Physics Embedded Graph Convolution Neural Network for Power Flow Calculation Considering Uncertain Injections and...

Area of Science:

• Electrical Engineering
• Computational Science
• Artificial Intelligence

Background:

• Probabilistic power system analysis requires efficient power flow calculations to quantify uncertainty impacts.
• Existing data-driven methods lack robustness against uncertain injections and topological variations.
• Computational burden of repetitive power flow calculations hinders real-time operational analysis.

Purpose of the Study:

• To propose a novel model-driven graph convolution neural network (MD-GCN) for enhanced power flow calculation.
• To improve computational efficiency and robustness in power system operations.
• To address limitations of traditional and data-driven power flow analysis methods.

Main Methods:

• Developed a model-driven graph convolution neural network (MD-GCN) by embedding a linearized power flow model into layer-wise propagation.
• Implemented a new input feature construction method with multiple neighborhood aggregations and a global pooling layer.
• Integrated physical connection relationships and system-wide impacts for comprehensive feature representation.

Main Results:

• MD-GCN demonstrated high computational efficiency and superior robustness to topology changes.
• The proposed method significantly outperformed existing approaches on IEEE standard systems (30, 57, 118, and 1354-bus).
• Enhanced feature extraction through neighborhood aggregation and global pooling improved accuracy.

Conclusions:

• MD-GCN provides an effective solution for accurate and efficient power flow calculations under uncertainty.
• The model's interpretability is enhanced by incorporating physical power flow principles.
• The approach is scalable and robust for analyzing large-scale power systems.