Dynamic Bayesian smooth transition autoregressive models applied to hourly electricity load in Brazil

Faria, Álvaro E. and Santos, Alexandre J. (2019). Dynamic Bayesian smooth transition autoregressive models applied to hourly electricity load in Brazil. In: ITISE 2018 International Conference on Time Series and Forecasting, 19-21 Sep 2018, Granada, Spain.

URL: https://itise.ugr.es/2018/program.php


Dynamic Bayesian Smooth Transition Autoregressive (DBSTAR) models are proposed for nonlinear autoregressive time series processes as alternative to both the classical Smooth Transition Autoregressive (STAR) models of Chan and Tong (1986) and the Bayesian Simulation STAR (BSTAR) models of Lopes and Salazar (2005). Unlike those, DBSTAR models are sequential polynomial dynamic analytical models suitable for inherently non-stationary time series with non-linear characteristics such as asymmetric cycles. As they are analytical, they also avoid potential computational problems associated with BSTAR models and allow fast sequential estimation of parameters.

Two types of DBSTAR models are defined here based on the method adopted to approximate the transition function of their autoregressive components, namely the Taylor and the B-splines DBSTAR models. A harmonic version of those models, that accounted for the cyclical component explicitly in a flexible yet parsimonious way, were applied to the well-known series of annual Canadian lynx trappings and showed improved fitting when compared to both the classical STAR and the BSTAR models. Another application to a long series of hourly electricity loading in southern Brazil, covering the period of the South-African Football World Cup in June 2010, illustrates the short-term forecasting accuracy of fast computing harmonic DBSTAR models that account for various characteristics such as periodic behaviour (both within-the-day and within-the-week) and average temperature.

Viewing alternatives

Download history

Item Actions