Cyclone trajectory and intensity prediction with uncertainty quantification using variational recurrent neural networks. (April 2023)

• Record Type: Journal Article
• Authors: Kapoor, Arpit; Negi, Anshul; Marshall, Lucy; Chandra, Rohitash
• Is Part Of: Environmental modelling & software. Volume 162(2023)
• Publication Date: 2023-04
• Subjects: Cyclone track prediction -- Recurrent neural networks -- Bayesian neural networks -- Long short-term memory -- Variational inference; Environmental monitoring -- Computer programs -- Periodicals; Ecology -- Computer simulation -- Periodicals; Digital computer simulation -- Periodicals; Computer software -- Periodicals; Environmental Monitoring -- Periodicals; Computer Simulation -- Periodicals; Environnement -- Surveillance -- Logiciels -- Périodiques; Écologie -- Simulation, Méthodes de -- Périodiques; Simulation par ordinateur -- Périodiques; Logiciels -- Périodiques; Computer software; Digital computer simulation; Ecology -- Computer simulation; Environmental monitoring -- Computer programs; Periodicals; Electronic journals; 363.70015118
• DOI: 10.1016/j.envsoft.2023.105654
• Languages: English
• ISSNs: 1364-8152

Abstract

Abstract: Cyclone track forecasting is a critical climate science problem involving time-series prediction of cyclone location and intensity. Machine learning methods have shown much promise in this domain, especially deep learning methods such as recurrent neural networks (RNNs) However, these methods generally make single-point predictions with little focus on uncertainty quantification. Although Markov Chain Monte Carlo (MCMC) methods have often been used for quantifying uncertainty in neural network predictions, these methods are computationally expensive. Variational Inference (VI) is an alternative to MCMC sampling that approximates the posterior distribution of parameters by minimizing a KL-divergence loss between the estimate and the true posterior. In this paper, we present variational RNNs for cyclone track and intensity prediction in four different regions across the globe. We utilise simple RNNs and long short-term memory (LSTM) RNNs and use the energy score (ES) to evaluate multivariate probabilistic predictions. The results show that variational RNNs provide a good approximation with uncertainty quantification when compared to conventional RNNs while maintaining prediction accuracy. Highlights: Cyclone trajectory and intensity forecast is a multivariate time-series prediction problem. RNN and LSTM models are excellent for modelling temporal correlations in the time-series. We use variational inference to train and quantify uncertainty in RNN and LSTM models. We use the energy score to assess the performance of probabilistic models. We find that variational inference-based RNNs provide a good approximation of their deterministic counterparts.