[Paper Review] PIDGeuN: Graph Neural Network-Enabled Transient Dynamics Prediction of Networked Microgrids Through Full-Field Measurement

Yu Yin, Xinyuan Jiang+2 more|arXiv (Cornell University)|Apr 18, 2022

Microgrid Control and Optimization4 citations

TL;DR

This paper proposes PIDGeuN, a physics-informed graph neural network that predicts transient dynamics in networked microgrids using only full-field, measurable bus states. By integrating physics-informed message passing and a Kirchhoff’s law-constrained loss, PIDGeuN achieves accurate, robust, long-term transient predictions without relying on unmeasurable internal states, outperforming baselines in both complete and initial transient response cases.

ABSTRACT

A Physics-Informed Dynamic Graph Neural Network (PIDGeuN) is presented to accurately, efficiently and robustly predict the nonlinear transient dynamics of microgrids in the presence of disturbances. The graph-based architecture of PIDGeuN provides a natural representation of the microgrid topology. Using only the state information that is practically measurable, PIDGeuN employs a time delay embedding formulation to fully reproduce the system dynamics, avoiding the dependency of conventional methods on internal dynamic states such as controllers. Based on a judiciously designed message passing mechanism, the PIDGeuN incorporates two physics-informed techniques to improve its prediction performance, including a physics-data-infusion approach to determining the inter-dependencies between buses, and a loss term to respect the known physical law of the power system, i.e., the Kirchhoff's law, to ensure the feasibility of the model prediction. Extensive tests show that PIDGeuN can provide accurate and robust prediction of transient dynamics for nonlinear microgrids over a long-term time period. Therefore, the PIDGeuN offers a potent tool for the modeling of large scale networked microgrids (NMs), with potential applications to predictive or preventive control in real time applications for the stable and resilient operations of NMs.

Motivation & Objective

• Address the challenge of predicting nonlinear transient dynamics in networked microgrids (NMs) with reduced inertia due to high DER penetration.
• Overcome limitations of conventional methods that depend on unmeasurable internal states like controller dynamics.
• Leverage graph neural networks to naturally represent microgrid topology and improve prediction accuracy through spatial dependencies.
• Integrate physical laws—specifically Kirchhoff’s current law—into the learning process to ensure model feasibility and improve robustness.
• Develop a data-driven, scalable framework suitable for real-time predictive or preventive control in large-scale NMs.

Proposed method

• Employ a graph-based architecture where buses are nodes and lines are edges, enabling natural modeling of microgrid topology.
• Use time delay embedding of measurable bus states (voltage, active/reactive power) to reconstruct system dynamics without requiring internal controller states.
• Design a physics-informed message passing mechanism that combines data-driven attention weights with physics-based admittance weights to model inter-bus dependencies.
• Incorporate a physics-informed loss term enforcing Kirchhoff’s current law at each bus, ensuring predicted currents sum to zero and improving prediction feasibility.
• Utilize Chebyshev polynomials in graph convolutional layers to approximate spectral propagation and enable efficient long-range information flow.
• Train the model end-to-end using transient response data from load changes, optimizing for both prediction accuracy and physical consistency.

Experimental results

Research questions

• RQ1 Can a graph neural network effectively predict long-term transient dynamics in networked microgrids using only measurable full-field states?
• RQ2 How does integrating physics-informed message passing improve prediction accuracy and robustness compared to purely data-driven GNNs?
• RQ3 To what extent does enforcing Kirchhoff’s current law in the loss function enhance the physical feasibility and generalization of the model?
• RQ4 How do hyperparameters such as polynomial order, hidden dimension, and input sequence length affect prediction performance and computational cost?
• RQ5 Can the model generalize to both complete transient responses and initial transient dynamics under frequent load fluctuations?

Key findings

• PIDGeuN achieves significantly lower C-RMSE than baseline models in predicting complete transient responses, accurately capturing system frequencies, damping ratios, and final equilibrium states.
• The model outperforms a standard GNN and a linear model (C1) by a large margin, with C-RMSE reduced by over 50% in the complete response case.
• The physics-informed loss term effectively constrains the parameter space, resulting in more feasible predictions and improved generalization, especially in long-term simulations.
• Increasing the Chebyshev polynomial order $K$ from 1 to 5 reduces prediction error, but $K > 3$ yields marginal gains, with $K=7$ causing overfitting and increased error.
• A hidden layer size $D=128$ provides a good balance between model capacity and prediction accuracy, with performance relatively insensitive to $D$.
• Including two time steps ($C=2$) in the input sequence already provides sufficient temporal context; further increasing $C$ yields only marginal improvements, indicating limited long-term temporal dependency in the system.