Multiple seasonal STL decomposition with discrete-interval moving seasonalities

Abstract

[EN] The decomposition of a time series into components is an exceptionally useful tool for understanding the behaviour of the series. The decomposition makes it possible to dis-tinguish the long-term and the short-term behaviour through the trend component and the seasonality component. Among the decomposition methods, the STL (Seasonal Trend decomposition based on Loess) method stands out for its versatility and robustness. This method, however, has one main drawback: it works with a single seasonality, and does not deal with the calendar effect. In this article we present a new decomposition method, based on the STL, which allows the use of different seasonalities while allowing the cal-endar effect and special events to be introduced into the model using discrete-interval moving seasonalities (MSTL-DIMS). To show the improvements obtained, the MSTL-DIMS technique is applied to short-term load forecasting in some electricity systems, and the results are discussed.