NARX parameter optimisation problem

I am trying to train a neural network with charge/discharge data from battery testing. My goal is to predict SOC (state of charge).

My input data is: [U, I, temperature, power]

My target data is: [SOC, energy]

The important target data (output) is SOC, however I included energy as an output to increase performance.

My question is how to choose input delay, feedback delay and hidden layer size.

Searching the mathworks forums, I've found the following:

1. Choose ID using target/input cross-correlation function and FD using target autocorrelation function

However I've no clue how to actually use the mentioned funcions.

Greg Heath suggested that I should search "newsgroup" for examples, however "newsgroup" doesn't exist anymore so I am unable to fund any examples. Any help with that would be great.

P.S. I've tried running my neural network multiple times with random weights and for some reason at the end of the calculations I get wildly fidderent results. Some are great, but others are very bad. Is there a way I could make them reliably good?