Prototype-Based Forecasting of Demand Curves in the Spanish Electricity Market

A. M. Alonso Fernández, A. El Kadi Lachehab, C. Ruiz Mora

This work proposes a cluster-based prediction framework for the matched demand curves of the Spanish daily electricity market. These curves are step functions relating price and traded energy, with the particular feature that each curve is defined on a different energy interval, from [0,Et], where Et varies hourly. To compare such heterogeneous objects, we use Dynamic Time Warping (DTW) which naturally handles different supports and non-uniform alignments. Based on the DTW dissimilarity, k-medoids clustering is applied to obtain representative prototypes (medoids). Each hourly curve is then represented by its vector of distances to the medoids, generating a multivariate time series. Day-ahead forecasting is performed on these distance vectors using direct multistep models. The predicted nearest medoid is used as the curve forecast. Results show that the proposed approach improves over the usual benchmark and that the medoid representation is highly informative for demand forecasting.

Keywords: functional data, clustering, dynamic time warping, mutivariate time series

Prediction and Classification
September 3, 2026  9:00 AM
Aula 24

T. Murillo Mosquera, C. C. Sánchez Zuleta

M. Ojaghi, P. Huidobro, I. Mariñas del Collado, S. Montes

M. C. Robustillo Carmona, A. Mateos Caballero, C. J. Pérez Sánchez, M. I. Parra Arévalo

J. I. Segovia Martín, S. Mazuelas Franco