GNGTS 2013 - Atti del 32° Convegno Nazionale

by a fixed exceedance probability in the considered time interval (g 0 ) . This implies that no comprehensive hazard curve can be actually computed from Eq. (1). Furthermore, no “ex- ante” scoring was provided by the proponents and thus, in this preliminary application, no comprehensive hazard estimate is attempted and the focus is on the empirical scoring only. Due to this, in the lack of data concerning future occurrences, testing can be only performed in a retrospective way on time-independent models. Selection of accelerometric data is a basic aspect of empirical scoring. To this purpose, the ITACA database (Luzi et al. 2008; Pacor et al. , 2011) of accelerometric registrations has been explored, searching for eccelerometric sites operated for long times. Kindly supported by INGV (F. Pacor and R. Puglia, personal communication) and by the Department of Civil Protection responsible for the Italian accelerometric network (A. Gorini, personal communication), we selected a number of 72 accelerometric sites continuously monitored for at least 25 years (Fig. 1). In particular, the accelerometric database relative these sites for the time span 1979- 2004 was taken into account. Since in most cases, site conditions of these accelerometric sites did not correspond to reference conditions considered in PSHA estimates, available PGA values were “corrected” by considering both soil stiffness and local topography by adopting coefficients provided in Eurocode 8 and National Seismic Code (NTC08). For each site, the maximum registered PGA value was considered only. For those sites where no registration was obtained, we fixed the trigger threshold (9.8 cm/s 2 ) as the “observed” maximum. These values where compared with the g 0 values and relevant exceedance probabilities P i ( g 0 |H i ) provided by considered PSHA procedures. However, the problem arose that exposure time considered by each model (30 y) was different from the one actually covered by accelerometric registrations (25 y). Furthermore, many of the considered PSHA procedures considered several exceedance probabilities P i ( g 0 ) = P i ( g 0 |H i ) to determine the reference g 0 value. In order to make the comparison possible, exceedance probabilities were re-assessed relying on the Poissonian character of the underlying PSHA models. In this case, for any PSHA procedure, one has (13) where Δ t is the exposure time considered for hazard computations in H i . When a different exposure time Δ T is considered, one has, for the same g 0 value, a different exceedance probability given by (14) These new exceedance probabilities were finally considered for testing. Since some models also provides g 0 value corresponding to different exceedance probabilities, some models were scored by considering these different “forecasts”. In general, since in the same model lower exceedance probabilities correspond to longer average return times and to higher g 0 values, different scoring can be attributed to different parts of the hazard curve. Tab. 1 reports outcomes of the scoring procedure applied to 9 PSHA models. For each model, different parts of the relevant hazard curve were considered (see the correspondent average return times). The counting test [Eq. (12)] indicates that one of the considered models (Mod 27) provides outcomes that appear not compatible with observations, since it provides significant underestimates of the observed hazard. In the case of Mod 28, hazard estimates relative to longer return time (higher g 0 values) are rejected by the counting test. As concerns lower g 0 values (30 of average return time), estimates are compatible with observations but also provide relatively low likelihood values. As a whole, if only PSHA models passing the counting test are considered, likelihood seem to provide quite different scores despite of the fact that a relatively small amount of control sites is considered and that hazard estimates are 13 GNGTS 2013 S essione 2.1