Active neural predictive control of seismically isolated structures

Abstract

An online identification and control scheme based on a wavelet neural network (WNN) and model predictive control (MPC) are presented. The WNN comprises a backpropagation neural network with wavelet activation functions and a parallel feedforward term. The WNN is used to identify the structural system, and the model is used to provide the predictions for MPC. The backpropagation network parameters and the controller are trained by the gradient descent algorithm to minimize performance indices. The feedforward component is trained using recursive least squares. The latter is found to drastically reduce the number of hidden layer neurons and significantly reduce the computational load of the neural network. Due to the general structure of the controller, its performance is satisfactory even under the strict condition imposed by a fixed learning rate. The efficacy of the control was demonstrated through a series of computational simulations of a 5-story seismically isolated structure with conventional lead-rubber bearings. Significant reductions of all response amplitudes were achieved for both near-field (pulse) and far-field ground motions, including reduced deformations along with corresponding reduction in acceleration response. In particular, the controller effectively regulated the apparent stiffness at the isolation level.