A REAL-TIME RECURRENT LEARNING NETWORK STRUCTURE FOR DATA RECONCILIATION

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 a re often used as modelling tools, since they easily map a variety of i nput-output patterns. Although they have a lot of advantages over othe r, more classic modelling techniques, neural networks also have a numb er of shortcomings. Training and collection of relevant training data is critical to obtain a good performance model and although they are s aid to be insensitive to the availability of sensor data, the practica l use of neural nets shows that this is hardly the case. Training of t hese networks becomes difficult and network performance reduces rapidl y owing to lack of sensor data. To cope with this kind of problem a ne twork structure for Real-Time Recurrent Learning Networks was develope d. Two recurrent networks, a model network and an identity network, ar e merged into one large, modular recurrent net, which combines robustn ess to lack of input data with a high modelling performance. This tech nique has been tested on a real-life modelling problem from the chemic al process industry. (C) 1998 Elsevier Science Limited. All rights res erved.