In progress: hyperparameter tuning, network architecture options (e.g., opt out edge features, gnn layer options), and data augmentations.

Extensions:

GINenergygrids

Code related to our paper: "Graph Isomorphic Networks for Assessing Reliability of the Medium-Voltage Grid." by Cambier van Nooten, C., van de Poll, T., Füllhase, S., Heres, J., Heskes, T., & Shapovalova, Y.

Graph Isomorphic Networks for assessing the N-1 principle on energy grids. Case study on a medium-voltage grid of a Distribution System Operator (DSO) in the Netherlands (Alliander).

Disclaimer : This code demonstrates the main algorithm (GIN) in a barebone manner. Please contact us if there are any questions.

Datasets

Real grid data (obtained from Alliander, DSO in the Netherlands) together with augmented data.

GIN framework

GIN layer details (supplementary material)

See the figures below for complementary illustrations of the GIN equations mentioned in the paper.

^{_{Fig 1. Overview of a single GIN block, example includes as input the first layer of embeddings ($k=0$), and results in the second layer of embeddings ($k=1$).}}

^{_{Fig 2. Representation of an example graph $G$ (1-hop neighbourhood), with target node $v$, highlighted in red. All neighbouring nodes $u$ of $v$, $u \in \mathcal{N}(v)$, highlighted in blue. The edges ($e$) included in this neighbourhood are highlighted in purple.}}

^{_{Fig 3. Apply the first MLPs in the first layer ($k=0$), for the edges. Update the embeddings using the previous layer.}}

^{_{Fig 4. Apply the first MLPs in the first layer ($k=0$). Apply aggregation (and combining in the second layer ($k=1$), for both the nodes and edges. For simplification, only the 1-hop neighbourhood of node $v$ is considered. In the real method, we considered these steps for all nodes and edges.}}

Training

Create virtual environment

virtualenv ENV -p python3 source ENV/bin/activate

Install requirements

pip install -r requirements