Applying Language Inspired Super-resolution Architecture (LISA) to Hub Height Wind Simulations Over the US West Coast Offshore

As the energy sector increasingly shifts towards renewable resources, accurate forecasting of wind speeds at subseasonal to seasonal (S2S) timescales has become essential for the effective operation and management of wind-energy systems. While short-term weather forecasting has seen significant advancements, predicting wind speeds beyond a 10-day lead time remains a challenge. Our previous analyses have highlighted the complex interactions between large-scale atmospheric circulations and local wind systems, revealing that the influence of these large-scale patterns on local wind energy varies significantly by region.

To address these challenges, we have developed a novel Generative AI model known as the Language Inspired Super-resolution Architecture (LISA). Designed to enhance wind speed forecasting along the U.S. West Coast offshore, LISA leverages large-scale atmospheric data from the ERA5 dataset to model wind conditions. The model applies principles from natural language processing (NLP) to achieve super-resolution performance, utilizing a transformer-based encoder-decoder architecture. In this architecture, linearized image patches are embedded by the encoder and then reassembled into high-resolution weather maps by the decoder, offering notable advantages over traditional convolutional networks.

A key feature of LISA is its use of Masked Auto Encoding, which enables efficient training and allows the encoder to be reused for downstream tasks, similar to the fine-tuning process in large language models. Preliminary evaluations comparing LISA’s output with High-Resolution Rapid Refresh (HRRR) data and lidar measurements from offshore buoys have shown promising results, particularly in capturing wind speed dynamics during significant atmospheric events, such as the atmospheric river on November 16, 2020.