Neural Network - Multi Step Ahead Prediction

I am very new to matlab.

Objective: To generate a future annual average, min and max temperature data (up to 2022) from existing data range (1983 - 2014) Existed Data arrangement in excel sheet:

           Year    Ta      Tmax    Tmin

           1983    28       34      23
           1984    28.1     34.2    24.2
           ....     ...     ....     ...
           2014     29       35      24

I tried to use nnstart tool (time series); unfortunately, I don't have the knowledge to generate future data (up to 2022)

   Please help me

Answer 1

[Greg Heath]

1. The final OL NARNET model should have 3 dimensional inputs and outputs.

2. The final CL NARNET model will have 3 outputs which are fed back to the input.

3. However, I would first see what happens with the 3 individual series

4 a.Use DIVIDEBLOCK with the default trn/val/tst ratios 0.7/0.15/0.15.

 b. Use the mean and std of the trn subset to normalize 
the whole series  

 c. Use the autocorrelation function of the trn subset 
to determine the statistically significant lags (i.e.,
the lags at which the absolute values of the 
autocorrelation function exceed the 95th percentile of
 Gaussian random noise)

 d. Use as few hidden nodes as possible to design a 
separate OL NARNET model with 

    NMSEtrn = mse(ytrn-ttrn)/mean(var(ttrn',1)) < 1e-4

 e. Use the training record tr  to obtain NMSE for the 
separate trn/val/tst subsets

 f. Close the loop and obtain the resulting CL 
performances 
(trn,val,tst and total)

 g. If the CL performance is unsatisfactory, train netc 
intialized with the final weights of neto.

 5. Once familiarized with the single series problem, 
try some of the 2-input/2-output problems. The 
techniques are similar EXCEPT 
there will be two feedback signals. Therefore, the significant 
lags of the corresponding CROSS-CORRELATION functions may have 
to be taken into consideration.

 6. It is a relatively complicated problem so let us know if 
there are difficulties.}

 Hope this helps.

 *Thank you for formally accepting my answer*

 Greg