A1011
Title: Probabilistic forecasting and forecast reconciliation for wind power production
Authors: Sven Pappert - TU Dortmund University (Germany)
Antonia Arsova - TU Dortmund University (Germany) [presenting]
Abstract: Forecast reconciliation is applied to ensure that forecasts for multiple time series on different levels of a hierarchy conform to the linear restrictions prescribed by the hierarchy. When reconciling probabilistic forecasts, this linear restriction is to be enforced on the distributional level. Building upon a previous study approach, where their reconciled forecasts are constructed as a linear function of the base-level forecasts, the reconciled forecasts are constructed using a feedforward multilayer perceptron (MLP) neural network. In order to justify the use of such non-linear reconciliation scheme, it is investigated under which circumstances the linear strategy is sufficient (and optimal) and when non-linear generalizations, such as the MLP, are needed. As an empirical application, the different reconciliation strategies are compared when applied to probabilistic one-step ahead forecasts for German wind-power production at three different spatial hierarchical levels.
