Influencia de los algoritmos de entrenamiento de RNAs en la predicción del nivel de embalse de agua en una estación hidroeléctrica.

Abstract

The present paper reports on the obtained results from the analysis of the influence of training algorithms, for artificial neural networks (ANNs), on the prediction error of the reservoir water level of a hydroelectric station. The studied algorithms are those forming the Keras library, which uses the back-end of TensorFlow. Data for this study are the historical records (2005-2016) of reservoir level, streamflow, and active power from an Ecuadorian hydroelectric plant. Such data wasdivided for the training, validation, and test stages. The hardware platform was a graphic processing unit (GPU) Nvidia 1050Ti, which allowed for exploitingthe highly parallel computing capability of TensorFlow. Seven algorithms were evaluated. The Tukey test revealed that the Nadam algorithm obtained the lowest significative difference respect to its counterparts, engaging it as the more efficient. The respective obtained RNA plant model reached effective prediction thresholds up to 48 hours. The obtained results allow for optimization of the planification of energy production on the hydroelectric station trough an accurate prediction of the hydric resources for quotas of desired production.