Ensemble of probabilistic hydrological models forecasts
Alana Renata Ribeiro1; Mauricio Felga Gobbi1; Eduardo Alvim Leite2; Diandra Akemi Alves Kubo3
1 PPGMNE-UFPR; 2 SIMEPAR; 3 PPGInf - UFPR
doi:10.20906/CPS/CILAMCE2017-0181
Resumo
In this paper probabilistic bayesian models for short-term streamflow forecasting are aplied on errors resulting from deterministic forecasts, with the main of explaining the uncertainties in the forecasting process. Bayesian inference was chosen because it allows this combination between prior information on probabilistic distributions parameters and data of errors committed by deterministic models. This way, MCMC (Markov chain Monte Carlo) sampling algorithms facilitate the retrieval of the parameters posterior distributions for which we wish to infer on. The errors used as input data for the probabilistic models come from four different deterministic hydrologic models. The first two are the rainfall-runoff conceptual models, Sacramento and 3R (Rainfall-Runoff-Routing); and the remaining are empirical models, the Multiple Linear Regressive and the MLP (Multilayer Perceptron) Neural Network. Despite the differences between them, the results obtained show that the approach efficiently handles errors of different model types. The time series data of streamflow and precipitation were obtained on an automatic monitoring station from the Paraná Meteorological System (SIMEPAR), located in a tributary of the Iguaçu River basin, in the state of Paraná, Brazil. Two probabilistic approaches, static and dynamic, were used for the inference and evaluated through statistical tests that allowed comparisons between the logarithms maximum values of the posterior probability distributions generated by each of them. The results demonstrated the superiority of the probabilistic model with the dynamic approach that considers the hydrological state at the moment when the prediction is being carried out for the sub-basin under study. Thus, a combination of dynamic probabilistic hydrologic models (ensemble) is developed to represent the uncertainties inherent in the deterministic process of streamflow forecasting. The results of the application of the forecast by set showed benefits mainly in cases of alerts of large floods, constant
Palavras-chave: Probabilistic; Forecasting; Bayesian inference; Markov Chain Monte Carlo