Aggregation of Multiscale Experts for Bottom-Up Load Forecasting

The development of smart grid and smart metering infrastructures induces an opportunity to improve electricity consumption forecasts. Bottom-up forecasting strategies aim at exploiting this new individual information to forecast electricity load at another resolution (a region, a portfolio of customers…). We suppose here to have access to a set of time series corresponding to the half-hourly consumption of a group of customers (typically all the customers of an electricity provider or a subset of it). We also suppose to have access to exogenous information on these customers like weather conditions in their area and survey data on their electricity equipment, social class, or building characteristics. In this context, we propose an online learning approach to forecast the total consumption of this group exploiting individual load measurements in real time.