Google Summer of Code 2025: "Adjuster this! TabPFN as a replacement for the adjuster" project discussion thread

[Sukh-P]

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 thread is for questions and clarifications, not for formal applications, formal applications are done via Google Summer of Code https://summerofcode.withgoogle.com/.

Project Description

For the OCF Quartz Solar forecasting model, we have a simple “adjuster” model. It currently looks at the pattern of errors in the last week, and adjust the new forecast. We would like to experiment with a new foundational timeseries forecasting model, TabPFN, with the planned outcome to dynamically adjust our PV forecasting model based on historical errors and improve the forecasting skill. Can compare to current rules based averages method used.

Expected Outcome

An answer to whether a pretrained ML model such as TabPFN can provide better adjustment of solar forecasts than rules based methods. If yes, then the plan would be to incorporate this into OCF's production system.

Other Key Information
Expected Size: 90hrs (small)
Skills: ML knowledge + familiarity with ML tools, timeseries forecasting, working with tabular data + some data processing skills required e.g. pandas
Difficulty level: Easy
Related Reading: https://github.com/openclimatefix/india-forecast-app/blob/main/india_forecast_app/adjuster.py, https://github.com/PriorLabs/TabPFN
Potential mentors: @Sukh-P

[PredictiveManish]

Hey @Sukh-P, this project is pretty good and I am interested in working and contributing to this project, I will be very grateful if you can share something which is must to go before familiarizing with codebase. Thanks in advance.

[Sukh-P]

Thanks for your interest @PredictiveManish, we are looking for someone who is familiar with timeseries forecasting and applying libraries such as https://github.com/PriorLabs/TabPFN to tabular data, in terms of familiarising yourself with our codebase there is not too be aware from apart from the above python file which gives the implementation of the current baseline (the adjuster) but that can be gone through during the project, I would say it would be good to get familiar with TabPFN if you are going to apply

[anshulg954]

Hi Sukh!
One of the contributors here from TabPFN-Client from ML Lab, Freiburg. Really intrigued by seeing this project in GSoC category as I've been working with TabPFN from since over an year now. I feel TabPFN-TS can help a lot in this use case where we can just make zero shot predictions by passing X_Train as previous history of errors and required adjustments while predict on the future adjustments.
Lets connect! Feel free to have a look at my resume :)

Resume_Anshul.pdf

[Sukh-P]

Hi @anshulg954 oh very cool! Yes we are interested in it's ability to make zero shot predictions in our applied case, are you applying to GSoC?

[anshulg954]

@Sukh-P Yup. I can be a part of the GSoC (since I am a Master's student as well at University of Freiburg pursuing Specialization in AI, and currently doing my Masters Project on enhancing Text capabilities for TabPFN) and can contribute individually if your team wishes to. Should I draft a proposal or a plan that can make this work - Maybe have some DFDs and how we can proceed?

[Sukh-P]

@anshulg954 I think the formal application through GSoC is the place where you can list your detailed plans, but feel free to shoot anymore questions my way, hopefully by looking at some of the other questions on this page you will have a better idea of what this project could entail, thanks!

[devojyotimisra]

Hi @Sukh-P,

I’m interested in contributing to this project, specifically in integrating TabPFN to improve the adjuster model. My plan includes:

• Understanding the current adjuster – Analyzing how it adjusts forecasts based on past errors.
• Benchmarking TabPFN – Running it on historical PV forecasting errors and comparing it with the existing rule-based method.
• Developing a dynamic adjustment approach – Using TabPFN to predict and correct forecast errors more effectively.
• Evaluating performance improvements – Measuring error reduction and forecast skill enhancement.
• Proposing integration strategies – Ensuring compatibility with the existing OCF pipeline.

I have experience in ML, and data processing (pandas, NumPy, Scikit-Learn) and would love to collaborate on refining this approach. Looking forward to feedback and next steps!

One more question which discord server, or similar communities should i join?

Best,
Devojyoti Misra
Devojyoti Misra.pdf

[Sukh-P]

Hi @devojyotimisra thanks for the detail, although just a reminder you will have to do a formal application through GSoC and that's where you can list your detailed plans, in terms of slack/discord I think GH discussions is the best place for now to ask questions

[SophiaLi20]

Hey @Sukh-P ,
This project sounds really exciting! I’ve worked with ML and data processing in Python, especially with Pandas, and I’d love to dive into how TabPFN could improve solar forecasting adjustments.
I had a few questions to get a better sense of the approach:

• What specific metrics will be used to compare TabPFN with the current rule-based method?
• How well does the existing adjuster model handle sudden weather shifts, and is TabPFN expected to improve that aspect?
• Are there any major challenges when it comes to integrating TabPFN into the current system?
• Will the model be retrained regularly, or is it more of a one-time setup with occasional tweaks?

I took a look at the repos, and the adjuster logic is pretty interesting. Looking forward to learning more and seeing where this goes!

[Sukh-P]

Hi @SophiaLi20 thanks for the interest, in terms of your questions:

What specific metrics will be used to compare TabPFN with the current rule-based method?

MAE, and some metrics for large errors e.g. the top and tail percentiles of errors will be interesting to see

How well does the existing adjuster model handle sudden weather shifts, and is TabPFN expected to improve that aspect?

We think the current method perhaps is a bit naive and can be improved upon, one of the questions to be answered is that true and does TabPFN do better than the metrics above

Are there any major challenges when it comes to integrating TabPFN into the current system?

One to think about during the project I think in terms of how you would deploy it but I imagine it shouldn't be too difficult to incorporate

Will the model be retrained regularly, or is it more of a one-time setup with occasional tweaks?

TabPFN is a foundational model so should not required retraining on our data, it would be interesting to see how it would perform without retraining first (zero shot prediction)

Thanks!

[SophiaLi20]

Hey @Sukh-P ,
Thanks for the detailed response!

The choice of MAE and percentile-based error metrics makes a lot of sense, especially for capturing extreme deviations. It'll be interesting to see if TabPFN improves performance in those edge cases.
For handling sudden weather shifts, I agree that the current adjuster might be a bit simplistic. It’ll be exciting to analyze whether TabPFN can dynamically adapt better, especially without retraining. I'll look into how well it generalizes across different scenarios.
For integration, I’ll keep an eye on how easily it can fit into the current system. If it works well with zero-shot predictions, that could make deployment much smoother.
Looking forward to diving in! I’ll start by getting familiar with the existing adjustments and TabPFN’s capabilities. Let me know if there’s anything specific you'd recommend exploring first.

[yuvraajnarula]

Hey @Sukh-P,

I'm intrigued by your TabPFN solar forecasting project.

Some questions I'm curious about:

1. Does your current "adjuster" perform differently during rapid irradiance changes (morning ramps, passing clouds) versus stable conditions?

2. Have you identified any cyclical patterns in the forecast errors that follow diurnal or seasonal cycles but aren't captured by simple averaging?

3. For feature engineering: are you considering using the gradient of recent errors rather than just absolute values? TabPFN might detect acceleration patterns in error trends.

4. OCF Quartz uses what forecast horizon? Is the error structure significantly different between short-term (1-3hr) and day-ahead horizons?

5. Have you explored whether certain weather types (clear, scattered clouds, overcast) have systematically different error characteristics that TabPFN could learn to recognize?

For implementation, you might consider framing this as multiple specialized TabPFN models for different atmospheric stability conditions, then ensemble their outputs. The project README suggests this kind of domain-specific approach often outperforms a one-size-fits-all model.

Would it be possible to share a sample distribution of your historical errors to better understand the patterns TabPFN would need to learn?

I'm particularly impressed by how this project tackles a critical problem in renewable energy integration - better solar forecasting directly impacts grid stability and energy market operations. TabPFN's approach seems uniquely suited to capture the complex, non-linear patterns in solar forecasting errors.

[Sukh-P]

@yuvraajnarula sorry for the delay in getting back to you, thanks for the questions!:

1. The current adjuster definitely prefers conditions where the error is consistent across time of the day/forecast horizon, I imagine it doesn't do that well for rapid ad-hoc changes.
2. Not investigated this greatly I believe
3. Might need a bit more detail here in terms of what you mean by gradient of recent errors
4. We forecast generally intraday (0-8 hours) and and day ahead (0-48 hours) with more focus on the 24-48 hours, so it would be interesting to see the impact on both of these
5. Not in detail

An ensembling technique could definitely be interesting to explore, although for what you describe we would definitely have some work to do to isolate those different weather conditions and create different dataset subsets for them. In terms of sharing data I will get back to you, we could share some perhaps in terms of shape/schema but I don't think we can share any data publicly sadly

[yuvraajnarula]

@Sukh-P Thanks for your detailed responses! Your work on TabPFN for solar forecasting is tackling a crucial challenge at the intersection of ML and renewable energy integration.
I've been thinking about some of these forecasting challenges in my own work and found your approach intriguing, particularly how you're handling the complex error patterns.

• For the "adjuster" issue with rapid changes, have you considered using a multi-resolution temporal analysis? This could help identify pattern changes at different time scales simultaneously.
• On error gradients, I'm specifically interested in how TabPFN might detect acceleration in error patterns before they manifest as large absolute deviations - potentially providing earlier warning signals for forecast corrections.
• For specialized models based on conditions, I've had success with automated condition clustering techniques that can identify regime shifts without manual labeling.

I've worked with irradiance modeling in variable weather conditions before and would be happy to help analyze your error distribution samples if you're able to share them. Perhaps we could identify condition-specific patterns that might inform your ensemble approach.
I'd be happy to discuss further or contribute to specific aspects of the project that align with your current development priorities. Would you be open to a more detailed conversation about potential collaboration opportunities?

[Samarpan1122]

Hey @Sukh-P,

This project sounds really exciting! I’ve worked with ML and data processing in Python, particularly with pandas, scikit-learn, and some PyTorch, and I’d love to explore how TabPFN could improve solar forecasting adjustments.

TabPFN’s ability to outperform traditional methods while being dramatically faster makes it a compelling candidate for dynamically adjusting forecasts based on historical errors. I’m really interested in contributing to this project—what would be the best way to get involved? Let me know how I can help!
Thanks!

[Sukh-P]

@Samarpan1122 Thanks for your interest, just to point out that this is a proposed Google Summer of Code project which we will be part of this year https://summerofcode.withgoogle.com/ so to be involved in the project you would need to apply to it!

[chaitra-samant]

Hi @Sukh-P ,
I am Chaitra Samant a Second Year Computer Engineering Student from VJTI Mumbai. I'd like to contribute to this project under GSoC 2025.

I am familiar with Machine Learning, Deep Learning and Data Preprocesssing, I have worked with Pytorch and Pandas as well. I have actively contributed to Open Source projects, completed Hacktoberfest and won a Open Source Contribution Contest. I am currently researching about TabPFN and how it can be incorporated in this project.

I'd like to know more about the GSoC preparation and selection process. Are they any tasks that need to be completed or contributions that need to be made?

Is there anything else which I have missed and should work on?

Thanks,
Chaitra

[chaitra-samant]

Thanks for the info! I will look into the codebase and explore contributions. Looking forward for working with the project

[chaitra-samant]

Hi @Sukh-P, I was working on a similar application with TabPFN based adjuster. The results look promising, TabPFN significantly improved MAE ,RMSE and R^2 score with very minimal retraining.

I would also love to get some insights on the proposal, and what all should be included in one.

Thank you so much,
Chaitra
Chaitra_Samant_resume

[chaitra-samant]

Hello @Sukh-P,
I have prepared a detailed draft proposal, could you please review it for me and give me feedback/any changes which are required according to project requirements?

I was wondering how could I share it with you

Thanks,
Chaitra

[Sukh-P]

@chaitra-samant we won't be reviewing any applications before the deadline closes out of fairness since we would not have time to do this for everyone sorry

[chaitra-samant]

That's understandable, thank you for your time

Chaitra

[Markusabel-est]

Dear @Sukh-P ,

My name is Markus and I’m a Computer Science student at Radboud University in Netherlands. I’m new to contributing to open source software, but this project sounds like an exciting challenge.

Since I’m new to GSoC, I’d appreciate any insights into the organizational side of the program. For example, how are project milestones, communication, and progress tracking typically managed?
Also, What specific forecasting errors does the current adjuster model correct?

Looking forward to collaborating!

Best regards,
Markus Abel

[Sukh-P]

Hi @Markusabel-est, thanks for your interest in this proposed GSoC project, so in terms of the organisation side of things:

• There are usually weekly/fortnightly (depending on the length of the project) check ins between mentors and GSoC participants
• We usually will try to have enough milestone markers in the project so we can see if the original goals can be met within the original time (or if an extension is needed)

The idea of the current adjuster is that it will correct (crudely) for any changes in the distribution of PV generation that may have occurred between training our models and now, it takes average errors over a week broken down by the forecast horizon and time of day (half hour period) part of this project will be analysing the differences in the correct distribution between the current rules based adjuster and outputs from a TabPFN model, hope that clarifies things a bit!

[Arya-GITh]

Dear @Sukh-P ,

I am Arya Datla, a freshman majoring in CS and Math at Cornell University. I'm deeply passionate about solving climate change problems using computing, and while exploring projects in that area, I saw your project strongly aligned with my interest. Apart from my previous experience working with Python, pandas, and time-sequence models, which I have proposed for smart energy grids. I am working on an AI Chatbot for the Cornell energy team working on improving Cornell's Grid Efficiency as part of the Engineering for a Sustainable World (ESW) Project team.

Specific to your project and the conversations I have seen, I was fascinated to explore TabPFN and utilize it to improve the performance of the error adjuster. Some of my initial thoughts and questions are mainly about the current adjuster's reliance on linear patterns via averaging, but TabPFN could help capture more complex relationships. Moreover, richer datasets, such as seasonal information, could be used to train the model beyond just relying on the MAE. Of course, as discussed heavily, the zero-shot capabilities of TabPFN would make it a great way to capture errors and improve the model.

Some of the initial questions I pondered were
If there would be any latency or computational constraints of using a pre-trained model and the extent of fine-tuning that would have to be done to achieve the desired accuracy
how would the micro-deadlines and project milestones be implemented to gauge progress throughout the program timeline?

I will continue to read through more on the topic and share more thoughts, but I also look forward to your guidance at your leisure,

Best Regards,
Arya Datla

[Sukh-P]

Hi @Arya-GITh, thanks for expressing your interest in this project, in terms of your questions:

If there would be any latency or computational constraints of using a pre-trained model and the extent of fine-tuning that would have to be done to achieve the desired accuracy

I think assessing whether using TabPFN as a pretrained model with no fine tuning is better than the incumbent adjuster method is definitely one of the first questions to answer, in terms of computational/latency constraints I don't envisage any but having decent profiling to figure out how long it takes to do inference with different amounts of resources would be useful to understand it's viability as a deployed model.

how would the micro-deadlines and project milestones be implemented to gauge progress throughout the program timeline?

I think we would try to plan at the beginning of the project what those micro-deadlines would be e.g. getting TabPFN working with a dataset which shows historical PV forecasts we have produced and actual PV generation, comparing this to the adjuser based values, profiling for this, then depending on results there could be work done to to try and fine tune the model etc, also would be good to work on a system design of how this could be deployed in our current production setup if the results were positive, perhaps even doing some of the implementation work if time permitted, I hope that is useful although I know a little open ended currently

[Arya-GITh]

Thank you so much for taking the time to answer my questions. The deadlines in place as of now are more than enough for me to understand and get a basic grasp of how the project would flow across the 12 weeks. As for the computational constraints, there likely would not be any significant constraints to work around beyond just efficiently utilizing the resources available.

[Arya-GITh]

Dear @Sukh-P ,
I had submitted my application over a week ago, and I was under the assumption that there would be one or more cycles of feedback, and then I would send another final proposal after that. Is that only for applicatints who submitted their own ideas.
Thank you for the clarification!

[TreTjy]

Dear @Sukh-P

I hope you are doing well. My name is Han, and I am a graduate student at University College London (UCL), studying Energy Systems and Data Analysis. I am very interested in this Google Summer of Code 2025 project.
.
I have a background in machine learning, data analysis, and simulation modeling. At the same time, I was an undergraduate physics student and had experience of processing and predicting planetary information at the undergraduate level, which I believe aligns with the goals of this topic.

Some of my relevant experience includes:

Machine Learning and Deep Learning: I have a background in machine learning, deep learning that I have worked extensively with time series forecasting, energy data analysis and prediction, and geo-spatial data processing and visualization and data visualization. I work with Keras, TensorFlow frequently in my model and am familiar with them.

I have read the guidance and some tutorial videos from the reference link on the GSOC Idea website, but I am still a little confused about which step I should finish before applying. like do I need to write a real model of my idea or just try to list our idea?

I wonder what contributions I can make beforehand to strengthen my application and how to file the application form of GSoC.

I appreciate your time and look forward to your reply.

Best regards
Han

[Sukh-P]

Hi @TreTjy, so I think in your applciation referencing anything that can demonstrate your familiarity with timeseries forecasting problems, using TabPFN or other foundational timeseries models would help, we are looking for someone with overall ML experience, preferably working within energy/power forecasting or similar domains.

Please see our recent post here: https://github.com/orgs/openclimatefix/discussions/36 for more general advice on how to write a strong application, all the best!

[Dakshbir]

Subject: Interest in Contributing to the TabPFN Integration Project

Hi @Sukh-P,

I’m Dakshbir Singh, a Computer Science student at NSUT with a strong background in Machine Learning, Deep Learning, and timeseries forecasting. I’m highly interested in contributing to the TabPFN Integration Project under OpenClimateFix for GSoC 2025.

To get started and better understand the codebase, I would love to take on some intermediate-level, high-priority issues that need to be solved. It would help me gain familiarity with the existing framework and contribute meaningfully.

Looking forward to your guidance on how I can begin contributing.
Thank you!

Best regards,
Dakshbir Singh

[Sukh-P]

Hi @Dakshbir, sorry for the delay in getting back to you, please see https://github.com/orgs/openclimatefix/discussions/35 for contribution advice in general.

[RishiP2006]

Hi @Sukh-P,

I am a first year CS students in BITS Pilani, and I am learning ML and python and this projects aligns with the the things I know and want to further work on. I get that the TabPFN will be a better model than the existing one as it would have been already trained on lots of data and I am really enthusiastic to work on this for you in GSoC. I just wanted to know what we can do now to increase our chance of getting selected in GSoC or anything you guys are looking in particular.

Best Regards,
Rishi Prasad

[Sukh-P]

Hi @RishiP2006 please read the guidance we have laid out here: https://github.com/orgs/openclimatefix/discussions/36, thanks

[amol-rangrao-patil]

Hi @Sukh-P,

I am a second year CS students in Maharashtra India and I am learning ML and python and this projects aligns with the the things I know and want to further work on. I get that the TabPFN will be a better model than the existing one as it would have been already trained on lots of data and I am really enthusiastic to work on this for you in GSoC. I just wanted to know what we can do now to increase our chance of getting selected in GSoC or anything you guys are looking in particular.

Best Regards,
Amol Patil

[RishiP2006]

Hi @Sukh-P ,

I just wanted to know one more thing, and that is what is your current adjuster model like, which prediction algorithm does it use (as not a lot of information is given on it )

Thanks,
Rishi

[Sukh-P]

@RishiP2006 if you take a look at the file linked in the Discussion original post you can see the logic, but it is an averages based approach based on calculating the average errors from the past 7 days based on both the forecast horizon and hour of the day.

[vsl366]

Hello @Sukh-P !

I’m Vishal S, a second-year Computer Science student at SSN College with experience in machine learning, data processing, and time series forecasting. I’m proficient in NumPy, pandas, scikit-learn, and transformers, and I’ve worked with ARIMA models for time series analysis. I’m excited about the opportunity to contribute to this project as part of GSoC 2025 and would love to understand how I can get started.

To familiarize myself with the project, I have a few questions. First, where can I learn more about the existing codebase and its structure? What’s the best way for new contributors to demonstrate their open-source experience and get involved? Would you recommend starting with a specific first issue or small contribution to gain hands-on experience with the project?

Regarding TabPFN integration and optimization, the current adjuster relies on historical error aggregation, interpolation, and smoothing. Beyond replacing these steps, are there other aspects where TabPFN could enhance accuracy or efficiency? From your perspective, what are some key challenges in transitioning from the rules-based adjuster to a TabPFN model? Are there any constraints related to interpretability or explainability that should be considered when integrating TabPFN?

On the topic of performance, experimentation, and evaluation, are there specific latency or computational constraints that the new model must adhere to? How do you manage dataset versioning and experiment tracking to ensure measurable improvements in forecasting performance? What are the key evaluation metrics or baselines used to compare TabPFN’s effectiveness against the existing adjuster?

I’m eager to contribute to this project as part of GSoC 2025 and appreciate any guidance on how to get started. Looking forward to your insights!

P.S Sorry if im asking stuff too late, i couldnt fixate on any orgs and i just discovered this a day ago :)

Best regards,
Vishal S

[Sukh-P]

Hi @vsl366, please see https://github.com/orgs/openclimatefix/discussions/35 for contribution advice in general.
In terms of your questions:

• I think exploring constraints of TabPFN would be part of this project
• No I don't think interpretability or explainability are ares we will be focussing on for this project
• I think I have answered this before on another thread in this discussion but specific latency or computational constraints that the new model must adhere to are to be explored as part of the project, we would specify what constraints there would be as the project goes on and when we have some early results

[amol-rangrao-patil]

Hello Sukh-P and Team,

Thank you for the detailed project description. I’m really excited about the opportunity to contribute to the OCF Quartz Solar forecasting model. I have a strong background in machine learning, time series forecasting, and data analysis. My experience with tools like Pandas, NumPy, and Scikit-Learn aligns well with the project requirements.

I’ve already gone through the adjuster.py code and explored the TabPFN repository to understand its capabilities. I’m particularly interested in analyzing how TabPFN can enhance the solar forecast adjustment compared to the existing rule-based method.

I’d be happy to discuss any further questions or clarifications you might have regarding my understanding of the project. Looking forward to contributing and making a meaningful impact on the project.

Best regards,
[Amol Patil]

[FawwazRaza]

Hi @Sukh-P ! Thanks for all the helpful info so far. Is there a small example dataset or something simple I can start with to test TabPFN before trying it on the full system?
How do you usually track experiment results? Is there a tool or format you prefer us to use to keep everything clear?

[FawwazRaza]

@Sukh-P Thank you so much for the detailed info. I have some small questions to clear the confusion:

1. Do you want me to also try to find new patterns or relationships in the data, or just focus on making the adjustment model as accurate as possible?

2. Will we compare TabPFN results with other models or just with the current rule-based method?

Regards

Fawwaz Raza

https://www.linkedin.com/in/fawwazraza

https://fawwazraza.github.io/fawwazraza-portfolio/

Fawwazraza2024@gmail.com

[FawwazRaza]

@Sukh-P One more question:

Which model are you currently using for adjustment? Is it based on regression or some other technique?

Thank you for your time and consideration.

[FawwazRaza]

@Sukh-P

[Sukh-P]

@FawwazRaza if you take a look at the file linked in the Discussion original post you can see the logic, but it is an averages based approach based on calculating the average errors from the past 7 days based on both the forecast horizon and hour of the day.

[FawwazRaza]

@Sukh-P Thank you for your reply.

Yeah! I have reviewed the code and noticed that we have to compare the accuracy using tabPFN and angular.py code and then come up with the best approach. Am I right?

[Kartik-C-Mehta]

Hi @Sukh-P,

I am excited to contribute to this project. After examining the code while considering potential implementation hurdles. I seek your professional guidance on several matters:

• What will the system do when users lack historical error data completely or when data points are spread out infrequently? Can we create backup strategies or conduct supplementary checks to measure system performance when such scenarios occur?

• The current adjuster system relies on a simple rolling mean combined with historical forecast error interpolation. How would the TabPFN ML model overcome the substantial challenges you found when using basic forecasting algorithms?

• We need to evaluate potential alignment problems between forecast and actual generation data which will impact our integration activities.

I want to gain further knowledge about these components to help achieve successful project outcomes. I look forward to collaborating!

[Sukh-P]

Hi @Kartik-Mehta-44, thanks for your interest and questions, you can assume there will always historical error available for this work, it is part of this work to explore whether TabPFN is more effective and how it performs over different weather conditions, thanks

[lokhans]

Hello Mr. @Sukh-P! I'm Sayura Lokhande, a freshman at Rose-Hulman Institute of Technology.

I wanted to express interest in this project and also ask (if this is the right place to) of what other ways we can contribute to Open Climate Fix besides GSoC. I feel that OPC's mission speaks to me and the organization aligns with my future goals, with the way it uses computational models to solve climate problems. So I hope to continue interacting with it beyond this program.

Have a nice day!

