Google Sumer of Code 2025: Improving Probabilistic Solar Forecasts

[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

Quantile regression approach in probabilistic solar forecasting directly estimates specific quantiles of the predictive distribution. Instead of direct quantity outputs, the proposition is modification of final layers to predict parameters of Gaussian Mixture Model(s) - for each component the mean, STD and coefficient. This should fundamentally capture more complex multimodal uncertainties inherent within probabilistic solar forecasting.

Expected Outcome

Complete continuous probability distribution as opposed to fixed quantiles. Better representation of uncertainties (such as transitions) and better capture of bimodal scenarios. Overall hopefully more accurate modelling and pattern comprehension - stronger uncertainty bounds perhaps.

Other Key Information

• Expected Size: 90hrs
• Skills: Academically - statistics / mathematics / engineering to Masters level. Experience implementing GMMs from scratch. ML/DL/PyTorch. Related architectural knowledge.
• Difficulty level: Easy-Medium
• Related Reading: https://github.com/openclimatefix/PVNet, https://scikit-learn.org/stable/modules/mixture.html https://github.com/ldeecke/gmm-torch
• Potential mentors: @felix-e-h-p

[arpity22]

Hi @peterdudfield and @felix-e-h-p ! 👋

I'm Arpit Yadav, a third-year undergraduate student at IIT Kanpur, majoring in Mechanical Engineering. My background includes computational modeling, machine learning, and sustainability research. I have contributed to cargo-semver-checks, an open-source linter in the Rust ecosystem, and have worked with data-driven forecasting techniques.

I am also active as a student researcher in the field of sustainable engineering (you can check out my work here ).

I'm particularly excited about this project because probabilistic solar forecasting is a crucial challenge in renewable energy, and using GMMs to capture multimodal uncertainties is an innovative approach. I'm looking forward to making meaningful contributions moving forward.

[yuvraajnarula]

Hi @peterdudfield and @felix-e-h-p,

The GMM approach to probabilistic solar forecasting is an elegant way to capture multimodal uncertainties. A few considerations that might be worth exploring:

• Mixture Component Selection & Temporal Consistency: Have you determined the optimal number of Gaussian components? In my experience, 2-3 components often suffice to distinguish clear-sky and cloud-obscured states, while more than 4 risks overfitting. Additionally, ensuring temporal coherence in the GMM parameters can be challenging—have you considered integrating a recurrent layer (e.g., LSTM) before the GMM output to improve consistency across forecast horizons?

• Loss Function & Computational Efficiency: For GMM parameter estimation, using negative log-likelihood with regularized covariance matrices has helped mitigate numerical instability, particularly during rapid weather transitions. Given the computational demands of real-time inference, are you exploring techniques like low-rank covariance approximations or variational inference to optimize performance, especially for edge deployment?

• Evaluation & Data Augmentation: Beyond CRPS, have you considered metrics like the energy or variogram score, which better capture multimodal distributions? Given the limited availability of high-resolution solar irradiance datasets, synthetic data augmentation (e.g., GANs or physics-informed perturbations) could also improve model generalization.

• Hybrid Approaches & Downstream Applications: Have you explored blending GMM with physics-based models for better interpretability and robustness? Also, how do you envision integrating uncertainty quantification into decision-making workflows, such as energy dispatch optimization or grid balancing?

I recently worked on a similar system using PyTorch's mixture module with custom loss functions that penalize physically implausible transitions between mixture states, improving forecast stability during dawn/dusk transitions and partly cloudy conditions. Would love to hear your thoughts on incorporating exogenous variables like cloud opacity or air mass into the mixture components.

[PragyaKum1]

Hello, @peterdudfield and @felix_h_p,
I really like your project! I am a student from Arizona State working in the Grand Challenge Scholars Program with a focus in Sustainability. So I'm just very interested in your organization in general! I would like to ask a couple questions.:

1. What kind of model architecture are you using before you transform the final layers for the GMM?
2. Are you primarily working with time related data, satellite imagery, or numerical weather predictions?
3. How is PVNet related to this particular project? Are you using it to extract spatial features for probabilistic solar forecasting, or is it included for a different reason?

I have experience working with TensorFlow and Keras, particularly in training deep learning models for classification tasks. I'm also minoring in Statistics. I'd love to understand how I can best prepare to contribute to this project effectively. Are there any specific resources or areas I should focus on?

Thanks!
Pragya

[felix-e-h-p]

Hi Pragya, hope you are well.

PVNet 2.1 utilises a late stage multi-modal fusion architecture. Separate encoder networks process each data type. These, along with GSP ID (via embedding) and solar position, are concatenated into a single vector. The project broadly focuses on replacing the current quantile output with a GMM.

Specifically regarding data modalities, we use all three: time-series data (historical PV output), satellite imagery and numerical weather predictions.

The project itself, fundamentally, is about improving PVNet. I.e. modifying PVNet's output layer to produce GMM parameters for probabilistic forecasting - replacing and/or augmenting the current quantile approach.

Thanks for your message - Felix

[PragyaKum1]

Hi Felix, thanks a lot for your reply!
I'd like to contribute to this project! I do not have much experience in GMMs in particular, though I do have some experience in Tensorflow and Keras and will be working with Pytorch in my current research this semester. Are there any specific resources or issues I can look at to get familiar with PVnet and GMMs?

Thanks,
Pragya

[chaitanya-rdY]

Hi @peterdudfield,@felix-e-h-p sir

I’m chaitanya from iit jammu ,India a machine learning and deep learning practitioner with experience in probabilistic modeling and optimization. I’ve worked on YOLOv8 pipelines for document classification, focusing on uncertainty estimation, as well as applying evolutionary learning techniques like NEAT. I’m particularly interested in how different modeling approaches handle uncertainty in real-world scenarios.

This project looks really exciting! Moving from quantile regression to a Gaussian Mixture Model approach seems like a strong way to better capture multimodal uncertainties, especially during transitions. Have you considered potential computational trade-offs with multiple GMM components, and would adaptive mixture weights be an option to dynamically adjust component importance based on context?

[victorgarcia98]

Hi @peterdudfield and @felix-e-h-p , hope you're well!

I love the idea and I would like to contribute.

A bit of myself: My name is Víctor and I'm currently studying a Master in Statistics and Operations Research at UPC (Barcelona, Spain) where I'm specializing both in statistical inference (Frequentist, Bayesian, Time Series) and optimization (Large Scale Problems, Stochastic programming). Previously, I've got some experience working with DL in satellite orbit determination. Moreover, I'm currently an Software/Data engineer working at Seita (maintainers of FlexMeasures - a Linux Energy Foundation project). I have quite a lot of experience developing forecasting/optimization systems, and handling data in general.

I'd be very interested in applying my current knowledge and expanding it to improve probabilistic forecast. In addition to GMM, we could try to use Cauchy Mixture Models to better model fat tails (which are common in renewables) and Bayesian Deep Learning using dropout layers.

Nonetheless, for a base demo, we can try something in the lines of [1], using a Negative Log-Likelihood cost function.

Please, let me know what do think :D

Cheers,
Víctor

---

[1] https://ieeexplore.ieee.org/document/374138

[victorgarcia98]

Hey @peterdudfield @felix-e-h-p,

Finally, I'm not applying to the project, but I worked on a PoC that might be useful for others.

It's a simplified and stylized version of the problem you are posing: learning the parameters of a normal distribution which vary with the input. An interesting follow-up would be to try with a gaussian mixture which requires adding more parameters and changing a bit the synthetic data generation.

