While significant progress has been made over the past few decades in forecasting tropical cyclone (TC) tracks, forecasting TC intensity has only shown modest improvements. To address this, researchers developed a deep-learning-based Multilayer Perceptron (MLP) TC intensity prediction model and automatically optimized it using a Bayesian search algorithm. MLP works by passing down and condensing information through many layers of mathematical transformations. The MLP model outperformed three of the most accurate operational models (by 5–22%) and correctly predicted more rapid intensification events than all studied models. The MLP’s predictions have different errors than the other models, making it a valuable addition to the field.
The challenge of reducing errors in TC intensity forecasting has interested the operational forecasting and research community for decades. The superior performance of the developed deep-learning-based predictive model for North Atlantic highlights the potential for applying deep learning to better forecast extreme weather. Improved TC intensity forecasting will allow agencies to alert the general public with higher precision, allowing for strategic evacuations that can save lives and money. This study also created a benchmark dataset and systematic approach that enable future advances in the development of machine learning algorithms for TC intensity forecasting.
Scientists have spent decades searching for breakthroughs in TC intensity modeling to provide more accurate and timely TC warnings. To this end, researchers developed a deep-learning-based predictive model for short-term (24- and 6-hour) intensity forecasts in the North Atlantic. They used TC data from 2019 and 2020 to simulate known storms as if they were in an operational forecast mode. When fed with climate model outputs, the intensity model coupled with a physics-based TC track model can easily generate millions of realistic synthetic TC events that are representative of the TC risks involved in a certain climate. The researchers found that the model’s 24-hour intensity forecast outperformed some of the most accurate operational models by 5–22%. The 6-hour intensity model produced realistic intensity labels for the modeled TC tracks. These results highlight the potential for using deep neural network based models to improve operational hurricane intensity forecasts and synthetic TC generation.
Dave Judi, Pacific Northwest National Laboratory, David.Judi@pnnl.gov
The operational forecast portion of this research was supported by the Deep Science Agile Initiative at Pacific Northwest National Laboratory (PNNL). It was conducted under the Laboratory Directed Research and Development program at PNNL. PNNL is a multiprogram national laboratory operated by Battelle for the U.S. Department of Energy.
The synthetic tropical cyclone portion of this research was supported by the Multisector Dynamics program area of the U.S. Department of Energy, Office of Science, Biological and Environmental Research program as part of the multiprogram, collaborative Integrated Coastal Modeling project.
K. B. acknowledges support from the Regional and Global Modeling and Analysis Program of the U.S. Department of Energy, Office of Science, Biological and Environmental Research program and from the National Oceanic and Atmospheric Administration's Climate Program Office, Climate Monitoring Program.
Published: September 14, 2021
W. Xu, K. Balaguru, A. August, N. Lalo, N. Hodas, M. DeMaria, and D. Judi. "Deep Learning Experiments for Tropical Cyclone Intensity Forecasts", Weather and Forecasting, 36, 1453-1470, (2021). [DOI: 10.1175/WAF-D-20-0104.1]