How to generate scenaio？(Is Monte Carlo or Latin hypercube sampling or other methods)

[Jiuhaoqiu]

Although this is a scene reduction program, I would like to ask everyone what method is used to generate the scenaio? (Is Monte Carlo or Latin hypercube sampling or other methods). Can you provide reproducible code or github links for wind power scenarios for reference？

[nepslor]

Hi Jiuhaoqiu,
unfortunately I cannot provide you with a link, since the generation of scenarios depends on the assumptions you do when modeling the uncertainties. For example, if you predict the scenarios by means of multivariate linear regression, under the assumption of Gaussian errors, you could simply retrieve the scenarios using Montecarlo simulations: just retrieve the standard deviation for each step ahead and draw independent observations. Unfortunately, this method do not take into account temporal interdependencies (you sample independently from different step ahead), hence you would need to reconcile the statistical properties of the scenarios using some techniques, e.g. empirical copulas. A more sound way could be to use a forecasting method which exploits bagging for the forecasting, and then directly use the response from single forecasters as the scenarios to be reduced. For example, you could fit a 200 trees random forest using TreeBagger class in Matlab https://it.mathworks.com/help/stats/treebagger.html and then predict the response of each tree in the forest, or use the estimators_ attribute of the skgarden.quantile.ExtraTreesQuantileRegressor class in python https://scikit-garden.github.io/api/#skgardenforest_1 to do the same in python.
Best,
Lorenzo

[Jiuhaoqiu]

Hi Jiuhaoqiu,
unfortunately I cannot provide you with a link, since the generation of scenarios depends on the assumptions you do when modeling the uncertainties. For example, if you predict the scenarios by means of multivariate linear regression, under the assumption of Gaussian errors, you could simply retrieve the scenarios using Montecarlo simulations: just retrieve the standard deviation for each step ahead and draw independent observations. Unfortunately, this method do not take into account temporal interdependencies (you sample independently from different step ahead), hence you would need to reconcile the statistical properties of the scenarios using some techniques, e.g. empirical copulas. A more sound way could be to use a forecasting method which exploits bagging for the forecasting, and then directly use the response from single forecasters as the scenarios to be reduced. For example, you could fit a 200 trees random forest using TreeBagger class in Matlab https://it.mathworks.com/help/stats/treebagger.html and then predict the response of each tree in the forest, or use the estimators_ attribute of the skgarden.quantile.ExtraTreesQuantileRegressor class in python https://scikit-garden.github.io/api/#skgardenforest_1 to do the same in python.
Best,
Lorenzo

Thanks a lot,I will try.