Regionalização bayesiana de vazões mínimas em condições de não estacionariedade

Abstract

Water has increasingly become an asset of interest to society due to the increase in water demand. The management of water resources, by means of grant establishments, becomes an important tool to avoid conflicts and guarantee the sustainable use of water. The streamflow monitoring system is limited to a few measurement gages and, therefore, hydrologists often use regionalization techniques to make estimates in unmonitored basins. In this context, regional estimates of low flows, grant indicators, and their predictive uncertainty become object of interest. This study aimed to develop a Bayesian hierarchical model for regionalization of low flows conditioned to temporal covariables, in this case, the sea surface temperature (SST) and to evaluate the benefits of this approach compared to a stationary reference model. Starting from a simpler stationary model, under the hypothesis of spatial independence of the parameters and conditional independence of the observations, the representation of the spatial variability of the model was initially captured by means of spatial covariates. Subsequently, possible gains in prediction were investigated by further describing the spatial description under the data and under the process. Then, the SST was introduced at the process level of the model in the form of a customized climate index, inferred from a field of SST values. Models were developed to assess the inclusion of the customized climate index in the parameters of the probability distribution employed. The model was applied in the Itajaí-Açu river basin (SC) and in the Doce river basin (MG / ES). The results showed that the non-stationary model performed better, in terms of the DIC criterion, than the reference stationary model and that the estimated quantiles (such as Q7,10) are strongly influenced by climatic variability. In addition, a more complex description of the spatial dependence of the process brings benefits to prediction in densely monitored regions, while interdependence in data observations, when considered, can also bring benefits to prediction.