Detection of remarkable values in Individual electric consumption´s series using non-parametric approach

The portfolio of the majority of the European subsidiary companies of EDF contains hundred of customers, large-scale consumer are generally non-thermo sensitive. The methods of forecasts of consumption applied to these portfolios are generally based on agglomerated individual forecasts, themselves based on models like exponential smoothing, ARIMA or Holt and Winter\´s methods. They can be even simpler with the "naive" method witch consists in reproducing the last known weekly data to envisage the week to come. These methods are particularly sensitive to the presence of remarkable or aberrant values, particularly for the method known as "naive". This paper aims to describe an automatic method to identify aberrant or remarkable values, abnormally large or small values present along individual series of consumption. This method, based on a nonparametric approach, is adapted to the training of a great number of individual curves because it makes it possible to be freed from the preliminary choice of a class of model.