Improving Situation-Dependent Uncertainty Estimation in Tropical Cyclone Track Forecasts Using Encoder-Decoder Neural Networks

Currently, the Monte Carlo Method is commonly used to estimate the uncertainty of tropical cyclone (TC) track forecasts. By performing random sampling of both along-track and cross-track errors, the potential range of official forecast errors is estimated (DeMaria et al. 2009; Tsai et al. 2011).

This study utilizes a Recurrent Neural Network (RNN) with an Encoder-Decoder architecture to represent situation-dependent track forecast uncertainty and the spatiotemporal correlations of forecast errors. The datasets used in this study include the Central Weather Administration’s (CWA) official TC track forecasts from 2018 to 2022, as well as deterministic and ensemble track forecasts from global numerical weather prediction models, specifically ECMWF and NCEP models.

Preliminary results indicate that the RNN-based approach reasonably reflects potential error ranges under different scenarios. For instance, TCs located at mid-to-high latitudes with higher translation speeds usually exhibit smaller cross-track forecast errors. Additionally, the prediction intervals (PIs) derived in this study can reasonably cover the proportion of observed data: the uncertainty ranges of the mean +/- one (two)  standard deviations encompass approximately 70% (95%) of observed data. Furthermore, large-scale environmental indices (e.g., steering flow and monsoon circulation) are considered to further reduce the uncertainty of TC track forecasts. More detailed findings will be presented during the meeting.