Machine learning and forecast models

[Ryan Pfeifer]

Ran into this paper, which sparked my interest:

http://aditya-grover.github.io/files/publications/kdd15.pdf

I'm not familiar very familiar with forecast models, and I imagine fluid dynamics will remain the main workhorse, but I also imagine there are opportunities in various stages of the models where machine learning techniques could potentially offer benefits. Maybe it's still too computationally expensive at the moment for the major models to explore.

Is anyone familiar with this topic?

 

[MClarkson]

There is usually a significant gain in accuracy when using statistical techniques in the post processing phase to improve upon pure fluid dynamic model output. Such techniques involve bias correction, regression, combining multiple models, a bit of observational nudging in the short range, perhaps some climo nudging in the long range. Generally they are referred to as "model output statistics (MOS) and have been around for a while. It is certainly not a new field, although there are often ways to improve upon existing work, especially for new variables/locations.

 

[Ryan Pfeifer]

I guess the part I was thinking of that might be new or different would be employing deep learning techniques, as the model in the paper does.

 

[Jeff Duda]

Machine learning is certainly an active area of research for severe weather forecasting. A colleague of mine at the University of Oklahoma is focusing on applying machine learning methods to storm-scale forecasts to improve tornado, hail, and heavy rain predictions. One paper recently published: http://journals.ametsoc.org/doi/abs/10.1175/JAMC-D-11-060.1

My understanding is that machine learning is essentially an advanced post-processing technique based on statistics and logic...very complicated logic. Ultimately, however, the success of [ul]any[/ul] numerical forecast is still rooted to the skill of the underlying numerical weather prediction model. Post processing techniques can only improve so much upon raw model output. But it's certainly a worthwhile research area. I suspect it's not as computationally expensive as running a model at "half-resolution" (half the horizontal grid spacing and time step, which requires at the very least, an 8-fold increase in computer power to be able to run in the same length of wall clock time), which could be considered a minimal improvement in the fundamental technology/build of any NWP model, which makes machine learning even more desirable. I suspect we will start seeing more machine learning algorithm output for storm-scale ensemble models in the near future, especially if what I'm seeing at conferences and seminars in the NWC is any indication of what's coming down the pike.