Analysis of training neural compensation model for system dynamics modeling

Incorporating a nominal model as a priori knowledge with a neural network to model a physical system has shown better performance than using only a conventional network model. However, there has been no explicit mathematical reason to describe why this technique, called neural compensation modeling, results in such improvement. The purpose of this paper is to analyze this concept by deriving a mathematical explanation of the neural compensation model compared to a convention network model. The explanation is derived based on normalization procedures of training sets and the properties of norms. In addition, the analysis is illustrated by using a motor coil thermal system