Machine Learning-assisted Agent-Based Modeling for Hurricane Track Prediction

Abstract:

Machine Learning-assisted Agent-Based Modeling (ML-ABM) develops surrogate models that can capture complex physical processes associated with natural hazards while minimizing computational costs. Specifically, ML-ABM leverages the strengths of conventional machine learning techniques such as Recurrent Neural Networks (RNNs), and augments them by modeling complex interactions using agent-based modeling. The ML-ABM approach was successful in modeling the complex interactions involved in wildfire propagation and suppression (Hu & Sun, 2007). This study explores the potential of using ML-ABM for hurricane trajectory prediction. Currently, it takes several hours to run accurate dynamic forecast models, which limits their utility in real-time applications. Statistical model runtimes can be reduced to seconds, but at a loss of complex representation. Our Hurricane Track Prediction ML-ABM aims to quickly model and predict hurricane tracks in only a few minutes, yet retains some of the complex physical process interactions of real storms through feature engineering and deep learning. This study adapts a particular RNN (Usmani, 2019) that employs bidirectional time distributed Long-Short Term Memory cells, accounting for positive and negative time direction in time series forecasting. The Hurricane Track Prediction ML-ABM uses the IBTrACS hurricane reanalysis database containing all major tropical cyclones from 2004 to 2020 to develop and validate the proposed RNN. A NetLogo ABM uses Agenthood to represent the observations and scaled predictions from the RNN for further emergent pattern analysis. The Hurricane Track Prediction ML-ABM compares model results to the official National Hurricane Center (NHC) 5-Year Average Forecast Errors from 2014-2019.