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

Yan Li, Yin Yu, Xinyuan Jiang, Daning Huang, Yan Li 0105, Meng Yue, Tianqiao ZhaoIEEE Access 2024 · DOI

TL;DR

This paper introduces PIDGeuN, a novel Graph Neural Network (GNN) approach for accurately predicting the transient dynamics of networked microgrids. By integrating physics-informed techniques into a GNN framework, PIDGeuN achieves robust, long-term predictions of microgrid behavior under disturbances, outperforming existing methods.

What Problem Does This Paper Address?

Modernizing the power grid to enhance resilience and reduce outages has led to the increasing integration of microgrids and networked microgrids (NMs). These systems, often relying on power-electronic interfaces from distributed energy resources (DERs), have significantly reduced inertia, making them highly susceptible to disturbances like PV fluctuations. This sensitivity leads to frequent and complex transient dynamics.

Traditional methods for studying these transients often require detailed system models, which become computationally challenging with the high dimensionality of NMs. Furthermore, microgrids are inherently nonlinear systems, and their operational states are constantly changing due to DER fluctuations and topology changes. While data-driven approaches offer a promising alternative, existing methods struggle with:

  • Local linearization: Linear models can't capture the full nonlinear operating envelope.
  • Complexity of nonlinear models: Identifying correct nonlinear terms is difficult, and the number of terms grows exponentially with system size.
  • Dependency on internal states: Many methods require access to internal system states (like controller variables) that are hard to measure in practice.
  • Ignoring network topology: Crucial inter-dependencies between buses are often overlooked.

This paper aims to bridge this gap by developing a data-driven method that accurately, efficiently, and robustly predicts microgrid transient dynamics using only practically measurable state information, while respecting the underlying physics of power systems.

Methods

The core of the proposed solution is the Physics-Informed Dynamic Graph Neural Network (PIDGeuN). It leverages a GNN architecture to represent the microgrid topology as a graph, where buses are nodes and connections are edges. Key novelties include:

  • Physics-Data-Infusion for Edge Weights: Instead of relying solely on predefined admittance matrices for edge weights, PIDGeuN dynamically learns and adjusts these weights by combining data-driven attention mechanisms with the physics-based admittance matrix. This better captures the inter-dependencies between buses.
  • Physics-Informed Loss Function: A loss term is incorporated to enforce the physical law of power systems, specifically Kirchhoff's law. This ensures that the model's predictions are physically feasible and improves long-term prediction accuracy by penalizing violations of power flow equations.
  • Time-Delayed Embedding: To capture temporal dependencies, PIDGeuN uses a time-delay embedding formulation, allowing it to reproduce system dynamics using past state information.
  • Hybrid Message Passing: The architecture employs a combination of Graph Attention Layers (GALs) and Graph Convolutional Layers (GCLs) in its "processor" component. GALs dynamically adjust edge weights based on node features, while GCLs perform spectral filtering on the graph Laplacian. This hybrid approach enhances expressiveness.
  • Encoder-Processor-Decoder Structure: The network follows a standard deep learning architecture: an encoder maps raw node states to a latent representation, the processor refines these representations through graph operations, and a decoder maps the final latent vectors to the predicted rate of change of system states.
  • Activation Functions: Batch Normalization (BN) and Parametric Rectified Linear Unit (PReLU) are used to handle potential differences in latent vectors caused by disturbances and improve training stability.

Datasets

The paper utilizes synthetic datasets generated from a detailed dynamic model of a typical 33-bus networked microgrid system. The dataset includes:

  • Training Data: 90 complete transient responses and 5000 initial transient responses with load changes.
  • Testing Data: 30 complete transient responses and 50 initial transient responses with load changes.

The disturbances are simulated as step load changes applied to randomly chosen nodes, with magnitudes ranging from ±10% of the nominal load, exceeding the regime of linear analysis.

Key Insights and Ideas

  • Topology Matters: The paper strongly emphasizes that the network topology of microgrids is crucial for understanding and predicting their transient dynamics, a factor often overlooked by non-graph-based methods.
  • Physics as a Constraint: Integrating physical laws (like Kirchhoff's law) directly into the training process is not just about accuracy but also about ensuring the feasibility of predictions, which is critical for long-term stability.
  • Dynamic Edge Weights: The idea of dynamically learning edge weights, rather than relying solely on static admittance matrices, allows the model to adapt to changing system conditions and better represent complex interdependencies.
  • Hybrid GNNs for Dynamics: Combining different types of GNN layers (GALs and GCLs) in a hybrid architecture appears to be more effective for capturing complex dynamic behaviors than using either type exclusively.
  • Long-Term Prediction is Hard (But Achievable): The paper highlights the significant challenge of long-term prediction in dynamic systems, where errors can accumulate. PIDGeuN's physics-informed approach is shown to be particularly effective in mitigating this error accumulation.

Key Results

  • Superior Accuracy: PIDGeuN consistently achieved the lowest Cumulative RMSE (C-RMSE) for both complete and initial transient responses, indicating significantly more accurate long-term predictions compared to baseline methods like subspace identification, LSTM, GCRN, and STGCN.
  • Robustness: The model demonstrated robustness in predicting dynamics under frequent load changes, mimicking real-world scenarios where systems are under continuous disturbance.
  • Reduced Error Accumulation: While other models diverged quickly (e.g., STGCN beyond 200 time steps), PIDGeuN showed a much slower growth in prediction error over time.
  • Efficiency: The PIDGeuN achieved its superior performance with a comparable number of parameters to some of the less effective baseline models.
  • Physics-Informed Loss Impact: Removing the physics-informed loss term (model C12) led to a significant increase in Kirchhoff's law violations and higher C-RMSE, demonstrating the importance of this component.
  • Hyperparameter Sensitivity: The study revealed that the number of message-passing steps (achieved through combinations of GALs and GCLs) is critical for performance, while parameters like latent size (D) are less sensitive.

Limitations and Future Work

  • Dataset Generation: The current study relies on synthetic data generated from a DAE model. Future work could involve validating PIDGeuN on real-world microgrid data.
  • Real-time Control: The authors explicitly state that their next step is to explore the application of PIDGeuN for real-time predictive or preventive control in microgrids, leveraging its accurate prediction capabilities.
  • Scalability: While GNNs are generally scalable, further investigation into the performance of PIDGeuN on even larger and more complex networked microgrid systems would be beneficial.
  • Disturbance Types: The current study focuses on load changes. Extending the model to handle other types of disturbances (e.g., generator faults, renewable source intermittency) would increase its applicability.

Summary updated

BibTeX
@article{li2024pidgeun,
  author = {Yan Li and Yin Yu and Xinyuan Jiang and Daning Huang and Yan Li and Meng Yue and Tianqiao Zhao},
  title = {{PIDGeuN: Graph Neural Network-Enabled Transient Dynamics Prediction of Networked Microgrids Through Full-Field Measurement}},
  journal = {IEEE Access},
  year = {2024},
  doi = {10.1109/access.2024.3384457},
}
Figure