The use of hybrid techniques has been applied to solve several problems, including convergence and precision. This paper proposes a hybrid methodology using SARIMA regression models and Multilayer Perceptron Neural Network trained by Levenberg-Marquardt algorithm applied to short term load forecasting. This forecasting is a notably important task for the planning and operation of electrical power systems to anticipate when and how much generation and transmission must be available to provide the required electrical load without interruption. The results are presented for a Brazilian electrical demand energy time series.