A new forecasting method for functional time series is proposed. The new model has been tested with the residual demand curves of the Spanish day-ahead electricity market and compared with other functional reference models. Electricity generators and retailers trading in electricity markets can take advantage of residual demand curves forecasts as tools for optimizing their bidding strategies. The model is aimed at extending the ARH model (Autoregressive Hilbertian model) to the SARH model in which the seasonality of the series is taken into account. This is a significant improvement, as high-frequency time series generated in the context of electricity markets show seasonal dynamics. Therefore, this model is built following a two steps procedure. In the first step, the structure of the model has to be identified. By means of a functional autocorrelation plot, significant autocorrelations in the functional time series are found, and initial values for the regular and seasonal autoregressive orders are inferred. Secondly, a seasonal autoregressive functional linear model is estimated using the inferred structure. While the functional parameter is usually estimated using a functional principal component basis, in this paper we propose a Gaussian estimator based on neural network techniques.
21st International Conference on Computational Statistics - COMPSTAT 2014
Publication date: August 2014.
J. Portela, A. Muñoz, E. Alonso, Forecasting residual demand time series in electricity markets: a functional approach, 21st International Conference on Computational Statistics - COMPSTAT 2014. ISBN: 978-2-8399-1347-8, Ginebra, Switzerland, 19-22 August 2014