From gitter thread:

Is your feature request related to a current problem? Please describe.
"I have multiple univariate target TimeSeries. I want to train models on different "levels" - e.g. some local models on the univariate TimeSeries and some global models on a higher level, combining multiple TimeSeries. Then I'd like to use learned ensembling to combine those models' predictions in a smart way."

Describe proposed solution
The EnsembleModel constructor should accept combination of Local and Global model. This will require some checks to make sure they are compatible (fit method already take care of raising exception for the training on multiple series and past-covariates functionnalities).

Describe potential alternatives
Cannot think of another approach...