Introduction
Predicting the path and intensity of tropical cyclones (TCs) is a high-stakes challenge for coastal communities and the insurance industry. Researchers often face a "small data" problem: reliable hurricane records, such as the HURDAT2 database, only span approximately 150 years. This limited historical record makes it difficult to prepare for rare "black swan" events or to accurately estimate risks in specific geographic locales.
While traditional numerical weather models provide high accuracy, they are computationally expensive and slow. Conversely, many existing machine learning approaches overlook the complex temporal and spatial variations of storms. HurriCast addresses this by providing a lightweight, automated framework that generates thousands of realistic, synthetic storm trajectories to augment limited historical data.
Methodology
The HurriCast framework utilizes a three-stage hybrid approach to create "plausible but hypothetical" storm scenarios by blending statistical forecasting with deep learning.
- Frequency Forecasting (ARIMA): The system uses an Autoregressive Integrated Moving Average (ARIMA) model to predict the expected number of hurricanes in a given period based on historical trends.
- Spatial Clustering (K-Means): To ensure geographic realism, K-Means clustering groups historical tracks with similar movement patterns. These clusters provide "representative seeds" that anchor the generative process in historical reality.
- Generative Engine (Perturbed Autoencoder): An Autoencoder is trained to compress hurricane tracks into a latent space. The core innovation is a perturbation module that adds Gaussian noise to these latent features. When the decoder reconstructs the track, this noise results in a new, synthetic trajectory that is physically plausible but distinct from any historical storm.
To maintain consistency, all tracks are pre-processed using spatial interpolation, condensing every hurricane into a set of 20 uniform data points (latitude, longitude, and minimum pressure).
Datasets and Evaluation
The study relies on the HURDAT2 (Hurricane Database 2nd generation) provided by the National Hurricane Center, covering tropical and subtropical cyclone activity in the Atlantic and East Pacific basins from 1851 to 2021. The model was trained on 1,134 tracks from 1875 to 2010, while data from 2011 to 2021 served as the ground truth for validation.
Key Results
HurriCast successfully replicates historical frequencies and spatial distributions, demonstrating its utility as a risk assessment tool.
- Spatial Fidelity: Heat maps of generated tracks show a high degree of agreement with historical data, accurately capturing the "peak frequency corridor" stretching from the mid-Atlantic through the Caribbean and into the Gulf of Mexico.
- Regional Accuracy: The model matched the peak annualized frequency of 0.36–0.50 cyclones per year in the Florida Keys, aligning with long-term climate patterns.
- Intensity Handling: The framework is capable of generating tracks across the full intensity spectrum. Synthetic Category 5 tracks—the most dangerous—closely mirror the clustering and trajectories of their real-world counterparts.
Limitations and Future Work
A primary physical limitation of the current model is land interaction. Synthetic tracks occasionally extend too far over continental land masses, whereas real hurricanes dissipate rapidly after landfall. Future iterations of HurriCast will incorporate land-interaction physics to simulate post-landfall dissipation more accurately. Additionally, the authors suggest exploring more advanced architectures, such as Generative Adversarial Networks (GANs) or Long Short-Term Memory (LSTM) networks, to further improve the temporal realism of generated tracks.
Summary updated