Modelos baseados em aprendizado de máquina para geração de séries sintéticas do setor elétrico

Abstract

Brazil’s main source of energy supply is hydroelectric, due mainly to its large hydro capacity. Understanding the flow behavior of its basins is a fundamental factor to optimize the production of this type of energy, but the present historical data are limited, becoming a hindrance to the study, given its importance in the planning of electric energy production. One solution that has been used in the recent literature is the generation of synthetic series. In this work, the following techniques were used for the synthetic generation of the flows of the Água Vermelha and Volta Grande stations: SynTise, model presented in citeonline Denaxas (2015), support vector machines (SVM), multilayer perceptron (MLP), random forest (RF) and the autoregressive model (AR). Synthetic series equivalent to 2000 years were generated for all these reservoirs. The work analyzed four different proposals for the selection of the random component of the AR, MLP, SVM and RF models, which are: through a symmetric probability distribution, through an asymmetric probability distribution, in chronological order and through the estimated residuals. The new random component proposals and the classical selection method, the random selection of the residues, were evaluated for the two stations, as well as SynTise, which was adjusted to generate synthetic monthly series for reservoir flow. The results showed that, for the two stations evaluated, models with random component over time were better options than the classic model of random component randomly selected in all the techniques evaluated. In the comparison between the best results of each technique, it was obtained that for the Volta Grande station, the SVM presented the best results, while for Água Vermelha, the MLP was better among all the models