Assessment of Particle Filter Technique for Data Assimilation in the Forecasting of Streamflows for the Tocantins River Basin in Brazil

Abstract

The Particle Filter (PF) technique is applied to forecasting of streamflows in the Tocantins River located in Brazil in this paper. This technique used as a data assimilation method is coupled to a semi-distributed hydrological model named MGB at hourly time intervals. The states variables are generated by computing rainfall forcing, considering time and spatial correlated errors. Sensibility tests were performed to highlight the importance of the precipitation error value and the particles number, as well as the low dependence on time and spatially correlated errors. The PF performance has been compared with streamflows predicting without assimilation, together with an empirical method. The resulting forecasts agreed well with the observations and maintained meaningful in terms of Nash–Sutcliffe at all stations analyzed even for long lead times. Also, PF technique performed well in greater lead time of forecasting when compared with empirical method.