Presentations

Abstract

The integration of deep learning into geoscience is driving groundbreaking innovations. Among these, super-resolution (SR) techniques using generative artificial intelligence (AI) have emerged as transformative tools, enhancing coarse datasets into high-resolution outputs at a fraction of traditional costs. In a study to be presented at WindEurope 2025 and published in the Journal of Physics Conference Series (March 2025), we introduced WindDM, a diffusion model-based SR framework for generating 100-m microscale wind resource maps from mesoscale inputs. This work utilized mesoscale data from the Weather Research and Forecasting (WRF) model at 800-m resolution and high-fidelity Large Eddy Simulation (LES) outputs from Whiffle at 100-m resolution. The training dataset spanned six diverse geographic domains, representing complex terrain and offshore environments, with extensive data augmentation to enrich the dataset. Built on an 18-layer U-Net architecture optimized for 8x SR, WindDM demonstrated the ability to reconstruct fine-grained wind patterns with accuracy comparable to LES but at significantly reduced computational costs. While these early results were promising, achieving an 8x SR capability for mesoscale-to-microscale transitions, this presentation highlights significant advancements and refinements that address key limitations and broaden the model’s applicability. KEY UPDATES AND ADVANCEMENTS * Expanded Training Dataset: The training dataset grew by 400% through additional high-resolution LES and mesoscale simulations, improving the model's ability to generalize across varied terrains and atmospheric conditions. This included full-year LES simulations across diverse geographic regions to capture seasonal variability. * Enhanced SR Range and Fidelity: Experiments with SR factors ranging from 8x to 32x allowed seamless transitions from global to microscale domains. Trade-offs between SR factors and model accuracy were explored. * Improved Loss Functions: New physics-based and probability-based loss strategies, including Physics-Informed Neural Networks (PINNs) and probabilistic constraints, ensured adherence to Navier-Stokes equations, significantly enhancing the physical realism of outputs. * Optimization Techniques: Advanced techniques like residual learning and adaptive noise schedules reduced training times and improved inference speeds by 25%, enabling faster mesoscale-to-microscale transitions. * Expanded Validation: Validation efforts extended to multiple U.S. sites with long-term met tower data, confirming WindDM’s superior performance over traditional methods. In this presentation, we summarize these recent advancements and present updated validation results, showcasing WindDM’s ability to accurately replicate wind speeds and directions on a timeseries basis. The refined WindDM framework represents a paradigm shift in how wind resource data can be generated and utilized. By combining advanced diffusion models with physics-informed constraints, we demonstrate how high-resolution microscale wind maps can be produced at unprecedented speed and cost efficiency. Future work includes further optimization of the SR framework, expansion to solar resource mapping, and integration into real-time forecasting systems. These efforts aim to democratize access to high-fidelity atmospheric data, advancing renewable energy optimization globally.