A deep learning approach for improving spatiotemporal resolution of numerical weather prediction forecasts

This study addresses the limitations of traditional numerical weather prediction models in wind forecasting for aviation operations by introducing a deep learning approach based on a spatiotemporal fusion model that enhances the temporal resolution and accuracy of wind forecasts. Specifically, the model integrates Global Forecast System (GFS) with European Centre for Medium-Range Weather Forecasts Reanalysis v5 (ERA5) data, employing a 1-dimensional convolutional layer for spatial data fusion and a bidirectional long short-term memory network for spatiotemporal pattern recognition. The presented approach considerably improves upon the numeric model, increasing temporal resolution from 3-hour to 1-hour intervals and reducing mean absolute error by over 50% for wind speed and direction forecasts. The proposed model achieves 82.85% accuracy in wind direction predictions within a 20° angle, compared to 64.46% for the GFS model forecasts. Case studies demonstrate the proposed model's superior performance in capturing wind variability, particularly in complex topographical settings like Madeira International Airport. These improvements have relevant implications for aviation safety, flight planning, and fuel consumption optimization. The geographic independence of the proposed approach suggests potential applicability across diverse regions.