Aggregation of Sleeping Predictors to Forecast Electricity Consumption

• Activity Number: 636
• Type: Contributed
• Date/Time: Thursday, August 5, 2010 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistical Computing
• Author(s): Yannig Goude*+ and Stoltz Gilles and Marie Devaine
• Companies: EDF R&D and CNRS and ENS ulm
• Keywords: forecasting ; electricity load ; combining ; sleeping experts

Abstract

We consider the issue of on-line combining a set of "sleeping experts". In this problem, the statistician has access to a set of experts to build his own forecast. This experts are said to be "sleeping experts" as they do not provide a forecast continuously. We propose new algorithms based on exponential weights and new oracle bounds in that case. This framework fits particularly well with a lot of industrial forecasting problems and especially to electricity load forecasting: due to software evolution, scientific innovation, data changes or even human intervention in this process, we are faced with the difficulty to observe a constant set of experts during a sufficently large period of time. We apply our algorithms on two data sets: the french and slovakian electricity load. We obtain significant improvement over the operational forecasts.