Quality Analysis of Probabilistic Hydrological Forecasts (PAP014794)

Código

PAP014794

Autores

Anne Caroline Negrão, Eduardo Alvim Leite

Tema

Flood forecasting and early warning systems

Resumo

Aiming at optimizing power generation and the increased reliability of flood warningsystems, decision makers are searching for hydrological models which are more accurate andrepresentative. Probabilistic models consider the uncertainty of forecasts, therefore they are more faithfulto the randomness of hydrological phenomena and more transparent with respect to the uncertainties ofthe forecast. However, there are still difficulties in the direct quantification of the quality of these type ofmodel. Katz and Murphy (1997) propose the analysis of forecasts? sharpness and discrimination. Both arebased on the concepts of climatological probability. The first is the forecast variability and the secondrelates the predictions overall average given a certain observation. In order to assess the predictionreliability, Jolliffe and Stephenson (2003) use a frequency histogram of the observation occurrences ineach ensemble of forecast probability distribution. This work presents a method of accounting for thesethree quality indices. Moreover, they are simultaneously assessed following the concept of multi criteria toquantify the quality of probabilistic hydrological forecasts. Two probabilistic models were developed andoperated by the Technological Institute SIMEPAR, Paraná, Brazil. One is based on the Bayesian theoryand SOM neural network; the other is based on historical errors of deterministic forecasts and Meta-Gaussian transformation. The indices allow the monitoring of forecasts? quality of both models and thecomparison to confirm which one has the best results.