Textual Environmental Context and Spatial Graphs for LLM-Based Regional SST Forecasting

2026-10-06

AI Summary

This study addresses the challenges faced by large language models (LLMs) in regional sea surface temperature (SST) prediction, specifically the difficulty of efficiently integrating spatiotemporal dependencies with environmental conditions and the prohibitive computational overhead of full-grid serialization. To overcome these limitations, this work proposes a continuous graph-derived prefix fusion mechanism that constructs textual environmental contexts and spatial graph prefixes. By leveraging static and dynamic graph neural networks to extract spatial features and inject them directly into the LLM, the approach circumvents exhaustive grid-to-text conversion. Furthermore, a rule-based module is incorporated to provide posterior contextual explanations. The proposed method enables multi-step conditional numerical generation, achieving superior mean absolute error (MAE) and R² scores compared to baseline approaches in ten-step SST forecasting for the South China Sea, thereby effectively balancing high predictive accuracy with model interpretability.

Abstract

Sea surface temperature (SST) forecasting depends on local temporal persistence, regional spatial dependence, and environmental conditions that evolve with the forecast date. We study how these heterogeneous conditions can be presented to a large language model (LLM) for regional multi-step forecasting without serializing the full SST grid as text. We formulate forecasting as conditional numerical generation: historical SST and anomaly sequences, date-aligned environmental records, and static ocean knowledge form a textual context, while regional spatial state is supplied through continuous graph-derived prefixes. A static graph encodes persistent geographic--climatological relations, and a dynamic graph encodes recent SST correlations and localized tropical-cyclone influence. Two graph neural networks produce a target-node representation that is mapped by a spatial-prefix fusion and injected into the LLM input. On SST forecasting in the South China Sea, the complete configuration achieves the best MAE and $\Rtwo$ among the compared methods over ten forecast steps. Alongside the numerical forecast, a rule-based module matches predicted trends and environmental-factor directions with knowledge entries to return source-linked, post-hoc contextual explanations.

Problem

• Sea surface temperature forecasting
• Large language model
• Regional multi-step forecasting
• Spatial dependence
• Environmental context

Innovation

• Large Language Model
• Sea Surface Temperature Forecasting
• Graph Neural Networks
• Spatial Prefix Fusion
• Explainable AI