Optimum aeroelastic control via iterative neural network training for wind-resistant cyber–physical buildings

Abstract

This research presents iterative optimum training (IOT), which integrates deep neural networks (DNNs) and population-based optimization techniques such as genetic algorithms (GAs). The proposed technique reduces the number of experiments needed for training without adding complexity compared with non-iterative DNN-GA techniques commonly used in the literature. In this work, IOT is used to train an optimal controller for minimizing wind-induced vibration (WIV) using distributed aerodynamic actuators. Wind tunnel experiments of a scaled cyber–physical aeroelastic building model are used to demonstrate a novel application of the technique. IOT trains a DNN to approximate building vibration at different wind conditions and actuator orientations using an initial set of experiments. After this initial training, a GA uses the DNN to predict actuator orientations that minimize WIV for the given wind condition. A group containing best orientations from the GA and uniform random orientations is used to perform additional experiments and training of the DNN to enhance exploitation and exploration. This process is repeated until the stopping criteria is achieved. This paper includes results of a benchmark study comparing IOT to GA and DNN-GA techniques. Experimental results show that IOT-based online control of the aeroelastic model reduces WIV acceleration amplitudes by up to 90% within 9.8 s upon controller activation.