Leveraging Lightweight Large Language Models for Hydrological Interpretation of Precipitation Time Series in Greece

As hydrological systems grow increasingly complex and data-rich, new tools are needed to support rapid interpretation and communication of climate and water cycle trends. This study evaluates the applicability of instruction-tuned large language models (LLMs) to interpret long-term precipitation records in the context of hydrological variability. Using a 23-year (2002–2024) ERA5 monthly precipitation time series from a location in Greece, the study tests whether lightweight, open-source models—including TinyLlama (1.1B) and Phi-2 (2.7B)—can generate semantically coherent summaries of seasonal dynamics, detect hydrological anomalies, and answer natural language questions relevant to water resource monitoring.

The time series is preprocessed into prompt-compatible text blocks, enabling models to produce narrative outputs describing dry and wet seasons, interannual shifts, and extreme events. Responses are evaluated against visual and statistical baselines to assess their hydrological fidelity. Phi-2 demonstrates stronger correlation with observed patterns, while TinyLlama provides fluent but less consistent outputs. All models show limitations in numerical reasoning and require tight prompt structuring to avoid hallucinated values.

Our findings suggest that even low-resource LLMs can serve as effective interpretive aids in hydrology, particularly for rapid diagnostics, stakeholder reporting, and data contextualization. When paired with physical constraints and structured input formats, LLMs could enhance exploratory analysis and decision-support capabilities in water management, climate services, and early warning systems. A reproducible Google Colab notebook and annotated model comparisons are included to support further hydrological application and refinement.