Peak electricity demand forecasting using time series regression models: An application to South African data

Forecasting of electricity demand requires the use of models which capture important drivers of demand. In this paper we use two time series regression (TSR) models for short term forecasting. South African hourly electricity data for the period, years 2000 to 2010 is used. The first TSR model is one in which the temperature effects are captured through heating and cooling degree days. We refer to this as TSR model 1. In the second TSR model the temperature effects are captured though regression splines. This is TSR model 2. The third model includes a component which captures the volatility in electricity demand. A comparative analysis is done with the first two models in out of sample predictions of up to four weeks. Empirical results show that the model in which temperature is incorporated through regression splines produces better forecasts. An analysis of under demand predictions is done by comparing model 3 with model 2. Results show that both models, i.e. 2 and 3 respectively are comparable. However model 3 seem to capture well the volatility in the residuals during the year 2008 when we experienced a world recession.