Determining the input time delay τ of a neural network for nonlinear time series prediction

Determining the input time delay τ of a feed-forward neural network (FNN) for nonlinear time series prediction is an aspect to be considered. Now there is a prevailing method to determine the input time delay τ of the neural network, which takes the τ as the optimal time delay τ_s in phase space reconstruction of the dynamic system. This paper presents a different idea that there is no direct relation between the input time delay τ of the neural network and the optimal time delay τ_s, in phase space reconstruction. Considering some other factors, the paper makes sure that taking the FNN input time delay τ as 1 is most reasonable. Some theoretic analyses and validation examples are given.