Under CopyrightBeinert, DominikDominikBeinertSchütz, Johannes Maximilian FranzJohannes Maximilian FranzSchützBraun, AxelAxelBraun2024-10-222024-11-112024-10-222024https://doi.org/10.24406/h-477892https://publica.fraunhofer.de/handle/publica/47789210.24406/h-477892The increasing integration of renewable energy, particularly wind power, necessitates enhanced forecast accuracy for effective grid management. Traditional methods that project wind power forecasts from a limited number of reference wind parks to larger grid areas often struggle with insufficient measurement data. We propose a zero-shot learning model utilizing publicly available wind park master data and local weather forecasts to generate locally distributed forecasts at new locations without prior training. This approach employs multi-task learning to embed parks based on master data, enabling comprehensive coverage when projecting forecasts across all wind parks. Our method, validated with field data, significantly improves forecast accuracy at the transmission system operator (TSO) control zone level compared to traditional methods and physical models, offering a robust solution for the growing demands of renewable energy forecasting.enzero-shot learningtransfer learningpower forecast aggregationmulti-task learningmaster dataconference paper