For anyone interested: I ended up merging the train and test set, then generating the dummies, and then splitting the data again at exactly the same fraction. That way there wasn't any issue with different shapes anymore, as it generated exactly the same dummy data.