Pruned-MLP neural network model of 350MW power generation

For modeling the nonlinear identification object of large thermal power generation unit, this paper present an NNARX model structure and corresponding system identification method based on pruned multi-layer preceptron (pruned-MLP) neural network. For the first time Lipschitz quotients method for identifying model orders of input-output system with unknown nonlinear dynamics is introduced into NNARX neural network model structure identification. As for the model parameter estimation, a modified Levenberg-Marquardt algorithm is proposed to train the network robustly and rapidly. Particular attention is paid to the model validation and neural network generalization error estimation, which is used as stop criterion of OBS network pruning. Optimal Brain Surgeon network pruning algorithm is discussed and used to obtain the final optimal network structure. Finally the identification method addressed above is applied successfully in a real 350MW power generation unit of Bao Steel power plant.