Multiple Time-Series with Different Intervals

I'm trying to create a TFT model that is supposed to predict the value of a certain time-series (the target) by getting a few other time-series as inputs (over the period defined by the encoder_length), The problem is that each time-series has a different interval :
time-series1 - interval of 1 second between every two values
time-series2 - interval of 1 minute
time-series3 - interval of 1 hour
time-series4 - interval of 1 day (this one is the target)

The problem is because the intervals are different the number of samples in each time series is different and the time_idx is not possible to synchronize.

Any idea on how could I feed the model with the last X samples for each time-series to create a prediction? (for example, predict next day's value by previous 5 days values, previous 5 hours, previous 5 minutes & seconds)
Thanks!

Start by reading the time_idx and encoder_length handling for TFT inputs, along with the dataset path that constructs synchronized time-series samples. Determine whether the existing input pipeline supports different sampling intervals and define the expected behavior for the four series described. Done means the supported input shape and prediction window are documented and covered by tests or an example.