R. Mitchell (UK)
Clustering, Forecasting, Principal Component Analysis
This paper describes an application of the hybrid mean tracking k-means cluster algorithm[1] for forecasting electricity demand. The overall concept is to cluster some training data, form linear models for each cluster, and to find the closest clusters to each unseen data record and use the associated linear models to forecast the demand. The data reported here were provided for a competition run as part of the EUNITE 2001 conference[2], where the aim was to forecast maximum daily electricity demand in East Slovakia, though other confidential data have also been processed successfully. This paper describes the systematic methodology used to determine suitable training data and pre-processing operations and how the parameters associated with the modelling were selected. The result is a set of forecasts having errors significantly smaller than those produced by the variety of methods used by the 26 entries to the competition.
Important Links:
Go Back
