Ultra-Short-Term Probabilistic Wind Forecasting: Can Numerical Weather Predictions Help?

Ultra-short-term wind forecasting (i.e. wind speed and power predictions issued for sub-hourly forecast horizons), are pivotal to the effective management and integration of wind farms into modern-day electricity systems. The dominant consensus in the forecasting literature and practice is that data-driven approaches may be best suited for such short-term horizons. This is in contrast to numerical weather predictions (NWP), or hybrid models thereof, for which the value is typically substantiated at relatively longer horizons (>1-3 hours). We propose a probabilistic data-driven model that actually makes use of NWP information (albeit indirectly) for ultra-short-term wind speed and power forecasting. Instead of directly using NWPs as input regressors (as in hybrid approaches), we indirectly invoke NWP information in selecting key parameters within the data science model, thereby guiding it to adhere to certain physical principles related to local wind field formation and propagation. We show that such indirect integration of NWPs within our data science model outperforms several prevalent forecasting methods, including but not limited to persistence forecasts, which are known to be highly competitive at ultra-short-term horizons. This work serves as an exemplar for leveraging the rich, yet coarser-resolution information of NWPs in benefiting data-science-based ultra-short-term wind forecasting models.