Experiment closes gap in weather forecasting | Hacker News Reader

Experiment closes gap in weather forecasting

phys.org

60 pointsdnetesn10 comments

Open articleView on HN

This would clearly be amazing. 25 days is approximately the amount of time needed to do these things:

* Make travel plans

* Grow a head of lettuce

* Schedule and reroof a home

* Ship goods from China to West Coast, USA

Seems unlikely that we'll suddenly leap from murky, mercurial 10 day forecasts to 25 days, but I'm looking forward to when we can.

Better planning of power prices and demand response for power grids with a high fraction of weather dependent renewable energy production.

> mercurial

Nice.

I think I found what it is talking about here: http://wxmaps.org/subx_custom.php

The UX is... primitive, but it does produce seemingly plausible results for the location I tried.

It's not clear from the article what is novel about subx forecasting beyond its time frame and potential applications. Are there other sources that dive deeper into its mechanics and clarify the distinguishing features that would allow it to produce usefully accurate predictions on these more distant futures?

This is the journal paper:

https://journals.ametsoc.org/doi/10.1175/BAMS-D-18-0270.1

It looks like they are focusing on "sources of predictability", i.e. weather phenomena on sub-seasonal scale that can be observed and then it can be predicted what will happen next.

So the methodology works better in certain areas of the world where well-defined sub-seasonal oscillations in the weather are found.

They have used an ensemble of different models, and have done predictions on several decades of weather data, as well as one year of "live" predictions and comparison to what actually happens.

From what I can tell, it seems like the best predictions can explain ~16% of the observed variation in weather.

>weather conditions 3-to-4 weeks out will soon be as readily available as seven-day forecasts

Get ready for a month of pre-hurricaine coverage before every storm.

This could be very valuable data for insurance companies and futures markets.

I know this might sound like I'm joining the hype train but has there been new AI models for weather forecasting that NOAA uses? If so, how does it compare to the traditional models?

I haven't been too close to this recently, but a common technique is called "Model Output Statistics" where the output from these computationally-intensive physical models are fed into another model, trained on historical observations and tuned to your specific risk profile and timeline. These secondary statistical models can be neural nets, or simpler models.

So you're not going to try to teach a CNN about Coriolis forces or boundary-layer mixing physics, but you can teach it that the physical model is more likely to be wrong in this situation, or that the physical model is consistently wrong in your specific location compared to your observations.

Reply on news.ycombinator.com