Assessment of data-driven modeling strategies for streamflow forecasting in a river of Sergipe

Abstract

Among several streamflow forecasting techniques, data-driven models (DDMs) are widely used. They employ a database formed by historical input (precipitation and streamflow) and output (streamflow) variables to perform the prediction. Artificial neural networks (ANNs) are the most common types of DDMs, and are typically more accurate than other empirical models, but have the disadvantage of not being sufficiently transparent. One of the machine-learning methods that do not have this problem is the so-called “instance based learning” (IBL). The KNN algorithm is an example of IBL. This research applies and proposes variations of the KNN model in order to forecast streamflows in a river of the state of Sergipe, Brazil. The results are compared to simulations carried out by the use of artificial neural networks and indicate the superiority of the ANNs, but also show satisfactory forecasts of the KNN.