Improving Narxnet Closedloop Performance

Hi,

I'm having some trouble developing a Naxrnet for streamflow prediction. I'm using a 1xN matrix of inputs (Rainfall) and a 1xN matrix of outputs (streamflow) where N = 61056. The data is on the 15 minute timestep. I have examined the Input-Output Cross Correlation & Output Autocorrelation via NNCORR (script and results below) and ACF,PACF & CCF plots.

[O,N] = size(x); % [1 36633]
[I,N] = size(t);
zt = zscore(t',1)'; % Standardised Z Scores
zx = zscore(x',1)';

% Average Target Variance etc

MSE00 = mean(var(t',1)); % 617.5553
MSEgoal = 0.005*MSE00; % 3.0878
Ntrn = N-2*round(0.20*N); % 36633
Ntrneq = Ntrn*O; % 36633
L = floor (0.95*(2*Ntrn-1)); % 39834

rng('default')

for i = 1:100

        zn=zscore(randn(1,Ntrn),1);
        autocorrn = nncorr(zn,zn,Ntrn-1,'biased');
        sortabsautocorrn = sort(abs(autocorrn));
        thresh95(i) = sortabsautocorrn(L);

    end

meanthresh95 = mean(thresh95); % 0.0101

autocorrt = nncorr(zt,zt, Ntrn-1,'biased');
sigflag95 = find(abs(autocorrt(Ntrn+1:2*Ntrn-1))>=meanthresh95);

diffflag = diff(sigflag95);
Temp = find([diffflag inf]>1);
lengths = diff([0 Temp]);
Ends = cumsum(lengths);

crosscorrxt = nncorr(zx,zt, Ntrn-1,'biased');
sigilag95 = -1 + find(abs(crosscorrxt(Ntrn+1:2*Ntrn-1))>=meanthresh95);

diffilag = diff(sigilag95);
Temp1 = find([diffilag inf]>1);
lengths1 = diff([0 Temp1]);
Ends1 = cumsum(lengths1);

Significant Feedback Lags:

    1  5688
    5751  6790
    6865  7594
    7672  8639
    8782  10347
    10371  10921
    11047  12965
    12988  13010
    13048  13553
    13588  13642
    13806  14050
    14190  14232
    14289  14342
    14400  14463
    14488  17685
    17867  21147
    21231  21253
    21405  21492
    21616  22158
    22293  23568
    23708  23718
    23775  23794
    23883  24095
    24166  24222
    24440  25285
    25358  26412
    26597  26646
    26687  26810
    26882  27160
    27256  36633

Significant Input Delays

   1  4302
  4323  5503
  5508  5786
  5893  6049
  6074  6164
  6210  6416
  6466  6581
  6711  6896
  6943  7012
  7069  7240
  7301  7338
  7352  7497
  7546  7566
  7572  7583
  7621  9214
  9234  14212
  14276  15829
  15880  18179
  18202  18227
  18235  18388
  18445  18550
  18625  18731
  18798  19526
  19555  19563
  19591  19925
  19962  20512
  20645  20686
  20920  21104
  21141  21838
  21944  22111
  22229  22263
  22348  22546
  22678  22695
  22803  23040
  23093  23473
  23513  23584
  23685  23808
  24108  25116
  25253  25273
  25312  25336
  25438  26003
  26115  26199
  26210  26618
  26663  26754
  26812  26946
  26993  27047
  27158  27189
  27215  27326
  27387  27757
  27956  27963
  28035  28243
  28398  28851
  28906  29110
  29220  29569
  29711  29798
  29829  29914
  29969  30089
  30115  30293
  30458  30582
  30650  31215
  31356  31610
  31630  31632
  31723  32068
  32168  32215
  32285  32330
  32628  32707
  32755  32820
  32858  32896
  33212  33475
  33582  34483
  34536  34590
  34662  34676
  34802  34841
  34888  35366
  35386  35785
  35812  35832
  36058  36425
  36492  36632

Based on this I have tried a large number of ID & FD combinations as well as varying hidden layer node. In general, I am happy with the open loop training performance (Example NMSE: 9.58e-5/5.72e-5/2.06e-4 train/val/test) but the closeloop performance is generally very poor even when the closeloop is trained. I have attached a graph of the raw data, training model response and closeloop performance (Model = ID: 1:24, FD: 1:2, H: 2)

At the moment, I'm thinking the issue could be non-stationarity in the data, poor model architecture or some other issue and I'd thus be interested in any ideas you all might have re: how to improve the performance, improve closeloop training etc.

Thanks for your time.

Best