Generation of Calibrated Power Output Scenarios for Offshore Wind Farms with Limited Measurements Using Generative Machine Learning

The increasing integration of offshore wind energy into modern power systems requires accurate probabilistic forecasts that capture uncertainty and temporal variability in order to support reliable and economically efficient operational decisions. This thesis investigates whether generative machine learning approaches can improve multivariate probabilistic wind power forecasting for offshore wind farms, with a particular focus on temporal dependency modeling and transfer learning in data-scarce settings. Two generative post-processing methods from recent literature are implemented and compared with established baseline approaches. In addition, a novel onestage framework, the One-Stage 4h Conditional Generative Model (CGM), is proposed, which directly generates calibrated power scenarios from numerical weather prediction ensembles and jointly learns marginal distributions and temporal dependencies.
Forecast quality is evaluated using the Continuous Ranked Probability Score (CRPS), Energy Score (ES), Variogram Score (VS), and an event-based ramp analysis. The practical relevance of the forecasts is assessed through a stochastic optimization application involving a wind-powered electrolyzer. Transfer learning strategies based on wind farm embeddings and feature-wise linear modulation (FiLM) are further examined for wind farms with limited historical data.
Results show that the One-Stage 4h CGM outperforms multi-stage approaches in multivariate calibration, temporal dependency representation, and economic performance, while generative post-processing methods do not provide substantial improvements over strong benchmark models. Transfer learning improves performance in low-data regimes but becomes less beneficial as more site-specific data becomes available.