Sea surface temperature (SST) forecasting relies on local temporal persistence, regional spatial dependence, and environmental conditions that evolve with the forecast date. This paper presents a way to feed these heterogeneous cues to a large language model (LLM) for regional multi‑step prediction without serializing the full SST grid as text. Forecasting is cast 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 via continuous graph prefixes. A static graph encodes persistent geographic‑climatological relations; a dynamic graph captures recent SST correlations and localized tropical‑cyclone influence. Two graph neural networks produce a target‑node representation, which is mapped by a spatial‑prefix fusion and injected into the LLM input. Experiments on South China Sea SST forecasting show that the full configuration achieves the lowest MAE and highest $R^2$ across ten forecast steps, outperforming all baselines. In addition to numerical forecasts, a rule‑based module matches predicted trends and environmental‑factor directions with knowledge‑base entries, delivering source‑linked post‑hoc explanations.
Review: The approach elegantly merges textual context with graph‑derived spatial cues, enabling LLMs to exploit both environmental and spatial dependencies for accurate regional SST forecasts while offering interpretability.