Development of a Machine Learning-Based Tropical Cyclone Track Prediction Scheme over the Western North Pacific

Tropical cyclones (TCs) rank among the most devastating natural disasters globally, with their accurate forecasting being paramount for disaster prevention in coastal regions. While TC track predictions are foundational within TC prediction parameters, advancements in observation techniques and numerical weather prediction (NWP) skills have notably enhanced their precision. The Korea Meteorological Administration (KMA) issues TC forecasts at 12-hour intervals, relying on post-processed observations and NWP forecasts updated every 4-8 hours. However, the long update cycles of NWPs pose challenges for real-time utilization by forecasters. This research introduces a ConvLSTM-based approach to TC trajectory forecasting, specifically tailored for real-time updates in the western North Pacific Ocean. ConvLSTM, which embeds convolutional recurrent cells within the LSTM framework, adeptly processes time series while retaining two-dimensional spatial characteristics. This trait makes it a preferred choice for constructing sequential spatial prediction models. The training dataset encompassed TC-related variables collected at 6-hour intervals from the current time up to the past 36 hours for each TC, sourced from RSMC-Tokyo Best Track and ERA5 reanalysis datasets spanning 1982-2022. We evaluated the ConvLSTM model across 6- to 120-hour forecast intervals at 6-hour increments and enhanced its performance through sensitivity analysis.

Key words: Tropical cyclone, TC track prediction, ConvLSTM

※ This work was funded by the Korea Meteorological Administration Research and Development Program "Developing Intelligent Assistant Technology and Its Application for Weather Forecasting Process" under Grant (KMA2021-00123).