Code

import torch
import torch.nn as nn
import torch.optim as optim
import numpy as np
import matplotlib.pyplot as plt
from torch.utils.data import DataLoader, TensorDataset

# Define the MLP model with one hidden layer
class MLP_Gaussian_NLL(nn.Module):
    def __init__(self, input_size=1, hidden_size=64):
        super(MLP_Gaussian_NLL, self).__init__()
        self.fc1 = nn.Linear(input_size, hidden_size)  # input -> hidden
        self.fc2 = nn.Linear(hidden_size, 2)           # hidden -> output (mean, variance)
    
    def forward(self, x):
        x = torch.relu(self.fc1(x))  # Apply ReLU activation
        output = self.fc2(x)         # Output mean and log variance
        mean = output[:, 0]          # First column is the mean
        log_var = output[:, 1]       # Second column is log(variance)
        return mean, log_var

# Negative Log Likelihood loss for Gaussian distribution
def gaussian_nll_loss(mean, log_var, target):
    var = torch.exp(log_var)
    nll = 0.5 * torch.log(var) + 0.5 * ((target - mean) ** 2) / var
    return nll.mean()  # Mean over batch

# Generate synthetic data (e.g., t = time, y = target variable)
t = torch.linspace(0, 10, 300).view(-1, 1)  # Time variable (input)

y_true = t + 0.5 * t * torch.randn_like(t)  # y = t + 0.5 * t* noise (target variable)

# Create DataLoader for batching
batch_size = 32
dataset = TensorDataset(t, y_true)
dataloader = DataLoader(dataset, batch_size=batch_size, shuffle=True)

# Create the model, loss function, and optimizer
model = MLP_Gaussian_NLL(input_size=1, hidden_size=3)
optimizer = optim.Adam(model.parameters(), lr=0.001)

# Training loop
num_epochs = 500
for epoch in range(num_epochs):
    model.train()
    total_loss = 0.0
    for batch_t, batch_y in dataloader:
        # Zero the gradients
        optimizer.zero_grad()

        # Forward pass
        mean, log_var = model(batch_t)

        # Compute loss
        loss = gaussian_nll_loss(mean, log_var, batch_y.view(-1))

        # Backward pass
        loss.backward()
        optimizer.step()

        total_loss += loss.item()

    # Print the loss every 50 epochs
    if (epoch + 1) % 50 == 0:
        avg_loss = total_loss / len(dataloader)
        print(f"Epoch [{epoch+1}/{num_epochs}], Loss: {avg_loss:.4f}")

# Test the model
model.eval()
with torch.no_grad():
    predicted_mean, predicted_log_var = model(t)
    predicted_var = torch.exp(predicted_log_var)

# Plot the results
plt.figure(figsize=(10, 6))
plt.plot(t.numpy(), y_true.numpy(), label='True Data', linestyle='--')
plt.plot(t.numpy(), predicted_mean.numpy(), label='Predicted Mean', color='r')
plt.fill_between(t.numpy().flatten(), 
                 (predicted_mean - 2 * torch.sqrt(predicted_var)).numpy().flatten(), 
                 (predicted_mean + 2 * torch.sqrt(predicted_var)).numpy().flatten(), 
                 color='r', alpha=0.2, label='Predicted 95% Confidence Interval')
plt.xlabel('Time (t)')
plt.ylabel('Target Variable (y)')
plt.legend()
plt.savefig("target.pdf")

[felix-e-h-p]

Thanks for this @victorgarcia98 - hope you are well.

[Taha2053]

Hello @peterdudfield , @felix-e-h-p and the Open Climate Fix team,

I am Almouthana Taha Khalfallah an Applied Mathematics and modelization Engineering student specializing in data science and AI, with a strong passion for coding and mathematical modeling. I am particularly excited about the opportunity to contribute to the "Improving Probabilistic Solar Forecasts" project as it aligns perfectly with my academic background. I have a little experience in machine learning, deep learning, and statistical modeling, but I have a strong foundation in the mathematical concepts required for this project, including Gaussian Mixture Models and probability theory. I am eager to apply my skills to improve solar forecasting models and help contribute to more accurate and reliable renewable energy predictions. I am excited to collaborate with your team and learn from this experience to advance this field.

Looking forward to hearing from you.

Best regards,
Almouthana Taha Khalfallah
My portfolio : https://taha2053.github.io/ATK-LOG.github.io/

[Dakshbir]

Hi @felix-e-h-p, @felix-e-h-p and OpenClimateFix Team,

I’m Dakshbir Singh, a Computer Science student from NSUT, India passionate about Machine Learning, Deep Learning, NLP, and Generative AI. I have a strong background in mathematics and statistics, along with expertise in Python, C++ (DSA), and ML/DL frameworks like PyTorch. Given my interest in probabilistic modeling and AI-driven forecasting, I’m excited about contributing to the "Improving Probabilistic Solar Forecasts" project for GSoC 2025.

Understanding of the Project
From my analysis, the project aims to shift from quantile regression to a Gaussian Mixture Model (GMM)-based approach for capturing uncertainties in solar forecasting. This modification should lead to better multimodal uncertainty modeling, stronger uncertainty bounds, and improved pattern comprehension.

I have a solid grasp of statistical modeling and probability distributions, but I am relatively new to Gaussian Mixture Models (GMMs). I'm eager to gain hands-on experience in implementing GMMs from scratch and applying them in real-world forecasting scenarios.

Initial Steps for Contribution
(I have already started contributing to https://github.com/openclimatefix/PVNet- 🔗 PR Link: openclimatefix/pvnet#341
Beyond exploring https://github.com/openclimatefix/PVNet, I would love to understand the specific areas I should focus on to start contributing effectively. Some questions I have:

1. GMM Implementation: Should I begin by implementing a small-scale GMM in PyTorch using gmm-torch and compare it against scikit-learn's mixture model?
2. Dataset & Evaluation Metrics: Are there any recommended datasets or metrics used in the current forecasting models?
3. Key Code Areas: Which parts of the existing repository are most relevant for understanding the transition from quantile regression to a GMM-based approach?
4. Early Contributions: Are there any small bug fixes, documentation improvements, or experimental implementations that would help me get familiar with the codebase?

I’m eager to start contributing and appreciate any guidance you can provide on the best way to proceed! 🚀

Looking forward to your insights!
Thanks!
Dakshbir Singh

[felix-e-h-p]

Please familiarise yourself via base_model and multimodal - transitionally related. Feel free to test what you think would be the most effective and efficient also GMM wise. Potentially further reading with soft clustering for unsupervised learning also may be beneficial. Thanks for your message!

[Dakshbir]

hey @felix-e-h-p , thanks for the reply :)
I want to contribute specific to this project (GMM). How can i do that? I am currently contributing towards other repositories of Open Climate Fix.
I am currently studying https://github.com/ldeecke/gmm-torch and https://scikit-learn.org/stable/modules/mixture.html and will study the further readings which you mentioned...

[Dakshbir]

Hi @felix-e-h-p and @peterdudfield,

I’ve been exploring the PVNet 2.1 architecture and the use of Gaussian Mixture Models (GMMs) for probabilistic solar forecasting. A few observations and questions I had:

GMM Component Optimization:

Are you considering adaptive selection of GMM components based on forecast complexity (e.g., clear vs cloudy conditions)?

This could prevent overfitting and improve distribution fit across diverse weather patterns.

Temporal Consistency:

Would adding a recurrent layer (LSTM) before the GMM output help enhance temporal coherence?

This could reduce forecast instability during rapid weather transitions.

Numerical Stability:

When training the GMM, are you using regularized covariance matrices or other stability techniques?

This could prevent instability during sudden weather shifts.

Looking forward to seeing how this project evolves!
Best,
Dakshbir Singh

[nii0708]

Hey @felix-e-h-p and OpenClimateFix Team,

I'm Nadhif, Data scientist with background in physical oceanography. I have experience in training model with python (pytorch), deploying model to production, data analysis (including marine and weather data), and ocean modelling (especially forecasting ocean current with Fourier Neural Operator and Graph Based model (GraphSage)).

