Machine learning may fail when time period is too small

[delhomer]

Some examples of failing commands:

python -m luigi --module jitenshea.tasks.city Clustering --city lyon --start 2017-09-01 --stop 2017-09-04 --local-scheduler
python -m luigi --module jitenshea.tasks.city Clustering --city bordeaux --start 2017-08-01 --stop 2017-08-04 --local-scheduler

They give following error:

ValueError: Input contains NaN, infinity or a value too large for dtype('float64').

In such a case, it seems that some stations will unsufficient amount of data are not filtered (see e.g. station n°44 in Bordeaux, in the last example: a value for hour 0, but NaN for every other hours).

Workaround: prevent from choosing too close start and stop dates
Fix: NaN management when computing df_norm