Predicting time series by a fully connected neural network trained by back propagation

Describes work carried out on the use of neural networks for time series prediction. The advantages of a neural approach are discussed, and some possible back propagation architectures described. Detailed evaluation is made of one particular architecture, based on a fully connected recurrent network. The network has been evaluated for both deterministically and stochastically generated time series, as well as real process data. Results are presented for the latter, and comparisons made with the performance achieved by a bespoke Kalman filter The article also discusses the use of a `spread encoding´ scheme for representing input and output data, which enables the network to learn local linearisations on the input data and allows for probabilistic interpretation of the output data