Abridged HCNN
In technical applications it might not be possible to unfold an HCNN across the whole history. We thus want to only choose a certain Memory Length. It has to be at least double the length of the forecast horizon. The final network should be applicable to new forecasts (not only on the current history) without retraining.
Learning
Learning is only done in the Architectural Teacher Forcing part of the network. It aligns the windows to the target series in the hidden state.
Partial Teacher Forcing would only degrade the alignment. Thus if you still want to use it you have to extend the past unfolding.
Initial State
For abridged HCNNs you must also use Adaptive Uniform Noise for the initial state to harden the model against this unknown state. We can always include future information by adding a second matrix that has the weights for external factors that we can use for future forecasts.