Vendredi 16 mars 2012
Adaptive GAM models for Day-Ahead and Intra-Day Electricity Consumption Forecasts
Generalized Additive Models have been investigated recently to forecasts day-ahead electricity consumptions at EDF R&D. These models achieve an interesting trade-off between accuracy of forecasts and adaptation to different data sets thanks to their semi-parametric structures. We propose here a new method based on QR decomposition (joined work with S. Wood) to learn this models on-line as we receive new data. This allows GAM models to react to smooth changes in the data generation process: economic crisis, loss or gain of customers… We illustrate it on different data sets and real forecasts.