Multi-step-ahead Cyclone Intensity Prediction with Bayesian Neural Networks

Abstract

The chaotic nature of cyclones makes track and wind-intensity prediction a challenging task. The complexity in attaining robust and accurate prediction increases with an increase of the prediction horizon. There is lack of robust uncertainty quantification in models that have been used for cyclone prediction problems. Bayesian inference provide a principled approach for quantifying uncertainties that arise from model and data, which is essential for prediction, particularly in the case of cyclones. In this paper, Bayesian neural networks are used for multi-step ahead time series prediction for cyclones in the South Pacific region. The results show promising prediction accuracy with uncertainty quantification for shorter prediction horizon; however, the challenge lies in higher prediction horizons.