Use of neural networks to predict the short-term behavior of chaotic time series, including effects of superimposed noise

The predictive capabilities of some simple backpropagation neural networks, as applied to chaotic time series, are investigated using time-series data generated from a three-dimensional numerical model of an electrochemical system. Regulated amounts of noise are superimposed on the originally “clean” chaotic data in order that effects of noise on predictive capabilities can be evaluated. The ability of the neural networks to make short-term predictions of time-series behavior is assessed in terms of network size, extent ahead in time of the prediction, and level of superimposed noise