A functional approach to nonparametric forecasting of electricity consumption

In a recent work (Antoniadis et al. (2012)), the authors proposed a prediction model for functional time series in the presence of nonstationarities. This model has been applied to the electricity demand of Electricite de France (EDF). The general principle of the forecasting model is to find in the past similar situations to the present and linearly combine their futures to build the forecast. The concept of similarity is based on wavelets and several strategies are implemented to take into account the various sources of non stationarity. We explore in this second article, three aspects of this model that complement the original methodology while highlighting the industrial usefulness of the method. First we discuss the construction of a confidence interval for the predictor function. Next, we examine the flexibility and simplicity of the model to provide, without extra effort, forecasts horizons further and further away. Finally, in the specific application context, we examine the ability of the method to provide good predictions in the presence of subtle signal nonstationarities induced by loss of customers coming from various scenarios.