Input Convex Neural Network-Assisted Optimal Power Flow in Distribution Networks: Modeling, Algorithm Design, and Applications
Rui Cheng, Yuze Yang, Wenxia Liu, Nian Liu, Zhaoyu Wang
TL;DR
This paper addresses distribution-network OPF when detailed network models are imperfect or unavailable by introducing ICNN-Assisted OPF, which embeds convex ICNN-based constraints into a traditional OPF. The key method combines a convex ICNN learning of the mapping from control inputs to voltages and powers with a regularized, fast primal-dual gradient algorithm, yielding proven convergence and optimality guarantees. The approach is demonstrated on meshed and unbalanced networks, showing accurate approximation of nonlinear power flows and safe, reliable operation under hard constraints, with robustness advantages over purely model-based or purely data-driven methods. The work offers a scalable, data-informed optimization framework that reduces modeling burden while maintaining feasibility guarantees, and it highlights directions for robustness and multi-period extensions.
Abstract
This paper proposes an input convex neural network (ICNN)-Assisted optimal power flow (OPF) in distribution networks. Instead of relying purely on optimization or machine learning, the ICNN-Assisted OPF is a combination of optimization and machine learning. It utilizes ICNN to learn the nonlinear but convex mapping from control variables to system state variables, followed by embedding into constrained optimization problems as convex constraints. Utilizing a designed ICNN structure, a fast primal-dual gradient method is developed to solve the ICNN-Assisted OPF, with the chain rule of deep learning applied to accelerate the algorithmic implementation. Convergence and optimally properties of the algorithm design are further established. Finally, different distribution network applications are discussed and proposed by means of the ICNN-Assisted OPF.