A Neural Network Structure for System Identification

Establishing a dynamic process model is the first step toward implementing a modern control algorithm. Because of the complexity of chemical processes, most models are identified, that is, determined from a known input/output sequence. Furthermore, models are usually linear and time invariant. This research focuses on the application of neural networks to the development of dynamic models. In particular, this paper presents a modification of the layered structure used most commonly with the Backward Error Propagation algorithm The modification is the addition of a set of weights connected directly from the input to the output layer, weights which contribute in a linear manner to the network output. This creates a number of advantageous compared to traditional structures, including initialization of network parameters based on process knowledge, additional insight to the leaning algorithm, and enhanced extrapolation outside of examples the learning data set.