PaperlayerView exact revision on arXiv

Source-linked AI summary

Unified Heterogeneous Graph Neural Network solver for Power Flow, Optimal Power Flow and State Estimation

Ferran Bohigas-Daranas, Hamid Latif-Martínez, Eduardo Prieto-Araujo, Oriol Gomis-Bellmunt, Pere Barlet-Ros

arXiv:2609.16738v1eess.SYcs.LG

TL;DR

Power-system analysis requires repeated, computationally expensive task-specific solvers for PF, OPF, and SE. The paper introduces one heterogeneous ResGated GCN backbone trained jointly across the three tasks. It reports comparable accuracy to task-specific GNNs and robustness to unseen loading and topologies.

  • Problem

    Existing GNN solvers separately address PF, OPF, and SE, leaving the shared physics and topology without a unified model.

  • Method

    A heterogeneous ResGated GCN jointly trains on PF, OPF, and SE using shared representations, topological perturbations, and task-conditioned multi-task learning.

  • Results

    The unified model achieves accuracy comparable to task-specific GNN solvers and generalizes to unseen loading conditions and topological configurations without retraining.

  • Takeaways & Limitations

    One model can simultaneously support three power-system analysis tasks while eliminating redundant task-isolated training pipelines.

Abstract

from arXiv · show

Power Flow (PF), Optimal Power Flow (OPF), and State Estimation (SE) are fundamental problems in power system analysis, but solving them is computationally expensive. Graph Neural Networks (GNNs) have been proposed as fast surrogates, yet existing solvers are trained for a single problem at a time, producing narrow models that must be rebuilt for each new task. We propose a more general approach: a single Heterogeneous Residual Gated Graph Convolutional Network that solves all three problems with one shared backbone. Rather than learning one mapping, the model learns a reusable representation of how the network behaves, from which PF, OPF, and SE can each be estimated. Trained jointly on the three problems across diverse topologies and loading conditions, and evaluated on the IEEE 14-bus and 118-bus systems, the shared model matches the accuracy of task-specific GNN solvers and stays robust on unseen loading levels and topologies. These results show that a single model can capture the basic operation of a power network and serve several analysis tasks at once, a first step toward a foundation model for power systems.

I. INTRODUCTION

Power-system analysis is computationally demanding, while existing GNN solvers typically address PF, OPF, and SE separately. The paper proposes a unified heterogeneous GNN trained across all three tasks and diverse operating conditions.

  • Traditional NR, IP, and WLS solvers are mathematically robust but computationally burdensome as system size grows, while DER variability increases the need for near-real-time solutions.MLPs and CNNs also fail to naturally exploit irregular transmission-network topology, motivating graph-based models.
  • GNN message passing aligns with power-system topology and the propagation of information governed by physical laws such as Kirchhoff’s Current Law.Power systems are represented by buses and lines or transformers, with heterogeneous bus types and electrical variables.
  • Existing approaches treat PF, OPF, and SE as isolated problems, creating redundant training pipelines and preventing a unified grid-aware representation.The paper identifies the absence of a peer-reviewed architecture demonstrating all three tasks in one trained model on standard benchmarks.
  • A unified heterogeneous GNN solves PF, OPF, and SE simultaneously, targeting accuracy comparable to task-specific solvers without separate training pipelines.The architecture includes topological perturbations to learn a generalized representation of power-system physics.
  • The proposed training uses a composite loss for the three tasks with problem-conditioned masking to prevent cross-problem gradient interference.The paper also reports systematic architectural hyperparameter tuning and ablations to quantify component contributions.

A. Power Flow

Power Flow computes the steady-state electrical state for specified generation and load conditions. Although Newton–Raphson is robust, its repeated Jacobian factorization motivates a one-pass GNN approximation.

  • PF determines bus voltages, angles, and branch power flows by solving nonlinear equations simultaneously across the network.It supports planning and real-time operation by checking line loadings and bus voltages against security limits.
  • Newton–Raphson iteratively refines the state from an initial operating point, with Jacobian formation and sparse factorization creating a super-linear computational bottleneck.Normal operation typically requires five iterations.
  • A trained GNN approximates the PF mapping in a single forward pass, retaining topology awareness without repeated matrix factorizations or sensitivity to initial conditions.The stated motivation includes avoiding divergence that can occur for Newton–Raphson under stressed conditions.

B. Optimal Power Flow

Optimal Power Flow selects operating controls and states to optimize an objective while satisfying electrical and operational constraints. Its formulation combines nodal balance equalities with limits on equipment and operating variables.

  • OPF optimizes generation dispatch, voltage settings, transformer taps, and other controls while enforcing physical and operational constraints.Unlike PF, which evaluates a specified operating condition, OPF searches for an optimal one.
  • OPF has evolved from economic dispatch into multi-objective optimization incorporating environmental factors, market operations, and renewable-energy integration.
  • The OPF formulation minimizes a scalar objective over state and control variables subject to equality power-balance constraints and inequality operating bounds.Bounds include generator outputs, nodal voltages, and branch thermal capacities.
  • State estimation reconstructs complete voltage magnitudes and phase angles from limited, noisy measurements, addressing incomplete observability, measurement noise, and bad data.The resulting state supports downstream functions including EMS, OPF, security assessment, and contingency analysis.
  • The operational SE workflow includes measurement validation, observability analysis, state computation, and normalized-residual bad-data identification.Weighted least squares is described as the most widely used SE method.

III. GRAPH NEURAL NETWORK BACKGROUND

GNNs model power networks directly as non-Euclidean graphs, but early architectures have limited structural discrimination and depth scalability. ResGated GCN addresses these issues through edge-aware vector gating and residual connections.

  • GNN message passing updates each bus representation by aggregating neighboring information, matching the graph structure of buses and electrical connections.This provides a topology-aware alternative to Euclidean architectures for power-system problems.
  • Early GCN and GAT models cannot distinguish some non-isomorphic graphs, a limitation that matters when a single line trip changes network behavior.
  • A. ResGated GCN: ResGated GCN combines dynamic gating and residual connections to support deeper, efficient graph learning.The design extends the layer transformation with explicit edge attributes.
  • A. ResGated GCN: Vector edge gating independently modulates each hidden-feature dimension, unlike scalar attention mechanisms.
  • A. ResGated GCN: The gate uses endpoint features and line parameters R, X, and B to produce direction-dependent message weighting and control neighbor information flow.The gate is applied element-wise through a sigmoid coefficient, while residual shortcuts help stabilize deeper training.

B. Heterogeneous Graph Neural Networks for Power Systems

Power-system buses and grid edges have distinct operational semantics and physical parameters, making homogeneous GNNs a poor architectural fit. Heterogeneous GNNs address this by assigning type-specific representations and transformations.

  • B. Heterogeneous Graph Neural Networks for Power Systems: Homogeneous GNNs flatten bus semantics and force type-specific behavior to be learned from data, increasing learning difficulty.This can conflate physically distinct operating modes.
  • B. Heterogeneous Graph Neural Networks for Power Systems: HGNNs use separate learnable parameters and message-passing logic for different node and edge types.This lets the architecture respect bus boundary conditions during feature aggregation.
  • B. Heterogeneous Graph Neural Networks for Power Systems: Slack, PV, and PQ buses differ in known and unknown electrical variables, requiring distinct boundary-condition handling.Slack buses fix V and θ, PV buses fix P and V, and PQ buses fix P and Q.
  • B. Heterogeneous Graph Neural Networks for Power Systems: Power-grid edges carry physical quantities such as resistance, reactance, shunt admittance, transformation relations, and phase shifts.These parameters distinguish electrical networks from generic relational graphs.

IV. PROPOSED METHODOLOGY

The proposed unified architecture extends an Encoder-Processor-Decoder HGNN to solve PF, OPF, and SE with type-specific bus encoders and decoders around a shared physics-informed backbone.

  • IV. PROPOSED METHODOLOGY: The architecture solves PF, OPF, and SE in one unified HGNN while treating slack, PV, and PQ buses as distinct node sets.Independent encoder and decoder parameters enforce type-specific boundary conditions, while the backbone is shared across tasks.
  • IV. PROPOSED METHODOLOGY: Type-specific affine encoders project each bus type’s raw features into a shared hidden space followed by Layer Normalization.The encoders use separate transformations for slack, PV, and PQ nodes.
  • IV. PROPOSED METHODOLOGY: Stacked heterogeneous graph layers use relation-specific transformations and incorporate branch impedance and shunt-admittance attributes into message passing.The design targets nonlinear relationships imposed by Ohm’s and Kirchhoff’s laws.

3) Global Context and Residual Integration

The model adds graph-level context and residual pathways so local message passing can account for global OPF constraints while preserving initial physical and task-specific information.

  • 3) Global Context and Residual Integration: A global context vector is computed at each layer by mean-pooling node embeddings across all bus types and applying a shared MLP.The pooled representation normalizes over the full graph regardless of type partition.
  • 3) Global Context and Residual Integration: The global vector is broadcast to nodes and combined with per-node updates through a multi-path residual connection.This integrates graph-wide information into local node representations.
  • 3) Global Context and Residual Integration: The residual skip to the initial encoding preserves task-specific and bus-type boundary conditions through network depth.It is intended to prevent oversmoothing from erasing physical constraints.
  • 3) Global Context and Residual Integration: Task-specific output heads decode final node embeddings for physically different quantities such as voltage magnitude and angle.Voltage outputs are initialized at 1.0 p.u. and 0.0 rad, matching classical flat-start values.

