Análisis del rendimiento de algoritmos de entrenamiento de redes neuronales artificiales, aplicadas al modelamiento dinámico de represas hidroeléctricas, mediante el error de predicción del nivel de embalse de agua.

Abstract

The objective of this investigation was to evaluate training algorithms from the Keras Python library; applying to the dynamic modelling of hydroelectric dams, by means of the prediction error of the water reservoir level. Historical dates were used (2005-2016) such as level, row and active power from the Agoyan Hydroelectric Power Plant stored in the database manager PostgreSQL which are divided into climatic seasons and are processed using signalling and standardization techniques. For the training of the models, a Python software platform was developed, with the use of the training algorithms of the Keras library plus the Tensorflow back-end. The processor used for these tasks was graphics processing unit GPU Nvidia 1050Ti. Through the variance analyze ANOVA a probability obtained was (p) of 6,02157E - 44 and 1.4024E - 42 for training error and validation error respectively, which rule out the null hypothesis since in both cases the probability was less than 0.05 this means (p<n) for a confidence level (n) of 95% the Tukey test determine that Nadam algorithm It has the least significant difference from other statistical efficiency found that the Nadam algorithm is the most efficient Var(Nadam)< Var(SGD). The model with training the algorithm Nadam reached a predictor of an effective level to a threshold 48 hours obtaining an RMSE minimum of 0.035876 [m] and a maximum of 0.344913 [m] of reservoir water level error of the dam. In this case study the backend Tensorflow was used but there are currently others like Theano and CNTK, it is recommended to train the algorithms in these structures and test their performance.