Atmospheric model results
The WRF model outputs of the hindcast simulations run in operational mode, i.e. forced with the IFS-FC products, are shown in Fig. 5 for all four configurations (panels a–d). In particular, the maps of the accumulated rainfall over 3 days (between the 4th and the 7th of November 1994) are shown. In the same figure, in panel e, the rainfall measured by the rain gauge networks operated by ARPA Piemonte and ARPAL is shown, accumulated over the same time period. This observational network was composed, at the time of the event, by roughly 150 rain gauges, with an average inter-distance of 15 km.
Qualitatively, all the model configurations produced a forecast in very good agreement with the observed rainfall field. In fact, they all simulated two areas of intense rainfall located over the western part of the boundary between Piedmont and Liguria (corresponding to the Tanaro catchment area) and in northern Piedmont.
Concerning the area in northern Piedmont, all the simulated fields show the accumulated rainfall peaks in the same locations, suggesting that it was indeed the geometry of the orographic barrier to determine the air ascent and, thus, the precipitation location. In terms of differences produced by the choice of the microphysical parameterization, only the WSM6 scheme (FC-W, panel a) stood out, producing a more intense precipitation field, which can not be ascribed to the fact that WSM6 is a single-moment scheme. In fact, the experiment FC-G, which was also run with a single-moment scheme with the same hydrometeors as WSM6, produced a rainfall field that is very similar to those of the FC-T and FC-Y runs, which are both based (at least partially) on a double-moment approach.
It is worth noticing that with respect to the numerical simulations that were performed in the years following the event, such as in Buzzi et al. (1998) and in Ferretti et al. (2000), the maxima of precipitation that occurred along the maritime Alps were here reproduced in a better way. In fact, both Buzzi et al. (1998) and Ferretti et al. (2000), that used domains with 10- to 30-km grid spacing, simulated these maxima on the southern side of the mountain chain and not on the northern one, as observed. Also Buzzi and Foschini (2000), that used a 4-km grid spacing setup with parameterized convection, were not able to correctly place the rainfall over the northern side of the maritime Alps. The fact that all the configurations of this study were able to model these rainfall maxima on the correct side of the maritime Alps suggests that it is the finer resolution associated with explicit convection that allowed for a correct rainfall spatial distribution. Note that also the good quality of the initial and boundary conditions may have contributed to the improved skills of the forecast simulations. The correct rainfall spatial distribution significantly impacts the hydrological forecast, as discussed in the following section.
To validate the atmospheric modelling experiments in a quantitative way, the Quantitative Precipitation Forecast (QPF) of WRF was compared with the rain gauge Quantitative Precipitation Estimate (QPE) by using the Method for Object-Based Evaluation (Davis et al. 2006a, 2006b, MODE) tool. MODE identifies precipitation structures in both forecast and observed fields according to a selected threshold on the accumulated rainfall field. By calculating a set of geometrical indices of the observed and forecast objects, listed in Table 1, MODE is able to perform a spatial evaluation of the model capability to reproduce the identified observed objects. As indicated in the table, MODE indices measure, for example, the objects’ relative position, their relative orientation, and overlap.
To calculate the MODE indices, both the observed and the simulated rainfall fields have to be on the same grid. Since the rain gauges are distributed with an average inter-distance of roughly 14 km, both the rain gauge QPE and the QPF of innermost domain were regridded on the 13.5-km model grid with a nearest-neighbour approach. The MODE tool was used with a 216-mm threshold, corresponding to a rainfall intensity of 3 mm/h for 72 h. MODE highlighted two main distinct clusters, as previously observed: the first one over the Tanaro basin (cluster 1) and the second one over northern Piedmont (cluster 2). As an example, Fig. 6 shows the main clusters identified by MODE in the observations and in the FC-W experiment.
