Situational Estimation of Tropical Cyclone Track Forecast Uncertainty Using an Autoregressive Encoder-Decoder LSTM Framework

Providing reliable information on tropical cyclone (TC) track forecast uncertainty is essential for effective disaster preparedness. Conventionally, the radius of the Cone of Uncertainty (CoU) is derived from historical distance errors, often resulting in a static, climatological value that fails to account for the characteristics of individual storms. While traditional approaches attempt to categorize scenarios based on factors like translation speed, they are often limited by sample sparsity and struggle to objectively incorporate complex environmental influences.

To address these limitations, this study proposes an autoregressive encoder-decoder Long Short-Term Memory (LSTM) framework to generate situation-dependent CoU estimates. We utilize a multi-source dataset comprising official forecasts from the Central Weather Administration (CWA) and global models (ECMWF and NCEP) from the past five years. By employing an autoregressive architecture, the model can also iteratively generate a large ensemble of potential track realizations to characterize the forecast error distribution while preserving serial correlations across lead times.

In this presentation, we compare traditional methods with the proposed LSTM approach to highlight the advantages of situation-dependent estimation. Our results also show that the LSTM-based CoU provides a robust representation of observed tracks, covering approximately 68% and 95% of observations within one and two standard deviations, respectively. Furthermore, integrating global numerical model information significantly reduces the uncertainty radius while maintaining reliable coverage. Overall, this work demonstrates how deep learning can offer context-aware uncertainty quantification, serving as a promising advancement for TC forecasting.