Redes Bayesianas Dinâmicas para Previsão de Séries Temporais: Aplicação ao Setor Elétrico
Publicações do PESC
This work presents an approach for time series prediction for which there are not many works in the literature: the prediction of a continuous value through the estimation of the probability density function of the continuous random variable meant to be predicted. For this nonparametric estimation we employed Bayesian Networks and Dynamic Bayesian Networks with discrete random variables by the discretization of the continuous data. This way we created several systems for prediction: Markov Model for Regression, Hidden Markov Model for Regression and Multi-Hidden Markov Model for Regression. The main contribution of this work was the generalization of these systems by the use of fuzzification instead of discretization: Fuzzy Bayes Predictor, Fuzzy Markov Predictor, Fuzzy Hidden Markov Predictor e Fuzzy Multi-Hidden Markov Predictor. We also developed some methods to make the partitioning of the space of continuous data in order to be used by our systems that make fuzzification. Our systems were applied to the tasks of single-step and multi-step forecasting of monthly electric load series. The employed time series present a sudden significant changing behavior at their last years, as it occurs in an energy rationing. We have obtained competitive results when compared with several known statistics techniques.