Operation policies for hydropower systems: using the unsupervised SONARX neural network

The optimal behavior of the hydropower plants (HP´s) depends as much on the relative position of each station throughout the cascade as on the relation between them. The main purpose of this work is to apply the Kohonen self-organizing map network (SOM) with dynamic system models, in the learning of these relations, objectifying its usage to simulate the optimal operation of a real hydroelectric power system: seven big HP´s placed on the Brazilian southeast system. This model is an unsupervised version of the NARX model; which has been called self-organized model or network (SONARX), capable of processing space-time patterns. The tests showed that the neural predictor has a high tendency towards the results of the deterministic optimization, optimizing the usage of the available water resources for electrical energy generation.