Transfer Learning and Domain Adaptation for PV Prediction Across Different Geographical Locations

Co-Supervised by: Fereshteh Jafari

If you are interested in this topic or have further questions, do not hesitate to contact fereshteh.jafari@unibe.ch.

Background / Context

Photovoltaic power prediction models trained in one geographical location may exhibit poor performance when deployed in different regions due to variations in climate conditions, solar irradiance patterns, seasonal characteristics, and local weather phenomena. Traditional approaches require collecting extensive local data and retraining models from scratch for each new installation site, which is time-consuming and expensive. Transfer learning and domain adaptation techniques offer promising solutions by enabling models to leverage knowledge gained from data-rich source domains to improve performance in target domains with limited data. This capability is crucial for scaling PV prediction systems globally and reducing the time and cost associated with deploying new installations.

Research Question(s) / Goals

The research aims to develop robust transfer learning frameworks for PV prediction that can:

  • Effectively transfer knowledge between different geographical locations with minimal target domain data
  • Handle various types of domain shift (climate zones, seasonal patterns, equipment differences)
  • Quantify and manage uncertainty when making predictions in new domains
  • Adapt to new locations with limited labeled data using few-shot learning approaches
  • Identify which meteorological features and patterns are most transferable across regions

Approach / Methods

The student will:

  • Collect and analyze PV generation and meteorological datasets from multiple geographical locations (datasets will be made available)
  • Implement and compare various transfer learning techniques
  • Develop domain adaptation methods specifically designed for time series forecasting
  • Create few-shot learning approaches for rapid adaptation to new installation sites
  • Investigate feature importance and transferability across different climate zones
  • Develop uncertainty quantification methods for out-of-domain predictions
  • Conduct comprehensive evaluation across diverse geographical locations and climate conditions

Expected Contributions / Outcomes

  • Novel transfer learning framework specifically designed for PV power prediction
  • Comprehensive analysis of domain shift effects in different geographical regions
  • Few-shot learning methods for rapid deployment in new locations
  • Publications in renewable energy and machine learning venues

Required Skills / Prerequisites

  • Machine learning knowledge with background in transfer learning and domain adaptation
  • Proficiency in deep learning frameworks (PyTorch, TensorFlow)
  • Background in time series analysis and forecasting methods
  • Interest for solar energy systems and meteorological data

Possible Extensions

  • Federated learning approaches for privacy-preserving knowledge transfer
  • Online domain adaptation for continuously changing environments
  • Multi-source domain adaptation combining knowledge from multiple regions
  • Real-time adaptation using streaming meteorological data

Further Reading / Starting Literature

  • Pan, S. J., & Yang, Q. (2009). “A survey on transfer learning.” IEEE Transactions on Knowledge and Data Engineering, 22(10), 1345-1359.
  • Ganin, Y., et al. (2016). “Domain-adversarial training of neural networks.” Journal of Machine Learning Research, 17(1), 2096-2030.
  • Wang, M., & Deng, W. (2018). “Deep visual domain adaptation: A survey.” Neurocomputing, 312, 135-153.