Surrogate models for high performance control systems in wind-excited tall buildings

Abstract

High-performance control systems (HPCSs), including active, semi-active, and hybrid systems, have been demonstrated as promising methods to mitigate a variety of excitations. However, their deployment in the field is still very limited, attributable to reliability concerns in the closed loop configuration. A solution to promote their applicability is the development of an uncertainty-based design procedure, but such solution comes at a high computational cost due to the large number of possible scenarios to consider on both the closed-loop configuration and external load sides. To alleviate the computational demand of such analysis, this paper investigates the use of data-driven surrogate assisted techniques for uncertainty quantification of HPCSs deployed in wind-excited tall buildings. Both a Kriging surrogate and an adaptive wavelet network (AWN) are investigated and compared to map the unknown relationship between structural inputs and responses. The Kriging model exploits an offline batch learning process while the AWN uses an online sequential process. The surrogate models are applied to a 39-story building equipped with semi-active friction devices exposed to wind load and are compared in terms of accuracy and computational time. Two applications of the surrogate models for uncertainty analysis of the case study building are presented. One is for uncertainty quantification and the other for identification of the most influential uncertain variables. Results show that Kriging provides a more accurate representation to map uncertainties to the system response and to quantify the uncertain performance of HPCSs, but that the AWN provides a significantly faster computational alternative. In particular, for a case containing 17 uncertain variables, Kriging found a representation in 3h20, while the AWN converged in 37 min. Under 41 uncertain variables, these metrics increased to 16h20 and 3h22 for Kriging and the AWN, respectively. These representations were leveraged to identify and remove the uncertainties from three key variables yielding high variance in structural response. Results showed that variables identified under Kriging yielded a 34.9% decrease in variance under 17 uncertain inputs and a 22.9% decrease in variance under 41 uncertain inputs, while AWN yielded a 29.0% and 19.8% decrease, respectively.