Identification of the downstream level dynamics of the UHE Tucuruí using LSTM neural networks

Abstract

This work investigates the application of recurrent neural networks LSTM to predict the downstream level of the Tucuruí hydropower, whose dynamic, non-linear and multivariable nature demands modeling formalisms based on deep learning. The trained networks seek to model the complex relationships between the upstream level and volume and the influent and effluent flows, with the target variable, the downstream level. Series with hourly sampling of real data of 13 years of operation of the Plant are used. The models are trained to predict up to 10 hours ahead, a window considered appropriate in real scenarios. The results show the feasibility of LSTM for downstream level prediction, with mean squared error accuracy to validation data ranging from 99.96% and 97.85% from the 1st to the 10th predicted sample.