Grid-Discretized Multi-Task Transformer for Joint Tropical Cyclone Path and Intensity Forecasting

Tropical cyclone (TC) forecasting has seen remarkable progress in track prediction over the past two decades, yet intensity forecasting remains a long-standing challenge. Recent comparative studies have shown that while Transformer architectures offer computational advantages, they can underperform recurrent and convolutional alternatives when applied directly to TC sequential data. In this work, we revisit the Transformer approach by introducing two key modifications: a coordinate grid discretization scheme that converts continuous trajectory regression into a structured spatial prediction task, and a multi-task learning formulation that jointly optimizes track and intensity predictions through a shared encoder. We train and evaluate our model on the HURDAT2 Atlantic basin dataset (1944–2022, 982 storms, 22,545 six-hourly records). Experiments show that grid discretization substantially improves Transformer-based path prediction, yielding 78.3% grid-cell accuracy (84 km mean track error) and 74.6% intensity classification accuracy across five Saffir-Simpson-based categories, outperforming LSTM, GRU, CNN-LSTM, and notably TCN—the architecture previously reported as superior to Transformers for this task. The entire model trains in approximately six minutes on a single Tesla T4 GPU, making it accessible for resource-constrained research settings