Physics-Informed Time‑Series Large Language Model for Extreme Typhoon Wave Height Forecasting

Abstract Typhoon-induced extreme waves pose severe threats to coastal communities, infrastructure, and maritime operations worldwide. Accurate and rapid forecasting of significant wave height is therefore critical for disaster response. Physics-based numerical wave models provide physically grounded predictions but can be computationally demanding for high-resolution forecasting, while purely data-driven models often struggle with rare, highly non-stationary extreme events. Recent time-series approaches based on large language models (LLMs) offer new opportunities for long-range forecasting but remain limited in explicitly representing coupled environmental variables and enforcing physically plausible predictions. Here, we develop the Typhoon Wave Large Language Model (TW-LLM), a physics-informed, domain-adapted framework for forecasting typhoon-induced coastal wave height. Building on Time-LLM, TW-LLM introduces cross-variable covariate interaction to explicitly represent dependencies between wave states and environmental forcing, structured domain-informed prompting for semantic conditioning, and a physics-informed training objective that combines magnitude-weighted Huber loss with wind-forcing consistency and temporal-rate constraints. Trained and evaluated using nine years (2017–2025) of hourly observations from eleven coastal buoys in China, TW-LLM achieves root-mean-square errors (RMSEs) of ~ 0.15 m for short-term (6-h) forecasts, 0.3 m for long-term (96-h) forecasts, and 0.23 m under extreme typhoon conditions. With an inference time as low as 5 s per forecast, TW-LLM provides a computationally efficient framework for near-real-time coastal wave forecasting and early warning, with potential to enhance coastal resilience to increasingly severe climate-related hazards.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-12
DOI
https://doi.org/10.21203/rs.3.rs-11525261/v1
Primary Topic
Ocean Waves and Remote Sensing
Type
preprint

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Physics-Informed Time‑Series Large Language Model for Extreme Typhoon Wave Height Forecasting

Zhiguo Hé, Yiting Zhao, Dongrui Han, Hui Shen et al.
Research Square
Ocean Waves and Remote Sensing
preprint

Physics-Informed Time‑Series Large Language Model for Extreme Typhoon Wave Height Forecasting

Zhiguo Hé, Yiting Zhao, Dongrui Han, Hui Shen, Xiangbing Kong, Ye Zhu, Li Li
preprint en

Abstract

Abstract Typhoon-induced extreme waves pose severe threats to coastal communities, infrastructure, and maritime operations worldwide. Accurate and rapid forecasting of significant wave height is therefore critical for disaster response. Physics-based numerical wave models provide physically grounded predictions but can be computationally demanding for high-resolution forecasting, while purely data-driven models often struggle with rare, highly non-stationary extreme events. Recent time-series approaches based on large language models (LLMs) offer new opportunities for long-range forecasting but remain limited in explicitly representing coupled environmental variables and enforcing physically plausible predictions. Here, we develop the Typhoon Wave Large Language Model (TW-LLM), a physics-informed, domain-adapted framework for forecasting typhoon-induced coastal wave height. Building on Time-LLM, TW-LLM introduces cross-variable covariate interaction to explicitly represent dependencies between wave states and environmental forcing, structured domain-informed prompting for semantic conditioning, and a physics-informed training objective that combines magnitude-weighted Huber loss with wind-forcing consistency and temporal-rate constraints. Trained and evaluated using nine years (2017–2025) of hourly observations from eleven coastal buoys in China, TW-LLM achieves root-mean-square errors (RMSEs) of ~ 0.15 m for short-term (6-h) forecasts, 0.3 m for long-term (96-h) forecasts, and 0.23 m under extreme typhoon conditions. With an inference time as low as 5 s per forecast, TW-LLM provides a computationally efficient framework for near-real-time coastal wave forecasting and early warning, with potential to enhance coastal resilience to increasingly severe climate-related hazards.

Research Square
Université du Québec à Rimouski (CA), Zhejiang University (CN)
National Natural Science Foundation of China, Natural Science Foundation of Zhejiang Province
Climate action
Ocean Waves and Remote Sensing
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon