Predictability of Wind Ramping Events in the Belgian Offshore Zone: Insights from NWP Models and Post-processing

Offshore wind capacity in the Belgian Offshore Zone (BOZ) is currently 2262 MW. The increasing reliance on wind energy highlights the need for accurate forecasting and effective energy dispatch. Rapid changes in wind power, known as wind power ramping events, pose particular challenges for grid management. To better support energy decision-making, in particular for the Belgian transmission system operator, the Royal Meteorological Institute of Belgium (RMI) has tested a Wind Farm Parameterization (WFP) in their numerical weather prediction (NWP) models to incorporate wake effects. Additionally, machine learning-based post-processing techniques, including a Multi Layer Perceptron and XGBoost, have been applied to enhance wind power forecast accuracy. While these efforts have led to noticeable decreases in overall power forecast errors, the improvements in power ramping predictability remain limited.

In many cases, models can capture the overall trend of power ramps but show timing shifts or slight under-/overestimation in ramp intensity. Rigid point-to-point verification may thus underestimate model skill. To apply a flexible verification framework, we investigated two approaches: the Buffer-Time approach, which allows a timing margin; and the Time-Window approach, which verifies event occurrences within a fixed interval. A power buffer is also introduced to account for small differences in ramp intensity. Using these approaches, we assess the predictability of 15 minute and hourly ramping events for a range of magnitudes of at least 15% of total BOZ capacity, for a 3-year period from June 2021 until June 2024. We compare various NWP models, including the operational RMI Alaro 4 km model, its WFP-enhanced version and the ECMWF HRES model (9 km). Results show that smaller and up power ramps are generally easier to forecast than larger and down ramps. Compared to the operational forecasts, improved modeling and machine learning-based post-processing helps reduce false alarms and better characterize the timing of ramping events. Meanwhile, the results suggest that precipitation has a notable impact on ramp forecast errors, given that many instances of false alarms correspond to precipitation events. In addition, the models, especially machine learning models, have difficulties in capturing extreme ramping events, particularly those caused by European windstorms.