Forecasting of daily streamflows downstream from reservoirs with streamflow regularization using machine learning methods

Abstract

Streamflow gauge stations (SGS) can show inconsistent daily streamflow estimates, due to the number of readings taken by the rating-curve method, throughout the recording time series. For SGS located downstream of streamflow regularization reservoirs (SRR), the use of time series for the outflow can serve as a reference for improving these records, since the daily data are estimated by the water balance method, with about 24 daily flow records. This work aims to fit machine learning (ML) models to the forecasting of daily streamflow data of SGS located downstream of an SRR. Besides indicating inconsistencies in streamflow data from the SGS, the results also showed that, for the SGS close to the SRR, the model based on Neural Networks was the most accurate. For the SGS most distant from the SRR, the Multiple Linear Regression model was the best fit.