I'm excited about this project, which aims to create continuous distributions using Gaussian Mixture Models (GMMs). This approach promises a more complete and accurate representation of uncertainty compared to fixed quantiles, especially for transitions and bimodal outcomes. The focus on uncertainty prediction is particularly appealing, as it has significant real-world and production applications. I have a few questions:

1. Will this project introduce a new loss function, different from the one used in the paper (https://www.climatechange.ai/papers/iclr2024/46)? If so, will retraining the model and benchmarking it (with a one-to-one correspondence to Figure 1 of the paper) be included? Additionally, will the project incorporate new evaluation metrics?

2. Given that GMMs require initialization of means and variances, will the model be updated to include priors for these parameters to improve results?

I'm particularly interested in the challenge of determining appropriate initial means and variances, especially considering the constraints involved (such as matrix positive definiteness, etc).

Best regards,

Nadhif

[ghost]

Hi @felix-e-h-p and @peterdudfield,

I’m Dakshbir Singh, an enthusiastic contributor to Open Climate Fix, and I’m incredibly excited about the Improving Probabilistic Solar Forecasts using GMMs project. I’ve been actively contributing to OCF over the past few weeks, and I’m eager to make more meaningful contributions as I prepare my GSoC 2025 proposal.

I find this project particularly compelling because Gaussian Mixture Models (GMMs) offer a more nuanced way to capture multimodal uncertainties in solar forecasting compared to fixed quantiles. This approach could significantly enhance the accuracy of uncertainty estimation, which is critical for real-world applications in energy dispatch and grid balancing.

After reviewing the project description and previous discussions, I had a few technical questions and would love to get your insights:

1. Optimal Number of GMM Components: Have you determined the ideal number of Gaussian components for the mixture model? For instance, do you anticipate 2-3 components being sufficient to represent clear-sky vs cloud-obscured states, or are you considering a more adaptive approach to prevent overfitting?
2. Model Architecture and Temporal Consistency: Since PVNet 2.1 uses a late-stage multimodal fusion architecture, have you considered adding a recurrent layer (e.g., LSTM) before the GMM output layer? This could improve temporal coherence across forecast horizons, reducing erratic uncertainty estimates during rapid transitions.
3. Loss Function and Stability: When estimating GMM parameters, are you exploring techniques like negative log-likelihood with regularized covariance matrices? This could mitigate numerical instability, particularly during sudden weather changes.
4. Immediate Contribution Opportunities: Are there any high-priority issues or areas within this project where I can contribute right away? I’d love to start working on relevant tasks and gain a deeper understanding of the project’s architecture.

I am genuinely excited to collaborate and learn from your guidance. Thank you for your time and dedication—I look forward to contributing further and helping OCF drive impactful, sustainable innovations! 😊

Best regards,
Dakshbir Singh

[AksAad]

Hi @peterdudfield and @felix-e-h-p and the OpenClimate Fix team , hope you're doing well!

I love the Open Data PVNet project and would love to contribute.

A bit about myself: My name is Akshat Soni, and I'm currently pursuing a B.Tech in Computer Science and Engineering at SRM University. My focus is on AI, Data Science, and Cloud Computing, and I have strong experience in ML model training, data analysis, and cloud-based workflows. I’ve worked extensively with PyTorch, NumPy, Pandas, and Xarray, making me well-equipped to handle large-scale Numerical Weather Prediction (NWP) datasets and train robust solar forecasting models.

I’m particularly interested in optimizing forecast accuracy while ensuring the model remains fully reliant on open data sources. I’d love to explore techniques like hyperparameter tuning, model benchmarking, and possibly integrating advanced feature selection to enhance predictive performance.

A couple of questions regarding the project:

1. Since we’re using only public NWP data, are there any known limitations in resolution or update frequency that could impact forecast accuracy?
2. Would it be beneficial to experiment with ensemble models, combining multiple open-data-based forecasts to improve reliability?

Looking forward to your insights!

