The operator LabelTrend2Classification does exactly this; but if your attribute is not the label you might need to designate it as such, using the ChangeAttributeRole operator, apply the LabelTrend2Classification operator, and then change back to the original label.
Actually there is also a ChangeAttributeType operator which appears to do the same as the ChangeAttributeRole, so two for the price of one!
Hi Andre, did you already applied a windowing function onto your time series? Then you could use feature construction and a UserBasedDiscretization to fullfill your goal. Lets say you have something like that as an example
time-3 time-2 time-1 time-0 30 40 45 52
You could then construct a new attribute difference using time-0 - time-1. time-3 time-2 time-1 time-0 difference 30 40 45 52 7
Using a UserBasedDiscretization you could wrap everything between -5 and 5 to stable and the rest onto down or up.