Work

Artificiall Intelligence for Tropical Cyclone Forecasting

Year: 2026
Type: dissertation
Abstract: Tropical cyclones (TCs) continue to be some of the most devastating weather events affecting human populations. Thus, significant efforts have been made to understand and model these phenomena in order to produce useful forecasts. Traditionally, these forecasts rely on General Circulation Models (GCMs) to produce a global forecast of the atmosphere and model the large scale phenomena that influence the behavior of these storms. To capture TC behavior, these models are then used to drive finer scale models, or alternatively the GCMs are run at a high enough resolution to sufficiently capture the behavior of these storms---both alternatives which carry significant computational costs. Recently, artificial intelligence weather prediction (AIWP) models have emerged as forecast-efficient alternatives to traditional GCMs---but as these models are not governed by our understanding of physics and instead rely on learning patterns in the data they are trained on, a need to understand if and how these models can be used in TC forecasting has arisen. This thesis aims to help answer that question. One objective of this thesis is to evaluate if the first generation of deterministic AIWP models provide information useful for TC intensity forecasting. It shows that the models do indeed provide information of value for predicting TC intensity, relying on a hierarchy of postprocessing algorithms to evaluate their use. Still, it finds that for the tested algorithms errors were of a magnitude comparable to the natural variability of the data after 3-4 days. Importantly, it also reports that in the tested model hierarchy spatially-informed algorithms (e.g., convolutional neural networks) did not necessarily provide better performance than simpler neural networks. Another objective is to provide benchmarking tools for TC forecasting. Given the challenges in interpreting the skill in TC forecasting reported by the various authors of AIWP models---e.g., variance in tracking, variance in test sets---it is important to find ways of fairly comparing the performance of these models, both to other AIWP models and to traditional numerical weather prediction models. This thesis puts forth such a benchmark, comprising deterministic and probabilistic evaluations for TC track and intensity forecasting. Finally, the thesis aims to provide tools for understanding the sources of uncertainty in TC forecasts made with AIWP models. Notably, it aims to separate the uncertainty in the forecast due to the variance in the predicted fields and the uncertainty stemming from the postprocessing of AIWP models. This remains an open question.
Source: IRIS
Language: English
Open Access status: green
OpenAlex ID: W7214419545
Cites: 200
Cited by:
Related to: 9