Design of a General-Purpose MIMO Predictor with Neural Networks

Abstract

A new multi-step predictor for multiple-input, multiple-output (MIMO) systems is proposed. The output prediction of such a system is represented as a mapping from its historical data and future inputs to future outputs. A neural network is designed to learn the mapping without re quiring a priori knowledge of the parameters and structure of the system. The major problem in de veloping such a predictor is how to train the neural network. In case of the back propagation algorithm, the network is trained by using the network's output error which is not known due to the unknown predicted future system outputs. To overcome this problem, the concept of updating, in stead of training, a neural network is introduced and verified with simulations. The predictor then uses only the system's historical data to update the configuration of the neural network and always works in a closed loop. If each node can only handle scalar operations, emulation of an MIMO mapping requires the neural network to be excessively large, and it is difficult to specify some known coupling effects of the predicted system. So, we propose a vector-structured, multilayer perceptron for the predictor design. MIMO linear, nonlinear, time-invariant, and time-varying systems are tested via simulation, and all showed very promising performances.