Hello everyone,
I am currently working on wind speed forecasting and conducting tests using Nvidia’s FourCastNet. I find this model incredibly powerful for global weather forecasting, but I have a few questions that I haven’t been able to find clear answers to. I would appreciate any insights from the community!
1) Interpolation Between 6-Hour Forecast Intervals
FourCastNet generates predictions for 20 atmospheric variables at 6-hour intervals on a 721×1440 grid. How can we analyze and reconstruct the missing hourly data between these predictions?
- Should we apply interpolation?
- If so, which method (linear, spline, etc.) would be most suitable?
- Or is there a better approach used in meteorological modeling?
2) Extracting Location-Specific Forecasts from the Grid
Since the model operates on a global grid, what is the best way to obtain forecasts for a specific geographical location (lat, lon)?
- Should we use bounding box interpolation (e.g., taking the four nearest grid points)?
- Or is there a more advanced method (bilinear interpolation, kriging, etc.) that is commonly used for this?
3) Ground Truth Data for RMSE and ACC Evaluation
In the inference script, RMSE and ACC metrics are calculated. But I am unsure about the ground truth data used for validation.
- If I run a forecast for February 28, 2025, at 12:00 UTC, predicting for the next 36 hours (prediction length: 6), how should I obtain the ground truth data for comparison?
- Do I need to fetch ERA5 reanalysis data for those future time steps?
- How is this process handled in a real-time/live forecasting setup?
I would love to hear from anyone who has worked with FourCastNet or similar models. Your insights would be incredibly valuable in improving the accuracy and application of these forecasts.
Thanks in advance for your help!