Using time-delay network to classify the actual output.

Is it possible to use a time-delay net to classify a set of features based on their actual and past values?

To be specific i am thinking to use a tapped delay line of [0:N] and modify the default time-delay net structure to implement a pattern recognition problem. In this particular case an observation at the instant t can be classified using the features at instant t but also their past values t-1, t-2, t-3.

I would like to know first if it is a viable implementation and if yes, if it is advisable to start from a time-delay net configuration or a patternnet.

Thanks in advance