Multitask Learning for Predicting Natural Flows: A Case Study at Paraiba do Sul River

Abstract

Forecasting the flow of rivers is essential for maintaining social well-being since their waters provide water and energy resources and cause serious tragedies such as floods and droughts. In this way, predicting long-term flow at measuring stations in a watershed with reasonable accuracy contributes to solving a range of problems that affect society and resource management. The present work proposes the MultiTask-LSTM model that combines the recurring model of Deep Learning LSTM with the transfer of learning MultiTask Learning, to predict and share information acquired along the hydrographic basin of Paraíba do Sul river. This method is robust for missing and noisy data, which are common problems in inflow time series. In the present work, we applied all 45 measurement stations’ series located along the Paraíba do Sul River basin in the MultiTask-LSTM model for forecasting the set of these 45 series, combining each time series’s learning in a single model. To confirm the MultiTask-LSTM model’s robustness, we compared its predictions’ results with the results obtained by the LSTM models applied to each isolated series, given that the LSTM presents good time series forecast results in the literature. In order to deal with missing data, we used techniques to impute missing data across all series to predict the 45 series of measurement stations alone with LSTM models. The experiments use three different forms of missing data imputation: the series’ median, the ARIMA method, and the average of the months’ days. We used these same series with imputing data in the MultiTask-LSTM model to make the comparison. This paper achieved better forecast results showing that MultiTask-LSTM is a robust model to missing and noisy data.