Short-term forecasting of the Italian load demand during the Easter Week

In electrical load forecasting the prediction of the demand during holidays is a challenging task because of the drift of the demand profile with respect to normal working days. Among holidays, the Easter Week is peculiar because it is a moving holiday: though the weekdays are always the same, it may fall anywhere between March and April. The main contribution of this work is to develop a short-term day-ahead predictor for the load demand during the Easter Week using the Italian data as benchmark. The proposed strategy uses a Gaussian Process (GP) estimator to track the difference between the target Easter Week and an average Easter Week load profile. Differently from usual GP approaches that employ 'canonical' kernels, we propose and validate the use of a tailored kernel based on the nonstationary autocovariance of the time series, whose estimation is made possible by the availability of historical load series starting from 1990. On the Italian data the novel approach outperforms both GP methods based on canonical kernels and the forecasts provided by the Italian Transmission System Operator (TSO) Terna. The scarce correlation between the prediction residuals of the novel technique and those of the Terna forecaster motivated the use of aggregation strategies that yielded a further improvement. Indeed, all the main error indexes exhibit a decrease in several tens percent over Terna. The proposed approach is of general validity if, thanks to the availability of historical datasets, the kernel can be tailored to the statistical properties of the time series.