add some randomness to multi-dimensional datasets

to work with different site/leadtime series and thus making sure that
there is no silent broadcasting resulting in the correct metric value
even if it worked with the wrong site/leadtime