Source-linked AI summary
Towards Hierarchical GNNs for multi-grid power flow: generalization across operating scenarios
Carmine Delle Femine, Leire Garin Atxaga, Asier Diaz-Iglesias, Juan Pablo Maroto Herrera, Ane Miren Florez-Tapia, Marco Quartulli. Izaro Goienetxea Urziku
TL;DR
Power-flow models must connect local electrical constraints to voltage fields that depend on the network as a whole. This paper adds two-reduction latent communication to a corrective GNN and compares Kron-derived and same-anchor Quotient transports with a flat backbone. In preliminary training, hierarchical models generalize to new operating scenarios on the training topologies, but do not yet demonstrate useful cross-topology generalization by the stated criterion.
Problem
Power-flow models must connect local electrical constraints with voltage fields that depend on the network as a whole, including when serving several topologies and operating conditions.
Source: §1 · p. 1Method
A two-reduction latent communication module within the GENCO corrective solver exchanges bus features through compact graphs using Kron-derived or same-anchor Quotient transports.
Source: §1 · p. 1Results
Kron reduces macro family error to 0.851, compared with 5.660 for Flat and 1.235 for Quotient; both hierarchical models beat the fitted bus-mean reference on every training grid in all three seeds.
Source: §1 · pp. 1–2, §7 · p. 10Takeaways & Limitations
Preliminary multi-grid trainings support hierarchical communication as a component for learning across operating conditions with shared model parameters on the studied topologies.
Source: §7 · p. 10, §6 · p. 10On two additional topologies, all three models remain above their fitted references in every seed, so these preliminary trainings do not yet demonstrate useful cross-topology generalization by the stated criterion.
Source: §5.2 · p. 9
Abstract
from arXiv · showhide
Hierarchical latent communication improves the generalization of a multi-grid power-flow model to new operating scenarios. The module exchanges information through two reduced graphs within a GENCO-based corrective network. We compare Kron-derived transports, a same-anchor Quotient construction and a flat backbone in preliminary trainings of 200 epochs on three grid topologies, with three initialization seeds per model. Evaluation uses 200 newly generated, preselected scenarios per grid. On the training topologies, Kron reduces the macro family-balanced voltage error from 5.660 +- 0.899 to 0.851 +- 0.110: an 85.0% reduction relative to Flat GENCO and 31.0% relative to Quotient, which reaches 1.235 +- 0.225. Both hierarchical models outperform a per-bus mean fitted on training solutions on every training topology in all three seeds. These results demonstrate generalization across operating scenarios within the studied topologies, with one set of learned parameters shared across grids. Evaluation on two additional topologies distinguishes this achievement from cross-topology generalization: the current models do not yet outperform the fitted reference in that calibrated- transfer setting. This preprint presents the architecture and preliminary evidence for hierarchical communication as a component of multi-grid power-flow learning, with generalization to unseen topologies as the next development objective.
1 Introduction
The introduction presents hierarchical latent communication as a way to exchange information across reduced graphs within a corrective power-flow solver. Preliminary results show scenario generalization on the training topologies, while unseen-topology generalization remains a separate objective.
1 Introduction
- Hierarchical exchange complements repeated local updates by compressing bus representations, exchanging information on a smaller graph, and returning context to the original buses.The introduction motivates this route for models serving multiple topologies and operating conditions.Source: §1 · p. 1
- The two-reduction module replaces two local corrective layers and gives bus features a compact-graph route for global attention while preserving local correction and decoding.The two transports are Kron-derived weights and a graph quotient with the same retained buses.Source: §1 · p. 1
- Kron reduces fresh-scenario error by 85.0% versus Flat GENCO and 31.0% versus Quotient; both hierarchical models beat the fitted bus-mean reference on every training grid in all three seeds.The evaluations use preliminary 200-epoch trainings on three grids.Source: §1 · pp. 1–2
- Scenario generalization evaluates new operating conditions on a fixed topology, whereas topology generalization evaluates a grid absent from optimization; the current trainings demonstrate the former.Evaluations on two additional grids identify topology generalization as a next objective.Source: §1 · p. 2
- The preprint reports a fresh-scenario comparison across three training topologies and three seeds, plus separate additional-topology assessment using preliminary fixed 200-epoch endpoints.Architecture and training design remain under development.Source: §1 · p. 2
2 Related work
Related work spans graph solvers for power flow, electrical and cell-based graph reductions, multiscale physical models, and shared grid representations. This paper positions its contribution as latent hierarchical exchange in a corrective backbone, evaluated separately for scenario and topology generalization.
2 Related work
- Graph power-flow solvers use Kirchhoff-law residuals, message passing, higher-order convolutions, and physics-based training across grid topologies and supply tasks.These studies motivate architectures combining network structure with physical information.Source: §2 · p. 2
- Kron reduction and graph-signal methods use retained nodes to represent interactions mediated by eliminated paths.This paper uses that construction for latent transports and coarse edge features, with learned feature updates.Source: §2 · p. 2
- Cell-based pooling ranges from multilevel partitioning and aggregation guarantees to learned soft assignments [9]; Quotient uses fixed one-hot cells sharing Kron’s anchors.The shared anchors make the cell-based and electrical transports comparable within one module.Source: §2 · p. 2
- Multiscale simulation and domain-decomposition GNNs motivate reduced-graph exchange, while grid-foundation models emphasize shared representations across conditions, tasks, and networks.This study applies common learned transforms through grid-specific geometry and assesses scenario and topology generalization separately.Source: §2 · p. 3
3 Kron and Quotient hierarchical communication
The GENCO backbone replaces two local corrective layers with a two-reduction latent module that routes bus features through compact graphs. Kron and Quotient share retained anchors but use different interior-to-anchor relations.
3.1 Backbone, notation and the poster schematic
- The module replaces two local corrective layers while retaining the backbone’s local correction and decoding machinery.The input graph contains buses, generators, and electrical connections; the final decoder produces the power-flow state.Source: §3.1 · p. 3
3.2 Kron reduction: eliminate interior variables
- Kron reduction uses a Schur complement to couple retained buses through eliminated paths and extend retained voltages into the interior when interior injections are zero.The construction uses sparse linear solves rather than a dense inverse of the fine-grid matrix.Source: §3.2 · p. 4
3.3 From electrical extension to latent transport
- Kron-derived latent transport uses sparsified extension weights to prolong coarse features to interior nodes and restrict interior features to anchors with a different normalization.The first-level extension retains at most sixteen coefficients per interior bus; the upper extension is untruncated.Source: §3.3 · p. 4
3.4 Quotient reduction: aggregate cells
- Quotient reduction partitions nodes into cells and combines interactions between cells in the coarse graph; interior nodes receive their cell’s anchor feature, while anchors receive the mean of their cell’s interior features.The module’s transports use interior rows, and an anchor with no interior members receives zero as input to the learned restriction map.Source: §3.4 · pp. 4–5
3.5 Why elimination and aggregation differ
- For fixed anchor values, harmonic extension minimizes dissipation, whereas cell-constant extension imposes an additional constraint; the resulting reductions differ before neural prediction error is measured.In the worked example with conductances a = 3 and b = 1, Kron uses weights (0.75, 0.25) and coarse conductance 0.75, while Quotient uses (1, 0) and coarse conductance 1.Source: §3.5 · p. 5
3.6 Two reductions and the learned return path
- Two reductions compress bus features, process them on coarser graphs, and return updated context to the original buses using shared learned maps across levels.The hierarchy uses two middle-level message-passing blocks and top-level dense attention; the 4,917-bus example has 984 and 197 nodes at its reduced levels.Source: §3.6 · p. 6
4 Experimental setup
The preliminary study trains three models jointly on three grids and evaluates them on fresh scenarios, using shared training policies and separate calibration for additional topologies. It compares learned voltages with fitted references under family-balanced error and prespecified descriptive criteria.
4.1 Preliminary multi-grid trainings
- Three models train jointly on 500-, 2,000-, and 4,917-bus grids for 200 epochs with three seeds; two additional grids are reserved for transfer evaluation.Training uses 1,856 examples per grid per epoch, and the 1,000 fresh test scenarios are shared by all final checkpoints.Source: §4.1 · pp. 6–7
4.2 Objective and normalization
- The objective combines masked bus mean-squared error with intermediate physics residuals, while additional topologies use separate calibration splits for scalar normalization.Validation and test use the full-strength objective; calibration scenarios support preprocessing and reference fitting, not optimization.Source: §4.2 · p. 7
4.3 Metrics, references and aggregation
- Evaluation averages voltage-magnitude and wrapped-angle RMSE into family-balanced errors, weighting grids equally and reporting means and sample standard deviations across seeds.References include a nominal predictor and a per-bus mean fitted on training or calibration solutions; prespecified criteria assess Kron against Flat, Quotient, and the fitted reference.Source: §4.3 · p. 7
5 Results
On fresh scenarios from the three training topologies, Kron outperforms Flat GENCO and Quotient, and both hierarchical models beat the fitted bus-mean reference. On two additional topologies, the models remain above their fitted references; validation results describe the preliminary 200-epoch training regime.
5.1 Generalization across operating scenarios
- 0.851 ± 0.110 for Kron versus 5.660 ± 0.899 for Flat GENCO and 1.235 ± 0.225 for Quotient in macro family error; reductions are 85.0% and 31.0%, respectively.These are fresh-scenario results after 200 epochs, averaged across three seeds; lower is better.Source: §5.1 · p. 7, Table 2 · p. 8
- Both hierarchical models outperform the fitted bus-mean reference on every training topology in all three seeds.Fresh and development-panel results are close, and Figure 4 displays errors relative to each grid’s fitted reference.Source: §5.1 · p. 8, Figure 4 · p. 8
5.2 Topology transfer as the next development stage
- All three models remain above the fitted references on both additional grids in every seed, so these trainings do not demonstrate useful cross-topology generalization by the stated criterion.Kron and Quotient also have higher physical residuals than Flat on both additional grids, leaving voltage accuracy and power-balance consistency as development targets.Source: §5.2 · p. 9
5.3 Learning in the preliminary training regime
- The hierarchical models attain lower final validation objectives than Flat under the common training policy, but the trajectories document only the current 200-epoch regime.The first twenty training epochs ramp the physics term while validation uses the full-strength objective throughout.Source: §5.3 · p. 9, Figure 6 · p. 10
6 Discussion and next steps
The discussion presents scenario generalization as a preliminary result and identifies generalization to unseen topologies as the next objective.
6 Discussion and next steps
- One shared hierarchical model generalizes to fresh operating conditions across the three training topologies, and both transport variants beat the bus-mean reference in every seed.Kron performs better than both Flat and Quotient in this setting.Source: §6 · p. 10
- Unseen-topology transfer changes the network geometry and remains a next-stage objective, including transfer of physical consistency.The additional-grid evaluation uses solved calibration data for normalization and reference fitting, so it concerns calibrated transfer.Source: §6 · p. 10
- The 200-epoch endpoints are provisional; broader topology coverage, depth-matched comparisons, and transport-operator interventions are proposed next steps.The three seeds describe initialization variation on the studied grids.Source: §6 · p. 10
- Recorded run times and peak memory describe the current workflow; a dedicated study is needed to assess resources at a specified accuracy, including geometry preprocessing.Times span 5.17–7.13 hours for Flat, 5.96–8.79 for Kron, and 5.94–8.83 for Quotient.Source: §6 · p. 10
7 Conclusion
Preliminary multi-grid training supports generalization to fresh operating scenarios on the training topologies. Kron has the lowest reported macro family error, while unseen-topology generalization remains the next objective.
7 Conclusion
- Kron’s macro family error is 0.851, versus 5.660 for Flat and 1.235 for Quotient, in preliminary fresh-scenario evaluations on three training topologies.Both hierarchical models improve on the fitted bus-mean reference on every training grid in all three seeds.Source: §7 · p. 10
- Both hierarchical models improve on a fitted bus-mean reference on every training grid in all three seeds, with shared parameters across operating conditions.The architecture is presented as a component for learning across operating conditions.Source: §7 · p. 10
- Generalization to unseen topologies remains the next objective, with a separate evaluation of voltage accuracy and physical consistency.The reported scenario results concern the three training topologies.Source: §7 · p. 10
A Additional numerical results
The additional numerical results comprise fresh-scenario error, transfer-channel, and earlier-training context tables, each with a distinct evaluation scope.
A Additional numerical results
- Table 4 reports final-checkpoint fresh-scenario macro family errors by seed across three training and two calibrated-transfer topologies, with means and sample standard deviations over seed macros.Every run completes 200 epochs.Source: Table 4 · p. 12
- Table 5 reports fresh-scenario voltage-magnitude and angle RMSE means and physical residuals on a common 100-MVA base.Values are averaged over 200 scenarios and then three seeds.Source: Table 5 · p. 12
- Table 6 provides training context from earlier twenty-epoch runs on ten topologies, not a controlled comparison of training duration.Its topology set and scheduler differ from the main study; the bus-mean macro reference is 1.412.Source: Table 6 · p. 12