Electricity consumption forecasting in Brazil: A structural Bayesian dynamic approach

Abstract

ABSTRACT The main aim of this paper is to project electricity consumption for the five Brazilian regions using the Bayesian Structural Time-series (BSTS) which presents the best prediction evaluated within the sample. This methodology estimates fourteen models, each with a different specification in terms of its functional form, which ensures flexibility and can vary in terms of levels, trends, and treatment of seasonal fluctuations, as well as dealing with different hypotheses on regression errors or disturbances. The forecast validation criteria calculated for a 24-month horizon showed very satisfactory results. The findings corroborated expectations regarding the suitability of this methodology for drawing up projections. In every case, the forecasting error was 5% or less, which is the reference figure for distribution utilities. In the Center-West, Southeast and South regions, this indicator fell below 3%. The model thus appears to be a robust tool for utilities to project electricity demands.