DS-PAH-GNN: Physics-Aware Hierarchical GNN for Microgrid Topology Optimization

This project implements a Data-Driven Digital Twin for a Dual-Bus DC Microgrid. It uses a Physics-Aware Graph Neural Network (GNN) to predict power losses and optimize grid topology in real-time, bypassing slow traditional physics solvers. This approach allows for rapid grid reconfiguration and optimization, enhancing stability and efficiency.

• Physics-Aware AI: Uses GINEConv layers that explicitly incorporate electrical conductance ($1/R$) into message passing.
• Heterogeneous Modeling: Distinctly models Main Buses, Local Stations, EV Chargers, PV Units, and Battery Storage.
• Topology Optimization: Automatically identifies the optimal switch configuration to minimize losses during load spikes.

1. Installation
2. Usage Pipeline
3. Documentation
4. File Structure
5. Configuration
6. License

1. Prerequisites: Python 3.8+, CUDA (optional for GPU support).
2. Install Dependencies:

   cd ds-pah-gnn
   pip install -r requirements.txt

The core workflow involves simulating a microgrid environment, training a GNN to predict system states, and then using the trained model to optimize grid topology in real-time.

Simulates 600,000 grid snapshots with stochastic EV loads and solar generation.

python generate.py

Trains the GNN to predict power loss from graph topology. Generates training_loss.png.

python train.py

Calculates $R^2$, MAE, and RMSE on unseen data. Generates inference_plot.png.

python inference.py

Finds the optimal topology for a stressed grid scenario (e.g., Station Overload).

python optimize.py

Research Use Only.