Using dividerand or divideblock for sequential inputs to be used in a dynamic NARXSP network

Hi, I am currently using Neural network toolbox in MATLAB R2009a. For my application, I need to create a dynamic NARX network. I am first creating a series-parallel network with newnarxsp and then converting with sp2narx function to a parallel network. This problem needs the inputs to be given as sequential inputs. So though I have 5 inputs with 3000 data points each, the function con2seq makes it 1 cell array with 3000 elements.(Each cell element has the values of 5 inputs of that particular instant.)

When I use dividerand function, it sees the input as 1 cell array and classifies it as the training set. So basically, all my 3000 data points comes inside the training set. So am not able to divide my input into validation and testing sets. Other divide functions like divideblock however deliver an error, the reason being the same. My network generalisation is poor for this reason. What can I do to improve generalisation for a dynamic network?

Thanks, Ramakrishnan.