Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Parameter setting and selection of variables #59

Open
graduatefast opened this issue Apr 18, 2024 · 2 comments
Open

Parameter setting and selection of variables #59

graduatefast opened this issue Apr 18, 2024 · 2 comments

Comments

@graduatefast
Copy link

Thanks for your invaluable work. I'm trying to train the models based on your pseudocode. However, I met some questions, and I will appreciate if you can provide some information:

  1. As you mentioned, the four models with different lead time steps are trained by using the same algorithms. Are the parameter setting of the training of these four models are totally same? I'm also wondering what the training loss values are after the 100 epochs.
  2. If all the variables (both in pressure levels and surface level) are necessary for training the model? If I only use surface variables, will the model performance will be greatly reduced?

Thanks very much

@graduatefast
Copy link
Author

graduatefast commented Apr 18, 2024

By the way, the area I selected is only around 15° * 20°, not the global area. Appreciate if you could kind provide some information.

@198808xc
Copy link
Owner

Hi,

To train good weather forecasting models, both the global data (rather than a region of 15 degs x 20 degs or so) is needed. This is because the future weather largely depends on the outer region -- imagine a winter storm that propagates from west to east. Without seeing weather data from the outer region, it is impossible for the model to "predict" when the storm will come to the region you are interested in.

Also, both upper-air and surface variables are needed. The reason is similar -- it is difficult to predict the surface variables alone without seeing the upper-air variables.

So, I think your training strategy needs to be changed, without which the observed loss values are less meaningful. BTW, we used the same hyper-parameters to train all four models.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants