Google Summer of Code 2025: Open Data PVNet Discussion Thread

[peterdudfield]

This space is for you to ask any questions you have about this project. We're here to provide clarifications and help you understand the project's goals, scope, and requirements. Feel free to ask about anything that interests you!

Please note that this discussion is for questions and clarifications, not for formal applications.

Project Description

We're building an open-source solar forecasting pipeline to integrate with OCF's PVNet model, using publicly available Numerical Weather Prediction data to forecast solar generation at the national level, starting with the UK. Currently, our main forecasting tool, Quartz Solar, is trained using a mixture of public and private datasets, and we want to create an effective model that uses 100% open data.

The data is ready to go, but we need a ML engineer to train the model. The aim will be to start with a UK forecast, but then extend to different countries.

Expected Outcome

A UK ML Solar forecast
trained on free NWP data with the accuracy benchmarked.

Other Key Information

• Expected Size: 170hrs
• Skills: ML knowledge, Pytorch, Python. NWP knowledge is a bonus
• Difficulty level: Medium
• Related Reading: https://github.com/openclimatefix/open-data-pvnet
• Potential mentors: @peterdudfield, @Sukh-P

[aayushyatiwari]

Hey!
I like what you guys are doing here. I'd like to contribute and learn. What type of model architecture are we looking at with regards to the open data model? Where should I go to, to understand the previous model and the accuracies it produced?

[peterdudfield]

Hi @aayushyatiwari , thanks for getting in contact. This repo has some good information for gettin started - https://github.com/openclimatefix/open-data-pvnet. There are some issues in there.

We've typically used PVnet and we'll be able to share some validation results soon.

[peach280]

I am interested in contributing to Open Climate fix for gsoc 2025.
I have a couple of questions regarding this project idea.

I'm eager to apply my Python and machine learning fundamentals. How much guidance will be available for the NWP specific aspects?
How much of the work will be focused on model building, versus data preprocessing?
How much of the work will require cloud computing knowledge?"

[peterdudfield]

1. We can provide guideance for getting new NWP data. We atcually have 2 NWPs almost ready to go, so it'll probably not focusing on that.
2. Hopefully lots of model building
3. We' hoping this can be done locally. The NWP data will be on the cloud, but hopefully accessible in a zarr format, so you dont have to download it all at once

[crs7617]

Hi, @peterdudfield

I'm Sairam, and I have a strong passion for machine learning, particularly with Python and LLMs. The Open Data Solar Forecasting Pipeline project has really caught my attention, and I'm excited about the possibility of contributing to it as part of GSoC '25.

I've begun looking into the PVNet repository to familiarize myself with the current work, and I've also checked out some "good first issues." Recently, I commented on an issue in the Analysis Dashboard repository and am now waiting to be assigned so I can make my first contribution to Open Climate Fix.

In addition to training the model using open NWP data, I would like to clarify what specific contributions you are looking for from me. Should I concentrate on enhancing model performance, integrating new datasets, or is there another area you’d prefer I focus on?

Moreover, since PVNet is a key component of this project and has piqued my interest, would it be more beneficial for me to tackle a "good first issue" within PVNet instead of working on another repository? If that's the case, do you have any suggestions for an issue that would align well with the project's goals?

I'm eager to begin with an initial contribution that supports the long-term objectives.

I look forward to your guidance!

[saurabh-3110]

Hey!!

I am highly interested in the intersection of climate science, machine learning, and sustainability, and I find the work being done here particularly intriguing. I have experience working with meteorological variables and applying machine learning techniques, and I would love to contribute to the ML modeling efforts.

After reviewing the repository, I understand that the initial focus is on predicting the target variable for the entire UK region. Please correct me if I am mistaken. I would appreciate any guidance on how I can best contribute to the project.

[vijayabhaskar78]

Hi @peterdudfield,

I'm S Vijaya Bhaskar, a machine learning enthusiast with strong Python and PyTorch skills, and I have a keen interest in climate data and sustainable tech. I'm really excited about contributing to the solar forecasting project. I'm currently also contributing to OpenClimateFix in the elexonapi, so I'm getting familiar with the ecosystem. I noticed that while there aren't any issues labeled as "good first issues" at the moment, there are several open issues. Could you recommend one that would be a great starting point for someone with my background? I'm eager to dive in—whether it's improving model performance, integrating new datasets, or any other area where support is needed.

Thanks for your guidance!

[ScarWar]

Hi! This GSoC 2025 project on transitioning PVNet to a public-data-only model is very intriguing. I've been following the advancements in Numerical Weather Prediction (NWP), particularly the application of diffusion models, hybrid models, and transformers, and I'm excited to see this applied to PVNet.

