MeteoEurope1km and ChatGPT integration: Chat with high-resolution historical meteorological data through advanced NLP
- Creators
-
Sekulić, Aleksandar1
-
Kilibarda, Milan1, 2
- 1. University of Belgrade, Faculty of Civil Engineering
- 2. GILAB DOO Beograd
Description
Daily gridded meteorological datasets are an important source of information for analysis of historical weather and many other research areas since they have no gaps in the spatio-temporal domain they cover. Most of the daily gridded meteorological datasets represent reanalysis or estimations from different remote sensing sensors or are generated by downscaling procedures.
We introduce MeteoEurope1km, a daily gridded meteorological dataset for Europe spanning the 1961–2020 period at a spatial resolution of 1 km. This dataset comprises five variables: maximum (TMAX), minimum (TMIN), and mean (TMEAN) temperature, sea-level pressure (SLP), and total precipitation (PRCP). Utilizing spatio-temporal regression kriging for temperature variables and ordinary kriging for SLP and PRCP, with an additional step for PRCP occurrence prediction using Indicator kriging, MeteoEurope1km integrates data from GHCN-daily, ECA&D, and SYNOP observations from OGIMET service after station duplication and outlier removal. Auxiliary variables such as geometric temperature trend, digital elevation model, and topographic wetness index enhance the accuracy of temperature datasets through multiple linear regression for trend modeling. Assessment via leave-one-station-out cross-validation reveals high models accuracy, with coefficients of determination exceeding 96% for temperature parameters and SLP, and over 76% for PRCP. Root mean square error is 1.3°C, 1.6°C, 1.8°C, 1.5 mbar, and 2.5 mm for TMEAN, TMAX, TMIN, SLP, and PRCP, respectively. Accessible through the dailymeteo.com portal and API, MeteoEurope1km stands as a valuable resource for historical weather analysis and diverse research endeavors.
Furthermore, we present an innovative approach to fine tune ChatGPT model, namely GPT-3.5 Turbo, on top of the dailymeteo.com API. By harnessing the natural language processing capabilities of ChatGPT, we created an intuitive interface that facilitates seamless access to meteorological insights. This novel dailymeteo.com application not only streamlines data retrieval but also enhances user engagement by providing a conversational interface for querying historical weather information. We will present the technical implementation of this solution, highlighting its potential to revolutionize the accessibility and usability of meteorological data for a wide range of users.
Future work will be oriented towards expanding the spatial coverage to encompass continents beyond Europe, interpolating additional daily meteorological variables, and enhancing model performance, particularly for PRCP, through the application of spatial machine learning techniques like Random Forest Spatial Interpolation. As dailymeteo.com users contribute new inquiries, the ChatGPT model trained on the dailymeteo.com API will undergo continual refinement, resulting in enhanced application functionality and accuracy.