A general purpose neural network architecture for time series prediction

The paper describes an innovative neural network architecture which is particularly suited to time series prediction applications. The system, which based on a fully connected recurrent 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 against performance achieved by a bespoke Kalman filter. The paper also describes 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