A real-time recurrent learning network structure for data reconciliation

Abstract

A lot of effort has been put into the modelling of non-linear dynamic systems owing to their presence 'in every day life'. Neural networks are often used as modelling tools, since they easily map a variety of input-output patterns. Although they have a lot of advantages over other, more classic modelling techniques, neural networks also have a number of shortcomings. Training and collection of relevant training data is critical to obtain a good performance model and although they are said to be insensitive to the availability of sensor data, the practical use of neural nets shows that this is hardly the case. Training of these networks becomes difficult and network performance reduces rapidly owing to lack of sensor data. To cope with this kind of problem a network structure for Real-Time Recurrent Learning Networks was developed. Two recurrent networks, a model network and an identity network, are merged into one large, modular recurrent net, which combines robustness to lack of input data with a high modelling performance. This technique has been tested on a real-life modelling problem from the chemical process industry. © 1998 Elsevier Science Limited. All rights reserved.