I would consider a max_train_size option to TimeSeriesSplit.

nice there's some info here:
http://topepo.github.io/caret/data-splitting.html#data-splitting-for-time-series
but this image cover 99% of timeseries cases (maybe one more variable about "rolling window step size" instead of rolling fixed +1, considering that each "line" ("step") of dataframe is one "timestamp"/"step") : http://topepo.github.io/caret/splitting/Split_time-1.png

i checked TimeSplit source code, there's a group not used
some dataframes are:

sensor_id,timestamp,value

maybe this could help split "groupying by" sensor_id, or something like it?
for example...

sensor_id,time
1,1
1,2
1,3
2,1
2,2
2,3

we should do:
[id=1,time=1] + [2,1] as "train", [1,2]+[2,2] as "test"
[1,2]+[2,2] as "train", [1,3]+[2,3] as "test"