Cheers,
Akshat
A little about me can be found here: ( https://www.linkedin.com/in/akshat-soni-393a02326/ )
I am currently at an intermediate stage in these topics: (https://scikit-learn.org/stable/modules/linear_model.html#polynomial-regression-extending-linear-models-with-basis-functions)
and using this (https://github.com/ldeecke/gmm-torch) to improve my skills on these topics.

[prajakta1321]

Hello @peterdudfield ,

I'm Prajakta, an aspiring GSoC contributor with experience in machine learning, time series forecasting, and probabilistic modeling. I’ve worked with Python, data preprocessing, and model evaluation, and I’m particularly excited about this project. I had a few questions:

1. What are the biggest challenges in improving the accuracy of probabilistic solar forecasts?

2. Are there any specific probabilistic models or frameworks currently preferred for this project?

3. How do we plan to validate and benchmark improvements—what metrics matter most?

Looking forward to your insights. Thanks for your time!

Best,
Prajakta
https://www.linkedin.com/in/prajakta-gurav-/

[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 @peterdudfield @felix-e-h-p !

Really excited about the "Improving Probabilistic Solar Forecasts" project in GSOC 2025 ! The idea of switching to a GMM approach to capture those complex uncertainties in solar data sounds super interesting. I've got some experience implementing GMMs from scratch, and I'm pretty comfortable with ML/DL/PyTorch.

To get a better handle on things, I was hoping you could clarify a few points:

• Could you give me a bit more detail on the specific structure of the PVNet final layers we'll be modifying?
• Are there any particular datasets or data formats we'll be working with?
• Any specific metrics you're most interested in seeing improved with the GMM approach?
• Are there any parts of the PVNet codebase you'd recommend I get familiar with ASAP?

Thanks for your time, and I'm really looking forward to potentially contributing to this project!

[felix-e-h-p]

Hi there @val2021-svg - thanks for your question.

Please familiarise yourself with multimodal.py within the models/multimodal directory and the late stage fusion process - alongside related encoders etc. This would be a very important preliminary step.

Thanks again - Felix

[adityasuthar]

Hi @peterdudfield, I am Aditya (GitHub), a final-year Computer Science student from India. I have some experience in Python, Django, Django Rest Framework, and contributing to open-source projects and working with APIs, databases, and deployment.

I have one question where should I submit my proposal I am looking for a review before submitting on the GSoC program site, thanks :)

[nasirabd]

Hi @peterdudfield and @felix-e-h-p , I am a fairly recent masters graduate from The Ohio State University in Geospatial engineering with a focus in photogrammetric computer vision. I am quite fascinated by the idea of addressing climate change with the use of renewable energy! I currently work part time for a startup company which shares similarities in this initiative.

Although my main focus of research DL techniques for urban scene understanding in remote sensing images, I have experience with creating GMM models for vision tasks like object detection and photo editing processes like lazy snapping. I was also curious if the team had explored some type of data fusion techniques like Bayesian uncertainty fusion to improve current model accuracies?

I believe this would be a great learning experience for myself in such an essential field of study like climate change.

[felix-e-h-p]

Ni @nasirabd - thanks for your message. This approach would certainly be an important additional scope to consider, alongside other potential methodologies for further validation.

[bassel-deg]

Hi @peterdudfield, @felix-e-h-p, and OpenClimateFix Team,

I hope this message finds you well.

My name is Bassel, currently a 3rd-year Computer Systems Engineering student (BUE/LSBU dual degree). I'm deeply passionate about leveraging technology, especially open-source tools, to address climate change challenges, which strongly aligns with Open Climate Fix's mission.

I'm writing to express my strong interest in the GSoC project "Improving Probabilistic Solar Forecasts" (#30). Having been involved in climate policy discussions (e.g., COP28, COP29) and sustainability initiatives (UNGC, Oxford Net Zero, IRENA), I understand the critical need for accurate renewable energy forecasting. The project's focus on improving uncertainty quantification using Gaussian Mixture Models, moving beyond simple quantiles, strikes me as a technically fascinating and impactful approach.

