An integrated methodology for post-processing ensemble prediction systems to produce more representative extreme water level forecasts: the case of the Río de la Plata estuary

The effects of weather extreme events can pose a threat to life and property, which is why proper prediction systems take on superlative importance. Despite the significant scientific advances in the field during the last decades, due to the intrinsic imperfections of prediction systems there will always be unavoidable uncertainties. To deal with them, deterministic prediction systems have been extended to “ensemble prediction systems” (EPS), defined as a composition of several simulations under different forcings, boundary conditions, parameters, models, etc., designed to represent the uncertainties. The mean of the EPS is often used for deterministic guidance to report the prediction but, in the presence of large differences among ensemble members, the average generates skewness that might underestimate the magnitude of the forecasts. In this paper, two techniques are revisited and readapted to improve the EPS forecasts. Firstly, it is proposed to partition ensemble forecasts into sub-ensemble forecasts, using cluster analysis to produce more representative predictions; this technique seeks to eliminate from the ensemble members which occurrence is considered unlikely. Secondly, it is suggested to associate to the ensemble forecast a complementary phase-aware ensemble (PAE) forecast, which computes the ensemble mean and spread separating the signal into carrier and modulated waves using the Hilbert transform. This integrated post-processing methodology was assessed with extreme storm surges that took place at the Río de la Plata estuary (Argentina) during this century with amplitudes exceeding $$\pm 2$$ ± 2 m (being the tidal range of about 0.75 m), for which the EPS presents large dispersion. Results show that, in the analyzed cases, the post-processing filters out the unlikely dynamical states, adjusts the mean ensemble to the observations and significantly reduces the uncertainty; the spread is reduced from 3 m to less than 1 m. The probability was also improved; for the analyzed cases, calibrated forecasts could anticipate peak events 4 days in advance with a relatively small uncertainty in both time and amplitude (one standard deviation). Additionally, the technique does not harm the forecast in cases when the dispersion between members of the ensemble is low. These results, together with the low computational cost of applying the technique, support incorporating our post-processing methodology as part of the EPS for storm surges, in which uncertainty is paramount for issuing warnings to face the effects of extreme weather events.