RNN LSTM that uses different timeseries as input and output

I’m currently working on simulating an optimization algorithm that involves using an hourly timeseries (input_df) to calculate 36 different hourly timeseries (output_df). My approach involves creating an RNN LSTM with input=1 and output=36. However, for simplicity, I’ve initially set the output to 1.

I’ve observed significant improvements in results during model training when incorporating the previous 24 observations of the input_df and the previous one observation of the output_df to predict the next input_df value (Y(t) = f(Y(t-1), X(t),…,X(t-24)).

My challenge arises when I attempt to make predictions for other timeseries. Specifically, I’m unsure of how to configure the RNN to utilize the previous observation as input. Additionally, I’m uncertain about how to initiate the entire process, especially considering that for X(1), I don’t have Y(1).

I would greatly appreciate any insights, advice, or suggestions on how to address these issues and optimize the RNN for predictions across various timeseries.