Exploring Deep Learning and Computer Vision in Solar Forecasting: Benchmarking Data-Driven Models

The Institute of Solar Research is dedicated to advancing the application of solar energy on Earth. Access to reliable solar resource information is crucial for our work. Within the solar energy meteorology group, we gather meteorological data and identify parameters essential for the design, operation, and qualification of solar power plants. Predicting future solar irradiance is a key aspect, with cloud dynamics significantly influencing shortwave radiation and, consequently, impacting solar power plant efficiency. The development of advanced solar forecasting models is essential. These models anticipate perturbations in solar irradiance induced by cloud dynamics. They contribute to the seamless integration of solar energy and thereby address the challenges of the energy transition for climate change mitigation.

In recent years, the field has seen significant growth, particularly with the advent of deep learning in computer vision, leading to the publication of numerous data-driven models. However, comparing these models is challenging as they are often trained and tested on proprietary data not publicly shared. To evaluate the true forecast performance, a benchmark study on a common and high-quality database is necessary.

This master's thesis aims to conduct a benchmark of data-driven solar forecasting models based on sky images from so called all-sky imagers and time series data. Utilizing high-quality data from the Plataforma Solar de Almería, the models will be trained on a standardized training set and validated on a versatile test set. In addition to exploring model architectures, the research will focus on data preprocessing and training methodologies of published results. The main tasks include:

• Conduct a comprehensive review of data-driven and machine learning-based methods for solar forecasting with all-sky imagers.
• Select and implement the most promising approaches from the literature, reproducing published results if data is available.
• Prepare training, validation, and test datasets for benchmarking the selected models.
• Validate the selected models, including our own model currently under development at DLR.
• Analyze benchmark results using standard forecasting metrics (e.g., RMSE) and investigate ramp event detection capabilities.
• Visualize results using explainable AI techniques to enhance understanding of model behavior.
• Compile findings and insights into a well-structured master's thesis.

• Location: Almería, Spain
• Minimum Period: 6 months (extendable)