Long-term electric load forecasting: A torus-based approach

Long-term forecasting of daily electric power load is investigated. After log-transformation and detrending of the data, the residual variability is decomposed as the sum of a “potential term”, accounting for seasonality and weekly periodicity, and intervention events accounting for consumption changes associated with holidays and other special events. The biperiodic potential term is modeled as a linear combination of basis functions obtained from the tensor product of 7-day and 365-day harmonics. The intervention events are modeled by searching for “similar dates” in the historical records. The new forecaster is tested through the prediction of the whole 2013 load profile based on historical data until December 31, 2012. The results prove the effectiveness of the proposed approach that achieves a Mean Absolute Percentage Error (MAPE) equal to 2.96% not far from state-of-art performances of one-day-ahead short-term forecasters.