Intradaily Smoothing Splines for Time-Varying Regression Models of Hourly Electricity Load

• Activity Number: 423
• Type: Contributed
• Date/Time: Tuesday, August 3, 2010 : 2:00 PM to 3:50 PM
• Sponsor: Business and Economic Statistics Section
• Author(s): Virginie Dordonnat*+ and Marius Ooms and Siem Jan Koopman
• Companies: EDF R&D and Vrije Universiteit Amsterdam and Vrije Universiteit Amsterdam
• Keywords: electricity demand ; dynamic regression ; state-space model

Abstract

We develop a multivariate model for daily vectors of hourly electricity demand within the linear Gaussian state-space framework. We model the daily pattern by means of cubic splines. Then we build dynamic regressions associated to the spline knots for all other components: yearly and weekly seasonals, weather effects. We therefore build a dynamic factor model with cubic spline weights as loading matrices. Given the large number of unknown parameters to estimate, we exploit recent developments for tractable simultaneous maximum likelihood estimation. We apply our methodology to a ten-year dataset for France and we investigate the components of demand and their evolution over time via signal extraction as well as we check forecasting accuracy. Results are compared with the 24 corresponding independent univariate models.