Stacked ensemble methods for short-term electricity demand forecasting

Judith Foster, Seán McLoone

Research output: Contribution to journalArticlepeer-review

12 Downloads (Pure)

Abstract

When developing load forecasting models a common strategy is to train a range of different models in order to determine the best model for prediction. Rather that retaining only the best model, employing stacked ensemble methods to combine the predictions from all the models can often yield superior prediction performance to any of the constituent models. In this paper we explore four different approaches for generating such stacked ensemble predictions: (1) simple heuristic rules; (2) conformal learning inspired model confidence weighting approaches; (3) optimal model classifiers, and; (4) non-linear regression based approaches. Short-term load forecasting for the Northern Ireland and New York State power systems are used as case studies. Our results show that the non-linear regression based stacked ensemble model yields the most consistent performance across the case studies, achieving reductions in mean absolute prediction error (MAPE) of between 10% and 14% relative to the best performing individual models.


Original languageEnglish
Pages (from-to)3100-3105
Number of pages6
JournalIFAC-PapersOnLine
Volume56
Issue number2
DOIs
Publication statusPublished - 22 Nov 2023
Event22nd World Congress of the International Federation of Automatic Control 2023 - Yokohama, Japan
Duration: 09 Jul 202314 Jul 2023
https://www.ifac2023.org/

Fingerprint

Dive into the research topics of 'Stacked ensemble methods for short-term electricity demand forecasting'. Together they form a unique fingerprint.

Cite this