Learning Geographical Constraints from Graph Electricity Data
Embedded in: grid-expansion planning at Alliander, building on GNN grid-reliability work (Cambier van Nooten et al., 2025). Supervisors: Yuliya Shapovalova, Tom van de Poll. Duration: 6 months.
1. Context and motivation
Alliander - one of the Netherlands' primary distribution system operators - manages over 90,000 km of electricity grid and faces a growing backlog of grid-connection requests, with waiting times of up to ten years. Meeting the demands of the energy transition requires rapidly expanding the physical grid. Expansion planning today relies heavily on manual expert knowledge to navigate geographic constraints: cables must follow roads, avoid protected nature areas (Natura2000), and cannot cross open water without cause.
This project builds directly on prior work in this group (Cambier van Nooten et al., 2025), which demonstrated that GNNs can reliably assess grid reliability on real Alliander data. The present thesis extends that foundation from reliability assessment to geographically-constrained expansion planning.
2. Objective and research questions
Central question: can a graph neural network learn, from existing grid and geographic data, which geographic features constrain cable routing - and use this to propose feasible new connections? The research questions below can be addressed based on the interest of the student.
-RQ1 (Encoding). How should heterogeneous geographic context - roads, Natura2000, waterways - be encoded as features in a spatial graph representation of the electricity grid? Comparing rasterised node features, auxiliary constraint graph layers, and coordinate-aware positional encodings.
-RQ2 (Learning constraints). Can a GNN trained on existing cable routes implicitly learn which geographic features act as routing constraints, and does this generalise to unseen regions? Existing cable placements encode decades of engineering decisions about geographic feasibility; this tests whether a model can recover these rules from data alone.
-RQ3 (Route proposal). Can the learned constraint representation propose feasible routes for new cable connections, and how do GNN-proposed routes compare to rule-based baselines? Framed as spatial link prediction; evaluated against shortest-path baselines and Alliander domain experts.
3. Background: state of the art
GNNs are well-suited to this problem: they operate natively on graph-structured data, learning representations that capture both grid topology and spatial relationships between assets. This thesis can build on the GIN-inspired GNN developed by Cambier van Nooten et al. (2025) for Alliander's medium-voltage grid, extending it to incorporate geographic context. Geographic constraint layers - roads, Natura2000 areas, and waterways - can be encoded as graph features using coordinate-aware positional encodings (Klemmer et al., 2023) and heterogeneous edge attributes (De Jonge et al., 2025). Grid expansion is framed as a spatial link-prediction task, drawing on the SEAL framework (Zhang & Chen, 2018) and adapted for flow-driven networks following Ludke et al. (2024). Related spatial-GNN work includes graph convolution for water-distribution systems (Ashraf et al., 2023).
4. Proposed approach
Data. Grid topology: nodes with (x,y) coordinates, connected by lines made of cable segments with connectors (also with (x,y) coordinates). Geographic context: flexible layers such as Dutch roads, Natura2000 areas, and rivers - all in the same coordinate system. Real Alliander grid data is accessed through the group's existing collaboration.
Method. Extend the GIN-inspired reliability GNN to ingest geographic context, encoding constraint layers via the encodings compared in RQ1. Frame grid expansion as spatial link prediction (SEAL-style) adapted for flow-driven networks. Train on existing cable routes so the model can recover implicit routing rules (RQ2), then use the learned representation to propose new routes (RQ3), evaluated against shortest-path baselines and Alliander domain experts.
5. Deliverables and student profile
Deliverables: 1) a geographic-context-aware GNN for the Alliander grid, with a comparison of encoding strategies (RQ1); 2) empirical evidence on whether routing constraints are learnable and generalise to unseen regions (RQ2); 3) a spatial link-prediction route-proposal method benchmarked against baselines and expert judgement (RQ3); (4) a thesis and code release.
Profile: MSc in CS / Data Science / AI; solid Python + PyTorch (PyTorch Geometric a plus); familiarity with graph neural networks; interest in geospatial / graph-structured data and applied ML. No prior energy-domain knowledge required.
Contact
Interested? Contact Dr. Yuliya Shapovalova — yuliya.shapovalova@ru.nl
Key references
[1] Cambier van Nooten, C., et al. "Graph neural networks for assessing the reliability of the medium-voltage grid." Applied Energy 384 (2025): 125401.
[2] De Jonge, M., et al. (2025). Generalizable Graph Neural Networks for Robust Power Grid Topology Control. arXiv:2501.07186.
[3] Klemmer, K., et al. (2023). Positional Encoder Graph Neural Networks for Geographic Data. arXiv:2111.10144.
[4] Zhang, M. & Chen, Y. (2018). Link Prediction Based on Graph Neural Networks (SEAL). NeurIPS 2018. arXiv:1802.09691.
[5] Ashraf, I., et al. (2023). Spatial Graph Convolution Neural Networks for Water Distribution Systems. AISTATS.