Important: if you want high-quality weather forecasting, please check out FourCastNet, which is much more accurate (not affiliated)

WeatherGPT Small

An experimental approach to predict weather using the transformers architecture.

Important

Disclaimer: I'm not a meteorologist! I'm just a random student who knows very little about meteorology/physics!

This project is experimental and does not take into account physics/wind/etc. TL;DR: Don't rely on this! Traditional weather algorithms are likely much more accurate!

How it Works

We propose a method of fine-tuning GPT-2 models of various sizes (and eventually larger models) to predict the weather. We propose to process each day's weather into something similar to the following format: [HIGH_TMP,LOW_TMP]. Example: [60,80] for a day with a low of 60° and a high of 80°.

Prompting Format

Example:

[35.0,56.0],[38.0,56.0],[33.0,55.0],[31.0,55.0],[34.0,57.0],[36.0,57.0],[48.0,61.0],[51.0,64.0],[53.0,64.0],[45.0,66.0],[42.0,62.0],[48.0,60.0],[49.0,56.0],[48.0,55.0],[44.0,56.0],[45.0,51.0],[38.0,55.0],[34.0,57.0],[33.0,55.0],[34.0,58.0],[32.0,57.0],[33.0,56.0],[33.0,62.0],[38.0,64.0],[35.0,57.0],[40.0,54.0],[32.0,57.0],[32.0,59.0],[40.0,58.0],[47.0,54.0],[38.0,54.0],[42.0,55.0],[37.0,58.0],[40.0,60.0],[39.0,54.0],[36.0,57.0],[45.0,52.0],[31.0,53.0],[35.0,57.0],[42.0,58.0],[40.0,65.0],[38.0,63.0],[43.0,61.0],[41.0,53.0],[37.0,49.0],[28.0,53.0],[45.0,49.0],[35.0,48.0],[32.0,49.0],[32.0,56.0],[32.0,55.0],[42.0,66.0],[37.0,69.0],[37.0,70.0],[45.0,68.0],[45.0,64.0],[46.0,61.0],[44.0,57.0],[44.0,55.0],