B. Unified Multi-Task Training Objective

The unified training objective uses task-conditioned masked losses and diverse physically consistent data to train one model across PF, OPF, and SE. Data augmentation targets both operating-condition and topology shifts.

  • B. Unified Multi-Task Training Objective: Mean squared error supplies the primary gradient signal across PF, OPF, and SE despite their different problem characteristics.PF and SE are regression tasks, whereas OPF additionally requires operational constraint satisfaction.
  • B. Unified Multi-Task Training Objective: Task-conditioned masking applies loss terms only to state variables that are outputs for each task and bus type.For PF, voltage magnitude loss applies to PQ buses, while voltage-angle loss applies to PQ and PV buses.
  • C. Data Generation and Pre-processing: Training data randomizes loads, generator setpoints, line impedances, network topology, and SE measurement quality to improve distributional robustness.The pipeline includes load ranges, ±5% parameter perturbations, branch additions or removals, Gaussian measurement noise, and 5% masking.
  • B. Unified Multi-Task Training Objective: A one-hot task indicator conditions initial node embeddings on the specific requirements of PF, OPF, or SE.This enables multi-task learning within the shared model.
  • C. Data Generation and Pre-processing: The high-topo scenario includes 40% randomly modified-topology samples and approaches the feasibility boundary, causing some candidate samples to be discarded.Rejections occurred when the solver failed to converge or produced infeasible operating points.

D. Evaluation and Generalization Metrics

The paper evaluates the unified solver with complementary metrics because average training error alone does not capture worst-case accuracy or constraint satisfaction.

  • MSE is the primary training objective but is insufficient alone because operational applications require worst-case and constraint-satisfaction assessment.
  • NRMSE normalizes root mean square error by the ground-truth range, enabling comparisons across outputs with different magnitudes.
  • R2 measures the proportion of target variance explained, with values near unity indicating high goodness-of-fit across the operating range.
  • MaxAE identifies the largest prediction error across buses and test samples, which is operationally important because one large error can violate security limits.

V. RESULTS AND DISCUSSION

The evaluation selects an efficient architecture through hyperparameter and ablation studies, then tests accuracy and generalization across loading conditions and unseen topologies. The unified model remains accurate outside its mid-scenario training distribution and performs comparably to task-specific GNN solvers.

  • Experimental setup: The evaluation combines 3000 PF, OPF, and SE samples from the mid scenario and tests the trained model on narrow, wide, and high-topo datasets.
  • GNN cell comparison: ResGated GCN achieves the lowest average NRMSE among seven tested GNN cells and is selected for subsequent experiments.The comparison uses the 118-bus mid-scenario dataset and averages NRMSE across predicted quantities.
  • Ablation study: Removing global mean pooling increases average NRMSE by approximately 16%, from 0.0031 to 0.0036, highlighting its importance for system-wide OPF constraints.
  • Hyperparameter study: Hidden dimensions improve accuracy up to 128, while deeper networks improve in-distribution accuracy but degrade unseen-scenario generalization and increase inference time.The selected four-layer, 128-dimensional configuration provides a favorable accuracy–efficiency trade-off.
  • Generalization evaluation: MSE ranging from 1.34e-7 to 3.55e-6 demonstrates acceptable generalization from mid-scenario training to unseen loads up to 250% and modified topologies.The tested modifications included branch additions and removals not seen during training.
  • Benchmark comparison: Accuracy is comparable to, and sometimes surpasses, task-specific GNN solvers despite joint training across PF, OPF, and SE, although cross-study comparisons use different metrics and networks.

VI. CONCLUSION

The paper concludes that one heterogeneous ResGated GCN can solve PF, OPF, and SE with task-comparable accuracy through a shared, physics-aware representation. It presents this capability as an initial proof of concept toward foundational power-system models rather than a completed foundation model.

  • The architecture solves PF, OPF, and SE with accuracy comparable to task-specific models while requiring only one training pipeline.
  • Integrating line impedance parameters and global context helps represent local KCL physics together with system-wide OPF constraints.
  • Average NRMSE below 0.004 across predicted quantities on the mid scenario accompanies generalization to unseen loads and topologies without retraining.The reported load scaling reaches 250%.
  • The work is characterized as an initial step toward a foundational model, with multi-task consolidation, topological invariance, and a physics-informed latent space identified as prerequisites.Future work is directed toward larger-scale pre-training on global utility datasets.
Loading 2609.16738v1…