional models. The change in accuracy is generally smaller since the optimal single-scale pool contains models with not very different errors. When the two pools are combined, the mmeS-GR achieves a better RMSE by 13–16 % compared to mmeS-G and by 2–3 % compared to mmeS-R. Further, the mean of the distributions of diversity, accuracy and RMSE from mmeS-GR differs from the corresponding mean of mmeS-G and mmeS-R (they passed the t test at the 5 % significance level). The same holds for the distributions (they passed the Kolmogorov-Smirnov test). Improvements are also revealed for the POD and the FAR, where the mmeS-R does better than mmeS-G, especially at high thresholds. The mmeS-GR generally improves the indices compared to mmeS-R, even though global models are included. Like before, the improvements are seen in all datasets, despite their temporal aggregation.

In Fig. 12 the spectra of the ensembles are presented. For the just-global- and just-regional-scale ensembles, and the rank 6 hybrid ensemble, the spectra of mme, mmeS, mmeW and kFO (Kolmogorov–Zurbenko first order) are shown in the figure. Figure 12 also shows the spectra of the four ensembles (mme_R6, mme_G6, mme_GR6 and mmeS_GR6) for which the six largest contributors from the regional models, the six largest contributors from the global models, and three regional plus three global models were used. From the picture we see that, regardless of the treatment, the ensemble data capture the ozone power spectrum with no notable deviation from the measured spectrum from one another. It is important to note that an ensemble treatment is a purely statistical treatment that does not consider any physics constraints. The deficiencies that were originally present in the individual model spectra are still present in the ensemble results, particularly the large power deficit in the range from 0.8 days to 100 days. The mme_GR spectrum appears to produce a slight improvement toward filling this energy gap, but the change is very small.