An application of neural network in identifying state space model of dynamic thermal behavior of building envelopes

A neural-network-based identification method to determine the state space model of dynamic thermal behavior of building envelope is presented first in this paper. A multilayered neural network is trained by the samples constructed from the experimentally measured input and output data of wall's thermodynamic system, and then the Markov parameters are produced, which are utilized to realize the reduced state space model expression by eigensystem realization algorithm. The training time is greatly shortened by the adaptive learning algorithm. The results show that this method has some advantages in programming and computational simplicity, very good properties of noise rejection and improved accuracy of the results.