Hi all.
I actually have two questions, the first one leads to the second.
So, I want to start a research project involving machine learning and weather forecasting. I wrote to some optional partners, but they showed no interest. Probably because I am out of my field here so I dont know where to find partners.
Can anyone direct me to people who might be interested in taking part in such project?
Naturally, the second question is how do you find partners for a research project.
Thanks for your responses
OK, let me present the idea briefly Current weather forecasting systems are not performing very well in short term forecasting This in part is due to the forecasting system wher high resolution models runs after the low resolution ones An experienced person can get better predictions htan the model
My idea is: lets use model forecast and station measurements and try to creat a mid range forecast
what do you think so far?
Do you mean something like https://en.wikipedia.org/wiki/Model_output_statistics ?
Thanks Antonio!
I did know this method.
But I mean something more like trying to train a neural network to make one hour prediction based on the last five hours measurements of a meteorology station.
Next stage will be to add more input such as model output remote sensing etc.
First, to find research partners, you basically have to keep asking until you find the right person. First, evaluate your own expertise & experience: this is what you have to bring to the party. If you realize you are missing some key expertise, find out who has it, and just ask them. You will need to persuade them that (a) your research idea is good, (b) you can contribute some key expertise, (c) they can contribute key expertise, and (d) it is practical for you to collaborate together (e.g., you can both get paid to do this...). A good place to do this is at a conference, especially a big one like the EGU annual meeting.
Second, I would say that short-term forecasting systems perform quite well on a wide-area basis; but as you suggest they can have systematic errors ("bias") for specific station locations, and indeed that bias can be removed by learning from previous forecast performance (machine learning, if you like). I would say that for an initial forecast, the laws of physics beat statistics any time, but then statistical post-processing can improve the forecast further. Here's a good paper that evaluates 7 different ("machine learning") methods for doing this: Article Reducing errors of wind speed forecasts by an optimal combin...
.Here's another paper describing a practical application of such machine learning - you may develop a system similar to, or better than, this: Article Operational Evaluation of a Wind-Farm Forecasting System
I suspect there are other opportunities to add value to NWP forecasts with machine learning post-processing, or to use machine learning in some completely new way - if you have the imagination and technical ability to find them!Good luck!
Enda
Machine learning employs statistical techniques to figure out patterns and based on the patterns established, it forecasts into the future. However, in most of weather forecasting stations, dynamical equations based on the basic and well established scientific facts are used. Because of this reason, still the dynamical method is better than the machine method. But weather is non-linear, chaotic and non-deterministic and the dynamical method is still not sufficient. So, perhaps, there are loopholes that can be patched by statistical or machine learning techniques. And I would recommend, you would use both techniques and try to improve the forecast period ahead of time.
My idea is not to replace the numerical weather forecasting but to imrove it where it can be improved. For instance short term forecasting can be highly improved. Specifically I think that numerical weather forcasting cycle causes short term and regional forecasts to be based on old measurements and I am trying to solve that.
- Badges
- Science topic