Forecasting electricity demand using generalized long memory

This paper studies the hourly electricity load demand in the area covered by a utility situated in the southeast of Brazil. We propose a stochastic model which employs generalized long memory (by means of Gegenbauer processes) to model the seasonal behaviour of the load. The proposed model treats each hour's load separately as an individual series. This approach avoids modelling the intricate intra-day pattern (load profile) displayed by the load, which varies throughout the week as well as through the seasons. The forecasting performance of the model is compared with a SARIMA benchmark using the years of 1999 and 2000 as the holdout sample. The model clearly outperforms the benchmark. Moreover, we conclude that long memory behaviour is present in these data.