Fast Diffusion with Physics-Correction for ACOPF

Generating large-scale, physically consistent AC Optimal Power Flow (ACOPF) datasets is essential for modern data-driven power system applications. The central challenge lies in balancing solution accuracy with computational efficiency. Recent diffusion-based generative models produce high-quality samples; however, their slow sampling procedures limit practical scalability. In this work, we argue that exact physical feasibility is ultimately enforced by power flow solvers or projection steps, and therefore the generative model only needs to produce good initializations rather than perfectly feasible solutions. Based on this insight, we propose a fast diffusion framework using Denoising Diffusion Implicit Models (DDIM) combined with physics-guided corrections during sampling. The proposed method replaces slow stochastic refinement with a small number of deterministic steps and explicit constraint guidance. Experiments on IEEE 6-, 24-, and 118-bus systems show that our approach achieves up to 20 times faster sampling than standard diffusion models while maintaining comparable statistical accuracy and physical consistency. This makes the method well suited for scalable OPF dataset generation and practical power system learning tasks. We release the implementation code at https://github.com/PSquare-Lab/DDIM_OPF.

💡 Research Summary

The paper addresses the pressing need for large‑scale, physically consistent AC Optimal Power Flow (ACOPF) datasets that underpin modern data‑driven power‑system applications such as deep‑learning‑based OPF proxies, security assessment, and contingency analysis. While diffusion‑based generative models (particularly Denoising Diffusion Probabilistic Models, DDPMs) have demonstrated impressive fidelity in capturing complex, non‑convex data manifolds, their inference cost is prohibitive because they require on the order of a thousand stochastic reverse‑diffusion steps. The authors argue that for ACOPF generation this level of stochastic fidelity is unnecessary: the final physical feasibility can be enforced by a power‑flow solver or a projection step, so the generative model only needs to provide a good initialization that lies within the basin of attraction of the feasibility operator.

Based on this insight, the authors propose a fast diffusion framework that replaces DDPM with Denoising Diffusion Implicit Models (DDIMs) and augments the deterministic reverse trajectory with physics‑guided constraint corrections. The method consists of three stages: (1) training a standard diffusion model on a dataset of feasible ACOPF solutions using the usual forward noising process; (2) at test time, generating samples via a non‑Markovian DDIM reverse process that can skip intermediate timesteps, dramatically reducing the number of steps (e.g., from 1000 to 30‑50); and (3) inserting a constrained‑guidance mechanism at each reverse step. Specifically, after the network predicts the clean estimate (\hat{x}_0), the authors map it to the physical variable space, compute differentiable residuals for the power‑flow equality constraints (R_H(x)=|H(x)|_2^2) and the inequality constraints (R_G(x)=|\max(G(x),0)|_2^2), and take a gradient step (\hat{x}’_0 = \hat{x}_0 - \lambda_t \nabla_x (R_H+R_G)). The guidance strength (\lambda_t) is larger in early, high‑noise steps and decays later, ensuring strong feasibility steering without sacrificing sample diversity. The corrected estimate is then fed back into the DDIM update, and the process repeats until the final latent is transformed back to the normalized space and denormalized to physical units.

Experimental evaluation is performed on IEEE 6‑bus, 24‑bus, and 118‑bus test systems. The authors compare their Constrained DDIM against a baseline DDPM (≈1000 steps) using three metrics: (i) distributional fidelity measured by 1‑Wasserstein distance and KL divergence, (ii) statistical similarity of marginal and joint distributions (active/reactive power, voltage magnitude, phase angle), and (iii) wall‑clock sampling time. Results show that Constrained DDIM achieves Wasserstein distances comparable to or slightly better than DDPM despite using 20‑30× fewer steps. Histograms and scatter plots reveal that the generated samples faithfully reproduce sharp peaks and boundary effects imposed by operational limits, and the joint distributions retain the physical correlations inherent to the AC power‑flow manifold. In terms of speed, sampling time drops from 1633 s to 57 s for the 6‑bus case, from 2637 s to 60 s for the 24‑bus case, and from 4423 s to 113 s for the 118‑bus case—representing up to a 20‑fold acceleration.

The paper’s contributions can be summarized as:

1. Introducing a DDIM‑based generative pipeline for ACOPF that dramatically reduces inference cost while preserving high‑quality samples.
2. Embedding ACOPF equality and inequality constraints directly into the reverse diffusion trajectory via gradient‑based guidance, effectively steering the generative path toward the feasible manifold.
3. Demonstrating scalability and statistical accuracy across small, medium, and large test systems, thereby providing a practical tool for rapid synthetic OPF dataset creation.

Future work suggested by the authors includes automatic scheduling of the guidance strength (\lambda_t), extension to more complex contingency constraints (e.g., N‑k failures), integration of real operational data for mixed‑domain training, and application of the same physics‑guided DDIM framework to other power‑system optimization problems such as voltage stability, renewable integration, and security‑constrained OPF. Overall, the study offers a compelling blend of diffusion‑model efficiency and physics‑informed guidance, opening a pathway toward fast, reliable synthetic data generation for the next generation of data‑centric power‑system tools.