From U-Net to Diffusion: Smarter Power Forecasts with Adaptive Uncertainty

We use a U-Net to make baseline power forecasts and train a diffusion model on its residuals to capture uncertainty. The diffusion samples naturally show low ensemble spread during stable atmospheric conditions and much wider spread when the atmosphere is more turbulent. This improves both reliability and interpretability compared to using the U-Net alone. The method provides a practical alternative to running full WRF simulations for uncertainty-aware wind farm power modelling.