I understand the core objective is to migrate PVNet from a mixed public/private dataset to a purely public dataset. This raises several key questions, and I'd love to gain a clearer understanding of the challenges:

• Performance Delta: How significant is the performance degradation observed when switching from the mixed dataset to a public-only dataset? Quantifying this delta would help define the scope of improvement needed.
• Previous Public-Only Attempts: Has PVNet been trained solely on public data before, and if so, what were the limitations encountered? Understanding past attempts can guide future strategies.
• Specific Challenges with Public Data: Are there specific characteristics of the public dataset that make it challenging for PVNet? (e.g., data sparsity, lower resolution, limited temporal coverage).
• Focus Areas: Beyond dataset migration, what are the primary focus areas for improvement? Are we aiming to:
  • Enhance model architecture to compensate for the reduced data richness?
  • Optimize training strategies to maximize the information extracted from public data?
  • Improve inference efficiency for practical deployment?
  • Address specific forecasting skill score metrics?
• Evaluation Metrics: Which evaluation metrics are most critical for this project? (e.g. RMSE, MAE, CSI, etc.)

I'm particularly interested in exploring potential architectural modifications

Thank you for your time, and I look forward to learning more about this project!

[SophiaLi20]

Hey ,
This project sounds super interesting! I have experience with Python and Machine Learning, and I’ve been diving into PyTorch recently, so I’d love to learn more about how the model will be trained and evaluated.
I had a few questions to better understand the scope:

• Is there a specific baseline model you’d like to compare against for benchmarking accuracy?
• Since the focus is on using 100% open data, are there any known challenges with public NWP datasets—like missing values or preprocessing steps—that we should be mindful of?
• Will this model be directly integrated into PVNet, or will it start as a separate module?
• What key metrics will you be using to measure forecast accuracy?

I checked out the repo, and it looks really cool! Looking forward to learning more and getting involved.

Best,
Unnati Bhardwaj

[ajitashwathr10]

Hi @peterdudfield

I'm Ajit Ashwath, and I’m really excited about this project. I’ve been working with ML for a while, mainly using Python, PyTorch and TensorFlow, and the idea of applying ML to renewable energy is really interesting to work on. I’d love to contribute to improving solar forecasting with open data. Before that, I do have some questions to ask.

Since this model will be trained solely on a public NWP dataset, are there any known issues or limitations when using it? Does it involve a lot of preprocessing we need to consider?

And, are there any particular, specific datasets available for expansion outside of the UK?

Looking forward to hearing more!

[Jiya873]

Hi @peterdudfield,

I'm Jiya Gupta, and I'm interested in contributing to this project and would love to get involved. Could you guide me on where to start? I'm particularly keen on understanding how the data pipeline is set up and how I can help in making the data samples more manageable for ML training.

Looking forward to your guidance!

[yeswanth244]

Hey @peterdudfield ,

I'm Yeswanth, and I'm really excited about contributing to the Open Data PVNet project as part of GSoC 2025. I have a strong background in Python and machine learning, and I'm eager to build the model from scratch using open Numerical Weather Prediction (NWP) data.

I've already started exploring the PVNet repo to understand the existing framework and get familiar with how things are structured. Before diving in, I wanted to clarify a few things about the scope and direction of this project:

Key Questions:

Core Model Goals – Since we’re building this from scratch, what’s the primary focus?

• Are we aiming to replicate PVNet’s performance using only public data?
• Or are we looking to redesign the model to work better with open datasets?

Challenges with Open Data – What are the biggest obstacles when training a model only on open-source NWP data?

• Are there known gaps in data coverage, missing variables, or resolution issues?
• Have previous models struggled with this transition?

Baseline Comparison – Is there an existing baseline we should aim to match or improve upon?

• Would it be useful to run an initial benchmark with PVNet using only public data to see the current gap?

Model Architecture Decisions – Should we reuse parts of the PVNet model or start fresh?

• If certain components from PVNet can still be useful, which ones?
• Are there any recommended ML architectures (CNNs, Transformers, hybrid models) that have worked well for this task?

Training and Deployment – What’s the expected workflow?

• Can this be trained locally, or will we need cloud-based training?
• Will the final model be integrated back into PVNet, or will it function as a separate module?

First Steps – What’s the best way to get started?

• Would you recommend contributing to a small issue in PVNet first to understand the codebase?
• Or should I focus directly on defining the dataset pipeline and preprocessing steps?