1. 需要加載模型 library(ensembleMOS)
2.模型訓練天數與性能分析
可以看出,由於總共有63天的數據,但並不是使用的數據天數越多越好,而是在40-55天的長度訓練是最佳的長度。
3.模型在擬合時總是出現了異常擬合,報錯信息如下:
Error in optim(pars, fn = objectiveFUN, method = control$optimRule, control = list(maxit = control$maxIter)) :
後來如模型的作者進行了溝通,過程如下:
Dear prof.
Sorry to disturb you.
I am a teacher of China. I want to use the EMOS model in R language, but I found it always presented some errors: " .Error in optim(pars, fn = objectiveFUN, method = control$optimRule, control = list(maxit = control$maxIter)) : ", I do not how to sovle it. Could you give me some suggestions? Please find the test code and data in the attachment.
Thanks a lot for your help.
Yours,
作者很友好,很快回復了:
Dear Colleague,
As you have 8 non-exchangeable ensemble members, your gev0 model has 12 parameters to be estimated. You want to do it from 30 forecast cases, that is data/parameter ratio is less than 3, which is voodoo, not statistics :) That is why the optimization procedure fails. The training length should be increased to at least 120, or you might also consider your members to be exchangeable resulting in just 5 parameters.
Best,
Sandor
增加樣本數量,效果立馬不一樣了。