All the spatial indices calculated by MODE for the two clusters are summarised in Table 2. Concerning the centroid distance for cluster 1, the best performing run is FC-Y with a mismatch in the centroid position of about 14 km. The other experiments have a centroid displacement lower than 27 km (FC-T), which is a very good value considering that the rain gauge inter-distance is roughly 13.5 km. Similar comments hold for cluster 2, where the best performing experiment is FC-Y with a centroid distance of about 19 km and the remaining experiments have a centroid distance lower than 20.5 km. The FG-G experiment has the best performance in terms of angle difference for cluster 1 and the FC-T experiment for cluster 2. The best area ratio is reached by FC-T for both clusters, meaning that it is the setup that models at best the extent of the rainfall area. The FC-G experiment shows the strongest agreement between the observed and the predicted near-peak (90th percentile) rainfall depth in cluster 1, corresponding to the Tanaro catchment, while FC-Y does so for cluster 2, corresponding to the northern Piedmont area.
As described in the “Introduction”, according to the Molini et al. (2011) criterion, severe rainfall events in the Mediterranean area can be classified in two categories: type I that are long-lived and widespread events and type II that are short and localised. The Piedmont 1994 case study was a type-I event, as indicated by the Monte Lema radar data (see Fig. 2), which show that the northern Piedmont Alpine arc was struck by persistent rain for roughly 2 days. However, as highlighted by previous works, such as Boni et al. (1996), the event was also characterised by intense convective activity, especially over central-western Liguria. This is confirmed by the fact that, in this area, the annual maxima of short-duration rainfall (accumulated over 1, 3, 6, 12, and 24 h) were attained during this event. Table 3 shows their values for five selected stations over western-central Liguria (as indicated by the black dots in Fig. 7). Note that in Fiorino the hourly accumulated rainfall maximum almost reached the exceptional value of 90 mm.
To gain a deeper understanding of the physical mechanisms responsible for this observational evidence, the FC-W experiment was further analysed, since it is the operational setup used at CIMA. The reflectivity vertical maximum intensity (VMI) field was mapped every 3 h between 12 and 21 UTC on the 4th of November 1994, together with the 10-m wind field. The different panels of Fig. 7 show the generation and the dissipation over the Ligurian Sea of a very localised and intense convective rainfall structure known as back-building mesoscale convective system (MCS). In particular, a well-defined surface convergence line between a cold northerly wind and a warmer south-easterly flow formed at 12 UTC (panel A). This was accompanied by an elongated structure of high VMI, that is generally indicative of intense convection. In the following hours, the interaction between the surface circulation and the slow-evolving large-scale dynamics forced the convective cells to develop in the same position for few hours, as indicated by the area of maximum VMI over mid-western Ligura at 15 UTC and 18 UTC (panels b and c). It is because of this dynamical feature that this kind of systems produces very high rainfall rates and they are called “back-building”. Finally, at 21 UTC (panel D), the convergence line over the sea completely disappeared and the convective activity inland diminished significantly.
These back-building MCSs are known to hit many regions of the Mediterranean, as southern France (Ducrocq et al. 2014; Nuissier et al. 2008), northern Italy (Fiori et al. 2014; Fiori et al. 2017; Lagasio et al. 2017), and eastern Spain (Millán et al. 1995; Pastor et al. 2010), to cite some. There is evidence that they have always been a meteorological hazard in this region (Parodi et al. 2017), and that their frequency is expected to increase because of climate change effects (Gallus et al. 2018). They generally develop over the sea (Nuissier et al. 2008; Fiori et al. 2014), which has been shown to strongly affect (1) the intensity of the precipitation through its mean sea surface temperature (SST) value (Lebeaupin et al. 2006; Pastor et al. 2001; Meroni et al. 2018b), and (2) the location of the rainfall band through the horizontal variations of SST (Meroni et al. 2018a; Cassola et al. 2016).
To conclude the meteorological result section, a qualitative discussion on the outputs of the IFS-AN-driven experiments (shown in Fig. 8) is done. In particular, all the configurations produced a rainfall field in agreement with the observations, with a higher simulated rainfall volume with respect to the IFS-FC-driven experiments. WSM6 was the one that produced, once again, the highest accumulated rainfall values. This suggests that it is indeed the numerical scheme to generate a higher value of rainfall. However, the differences among the IFS-FC-driven experiments and among the IFS-AN-driven ones are smaller than the difference between the IFS-FC run and the corresponding IFS-AN one for a given setup. This confirms that, as previously discussed, this event was strongly driven by the large-scale conditions (especially for the northern Piedmont area). Thus, by changing the initial and boundary conditions (from IFS-FC to IFS-AN), larger differences are introduced with respect to changing the numerical description of the microphysical processes.
