J9B.1 Using a Large Language Model to Generate Text Weather Forecasts

Wednesday, 31 January 2024: 8:30 AM
338 (The Baltimore Convention Center)
Neil David Gordon, ScD MIT, Neil Gordon Consulting, Otaki, WGN, New zealand

Handout (8.9 MB)

Text weather forecasts continue to be important as a means of clearly and succinctly communicating weather information, complementing graphical presentations. Computer-worded forecasts have been produced for more than fifty years, starting with bespoke computer programs, then simple templates, and evolving to some quite sophisticated and expensive Natural Language Generation systems.

This all changed with the recent availability of Large Language Models (LLMs). Systems such as OPENAI’s GPT-3.5 and GPT-4, the models behind ChatGPT, are based on so much text, including weather forecasts, that they are capable of generating text weather forecasts without needing special training or fine-tuning (though fine-tuning may become helpful in future).

To generate a good text weather forecast based on model forecasts, it is important to get the “prompt” or instructions right, use appropriate system settings, and carefully choose what forecast weather elements should be provided, and in what format. Hallucinations must be avoided, with the text forecast based only on the model information.

This talk will cover the choices made for a system which can produce automated text weather forecasts for any place in the world, based on open source model forecasts (from https://open-meteo.com/), using GPT-4 with a ChatGPT plugin. Although designed to produce English output in any combination of the usual units, outputting the forecasts in any language is a trivial extension. The system will be demonstrated, and implications for forecast production and individual customisation of forecasts discussed.

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner