A Hybrid LSTM--Vision Transformer Architecture for Predicting HRRR Forecast Errors
Researchers have developed a novel hybrid LSTM-Vision Transformer (LSTM-ViT) architecture to improve the prediction of forecast errors in high-resolution numerical weather prediction (NWP) systems. This new framework integrates temporal sequence learning from surface observations with atmospheric profile data, outperforming previous LSTM-only models. The LSTM-ViT demonstrates a twofold increase in predictive skill for precipitation forecast errors and better captures complex atmospheric phenomena like PBL activity and convection. AI
IMPACT This hybrid architecture could lead to more accurate weather forecasts by better predicting model biases and forecast confidence.