No, that was because a data shift of 1 hour creates much larger data frames, and my laptop at the time did not manage that well.

Of course, in a real live setting, the prediction timeframe is set by the demand of the task and not your laptop power :-).

Also, the longer you predict in the future, the more difficult it becomes if the data-changes of a failure are not of low frequency. Your model needs to become more sophisticated and thereby error-prone. Most likely, you also have to put more effort in the data preprocessing. For the article here, that would have been too much.