A0343
Title: Probabilistic forecasting and reconciliation of wind turbine power
Authors: Sven Pappert - TU Dortmund University (Germany)
Antonia Arsova - TU Dortmund University (Germany) [presenting]
Abstract: New models are explored for probabilistic forecasting of hierarchical time series with an application to wind turbine power production. Considering different levels of (cross-sectional) aggregation of individual time series may yield a high-dimensional hierarchy where forecasts at each aggregation level are required. A desirable property of such forecasts is that they obey the same linear restrictions prescribed by the hierarchy as the original time series: e.g., the sum of wind power forecasts for all districts in a given region has to sum up to the forecast for that whole region. One way to achieve this coherency is forecast reconciliation. Base probabilistic forecasts for the individual time series are obtained using the classical ARIMA-GARCH model and the recently introduced MAGMAR-Copula model. Existing approaches (score optimal and bottom-up) to reconcile the base forecasts are compared, and the results are evaluated by the CRPS and the energy score. Furthermore, new reconciliation approaches are explored by incorporating nonlinear instead of linear transformations of the base forecasts to form the reconciled ones. The question of optimality of reconciliation methods for probabilistic forecasts is also explored, and different distance-based criteria are considered.
