I am currently working with Gated Recurrent Units for time series prediction. Anyhow, my question should apply to all kinds of such recurrent networks.
In my data, I have timestamps of 15 Minutes, meaning 96 datapoints represent a day. The goal is to predict the progress of the next two days of a time series, meaning 96*2=192 data points.
As far is I know for now, in order to predict these 192 datapoints I have to set 192 output neurons at the end of my architecture each representing a datapoint (for example using Keras this would be a dense layer with 192 neurons on top of a GRU-layer).
The questions are now:
- Is using 192 output-neurons the only chance, or is there some kind of "sliding" functionallity within recurrent networks which I can use to, for example, consecutively predict each day (meaning 96 datapoints for the first and then 96 for the second day)
- In case I have to use 192 output-neurons, is it then mostly appropriate to split my time series in sequences of 192 datapoints in order to keep my input-data the same sice as my output?