A comparison of periodic autoregressive and dynamic factor models in intraday energy demand forecasting

We suggest a new approach for forecasting energy demand at an intraday resolution. The demand in each intraday period is modeled using semiparametric regression smoothing to account for calendar and weather components. Residual serial dependence is captured by one of two multivariate stationary time series models, with a dimension equal to the number of intraday periods. These are a periodic autoregression and a dynamic factor model. We show the benefits of our approach in the forecasting of (a) district heating demand in a steam network in Germany and (b) aggregate electricity demand in the state of Victoria, Australia. In both studies, accounting for weather can improve the forecast quality substantially, as does the use of time series models. We compare the effectiveness of the periodic autoregression with three variations of the dynamic factor model, and find that the dynamic factor model consistently provides more accurate forecasts. Overall, our approach combines many of the features which have previously been shown to provide high quality forecasts of energy demand over horizons of up to one week, as well as introducing some novel ones.