Graph-Structured Kernel Design for Power Flow Learning using Gaussian Processes

Parikshit Pareek, Deepjyoti Deka, Sidhant Misra

Abstract

This paper presents a graph-structured Gaussian process (GP) model for data-driven risk assessment of critical voltage constraints. The proposed GP is based on a novel kernel, named the vertex-degree kernel (VDK), that decomposes the voltage-load relationship based on the network graph. To estimate the GP efficiently, we propose a novel active learning scheme that leverages the additive structure of VDK. Further, we prove a probabilistic bound on the error in risk estimation using VDK-GP model that demonstrates that it is statistically comparable to using standard AC power flow (AC-PF), but does not require computing a large number of ACPF solutions. Simulations demonstrate that the proposed VDK-GP achieves more than two fold sample complexity reduction, compared to a generic GP on medium scale 500-Bus and large scale 1354-Bus power systems. Moreover, active learning achieves an impressive reduction of over 15 times in comparison to the time complexity of Monte-Carlo simulations (MCS), and have risk estimation error of order 1E-4 for both 500-Bus and 1354-Bus system, demonstrating its superior efficiency in risk estimation.

Introduction

In safety-critical power systems, Voltage Risk Assessment (VRA) is essential for determining the likelihood of bus voltages exceeding safe limits due to load fluctuations and renewable energy variability. While the Alternating Current Power Flow (ACPF) model is the industry standard, its computational cost—often requiring millions of Monte Carlo simulations (MCS)—makes it impractical for real-time applications. Machine learning alternatives like Deep Neural Networks (DNNs) offer speed but lack the formal reliability and safety guarantees required by grid operators. This paper introduces a Gaussian Process (GP) framework that bridges this gap by incorporating grid topology into the model architecture.

The Vertex-Degree Kernel (VDK)

To address the "curse of dimensionality" inherent in large-scale power grids, we propose the Vertex-Degree Kernel (VDK). The VDK exploits the physical structure of the network by assuming that a bus's voltage is primarily influenced by its immediate neighbors. By decomposing the global kernel into local, additive sub-kernels based on the network graph, the model becomes significantly more data-efficient and scalable.

Graph-structured decomposition of the Vertex-Degree Kernel (VDK) based on network topology.

This topology-aware approach allows the GP to learn complex voltage-load relationships using a fraction of the data required by generic kernels.

Network-Swipe Active Learning and Probabilistic Guarantees

We developed a "Network-Swipe" active learning (AL) algorithm to further optimize the training process. Instead of relying on random sampling, the algorithm uses the GP’s predictive variance to identify the most informative operating points across the network. This eliminates the need for expensive validation datasets and ensures the model is trained where it is most uncertain.

Furthermore, we provide theoretical probabilistic bounds proving that VDK-GP risk estimates are statistically equivalent to MCS. Crucially, the model is designed to be conservative, mathematically ensuring it is more likely to slightly overestimate risk than underestimate it, providing a "safety-first" tool for grid operators.

Experimental Results

The framework was evaluated on standard benchmarks, including IEEE 118-bus, 500-bus, and 1354-bus systems. The VDK-GP consistently achieved mean absolute errors (MAE) below 1E-03 p.u., which is the industry standard for high-fidelity modeling.

MAE performance comparison across different nodes, highlighting the efficiency of VDK-GP and Active Learning.

Key performance metrics include:

  • Efficiency: The model reproduced MCS-level risk estimates using 15x fewer ACPF computations.
  • Speed: Once trained, the GP surrogate provided a 120x reduction in evaluation time compared to traditional iterative solvers.
  • Data Savings: VDK-GP required less than 50% of the samples needed by standard GP models to reach target accuracy.

Conclusion

The proposed VDK-GP framework offers a scalable and theoretically grounded solution for real-time voltage risk assessment. By baking physical topology into the machine learning architecture, we achieve high accuracy and significant computational speedups without sacrificing reliability. Future work will explore non-Gaussian uncertainty models and time-coupled stochastic processes to further enhance grid resilience.

Summary updated

Abstract

This paper presents a graph-structured Gaussian process (GP) model for data-driven risk assessment of critical voltage constraints. The proposed GP is based on a novel kernel, named the vertex-degree kernel (VDK), that decomposes the voltage-load relationship based on the network graph. To estimate the GP efficiently, we propose a novel active learning scheme that leverages the additive structure of VDK. Further, we prove a probabilistic bound on the error in risk estimation using VDK-GP model that demonstrates that it is statistically comparable to using standard AC power flow (AC-PF), but does not require computing a large number of ACPF solutions. Simulations demonstrate that the proposed VDK-GP achieves more than two fold sample complexity reduction, compared to a generic GP on medium scale 500-Bus and large scale 1354-Bus power systems. Moreover, active learning achieves an impressive reduction of over 15 times in comparison to the time complexity of Monte-Carlo simulations (MCS), and have risk estimation error of order 1E-4 for both 500-Bus and 1354-Bus system, demonstrating its superior efficiency in risk estimation.
BibTeX
@misc{pareek2023graph,
  author = {Parikshit Pareek and Deepjyoti Deka and Sidhant Misra},
  title = {{Graph-Structured Kernel Design for Power Flow Learning using Gaussian Processes}},
  year = {2023},
  eprint = {2308.07867},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2308.07867},
}
Figure