Talk / Overview

Materializing deep decarbonization scenarios requires robust insight into the potential of renewable technologies to meet ambitious sustainability goals. As a key enabling technology, wind power is expected to play a prominent role in energy transition. Exploration of the role of technology innovation will be critical for identifying viable solutions at scale, however, conventional high-fidelity methods for evaluating wind power generation are computationally expensive and therefore intractable for national scale applications and for comprehensive evaluation of design space. As a cost-efficient surrogate, we propose a flexible graph neural network (GNN) model for describing turbine wake interactions and predicting power generation for arbitrarily designed wind farms under a broad range of atmospheric inflow conditions and for innovative control strategies. The GNN perspective represents the wind farm as a graph with individual turbines representing nodes and directed edges encoding downwind wake effects. Previous wind farm surrogate models limit the degree of design flexibility by relying on parameterized embeddings to describe turbine layouts. By leveraging emerging GNN technologies, this approach enables rapid modeling of generalized wind farm layouts as well as heterogeneous technologies. We demonstrate the generalizability of the trained model for diverse layout configurations and for divergent atmospheric conditions at millions of potential locations. The trained model can inform diverse applications including national technical potential, localized wind farm siting through layout optimization and real-time operations through yaw controls.

Talk / Speakers

Dylan Harrison-Atlas

Senior Researcher, National Renewable Energy Laboratory

AMLD / Global partners