GSoC_2025_report.md

Google Summer of Code 2025 Final Report

Summary

I contributed to Open-Data-PVNet, a project under OpenClimateFix, where we are adapting the PVNet solar generation forecasting model to work entirely with open, publicly available data. The aim is to make high-quality solar generation forecasts accessible to anyone, anywhere, without requiring proprietary datasets. You can learn more about the base PVNet model here.

Goals

The main project goals were:

• Identify and integrate a freely available, long-term NWP dataset suitable for training and inference.
• Collect ground truth generation data to validate forecasts.
• Reduce friction for new users by simplifying setup across multiple repositories.

My Work

1. Integrating GFS Weather Data

After evaluating options, we selected GFS (Global Forecast System) as the primary NWP dataset. I implemented GFS support in ocf-data-sampler, enabling the pipeline to transform GFS forecasts into training samples.

• PR: #199

2. S3-Based Training Workflow

A major bottleneck was the requirement to download large volumes of weather data locally. To solve this, I added support for streaming training directly from S3 buckets, removing the need for local storage of terabytes of data.

• PR: #267

3. Ground Truth Data Collection

I reviewed and tested the addition of a script to collect UK solar generation ground truth data (2020–2024). This dataset is critical for benchmarking and validating PVNet forecasts.

• PR: #95

4. Simplifying Model Setup

Previously, training required cloning the model repository separately. I worked on reducing this friction by allowing configuration and training directly within open-data-pvnet, making onboarding much easier.

• PR: #101

5. Pipeline Testing

We verified the full pipeline by running experiments on subsets of data, ensuring the integration worked end-to-end. Some of these experiments were logged on Weights & Biases:

Test run 1	Test Run 2

6. Documentation

I am preparing training instructions and documentation to guide new contributors and users through the setup and training process. (TBD)

Future Work

• Extend support beyond GFS to other open NWP datasets and incorporate satellite data for improved forecasts.
• Expand PVNet training to datasets outside the UK, unlocking applications across Europe and globally.
• Updating compatibility to latest versions of PVNet and ocf-data-sampler.
• Continue refining documentation and usability to grow adoption.

---

This completes my GSoC 2025 contribution. Links to my main work:

• ocf-data-sampler PRs
• open-data-pvnet PRs