Using Exponential Smoothing Methods for Modelling and Forecasting Short-Term Electricity Demand

In this paper different exponential smoothing methods are considered for modelling and forecasting short-term electricity demand in England and Wales. The time series contains half-hourly time periods and two seasonalities can be observed – one within each day and one within each week. Both seasonalities are modelled separately with the Holt-Winters methods. The double seasonal Holt-Winters method introduced by Taylor (2003) accommodates both seasonal patterns. An autoregressive model is fitted to the residuals in order to deal with first-order autocorrelation. In addition, a model for multiple seasonal (MS) processes introduced by Hyndman et al. (2008) is used, which divides the seasonal component into several sub-cycles and allows for the seasonal sub-cycles to be combined. This reduces the amount of seed values to be estimated. Moreover, the MS process allows for the seasonal terms to be updated more than once during a seasonal cycle. Although the MS process with three combined sub-cycles reduces the amount of seed values to be estimated and gives accurate forecasts, the double seasonal Holt-Winters method outperforms the other methods in forecasting the short-term electricity demand.