I have a continious variable I wish to predict, for which I have sensory data on a fixed time interval t, denoted as y_t. In addition, I have a set of features which are consistent across time, called f_t, but also time itself is a feature, which naturally changes over time. The input could thus look as follows
t1| feature1, feature2, feature3, timevalue1 t2| feature1, feature2, feature3, timevalue2 t3| feature1, feature2, feature3, timevalue3
My first question is whether LSTM is suitable for such problem? I have already obtained satifactory results using tradditional ML methods. So this is more 'extra' to see if I can extract additional performance.
Secondly, could a trained LSTM generate the entire time-series without any initial sequence? In other words, provided only the input data at each time t, could it re-create the entire timeseries 'graph' from t1 to tn?