[Sukh-P]

Hi @lokhans, thanks for your interest, please see https://github.com/orgs/openclimatefix/discussions/35 for general contribution advice for Open Climate Fix

[abelsaj2005]

Hi @Sukh-P, I'm interested in working on this project! I have experience in ML, Python, and data processing, and I’ve worked on projects involving predictive modeling. I previously built an LSTM model for time series forecasting where I predicted the NVIDIA stock price, which gave me hands-on experience with sequence-based prediction techniques. I'm also a Computer Science and Statistics major at UNC, which allows me to analyze and model complex datasets effectively. I'm excited to explore whether TabPFN can improve solar forecasting adjustments compared to rules-based methods.

I'm looking forward to getting involved!

[prajakta1321]

Hello @Sukh-P ,

I'm Prajakta, an aspiring GSoC contributor with experience in machine learning, Python, and data-driven modeling. I’ve worked with tabular data, predictive modeling, and model evaluation, which makes me excited about this project. I had a few questions to better understand its scope:

1. Are there specific scenarios where TabPFN struggles compared to traditional methods, and how can we mitigate them?

2.Would optimizing TabPFN’s inference time be a priority, and what constraints should we consider?

3. How much flexibility do we have in modifying TabPFN’s architecture to better fit the adjuster’s needs?

Looking forward to your insights. Thanks for your time!

Best,
Prajakta

[Sukh-P]

Hi @prajakta1321, thanks for your interest and questions:

1. Very much will be part of this work to explore that
2. It would depend on how long inference takes, part of this work will be to evaluate that
3. I don't think the scope of this project will allow for changes in TabPFNs architecture, however there could be some training/fine tuning potentially rather than just zero shot predictions
   Thanks

[prajakta1321]

Hi @Sukh-P ,

Thanks for your response and for clarifying the project scope! I appreciate the insights.

1. Since part of the work involves evaluating where TabPFN struggles, do we have a benchmark or specific datasets in mind to test its performance? I'd love to explore how it compares across different data distributions.

2. Regarding inference time, do we anticipate any trade-offs between speed and accuracy? If so, what strategies could be viable for maintaining robust performance while optimizing inference?

3. You mentioned fine-tuning/training adjustments rather than architectural changes. Would meta-learning techniques or adaptive hyperparameter tuning be something we can explore to enhance performance in different scenarios?

Looking forward to your thoughts! Thanks again for your time.

Best,
Prajakta

[Shalvigour]

Respected sir/mam,
I hope you're doing well. I recently came across the Open Climate Fix project, and I genuinely appreciate its concept and impact. I was keen on contributing earlier, but due to health issues and university exams, I couldn’t start on time. However, I’m now fully available and eager to contribute. I understand that an early start is ideal, but I wanted to check—if I begin contributing now, would I still have a fair chance of getting selected for GSoC ? If yes, I’d really appreciate any guidance on how I can make a meaningful contribution within the remaining time.

[val2021-svg]

Hello @Sukh-P !

Super stoked about the "Adjuster this! TabPFN" project for GSoC 2025! Really digging the idea of using TabPFN to give your solar forecasting a boost.

I've been playing around with time series forecasting and even got some experience using TabPFN on tabular data, so I'm keen to see how we can make it work here.

Just a few quick questions to get the ball rolling:

• Any pointers on where to find the data and what format it's in?
• Are there any specific metrics you're most interested in for measuring forecast accuracy?
• Anything I should read up on in your existing codebase before diving in?
• If we get TabPFN working, are there any particular things to keep in mind for production integration?

Looking forward to hearing from you and hopefully getting started!

[bgprudhomme]

Hello @Sukh-P,

I am a second-year CS PhD student and am very interested in AI and machine learning, especially for contributing to important issues such as climate change. I have taken Andrew Ng’s Machine Learning and Deep Learning online course specializations, and am currently taking ML in graduate school, where I am actively working on a project focused on predicting household energy usage based on historical consumption patterns. Given my familiarity with Python, and tools like pandas, sci-kit learn, and PyTorch, I am very interested in contributing to this project.

I’ve been digging into the current rules-based adjuster for the OCF Quartz solar forecasting model and I really like the idea of improving it with a learned model like TabPFN. I understand that the current method averages errors over the last week by horizon and time of day, which is solid for steady patterns but might underperform during fast-changing conditions like sudden cloud cover. I think it will be very interesting to see if TabPFN can better handle those rapid changes in a zero-shot setting.

