I have a problem with how to prepare a dataframe in Keras for LSTM with multiple parallel, multivariate time series to predict a single value. So currently I create samples using only Time series A, with a rolling window with given time step, using the column1, column2, column3 from previous steps to predict single output for the most recent value from column3. Below timeseries examples filled with dummy data.
Time series A
| time | column1 | column2 | column3 |
|---|---|---|---|
| 1 | 80 | 1.7 | 4 |
| 2 | 70 | 1.2 | 5 |
| 3 | 87 | 1.0 | 3 |
Time series B
| time | column1 | column2 | column3 |
|---|---|---|---|
| 1 | 100 | 1.5 | 5 |
| 2 | 120 | 2.0 | 8 |
| 3 | 105 | 2.2 | 3 |
Now I would like to add to time series A data from the time series B (or 50 other similar time series) and predict similarly predict single-output (most recent value from column3 from time series A). How can I join multiple parallel multivariate time series together, when using a rolling window?
EDIT:
I've read the great answer of Daniel Möller. The section with "Inputs for sliding windows" for multiple tanks, looks exactly what I need. I am still confused how the different data from multiple tanks are interpreted in this approach.
