IMPROVED NEURAL NETWORK PREDICTIVE CONTROL FOR DYNAMIC RESPONSE OF SPATIAL RETICULATED SHELL STRUCTURES

The constitution of neural networks was optimized by empirical equations and experimental comparison, and Levenberg-Marquardt (LM) Algorithm was used to improve the traditional Back-Propagation (BP) Algorithm. Based on the combined application of two neural networks, an improved neural network predictive control strategy was introduced to predict the effective control forces for structural dynamic responses. One network was used to identify structural dynamic responses and the other to predict the effective control forces. The numerical simulation shows that Levenberg-Marquardt algorithm has one order higher rate of convergence than the steepest descent method based on the traditional BP network, and that the introduced control strategy succeeds in identifying structural dynamic responses and predicting effective control forces except for several time steps.