direct multi-step (DMS) forecasting

[kaya0007]

Thank you for your excellent work!
Is the aggregated feature z (t) obtained from formula (5) for a single time t, or does it include all past historical times? Is it theoretically feasible to directly map this feature to future multi-step outputs through MLP? How to explain interpretability?
Thank you for your answer!