Authors: Elisa Raspanti; Andrea Marziali
Addresses: Dipartimento di Matematica, Università di Bologna, Piazza di Porta San Donato, 5, 40126, Bologna, Italy ' Dipartimento di Matematica, Università di Bologna, Piazza di Porta San Donato, 5, 40126, Bologna, Italy
Abstract: This paper deals with one-day-ahead forecasting of Italian electricity demand (IED). The problem is addressed through machine learning techniques, nine base models (ridge regression, LASSO, elastic net, support vector machine, Gaussian process, k-nearest neighbour, random forest, artificial neural network and torus model) and five aggregation models based on base model predictions whose errors are automatically corrected by a SARIMA model. In addition to the ensemble models, we analyse also the time and spatial-time aggregations; indeed, the models first are applied to the daily IED time series, then repeated for the daily time series of each of the 24 hours and finally extended to each hour of each Italian zone for a total of 144 time series with daily frequency. Dimension reduction by the PCA is also pursued in order to reduce the computation times and investigate the possible risk of overfitting.
Keywords: electricity demand; time series forecasting; statistical learning; neural networks; SARIMA; principal component analysis; PCA.
International Journal of Energy Technology and Policy, 2021 Vol.17 No.6, pp.590 - 618
Received: 23 Feb 2021
Accepted: 04 May 2021
Published online: 23 Feb 2022 *