A Modular LLM-Agent System for Transparent Multi-Parameter Weather Interpretation

Weather forecasting is not only a predictive task but an interpretive scientific process requiring explanation, contextualization, and hypothesis generation. This paper introduces AI-Meteorologist, an explainable LLM-agent framework that converts raw numerical forecasts into scientifically grounded narrative reports with transparent reasoning steps. Unlike conventional forecast outputs presented as dense tables or unstructured time series, our system performs agent-based analysis across multiple meteorological variables, integrates historical climatological context, and generates structured explanations that identify weather fronts, anomalies, and localized dynamics. The architecture relies entirely on in-context prompting, without fine-tuning, demonstrating that interpretability can be achieved through reasoning rather than parameter updates. Through case studies on multi-location forecast data, we show how AI-Meteorologist not only communicates weather events but also reveals the underlying atmospheric drivers, offering a pathway toward AI systems that augment human meteorological expertise and support scientific discovery in climate analytics.

💡 Research Summary

The paper presents AI‑Meteorologist, a modular large‑language‑model (LLM) agent framework that converts raw numerical weather forecasts into scientifically grounded, explainable narrative reports. The system operates without any model fine‑tuning; instead it relies on carefully crafted in‑context prompts and a pipeline of specialized agents that together perform data acquisition, reasoning, text generation, and visualization.

Data acquisition and preprocessing: The pipeline queries the OpenWeather One Call API to retrieve hourly forecasts for up to five days, covering more than ten meteorological variables (temperature, humidity, wind speed and direction, precipitation, visibility, etc.). Simultaneously it gathers geographic metadata from OpenStreetMap (urban/rural classification, coastal vs. inland) and long‑term climatological normals (20‑year monthly averages of temperature and precipitation) from the Meteostat API. All information is serialized into a single “EXTERNAL INFO” block that is passed downstream.

Meteorologist agent: This core reasoning component receives the serialized input and produces a structured output with four fields: (1) a concise natural‑language summary of the expected weather evolution, (2) a causal explanation that links observed parameter changes to atmospheric processes (e.g., pressure gradients, wind‑direction rotation, rapid temperature drops), (3) a self‑assessed confidence score derived from model‑internal uncertainty and recent observations, and (4) warnings that flag hazardous or anomalous conditions by comparing short‑term forecasts to climatological percentiles. The causal explanation is expressed as a step‑by‑step reasoning chain, making the model’s logic transparent.

Writer agent: Building on the meteorologist’s structured output, the writer formats the content into a conventional weather‑bulletin layout: a title, an introductory paragraph that synthesizes the summary, proof, and warnings, followed by a detailed “weather parameters” section that describes each key variable in plain language. User‑specific preferences (e.g., emphasis on certain variables, language choice) can be injected via the prompt, demonstrating the system’s flexibility.

Illustrator agent: For the selected variables, this agent automatically generates Python / matplotlib code that plots temperature, wind, precipitation, and other time‑series. The resulting figures are embedded in the final PDF, providing visual corroboration of the textual explanations.

Report compilation: The final PDF integrates the title, narrative, confidence score, warnings, and visualizations into a single, professionally styled document. The report therefore delivers both a textual interpretation and a quantitative visual representation of the forecast.

Evaluation: The authors tested the system on three representative locations—a coastal site, an inland region, and a high‑altitude area—using real‑time multi‑parameter forecasts. In the coastal case the agents correctly identified an approaching cold front by jointly analyzing pressure gradients, wind‑direction rotation, and a rapid temperature decline, and they supplied a scientifically accurate description of frontogenesis involving maritime and continental air masses. In the inland case the system detected anomalously heavy precipitation by comparing forecast values to long‑term climatological percentiles, cited historical analogues, and issued a data‑driven flood‑risk warning. In each scenario the reasoning chain was explicit, and the confidence scores aligned with the underlying forecast model’s uncertainty estimates.

Discussion and limitations: The work demonstrates that a general‑purpose LLM (GPT‑4o) can perform sophisticated meteorological reasoning when supplied with properly serialized inputs and modular prompting, eliminating the need for domain‑specific fine‑tuning and reducing computational overhead. The modular architecture allows individual agents to be upgraded or replaced without disrupting the overall workflow, facilitating future extensions such as multilingual support, probabilistic ensemble integration, or additional data sources (e.g., radar imagery). However, the system’s performance is bounded by the quality and coverage of the external APIs, and the paper does not provide a quantitative benchmark of forecast accuracy or explanation quality. Future research directions include systematic human‑vs‑AI bulletin comparisons, incorporation of ensemble forecasts, and development of self‑refinement loops to improve robustness.

Conclusion: AI‑Meteorologist offers a transparent, extensible, and domain‑agnostic solution for turning dense, multi‑parameter weather forecast tables into accessible, scientifically justified narratives and visualizations. By showcasing that in‑context prompting and agentic decomposition can yield explainable AI in a complex scientific domain, the paper paves the way for broader adoption of LLM‑driven decision support tools in meteorology and related fields.