Sager for Meteotemplate?

[adelrosso]

It would be really nice to have a plugin for Meteotemplate that implement Sager forecasting algorithm.
I would like to use my station data not only for statistics but also to have an approximate idea of how time will change locally. Obviously I do not claim to have the complexity of WxSim, I would be happy with a very simple thing like Sager

[Jachym]

WxSim has been in development for decades

Now... I have to say Im not a fan of this because these forecasts tend to be very inaccurate and the way I look at things is rather less, but accurate than more, but inaccurate.

There simply is NO way to predict the weather just based on your station data. You would also need to use information about fronts etc. loaded from some external source. Not to mention the fact that these equations tend to be almost hyperlocal, so the coefficients change from location to location

[Jachym]

I actually remember I was looking at this about half a yr ago. I did some very basic tests using my station and the accuracy was close to what you would expect when doing it purely randomly.

[adelrosso]

Consumer weather stations like Oregon Scientific, that implement sager-like algorithms for forecast the weather 12-24 hours in advance using graphical icons, declare an accuracy of 70-75% Ok, I know it is all crap

[Jachym]

Yes, but to be quite honest I dont care what Oregon does I never look at who does what.

Look at how WU or NWS calculates daily average temperature....

[Jeffm5690]

Probably more accurate to place a rock outside and see if it is cold, wet, etc... lol

[iatros]

....or, since the weather is not changing every day, saying that tomorrow's weather will be the same as today's gives you a chance of at least 75% to be right...

[Jachym]

The problem is that if for example you have a cold front coming, you can see it on satellite images, you can forecast the change, but the pressure is only going to start changing several hours in advance so any forecast based on just measurements would be totally useless.

[wx_jon]

Right now I'm working on some machine learning algorithms for forecasting weather. My first attempt does have the benefit of knowing what a weather model predicts, but I have the goal of making a version based on observations only. The idea would be to have a station's observations and a couple of nearby radiosonde measurements to predict tomorrow's high, low, peak wind, and rain. With the model, I can almost always get within 5 ºF (3 ºC) and 5 knots wind, according to my preliminary testing for some sites in the US. I expect accuracy to be lower with only observations, but who knows? It's a machine!

[Jachym]

Right now I'm working on some machine learning algorithms for forecasting weather. My first attempt does have the benefit of knowing what a weather model predicts, but I have the goal of making a version based on observations only. The idea would be to have a station's observations and a couple of nearby radiosonde measurements to predict tomorrow's high, low, peak wind, and rain. With the model, I can almost always get within 5 ºF (3 ºC) and 5 knots wind, according to my preliminary testing for some sites in the US. I expect accuracy to be lower with only observations, but who knows? It's a machine!

It will work for some scenarios, usually for those where there are no significant changes. For the rest you need satellite data

[wx_jon]

It will work for some scenarios, usually for those where there are no significant changes. For the rest you need satellite data

Well, the nice thing about a machine is that you can give it satellite data too...

[Jachym]

yes, but it would have to be extremely sophisticated then, interpreting graphical images is very complicated, especially because you have to interpret the changes, how the fronts move, how fast etc. not even a static image is enough. It is of course possible and this is how the models work, but it is thousands of pages of code to do that

[Jachym]

or you could use a neural network, that would not be so difficult, but you would need an extremely large set of data and images to train it properly