Dynamic Modeling of Forecast Errors for Electric Consumption

Despite the high standard of forecast regarding electricity consumption, forecast errors are inevitable in a predictive excercise. These errors are not time-independent noise, since they are caused by unanticipated physical phenomena. Modeling the evolution of these forecast errors could therefore help improve the way in which they are compensated for in real time. After highlighting the complex temporal structure of day ahead forecast errors, this contribution proposes several time series models whose statistical quality is discussed. These models are then used in an value-oriented approach to solve the optimal control problem of a storage aiming at compensating for forecast errors. Although simplistic, this application highlights that the statistical quality of a model is not systematically linked to its added value in an optimal control application context.