An easily trained neural model of a distributed parameter system

This paper is concerned with black-box modelling of a distributed parameter thermal system (a long duct) by means of neural networks. A new model structure is discussed which consists of a set of local neural sub-models and a neural interpolator. The local sub-models calculate temperatures for a number of predefined locations of sensors. They are trained independently, using limited data sets. Next, the neural interpolator, using the local temperatures modelled by the sub-modes, calculates the value of the temperature for any sensor location. The interpolator is also trained independently. This paper also discusses the method of choosing which local sub-models should be actually used. It is shown that for the initial structure with 10 sub-models as many as 6 or 7 of them may be removed without significant deterioration of overall model accuracy.