Towards Hierarchical GNNs for multi-grid power flow: generalization across operating scenarios

Problem

This work addresses the challenge of generalizing multi-grid power-flow models to new operating scenarios, which is critical for the adaptability of power systems. The authors highlight that existing models struggle with this generalization, particularly in the context of varying grid topologies and operational conditions. The paper is a preprint and has not undergone peer review.

Method

The authors propose a hierarchical latent communication module integrated within a GENCO-based corrective network. This architecture utilizes two reduced graphs to facilitate improved communication and representation learning across different grid configurations. The training dataset comprises three distinct grid topologies, with 200 newly generated preselected scenarios for each topology. The models were trained for 200 epochs, employing three different initialization seeds to ensure robustness in the results. Evaluation was conducted using Kron-derived transports, a same-anchor Quotient construction, and a flat backbone for comparative analysis.

Results

The proposed hierarchical models demonstrated significant improvements in performance metrics. Specifically, the macro family-balanced voltage error was reduced from 5.660 ± 0.899 to 0.851 ± 0.110, achieving an 85.0% reduction compared to the Flat GENCO baseline and a 31.0% reduction compared to the Quotient model, which reported an error of 1.235 ± 0.225. Furthermore, the hierarchical models consistently outperformed the per-bus mean fitted on training solutions across all training topologies and initialization seeds, indicating their superior generalization capabilities.

Limitations

The authors acknowledge that the current models do not outperform the fitted reference in a calibrated-transfer setting when applied to additional topologies. This limitation suggests that while the hierarchical approach shows promise, there are still scenarios where it may not generalize effectively. Additionally, the reliance on specific grid topologies for training may restrict the applicability of the findings to more diverse real-world scenarios.

Why it matters

The implications of this research are significant for the development of adaptive power system models. By enhancing the generalization capabilities of multi-grid power-flow models, this work paves the way for more resilient and flexible power systems that can better accommodate varying operational conditions. This could lead to improved efficiency and reliability in power distribution networks, ultimately benefiting energy management and sustainability efforts.