High-frequency ensemble wind speed forecasting using deep learning

In this study, we present a deep learning-based method to provide seamless high-frequency wind speed forecasts for up to 30 hours ahead. For each selected site, our method generates an ensemble forecast with an update frequency of 10 to 15 minutes(depending on the observation site’s update-frequency). The main objective in this machine learning based post-processing method is to optimally exploit highly resolved NWP models and particularly utilize their multi-level meteorological parameters to integrate the three-dimensionality of weather processes. Further key objectives of this research are to consider different spatial and temporal resolutions and different topographic characteristics of the selected sites. We evaluate the best praxis for efficiently post-processing both the 10-meter wind speed at selected Austrian meteorological observation sites and wind speed on hub height of wind turbines in wind farms.

The method is based on an artificial neural network (ANN), particularly a long-short-term-memory (LSTM) adopted to process several differently structured inputs simultaneously (i.e., different gridded inputs along with observed time-series) and generate ensemble output. An LSTM layer models recurrent steps in the ANN and is, thus, useful for time-series, such as meteorological observations.

Our ensemble forecast method is evaluated for a case study in 2021 using several years of training, including extreme weather event for the selection of sites. The utilized data includes the meteorological observations, gridded nowcasting data as well as NWP data from ECMWF IFS and AROME at several pressure/altitude levels. Hourly runs for 12 test locations (selected TAWES sites covering different topographic situations in Austria) and two wind turbine sites in different seasons are conducted. The obtained results indicate that the model succeeds in learning from inputs while remaining computationally efficient. In most cases the ANN method yields high forecast-skills and is compared to available methods such as the raw NWP model output, climatology, and persistence.