The added value of the IFS-AN runs might be seen in the convective rain phase that hit the mid-western part of Liguria. In fact, all the IFS-AN runs produced a similar v-shape pattern over the sea (which is almost touching the southern border of the figure, at around 9° E) of accumulated rainfall above roughly 80 mm (yellow shading in Fig. 8), which is characteristic of the back-building MCSs, as discussed above. In the IFS-FC-driven runs (Fig. 5), instead, the four configurations produced different elongated rain bands, with rainfall covering larger areas with a weaker intensity, suggesting that the large-scale conditions were not properly driving the atmospheric dynamics in the generation of the back-building MCS.
Hydrological model results
The hydrological chain was then applied starting from all the IFS-FC experiments, simulating an operational framework. The simulated rainfall fields were downscaled to the Continuum grid using RainFARM, as discussed in Section 2.2. The results in terms of peak flow prediction are summarised in Table 4 in which the following quantities are reported:
The observed peak (only the peak estimates were available (Piemonte 1998));
The peak flow of the hydrological modelling fed with the observed rainfall field (named “Run Observation”);
The minimum and maximum peak discharge values from the hydrological ensembles of each rainfall forecast.
All the analyses were performed on the seven stations presented in Fig. 4. Here, for the sake of brevity, the results of four representative stations opportunely chosen are reported: Fossano and Monte Castello are chosen for cluster 1 (Tanaro basin), while Tavagnasco and Torino stations are chosen for cluster 2 (northern Piedmont). In Fig. 9, the hydrographs for the same four-level gauges are shown. Note that at the time of the event, no official hydrological forecast was available.
The Run Observation experiment (black lines with empty circles) reproduced quite well the observed peak flows (full red circles) in the selected sections. The sources of error can generally be related to the uncertainties of the hydrological modelling, such as the parameterisations and the soil moisture initial conditions, and to a bad representation of the rainfall input over part of the basin. This was indeed the case, as the rain gauge stations used to generate the interpolated rainfall maps were not very dense, meaning that the rainfall pattern was not very well represented. The Run Observation experiment in the Tavagnasco and Monte Castello sections evidenced a low underestimation of the observed peak discharge, while for Torino and Fossano sections, the simulations were good.
Concerning the FC experiments, on the Tavagnasco station, all the configurations overestimated the peak discharge, probably because the simulated rainfall exceeded the observations in northern Piedmont (cluster 2). The FC-W experiment had the best performances on the Tavagnasco section, while it slightly overestimated the peak flows on Torino, Fossano, and Monte Castello with all members. This overestimation can be probably ascribed to the larger amount of rainfall predicted by FC-W simulation over cluster 1 (refer to panel a of Fig. 5). The FC-G showed good skills from an early warning perspective on Torino and Monte Castello sections for all members. The best performances in cluster 1 sections were achieved by FC-T and FC-G, which both produced a good forecast in an early warning perspective. The good agreement between the Run Observation peak flow forecast (black lines with empty circles) and the actual observed peak flow values (full red circles) suggests that the hydrological model was correctly calibrated and it was reliable on the analysed sections. The mismatch between FC runs and the peak flow observations, thus, can be mainly ascribed to the uncertainties in the location and intensity of the rainfall prediction. In fact, neither a bad positioning of the rainfall field on spatial scales larger than Srel nor large errors in the accumulated rainfall depth can be managed by the downscaling algorithm.
Concerning the forecast timing, in Monte Castello and Torino, all the simulated peak flows preceded the observed ones (in Monte Castello, this holds also for the Run Observation experiment), but they would still be good to issue a warning. A very good timing was reported for Fossano, which had just a slight delay, and Tavagnasco section, characterised by a peak discharge overestimation. Globally, in an early warning system perspective, even with some uncertainties in the peak timing and intensity, all the FC experiments revealed very good performances and they would have led to issue an alert. This is particularly relevant for the sections hit by cluster 1 (Monte Castello and Fossano), because, as already discussed in the previous section, past numerical works were not able to correctly place the heavy rainfall of cluster 1 on the northern side of the Ligurian Alps, which would have led to completely miss the strong hydrological response.