Regionalization of GR4J model parameters for river flow prediction in Paraná, Brazil

Abstract

Regionalization methods dependent on hydrological models comprise techniques for transferring calibrated parameters in instrumented watersheds (donor basins) to non-instrumented watersheds (target basins). There is a lack of flow regionalization studies in regions with humid subtropical and hot temperate climates, and one of the main novelties of this research is to assess the regionalization of low flows in Paraná in the south of Brazil. In addition to filling this gap, this research presents innovative artificial-intelligence techniques for transferring parameters from hydrological models. This study aims to evaluate regionalization methods for transferring GR4J parameters and predicting river flow in catchments from the south of Brazil. We created a dataset for the state of Paraná with daily hydrological time series (precipitation, evapotranspiration, and river flow) and watershed physiographic and climatological indices for 126 catchments. Rigorous quality-controlling techniques were applied to recover data from 1979 to 2020. The regionalization methods compared in this study are based on simple spatial proximity, physiographic–climatic similarity, and regression by random forest techniques. Direct regression of Q95 was calculated using random forest techniques and compared with indirect methods, i.e. using regionalization of GR4J parameters. A set of 100 basins was used to train the regionalization models, and another 26 catchments (pseudo-non-instrumented) were used to evaluate and compare the performance of regionalizations. The GR4J model showed acceptable performances for the sample of 126 catchments, with 65 % of watersheds presenting a log-transformed Nash–Sutcliffe coefficient greater than 0.70 during the validation period. According to the evaluation carried out for the sample of 26 basins, regionalization based on physiographic–climatic similarity was shown to be the most robust method for the prediction of daily and Q95 reference flow in basins from the state of Paraná. When increasing the number of donor basins, the method based on spatial proximity has comparable performance to the method based on physiographic–climatic similarity. Based on the physiographic–climatic characteristics of the basins, it was possible to classify six distinct groups of watersheds in Paraná. Each group shows similarities in forest cover, urban area, number of days with more than 150 mm of precipitation, and average duration of consecutive dry days. Although the physiographic–climatic similarity method obtained the best performance, the use of machine learning algorithms to regionalize the model parameters had good performance using climatic and physiographic indices as inputs. This research represents a proof of concept that basins without flow monitoring can have a good approximation of streamflow if physiographic–climatic information is provided.