Fresh off the press in Data-Centric Engineering!
Flexible multi-fidelity framework for load estimation of wind farms through graph neural networks and transfer learning: work led by Gregory Duthé of our Chair of Structural Mechanics and Monitoring at ETH Zurich. The work is co-authored with RTDT Laboratories AG, Vrije Universiteit Brussel & OWI-Lab.
Our latest work, available in opeen access in the Data Centric Engineering journal, delivers a flexible multi-fidelity framework for load estimation of wind farms through graph neural networks and transfer learning.
This work is led by Gregory Duthé of our Chair of Structural Mechanics and Monitoring at ETH Zurich. The work is co-authored with Francisco de Nolasco Santos (VUB), Imad Abdallah (RTDT Laboratories AG), Wout Weijtjens & Christof Devriendt (VUB & OWI-Lab), and Eleni Chatzi.
🔗 Read the paper: external page https://doi.org/10.1017/dce.2024.35
🐍 Access the code: external page https://github.com/gduthe/windfarm-gnn
The work tackles the challenging problem of predicting wake-induced fatigue loads in wind farms—bridging the gap between costly high-fidelity simulators and less accurate engineering models.
💡 Our Approach:
Graph Neural Networks (GNNs): These ensure our model is layout-agnostic and extremely fast!
Multi-fidelity Learning: We pretrain a "generalist" model on large pools of low-fidelity simulations to capture basic physics.
Transfer Learning (LoRA): With minimal high-fidelity data (~100 high-res simulations or real data), we fine-tune the model to accurately capture complex wake physics.
Our pretrained GNN model, combined with LoRA transfer learning, offers a practical, scalable surrogate tool for high-fidelity wind farm load predictions to support design and monitoring tasks in real-world applications 🚀
Why it matters:
🔄 Wind farm layout optimization – Quickly estimate fatigue loads to test new candidate farm designs.
🎯 Wake steering control strategies – Improve fatigue-aware controls (e.g., WakeWISE RL environment developed by Bas van Berkel).
📡 Instrumentation placement – Optimize sensor layouts for data collection.
⚙️ Gradient-based optimization – Support load-dependent engineering designs.
This work, developed in collaboration with OWI-Lab, builds on our previous efforts:
- Which GNN to use? 👉 external page https://doi.org/10.1088/1742-6596/2647/11/112006
- What kind of graph connectivity? 👉 external page https://iopscience.iop.org/article/10.1088/1742-6596/2505/1/012014/meta
- Can GNNs provide uncertainty estimates? 👉 external page https://doi.org/10.58286/29744