Short-Term Streamflow Forecasting for Paraíba do Sul River Using Deep Learning

Abstract

Water resources are essential for sustainable economic and social development, as well as be a vital element for the conservation of ecosystems and the life of all beings on our planet. On the other hand, natural and anthropic disasters from floods and droughts may occur. The modeling of hydrological historical series has extensively been studied in the literature for important applications involving the water resources’ planning and management. There are several temporal series prediction’s techniques in the literature. Some of them are characterized as classical linear methods whose adjusts for multivariate or multi-input prediction problems can be difficult. On the other hand, artificial neural networks can learn complex nonlinear relationships from time series, and the deep learning model LSTM is considered the most successful type of recurrent neural network capable of directly supporting multivariate prediction problems. This work presents a comparison between two forecasting’s models of time series: ARIMA, a classical linear model, and an LSTM neural network, a nonlinear model. As a case study, we used the time series of four measurings’ substations of one of the very important Brazilian rivers - the Paraíba do Sul river. These time series are difficult to predict since their history series has flaws and high oscillation in the data. The LSTM, which is a robust model, performs better in analyzing the behavior of this type of time series.