GridFM: A Physics-Informed Foundation Model for Multi-Task Energy Forecasting

GridFM is a physics-informed foundation model designed for multi-task energy forecasting using real-time data from Independent System Operators (ISOs). It combines pre-trained time series foundation models with domain-specific adaptations that respect power system physics.

🌟 Key Features

• FreqMixer Adaptation Layer: Novel frequency-domain mixing mechanism that transforms general-purpose foundation model representations into power-grid-specific patterns
• Physics-Informed Constraints: Integration of power balance equations and DC power flow approximations
• Multi-Task Learning: Simultaneous forecasting of load, prices, emissions, and renewable generation
• Graph Neural Networks: Zonal topology encoding via GCN for spatial dependencies
• Explainability: SHAP-based feature attribution and attention visualization

📊 Performance

GridFM achieves state-of-the-art performance on NYISO, PJM, and CAISO datasets:

🚀 Quick Start

Installation

# Clone the repository
git clone https://github.com/GridFM/GridFM.git
cd GridFM

# Create conda environment
conda create -n gridfm python=3.10
conda activate gridfm

# Install dependencies
pip install -r requirements.txt

# Install GridFM
pip install -e .

Basic Usage

from gridfm import GridFM
from gridfm.data import NYISODataLoader

# Load pre-trained model
model = GridFM.from_pretrained("gridfm-base-nyiso")

# Load data
dataloader = NYISODataLoader(
    start_date="2023-01-01",
    end_date="2023-12-31",
    zones=["WEST", "CENTRL", "NORTH"]
)

# Generate forecasts
forecasts = model.predict(
    dataloader,
    horizon=24,  # 24 steps ahead (2 hours at 5-min resolution)
    tasks=["load", "lbmp", "emissions"]
)

# Access results
print(f"Load MAPE: {forecasts.metrics['load_mape']:.2f}%")

Training from Scratch

from gridfm import GridFM, GridFMConfig
from gridfm.training import Trainer

# Configure model
config = GridFMConfig(
    backbone="moirai-moe-base",
    num_zones=11,
    hidden_dim=256,
    num_freq_components=64,
    physics_weight=0.1,
    coupling_weight=0.05
)

# Initialize model
model = GridFM(config)

# Train
trainer = Trainer(
    model=model,
    train_dataloader=train_loader,
    val_dataloader=val_loader,
    learning_rate=1e-4,
    num_epochs=100
)
trainer.train()

📁 Repository Structure

GridFM/
├── gridfm/                    # Main package
│   ├── models/               # Model architectures
│   │   ├── gridfm.py        # Main GridFM model
│   │   ├── backbone.py      # Foundation model backbone
│   │   └── task_heads.py    # Task-specific output heads
│   ├── layers/              # Custom layers
│   │   ├── freqmixer.py     # FreqMixer adaptation layer
│   │   ├── gcn.py           # Graph convolutional network
│   │   └── attention.py     # Attention mechanisms
│   ├── physics/             # Physics constraints
│   │   ├── power_balance.py # Power balance equations
│   │   └── dc_power_flow.py # DC power flow approximation
│   ├── data/                # Data loading utilities
│   │   ├── nyiso.py         # NYISO data loader
│   │   ├── pjm.py           # PJM data loader
│   │   └── caiso.py         # CAISO data loader
│   └── utils/               # Utility functions
├── configs/                  # Configuration files
├── scripts/                  # Training and evaluation scripts
├── examples/                 # Example notebooks
├── docs/                     # Documentation
├── paper/                    # Paper and supplementary materials
└── tests/                    # Unit tests

📖 Documentation

• Installation Guide
• Data Preparation
• Model Architecture
• Training Guide
• API Reference

🔧 Configuration

GridFM uses YAML configuration files. Example:

# configs/gridfm_nyiso.yaml
model:
  backbone: moirai-moe-base
  hidden_dim: 256
  num_freq_components: 64
  freq_delta: 2
  
physics:
  enable_power_balance: true
  enable_dc_power_flow: true
  physics_weight: 0.1
  
training:
  batch_size: 32
  learning_rate: 1e-4
  num_epochs: 100
  early_stopping_patience: 10

📈 Reproducing Paper Results

To reproduce the results from our paper:

# Download and preprocess data
python scripts/download_data.py --iso nyiso --start 2019-01-01 --end 2023-12-31

# Train model
python scripts/train.py --config configs/gridfm_nyiso.yaml

# Evaluate
python scripts/evaluate.py --checkpoint checkpoints/best_model.pt --test_year 2023

# Generate figures
python scripts/generate_figures.py --results results/evaluation/

🧪 Testing

# Run all tests
pytest tests/

# Run specific test module
pytest tests/test_freqmixer.py -v

# Run with coverage
pytest --cov=gridfm tests/

📝 Citation

If you use GridFM in your research, please cite our paper:

@article{gridfm2024,
  title={GridFM: A Physics-Informed Foundation Model for Multi-Task Energy Forecasting Using Real-Time NYISO Data},
  author={Sayghe, Ali},
  journal={MDPI Energies},
  year={2024}
}

🤝 Contributing

We welcome contributions! Please see our Contributing Guide for details.

1. Fork the repository
2. Create your feature branch (git checkout -b feature/amazing-feature)
3. Commit your changes (git commit -m 'Add amazing feature')
4. Push to the branch (git push origin feature/amazing-feature)
5. Open a Pull Request

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

🙏 Acknowledgments

• NYISO, PJM, and CAISO for providing open access to grid data
• The Moirai team for the foundation model backbone
• PyTorch Geometric team for GNN implementations

📧 Contact

For questions or feedback, please open an issue or contact [[email protected]].

---

Made with ❤️ for sustainable energy systems