NARX closeloop for prediction: how to select the input

I am using the NARX Neural Network for predicting the wind speed. I have the hourly historical value of the temperature, pressure and wind direction (my INPUT) for the last 5 years and the same for the wind speed data (my TARGET).

I designed an open loop with the following input and target:

inputOL=X(1:end-prediction_horizon)
targetOL=T(1:end-prediction_horizon)

Then, I designed a closed loop (trained using inputOL and targetOL).

My question is: which is the input series that I should use in order to predict my target multistep ahead?

I read several answers which used:

 inputCL=X(end-prediction_horizon+1:end) %used as input of the netc
targetCL=T(end-prediction_horizon+1:end) %used to compare the performance of the netc

But, THIS SOLUTION IS NOT APPLICABLE TO MY PROBLEM BECAUSE, IN THE REALITY, I CANNOT KNOW THE FUTURE INPUT AS I DON'T KNOW TH FUTURE TARGET.

Thus, I tried to use the LAST VALUES KNOWN OF THE INPUT:

newinput=X(end-2*prediction_horizon+1:end-prediction_horizon)

IS IT CORRECT?

DO I NEED TO TAKE INTO ACCOUNT THE DELAY,TOO?

DOES THE prediction_horizon NEEDTO BE EQUAL TO THE DELAY?

Thanks, Cheers

Ilaria