My engineering background provides a solid foundation in programming (Python) and analytical thinking, complemented by foundational ML knowledge. While implementing GMMs from scratch will be a key learning area, I'm very eager to dive into this challenge within the PyTorch framework.

I'm very enthusiastic about the possibility of contributing to this project and learning from the team. Thanks for considering my interest!

Best regards,
Bassel Deghaidy

[ferro-gianluca-29]

Hello @peterdudfield and @felix-e-h-p,

I’m Gianluca Ferro, a Master’s graduate in Electronic Engineering and currently a research fellow in Artificial Intelligence at ciparlabs , applied to smart grids and renewable energy communities. My Master’s thesis focused on time series forecasting for both energy consumption and renewable generation, where I explored gradient boosting methods, statistical models, and neural networks. I find your project on transitioning from quantile regression to Gaussian Mixture Models especially compelling, and I’m eager to experiment with these data-driven approaches and refine them for more accurate, multimodal forecasting.

I do have a couple of questions:

1. Have you considered integrating physics-based constraints or domain-specific rules within the GMM framework? (capturing, for example, the transitions between clear-sky and cloudy conditions)

2. How are you planning to address the potential computational overhead introduced by multiple mixture components, particularly for real-time forecasting or large-scale deployment scenarios?

I look forward to your thoughts on these points and to potentially contributing to this initiative.

Thank you,
Gianluca

[felix-e-h-p]

Hi @ferro-gianluca-29 - thanks for your questions.

Regarding physics-based constraints within the GMM framework - potentially say whereby a singular component could model clear sky conditions whilst others capture different cloud coverage scenarios. Definitely.

Regarding computational efficiency, it would be recommendable (largely speaking) to use a limited number of mixture components with param sharing.

Thanks again

[ClaudioGlez21]

Hi everyone,

My name is Claudio Gonzalez. I’m currently a third-year Data Science student at Tec de Monterrey in Guadalajara, Mexico, with a strong background in machine learning and deep learning, particularly using PyTorch. I’ve worked on various projects involving statistical modeling and have hands-on experience implementing advanced algorithms such as Gaussian Mixture Models. I'm excited about the potential of applying these techniques to improve probabilistic solar forecasting and contribute to this innovative project.

I have a couple of questions:

1. Integration Challenges: Could you elaborate on the main challenges you anticipate when integrating the GMM-based output layer into the existing model architecture (e.g., PVNet)? Are there specific pitfalls to look out for when transitioning from quantile regression outputs to a continuous probability distribution?
2. Evaluation Metrics: What evaluation metrics or methodologies are currently used to assess the performance of the probabilistic forecasts? Additionally, will there be any exploratory work or benchmarks provided for comparing the current approach to the proposed GMM-based approach?

Thank you in advance for your insights. I’m really looking forward to learning from your experiences and contributing to this project!

Best regards,
Claudio

[felix-e-h-p]

Hi @ClaudioGlez21 - thanks for your questions.

Regarding integration challenges - The main challenges likely will include managing increased parameter space and ensuring training stability with unbounded variance estimates.

Regarding evaluation metrics - For direct GMM comparison, sample from the mixture distribution for metric compatibility. Benchmarks can hence compare approaches, i.e. focusing on component configurations against observed error distributions.

[prajakta1321]

Hello @felix-e-h-p ,

I'm Prajakta, an aspiring GSoC contributor with experience in machine learning, time series forecasting, and probabilistic modeling. I’ve worked with Python, data preprocessing, and model evaluation, and I’m particularly excited about this project. I had a few questions:

What are the biggest challenges in improving the accuracy of probabilistic solar forecasts?

Are there any specific probabilistic models or frameworks currently preferred for this project?

How do we plan to validate and benchmark improvements—what metrics matter most?

Looking forward to your insights. Thanks for your time!

Best,
Prajakta
https://www.linkedin.com/in/prajakta-gurav-/

[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!

[peterdudfield]

I'm closing this discussing now as GSOC 2025 is nearly over. Thank you for everyones input and help.

We hope to take part next year and we'll be posting info here