Development of improved adaptive approaches to electricity demand forecasting

Abstract

This paper develops a short-term forecasting system for hourly electricity load demand based on Unobserved Components set up in a State Space framework. The system consists of two options, a univariate model and a non-linear bivariate model that relates demand to temperature. In order to handle the rapidly sampling interval of the data, a multi-rate approach is implemented with models estimated at different frequencies, some of them with ‘periodically amplitude modulated’ properties. The non-linear relation between demand and temperature is identified via a Data-Based Mechanistic approach and finally implemented by Radial Basis Functions. The models also include signal extraction of daily and weekly components. Both models are tested on the basis of a thorough experiment in which other options, like ARIMA and Artificial Neural Networks are also used. The models proposed compare very favourably with the rest of alternatives in forecasting load demand.

Keywords:

• neural networks
• electricity markets
• forecasting comparisons
• state space models
• unobserved components models

Acknowledgements

We are most grateful to the UK Engineering and Physical and Science Research Council (EPSRC: Grant ESA7088) and two British Companies that provided the data used in this project. We also would like to thank two anonymous referees and the editor's comments on a previous version of this paper. Readers interested in further technical details may ask the authors for an expanded version of this paper.