Quantitative precipitation forecasts:: a statistical adaptation of model outputs through an analogues sorting approach

Medium-term quantitative precipitation forecasts (QPFs) up to several days ahead are required to issue early flood warnings and to allow optimum operation of hydraulic structures or reservoirs. This paper describes an approach which can be seen as an adaptation of deterministic meteorological model outputs. It involves searching for a sample of past situations similar to the current one from a long meteorological archive. The analogy is considered in terms of general circulation patterns over a window covering western Europe. For this restricted sample of days similar to the day at hand, the corresponding sample of observed daily precipitation is extracted for each catchment. The rainfall to be observed during the current day is assumed to follow the same distribution, known from this empirical sample. This provides a probabilistic forecast expressed, for example, by a central quantile and a confidence range. This paper describes the many choices underlying the optimisation of this approach: choice of predictor variables to characterise a meteorological situation, choice of similarity criterion between two situations, criterion for performance evaluation between two versions of the algorithm, etc. This method was calibrated over about 50 catchments located in France, Italy and Spain, using a meteorological and hydrological archive running from 1953 to 1996. Comparisons carried out over a validation sample (1995-1996) with three poor-man methods prove the interest of this approach, in a perfect prognosis context. In real-time operation, the use of forecast instead of observed predictor variables, essentially geopotential fields, produces only a minor decrease in performance. The use of the single-valued central quantile supplemented by the confidence interval provided a QPF that has proved effective and informative on the potential for extreme values. (C) 2002 Elsevier Science B.V All rights reserved.