If it’s not too late for me to ask, I’d love to hear what you think about the following questions I have:

1. In preparing my proposal, I’d like to make sure the project milestones reflect a structure that works well for OCF. I saw your note about regular check-ins and milestone markers — are there particular phases you’d expect to see, or should I treat milestone setting as entirely flexible depending on the approach I propose?

2. I noticed that the system doesn’t explicitly handle noisy or missing data — it skips any forecast-generation pairs that don’t align on timestamp, and defaults to a zero adjustment if no ME is available. This seems like a reasonable conservative fallback, but I wonder: have you found that missing or erratic data meaningfully affects the current adjuster in live usage, and would this become more of a concern when using a model like TabPFN?

Thanks and cheers,
Benjamin

[ben002mr]

Hi Sukh,

I'm very interested in working on the "Adjuster this! TabPFN as a replacement for the adjuster" project for GSoC 2025. I'm at the start of a long journey to transition into data science. I'm comfortable with Python, Pandas, and learning new tools quickly.
This project looks like a great fit for me as I begin contributing to ML-based solar forecasting, and I’m especially excited to explore the use of TabPFN for improving forecast adjustments. It'll be great if you can share something which is must to go before familiarizing with codebase.
I'd love to begin contributing to this project.
Looking forward to your guidance!

Thank you,
Benhein

[ferro-gianluca-29]

Hi @Sukh-P,

I’m Gianluca Ferro, and I graduated with a master’s degree in Electronic Engineering. I’m now a research fellow at ciparlabs, working on artificial intelligence for smart grids and renewable energy communities. My master’s thesis focused on time series forecasting, where I developed a framework of models (including gradient boosting, statistical methods, and neural networks) to predict energy consumption and production from renewable sources.

I find your TabPFN-based adjuster project extremely interesting because it aligns perfectly with my interest in applying advanced ML solutions to improve solar forecasting. I’m eager to experiment with integrating TabPFN into an existing pipeline and train models using your data, aiming to enhance the overall forecast accuracy.

A couple of questions I had in mind:

Since TabPFN is a tabular-focused model, do you have any specific strategy in place for capturing the temporal structure of forecasting errors, or would we primarily rely on feature engineering to introduce time-related variables?

Are there any real-time or near real-time constraints for the adjuster model’s inference process, and if so, how might we optimize TabPFN’s performance for such scenarios?

I look forward to your feedback and the possibility of contributing to this project.

Best regards,
Gianluca

[ben002mr]

Hi Sukh,

I’ve just submitted my GSoC 2025 proposal for the project titled “Improving Solar Forecast Accuracy Using TabPFN for Dynamic Adjustment.” The project explores using TabPFN to replace the current rules-based adjuster in the Quartz solar forecasting pipeline.

Here’s a brief summary:
The proposal aims to leverage TabPFN to dynamically adjust PV forecasts based on recent errors. It includes adapting TabPFN for this task, comparing it to the existing method, and integrating it into the Quartz pipeline if it proves effective.
I've really enjoyed digging into the existing codebase and learning more about OCF’s work. If you have any feedback or suggestions on the proposal, I’d love to hear it!

Looking forward to the opportunity to contribute.
Best Regards,
Benhein Michael Ruben L (Ben)
LinkedIn | Email: harihusseler@gmail.com

[Divyateja2709]

Hey Sukh P Sir!
Just wanted to share an update — I’ve submitted my proposal for GSoC 2025! The project I’m working on is titled “Adjuster this! TabPFN as a replacement for the adjuster.”

I’m also actively contributing to the project right now, exploring different approaches to improve forecast adjustments using machine learning. It's been an exciting learning experience so far!

Would love to hear your thoughts or any advice you might have!

[emlweb]

Google Summer of Code 2025 applications are now closed.

We are currently reviewing all applications. Contributors will be announced 8 May 2025. Thank you!

[debarghya18]

Hi @Sukh-P

I am probably late to the party. But hi. I am very interested in this project and I'm considering contributing to it in GSOC'26. So, can you give me the 101s about the project and where I should start

[AnonymousVX]

Hey @Sukh-P , this project is really nice and I am keen towards working for this project, I will be thankful if you can share something which is must knowing before understanding with the code.
Thankyou!