METHODOLOGY DOSSIER
Cascade-GNN
Cascading failures in power networks: a graph-classification pipeline and validation methodology.
In this document
01 · Problem#
How can a power network and its failures be represented for graph classification? The prototype explores a research method without claiming operational forecasting.
02 · PowerGraph data#
PowerGraph is the dataset used in the pipeline described by my CV and repository. The exact version, subset, and scenario identifiers must accompany any published results.
03 · Graph representation#
The network is represented by nodes and connections. Input variables, labels, and their observation time must be specified in an experiment configuration before predictions are interpreted.
04 · Pipeline#
- PowerGraph
- Import & preparation
- Splits
- GNN / XGBoost
- Metrics & attribution
The existing scope includes data imports, GNN models, an XGBoost baseline, metrics, line-level attribution, and tests.
05 · Train, validation, test#
Consolidation protocol: identify scenarios before splitting; separate related groups; reserve the test set for final evaluation. Fitting transformations and selecting models use only the permitted partitions.
- Identified scenarios
- Group-aware split
- Train / validation
- Held-out test
This protocol describes validation requirements. Split files, configuration, and the commit for the next result must make them verifiable.
06 · Leakage prevention#
Related scenarios in multiple splits, normalization fitted on the entire dataset, or variables observed after the event can make an evaluation misleading.
Read the methodology note07 · XGBoost baseline#
Compare the GNN with XGBoost on the same splits and target. The baseline is documented; no numerical superiority is published here.
08 · GNN architecture#
The prototype uses PyTorch and PyTorch Geometric. The repository is the reference for implemented layers and parameters; this page does not assign an unverified architecture or version.
09 · Metrics#
For the next evaluation: publish class counts, a confusion matrix, precision, recall, and F1, including the averaging convention. Final choices depend on the target and the cost of errors.
10 · Explainability#
Line-level attribution is part of the prototype scope. Attribution describes model sensitivity; it does not prove physical causality or a future failure.
11 · Reproducibility & tests#
The pipeline and tests are described in the repository. For each publishable result, archive the command, environment, seeds, configuration, splits, and commit. This page does not attest to the latest CI outcome.
12 · Results#
A results table will be added when dataset/version, splits, baseline, metrics, seeds, mean/spread, and commit can be supplied together.
13 · Limitations#
A result on simulation data does not establish generalization to an operating grid. Scenario distribution, target definition, and potential leakage remain central to the analysis.