Short-term Prediction of Wind Velocity on Bridge Deck

Abstract

In this study, method to predict strong winds on bridge deck using deep learning algorithm, Long Short-term Memory (LSTM) is investigated. The prediction performance of LSTM is evaluated. The prediction accuracy increases when wind velocity data similar to wind velocity data, which is target to be predicted, is used as training data. In order to improve the accuracy of LSTM, the prediction target is specified to strong winds generated during typhoons. As a method to apply similar training data, a method to select similar typhoons to target typhoon is proposed. Similar typhoons are selected based on two criteria, which are the similarity of typhoon path and central pressure. Similar typhoon selection techniques and LSTM are applied to predict wind velocity data, which are measured on actual bridge. The prediction accuracy of LSTM is improved when wind velocity data at the time of similar typhoons are used as training data. The prediction accuracy of LSTM is examined while changing the number of similar typhoons used for training. The prediction accuracy is maximized in a certain number of training data, then decreases, and finally converges to a certain level. This means that it is important to determine the number of similar typhoons to be used for training.