APPLICATION OF A RECURRENT NEURAL NETWORK IN ON- LINE MODELLING OF REAL-TIME SYSTEMS ∗
Jorge M. O. Henriques, Paulo Sousa Gil, Antonio Carlos Dourado, Hermínio Duarte-Ramos · 1999
ABSTRACT: Given the universal approximation properties, simplicity as well its intrinsic analogy to the non-linear state space form, a recurrent Elman network is derived and applied for modelling non-linear plants. Learning is implemented on-line, based on input and output data and using a truncated backpropagation through time algorithm. Regarding its structural simplicity, previous knowledge based on a linear description of the plant to be modelled might be used for initialising the network weights. The main goal of this work is to emphasise the potential benefits of this architecture for real-time identification. Experimental results collected from a laboratory heating system, for several operating conditions, confirm the viability and effectiveness of the proposed methodology.