Turbulence Prediction System from Wind Distribution using CNN

Abstract

In order to operate aircraft safely, it is necessary to predict the occurrence of turbulence and take appropriate measures to prevent injuries to passengers and crew. However, there has been no established method to forecast turbulence at this point. In this paper, we propose a turbulence prediction system using convolutional neural network (CNN). The input of the CNN model is two-dimensional wind forecast data, which was prepared and delivered by the Japan Meteorological Agency. The labeled train dataset was prepared based on turbulence reports by ANA pilots. In order to consider meteorological differences among seasons, four models corresponding to four seasons were trained. As a result of evaluation, the trained CNN models achieved 70-80% accuracy on average, which is higher accuracy compared to the conventional method based on point-based prediction. The prediction model was implemented as an auto-execute system, which yields turbulence prediction results every three hours.