Parts | Days | Selection | Search | Updates | Downloads | Help

AKPIK: Arbeitskreis Physik, moderne Informationstechnologie und Künstliche Intelligenz

AKPIK 1: Poster

AKPIK 1.2: Poster

Monday, March 20, 2023, 16:00–18:00, HSZ OG2

Predicting volatile wind energy: Stochastic forward modeling and machine learning — Juan Medina, •Marten Klein, Mark Simon Schöps, and Heiko Schmidt — BTU Cottbus-Senftenberg, Cottbus, Germany

Forecasting power output from wind farms is a standing challenge due to complex dynamical processes in the atmospheric boundary layer that manifest themselves by a strong spatio-temporal variability of the wind field. Statistical postprocessing of numerical weather prediction (NWP) ensemble data using machine learning, e.g., by multivariate Gaussian regression, has been utilized to estimate the probability of power ramp events for near-future power grid regulation. However, predictions on the scale of single turbines are not possible demonstrating that there is a lack in modeling for short-term forecasting. In this contribution, this lack is addressed by an economical stochastic modeling approach that autonomously evolves vertical profiles of the wind velocity and temperature. The model aims to reproduce turbulent cascade phenomenology by a stochastic process, respecting fundamental physical conservation principles in a dimensionally reduced setting. As a first step, standalone model predictions of wind field fluctuations in weakly and strongly stratified atmospheric conditions are analyzed by conventional and event-based statistics, including clustering and regression of model output. Forthcoming research aims at developing an economical tool for physics-informed downscaling of NWP data. Coupling with wind power plant models and abstraction by neural networks might hence provide additional physical details to power grid models.