[Paper Review] Predicting Hurricane Trajectories using a Recurrent Neural Network

Sheila Alemany, Jonathan Beltran+2 more|arXiv (Cornell University)|Feb 1, 2018

Tropical and Extratropical Cyclones ResearchEarth and Planetary Sciences20 citations

TL;DR

This paper proposes a fully connected recurrent neural network (RNN) over a fine spatial grid to predict hurricane trajectories with high accuracy. By modeling temporal dynamics using latitude, longitude, wind speed, and pressure data at 6-hour intervals, the RNN achieves a root-mean-squared error of 0.11, outperforming traditional NHC methods and effectively capturing nonlinear, looping storm behaviors that prior models miss.

ABSTRACT

Hurricanes are cyclones circulating about a defined center whose closed wind speeds exceed 75 mph originating over tropical and subtropical waters. At landfall, hurricanes can result in severe disasters. The accuracy of predicting their trajectory paths is critical to reduce economic loss and save human lives. Given the complexity and nonlinearity of weather data, a recurrent neural network (RNN) could be beneficial in modeling hurricane behavior. We propose the application of a fully connected RNN to predict the trajectory of hurricanes. We employed the RNN over a fine grid to reduce typical truncation errors. We utilized their latitude, longitude, wind speed, and pressure publicly provided by the National Hurricane Center (NHC) to predict the trajectory of a hurricane at 6-hour intervals. Results show that this proposed technique is competitive to methods currently employed by the NHC and can predict up to approximately 120 hours of hurricane path.

Motivation & Objective

• To improve hurricane trajectory forecasting accuracy by leveraging the nonlinear temporal modeling capabilities of recurrent neural networks.
• To address the limitations of traditional statistical and dynamical models in capturing complex, erratic hurricane movements, including looping behaviors.
• To reduce truncation errors and improve spatial resolution by employing a fine-grid representation of storm locations.
• To develop a lightweight, fast-predicting model that outperforms computationally intensive NHC statistical-dynamical methods.
• To enable accurate conversion from grid-based predictions to real-world geographic coordinates (latitude/longitude) through future deep learning enhancements.

Proposed method

• A fully connected recurrent neural network (RNN) is trained on historical hurricane data including latitude, longitude, wind speed, and central pressure at 6-hour intervals.
• The model operates on a spatial grid system, with each grid cell representing a discrete location in the storm's path, reducing truncation errors.
• The RNN learns temporal dependencies between sequential storm positions, modeling direction and distance traveled as dynamic features.
• The model is trained and evaluated using mean-squared error (MSE) and root-mean-squared error (RMSE) metrics on both training and testing sets.
• Grid-based predictions are later converted to geographic coordinates (latitude/longitude), though current conversion introduces up to 50 km error.
• Future work includes applying deep learning to refine the grid-to-geographic coordinate transformation and potentially integrating Bayesian neural networks for uncertainty quantification.

Experimental results

Research questions

• RQ1 Can a recurrent neural network effectively model the complex, nonlinear temporal dynamics of hurricane trajectories?
• RQ2 Does a grid-based RNN approach reduce truncation errors and improve prediction accuracy compared to traditional NHC methods?
• RQ3 Can the model accurately predict hurricane paths beyond 120 hours with high spatial precision?
• RQ4 How does the RNN perform on non-monotonic storm paths, such as those that loop or reverse direction?
• RQ5 To what extent can grid-based RNNs generalize to diverse hurricane behaviors compared to sparse or non-grid RNNs?

Key findings

• The RNN achieved a root-mean-squared error (RMSE) of 0.11 on both training and testing sets, indicating strong generalization and predictive accuracy.
• The model successfully predicted hurricane trajectories up to approximately 120 hours into the future with high consistency.
• The grid-based RNN outperformed traditional NHC methods in accuracy, particularly in capturing nonlinear and looping storm behaviors.
• The model demonstrated robustness to complex storm dynamics, unlike prior 'sparse RNN' approaches that failed on non-monotonic paths.
• The current 50 km margin of error in latitude/longitude conversion is attributed to the 1x1-degree grid resolution, which is still competitive with existing NHC methods.
• The approach is computationally efficient, enabling fast predictions—critical for real-time disaster preparedness—compared to the hours-long processing of current NHC statistical-dynamical models.