Introduction
Natural disaster management currently faces a critical challenge: traditional forecasting models are often computationally expensive or prone to false positives, leading to inefficient evacuation plans. These inefficiencies disproportionately affect underserved communities, who face higher risks of financial and welfare losses during crises. This research introduces a Machine Learning Agent-Based Model (ML-ABM) designed to bridge this gap by predicting hurricane trajectories and intensities with high speed and accuracy. By utilizing bidirectional Long Short-Term Memory (LSTM) networks, the model provides a computationally efficient framework to forecast storm paths in minutes, aiming to improve real-time disaster management and evacuation planning.
Methodology
The proposed ML architecture centers on a Recurrent Neural Network (RNN) integrated with an Agent-Based Modeling (ABM) environment. The core components include:
- Bidirectional Time-Distributed LSTMs: Unlike standard LSTMs that process data linearly, this model accounts for both forward and backward time directions. This approach is particularly effective for time-series forecasting in dynamic systems like weather, where past and future states are deeply interconnected.
- Multivariate Regression: A regression layer is employed to predict three critical variables: latitude, longitude, and wind intensity.
- Feature Engineering: Input features incorporate comprehensive tropical cyclone records, including ISO time, maximum sustained wind, minimum central pressure, and wind radii across four quadrants.
- NetLogo Integration: The system represents observations and predictions as a multi-agent system within the NetLogo environment, facilitating the analysis of emergent patterns and spatial dynamics.
Data and Experimental Setup
The study utilizes the IBTrACS (International Best Track Archive for Climate Stewardship) dataset from NOAA, focusing specifically on the North Atlantic basin. The data is processed in 3-hour increments for major storms. For model development, data from 2004–2016 was used for training (utilizing 5-fold cross-validation), while the period from 2017–2020 served as the evaluation set to test the model against recent, high-impact events.
Results and Analysis
The ML-ABM demonstrated strong predictive performance, achieving 85.3% accuracy for trajectory (latitude and longitude) and 73.7% accuracy for wind intensity. A key finding is the model's ability to generate these predictions in just a few minutes, suggesting that deep learning can bypass the heavy computational requirements of traditional fluid dynamics-based meteorological models without sacrificing essential physical realism.
When compared against the National Hurricane Center’s BCD5 Error Model, the researchers reported an error margin of +187 nautical miles for distance and +2 knots for wind intensity. The model successfully translated these predictions into physical measurements across 24-hour intervals for a full 5-day forecast window.
Analysis of the results revealed a moderate positive correlation between a storm's overall movement direction and its strongest wind direction. Interestingly, movement speed exhibited no meaningful static correlations; short-term velocity changes were found to be better gauged by average acceleration in prior timesteps rather than static features.
Discussion and Future Work
While the results are promising, the authors identified biases in the initial RNN results, particularly after a storm makes landfall. Future iterations may incorporate an ABM decision tree to correct these post-landfall discrepancies. Additionally, the integration of Voronoi Diagrams is proposed to better capture "micro-interactions" at the cell level, allowing the system to evaluate spatial boundaries more dynamically. While this study focused on the North Atlantic, the framework is designed to be adaptable for other basins and different disaster types, such as wildfires, providing a scalable tool for global disaster resilience.
Summary updated