Aplicação de técnicas de aprendizagem de máquinas na previsão de vertimento em usinas hidrelétricas

Abstract

Operation at a hydroelectric plant is dependent on several factors such as the schedule of power generation, the volume of water available in its reservoir, the conditions of the river downstream, and the safety of the dams. A major challenge of the operation is to control the spillage from the reservoir. Although the spillage action represents the loss of energy resources, this action is also a powerful strategy to control the level of the reservoir, ensuring the safety of the dam. Decision-making regarding this operation is carried out in advance and is generally based on estimated level and demand information. In this context, this work applies supervised machine learning techniques to predict, for five hours to come, the operating condition of pouring in a hydroelectric plant. Intending to be used in real-time, this method aims to assist the operator, so that he can make more assertive and safer decisions, preserving energy resources and promoting increased safety of dams, and consequently, of workers and the population that resides river banks downstream of the plant. Random Forest, Multilayer Perceptron and the combination of these learning algorithms are adopted and compared in this work. The proposed methodology was implemented and tested with a hydroelectric plant located on the Tocantins River, Brazil, with a generation capacity of 902.5MW. The results of the methodology demonstrated that the tool has the capacity to be an efficient aid to the operators of a plant in decision making, since the forecasting models reached levels above 99% of correctness in the spillage forecasts.