Overview

DiffOPF treats solving the Optimal Power Flow (OPF) problem as conditional sampling from the underlying distribution of historical OPF data points.

Given active and reactive load demands $(P_d, Q_d)$, the model learns a conditional distribution over generator dispatches $(P_g, Q_g)$.

The pipeline consists of two main stages:

1. Data generation (Julia + PowerModels)
   Generate AC-OPF data points under randomized load and cost perturbations.

2. Training + Sampling (Python + Diffusion Model)
   Train a diffusion model on OPF solutions and sample generator outputs conditioned on loads.

---

Repository Structure

DiffOPF/
├── main.py                              # Entry point: sampling/inference
├── training/
│   └── train.py                         # Diffusion model training script
├── sampling/
│   └── sample.py                        # Reverse diffusion sampler (DPS-guided)
├── models/
│   └── model.py                         # Score network (SimpleNN + time embedding)
├── utils/
│   ├── diffusion.py                     # Noise schedule + diffusion utilities
│   └── normalization.py                 # Min-max normalization / denormalization
│
├── data/
│   ├── train/                           # Training dataset (generated via Julia script)
│   ├── test/                            # Test dataset (generated via Julia script)
│   └── test_case/                       # Test case inputs 
│
├── configs/
│   └── IEEE_118_Parameters.json         # System parameters (buses, lines, generators)
│
├── checkpoints/
│   └── trained_model.pth                # Trained diffusion model weights
│
├── outputs/
│   └── DiffOPF_Solution.csv             # Generated OPF samples
│
├── data_generation/
│   └── opf_data_generation.jl           # Julia script for AC-OPF dataset generation
│
├── requirements.txt                     # Python dependencies
└── README.md

---

Installation

git clone https://github.com/<your-username>/DiffOPF.git
cd DiffOPF
pip install -r requirements.txt

Requirements

• Python ≥ 3.8
• PyTorch ≥ 1.12
• Julia ≥ 1.8 (for data generation)
• PowerModels.jl, JuMP, Ipopt

---

Data Generation (Julia)

Before training the diffusion model, generate OPF datasets using:

julia data_generation/opf_data_generation.jl

Notes

• The script solves AC-OPF for randomized load and cost perturbations.
• Inputs are read from:

data/test_case/

• Output datasets are saved into:

data/train/
data/test/

Output format

Each row contains:

Pd_1 ... Pd_n, Qd_1 ... Qd_n, Pg_1 ... Pg_n, Qg_1 ... Qg_n

---

Training

Train the diffusion model:

python training/train.py

What it does

• Loads training dataset from data/train/
• Applies Min-Max normalization
• Trains a score-based diffusion model
• Saves:
  • model checkpoint → checkpoints/
  • loss curve → outputs/

---

Inference / Sampling

Run conditional sampling:

python main.py --n_instances 5 --num_samples 2 --output outputs/DiffOPF_Solution.csv

Arguments

Argument	Default	Description
--n_instances	5	Number of OPF test instances
--num_samples	2	Samples per instance
--output	outputs/DiffOPF_Solution.csv	Output CSV path

---

Output Format

Each generated sample contains:

instance_id, sample_id,
Pd_1 ... Pd_n,
Qd_1 ... Qd_n,
Pg_1 ... Pg_n,
Qg_1 ... Qg_n

---

Pretrained Model

A pretrained checkpoint is provided:

checkpoints/trained_model.pth

---

Citation

@misc{hoseinpour2026diffopfdiffusionsolveroptimal,
      title={DiffOPF: Diffusion Solver for Optimal Power Flow},
      author={Milad Hoseinpour and Vladimir Dvorkin},
      year={2026},
      eprint={2510.14075},
      archivePrefix={arXiv},
      primaryClass={eess.SY},
      url={https://arxiv.org/abs/2510.14075}
}

License

This project is licensed under the MIT License. See LICENSE for details.