Talk
Combining various dissimilarity measures for clustering electricity market prices
Margarida G. M. S. Cardoso (Cardoso, M. G. M. S.); Ana Alexandra A. F. Martins (Martins, A. A. A. F.); João Lagarto (Lagarto, J.);
Event Title
O XXIV Congresso da Sociedade Portuguesa de Estatística
Year (definitive publication)
2019
Language
English
Country
Portugal
More Information
--
Web of Science®

This publication is not indexed in Web of Science®

Scopus

This publication is not indexed in Scopus

Google Scholar

This publication is not indexed in Google Scholar

Abstract
The development of an internal market of electricity has long been a goal of the European Union, since it enables European citizens and businesses to choose their supplier, creates new business opportunities and enhances cross-border trade with the purpose of ensuring effciency gains and competitive prices and, contribute to security of supply and sustainability. In order to better understand the degree of integration of the electricity markets of different European countries, we propose to cluster time series regarding each country's day-ahead electricity market prices, investigating similar market prices behaviours. The electricity markets in study are the MIBEL, the Italian, the Nordpool, the French and the German markets. Hourly data are considered and were obtained from the market operators' websites. When clustering time series, the selection of the dissimilarity or a distance measure is a key issue. In fact, different dissimilarity measures between time series have been proposed in the literature (e.g [1]), each one revealing specific characteristics of the data. For example, the Euclidean distance stresses the scale differences, measures based on Pearson's correlation, periodogram or autocorrelation compare the time series regarding their dynamic structure. In order to integrate, in the clustering task, the diversity of available information, we propose to use a linear combination of four dissimilarity measures: Euclidean, Pearson correlation, Periodogram and Autocorrelation based measures. A sensitivity analysis, regarding the weights of the various dissimilarity measures in the linear combination, is conducted to provide a clear view of the weights impact in the clustering results. The K-Medoids algorithm is used to form the clusters. The clustering solutions are selected by simultaneously taking into-account visual representations of partitions (e.g. [3]) and also resorting to cohesion-separation measures. In the implementation we resort to software R and also use specific packages like cluster , TSclust and fpc .
Acknowledgements
--
Keywords