Everything in one place
Resources
Every document, figure and source-code archive the project produced, with the data statement and citation details.
Joint deliverables
Documents produced by the team as a whole.
Final papers
Each paper reports its author's own workstream against the two shared baselines.
Literature reviews
Submitted earlier in the project, setting out the literature each workstream builds on.
Source code
WeatherFLOW follows a layered design, separating shared data and evaluation utilities from the model integrations above them. Upstream model repositories are vendored at pinned commits and never modified; each integration is an adapter that wraps the published code behind a common interface for model construction, forward computation, prediction and parameter counting — so every experiment runs the code the original authors released, over identical preprocessed tensors, splits, loss, optimiser, seeds and evaluation suite.
One archive holds the whole project — the shared framework and all three workstreams. Each workstream's models sit in their own directory within it:
| Workstream | Models | Directory |
|---|---|---|
| Shared baselines | AGCRN, CLCRN | models/baselines/ |
| Continuous-time | MTGODE, STGODE | models/continuous_time/ |
| Diffusion-based | DiffSTG, USTD | genevieve/Diffusion_Based_STGNNs/ |
| Lightweight | STLGRU, LiteSTGNN | models/lightweight/ |
What the archive excludes. Raw and preprocessed station data, trained model checkpoints, the Singularity container image and raw run outputs are omitted — together they exceed 19 GB, and the station data is not ours to redistribute (see Data availability). The archive contains the code required to reproduce the pipeline once the data is obtained from SAWS.
Live repositories
The archive above is a snapshot; the repositories below are the working codebases, with full commit history. These are external links to GitHub and are not guaranteed to remain available indefinitely.
Data availability
The station observations used throughout this project were supplied by the South African Weather Service for research use, under the SAWS data policy. They are not redistributed on this website. The dataset is available on request from SAWS; the list of stations used, with their coordinates and elevations, is given in the appendices of the final papers, so the same network can be requested and reconstructed.
The data contains no personally identifiable information — it comprises physical measurements from automatic weather stations. Station identities and coordinates are reported in the papers because they are a matter of public record and are necessary to interpret the geographic results.
A note on intended use. Forecasts of extreme weather carry asymmetric risk: a missed heatwave or storm costs more than a false alarm. The models evaluated here are not calibrated for operational warning, and these results are reported as a benchmark comparison. They should not be read as evidence of readiness for deployment in a public warning system.
Figures
Every figure on this website is reproduced directly from the final papers, at the resolution the papers embed, so that what is shown here and what was submitted are the same artefact. Each figure carries its source paper and figure number in its caption.
Figures appear on Results, Continuous-time, Diffusion and Overview. The full set, including those not reproduced here, is in the papers themselves.
How to cite
Cite the individual paper for the workstream you are referring to.
BibTeX — continuous-time workstream (Jaravani)
@mastersthesis{jaravani2026continuous,
title = {Continuous-Time and Discrete Spatio-Temporal Graph Neural Networks
for Normal and Extreme Weather Prediction on South African
Station Data},
author = {Jaravani, Praise},
school = {University of Cape Town},
type = {Honours project},
address = {Cape Town, South Africa},
year = {2026}
}BibTeX — diffusion workstream (Chikwanha)
@mastersthesis{chikwanha2026diffusion,
title = {Weather Flow Prediction using Diffusion-Based Spatial-Temporal
Graph Neural Networks},
author = {Chikwanha, G{\'e}nevieve},
school = {University of Cape Town},
type = {Honours project},
address = {Cape Town, South Africa},
year = {2026}
}BibTeX — the project as a whole
@misc{weatherflow2026,
title = {WeatherFLOW: Weather Prediction using Spatio-Temporal Graph
Neural Networks},
author = {Chikwanha, G{\'e}nevieve and Fuyana, Zoleka and
Jaravani, Praise},
howpublished = {Honours project, Department of Computer Science,
University of Cape Town},
year = {2026}
}About this site
Static HTML, one hand-written stylesheet and one small script. No framework, no build step at view time, and no external requests — every asset, including the two typefaces, is served from this archive by relative path, so the site renders identically whether it is opened from a web server or straight off disk.
The script is progressive enhancement only: it adds scroll reveals, a reading-progress indicator and the counting statistics. With scripting disabled the site is fully readable and every link works — nothing is hidden behind JavaScript. All motion is suppressed for visitors whose system requests reduced motion.
- Markup
- Hand-written HTML, validated against the W3C Nu HTML Checker
- Styling
- One stylesheet — CSS custom properties, grid, flexbox; no preprocessor
- Scripting
- One 4 KB script, no dependencies; enhancement only, honours reduced-motion
- Typefaces
- Source Serif 4 and Inter, both SIL Open Font License 1.1, subset to Latin and vendored locally
- Figures
- Extracted from the final-paper PDFs at 240 dpi
- Hero graphic
- SVG generated from the project's own station coordinates
- External links
- Three — the departmental project archive and the two live GitHub repositories, each labelled where it appears