Estimação bayesiana dos parâmetros de modelos hidrológicos a partir do emprego de assinaturas hidrológicas

Abstract

Despite the technological advances in the last decades, modelling hydrological processes still faces several challenges. Among them, we highlight uncertainty analysis and prediction in ungauged basins. This work evaluates the applicability of streamflow signatures and ABC methods to estimating parameters of hydrological models. The final purpose was to verify the viability of hydrological signatures to allow prediction in ungauged basins. To do so, simulations were realized on hourly and monthly timescales, using GR4H (Génie Rural à 4 paramètres Horaires) and GR2M (Génie Rural à 2 paramètres Mensuel) models, respectively, intending to approximate the signatures calculated from the simulated series to those estimated for the monitoring series of selected gauges. The combination of different signatures was also evaluated. Model performance and posterior distributions of model parameters were evaluated concerning the results of a “classical” Bayesian approach, using a MCMC sampler (Diffential Evolution Adaptive Metropolis – DREAM) and specification of the likelihood function (generalized likelihood function), in the time domain. The main result, in the monthly timescale, was that similar performance measures were found both in the signature and in the time domain. In the hourly model, there was a reduction in the values of these performance indices. The combination of different signatures, in this timescale, proved to be advantageous, improving the ability of the model to reproduce the peaks and the recession of the hydrograph. Furthemore, the posterior distributions of model parameters suggest the problems of equifinality and identifiability in both timescales. Finally, a simplified assessment of the potential for regionalization of signatures was conducted on an hourly scale, which showed worse representations of the variability of the hydrograph, but performance measures close to those calculated to local signatures. As a conclusion, it was observed great potential for using the proposed methodology, as long as there are a careful selection of signatures and compatible computational power.