Multi-step prediction of a dynamic time series with jumps and decay curves using NARX networks

Hello everybody,

I am trying to predict a dynamic time series with a sequence of sloping curve sections with the help of a NARX network. The aim is to predict the further course of the curve segments for 60 minutes each. The following shows three diagrams with the target variable T and the 3 input variables X1 to X3 for 25,000 minutes:

X1 represents the time after the peak, while variables X2 and X3 are measured process variables. The length of the entire time series is 150,000 minutes.

My basic approach is this:

1) Splitting the time series into a training data set (TD, 120,000 minutes) and a prediction data set (PD, 30,000 minutes).

2) Using divideind to split TD into training, validation and test data sets (84000/18000/18000 minutes)

3) Open-loop training with TD with variation of the hyperparameters

4) Closed-loop re-training with TD

5) Test the MSP with PD.

Does anyone have any suggestion how best to solve this problem?

Kind regards

Torsten