Towards Real-Time Machine Learning Approximations of AC Optimal Power Flow

This repository provides open-source code and experiments accompanying the paper:

"Towards Real-Time Machine Learning Approximations of AC Optimal Power Flow"
IEEE Conference on Artificial Intelligence (CAI), 2025

📄 Overview

This project explores how machine learning can approximate solutions to the Alternating Current Optimal Power Flow (ACOPF) problem — a fundamental but computationally intensive task in power system operation. The work compares Neural Networks (NNs), Gaussian Mixture Models (GMMs), and Linear Regression (LR) across IEEE standard test cases.

The goal is to provide insights into the feasibility, accuracy, and runtime of data-driven OPF surrogates for near real-time applications, especially under varying grid sizes and complexity.

🚀 Features

• ✅ ACOPF solved using pandapower on multiple IEEE benchmark networks
• ✅ Clean modular codebase with support for:
  • Neural Network model training and tuning
  • GMM clustering with error estimation
  • Baseline Linear Regression
• ✅ Hyperparameter grid search with early stopping
• ✅ Evaluation across five test systems (30-bus to GB-reduced)
• ✅ Extensible framework for future researchers

📘 Paper Summary

Main Contributions:

• A reproducible pipeline to extract input-output mappings from ACOPF solutions:
  (Pd, Qd) → (Pg, Vg)

• Comparative analysis of NN, GMM, and LR across 5 power networks:

  • 🏆 NNs perform best in large/highly nonlinear grids
  • 📈 LR is surprisingly strong in small systems
  • ⚠️ GMMs struggle with high-dimensional data

• MSE benchmarking on scaled data

• Discussion of feasibility, generalization, and real-time deployment risks

📄 Full paper: (link to proceedings will be added, now use this DOI)

🧪 Used Power Systems Test Cases for OPF Scenarios

| Case | Buses | Generators | Scenarios Used | Description | |-------------------|-------|------------|----------------|--------------------------------------| | case30 | 30 | 6 | 198 | Classic IEEE 30-bus system | | case_ieee30 | 30 | 6 | 300 | Variant with minor topology tweaks | | case39 | 39 | 10 | 299 | New England benchmark | | case118 | 118 | 54 | 300 | Large IEEE network | | GBreducednetwork| ~150 | Many | 300 | UK transmission-level abstraction |

🧠 Model Architecture

The neural network model learns a mapping from load demands to optimal generator setpoints using a feedforward structure.

The core approximation learned is:

(Pd, Qd) → (Pg, Vg)

Where: - ( Pd, Qd ) are the active and reactive power demands at load buses (inputs) - ( Pg, Vg ) are the generator real power outputs and voltage magnitudes at generator buses (outputs)

Each architecture is tuned using grid search on:

• Learning rate
• Batch size
• Hidden layer width/depth
• Dropout rate
• Patience for early stopping

---

⚙️ Installation

1. Clone this repository: bash git clone https://github.com/LIMESS-Sustainability-Lab/Towards-ML-for-OPF.git cd Towards-ML-for-OPF

2. Create and activate a virtual environment (recommended): ```bash

   On Windows

   python -m venv venv .\venv\Scripts\activate

On Linux/Mac

python -m venv venv source venv/bin/activate ```

1. Install the required dependencies: bash pip install -r requirements.txt

🚀 Usage

The main script opf-runner.py can be run directly:

bash python opf-runner.py

This will: 1. Generate up to 300 OPF scenarios for multiple OPF scheme test cases 2. Train and evaluate different models (Neural Network, GMM-GMR, Linear Regression) 3. Output the results showing performance on test sets

The script includes several test cases: - case30 - case_ieee30 - case39 - case118 - GBreducednetwork

📃 Requirements

The project requires Python 3.8 or higher and the following packages: - pandapower>=2.14.0 - scikit-learn>=1.3.0 - matplotlib>=3.7.0 - numpy>=1.24.0 - torch>=2.0.0 - pandas>=2.0.0

All dependencies are listed in requirements.txt and can be installed using pip as shown in the installation instructions.

📌 Known Limitations

• Only a single ACOPF sample is used per case → no time-series/load diversity
  
• No physics-aware features beyond load inputs
  
• NN feasibility violations possible without post-processing
  
• GMM scalability limited in high-dimensional settings
  

---

🧩 Citation

If you use this codebase or find it helpful in your work, please cite it using the CITATION.cff file located in the root of this repository.

To automatically generate a citation, you can use GitHub's "Cite this repository" feature or reference the metadata in CITATION.cff.

🤝 Acknowledgments

This work was supported by the grant SGS24/093/OHK5/2T/13 and the CTU Distinguished Co-Supervisor Grant.

We gratefully acknowledge the contributions of the open-source community, whose tools made this research possible:

• pandapower — for flexible and user-friendly power system modeling
• PyTorch — for building and training machine learning models
• scikit-learn — for baseline models and preprocessing utilities

We also thank the maintainers of the IEEE benchmark networks for providing standardized test cases used in this study.