A deep-learning based method for accelerating dynamic reconfiguration of distribution networks
Yanqing Wu, Tao Qian, Jingwen Ye, Qinran Hu, Qiangsheng Bu, Zhigang Ye
Abstract
Distribution network dynamic reconfiguration (DNDR), formulated as a mixed-integer quadratic programming (MIQP) problem, is computationally intractable for large-scale systems due to combinatorial complexity and temporal coupling. To address this issue, this paper presents a deep learning-based framework that converts the MIQP problem into a solvable quadratic program (QP), enabling accelerated DNR. The framework uses the Informer model with ProbSparse self-attention to identify temporal dependencies in time series data and predict the status of line switches. It then uses a threshold decision mechanism to determine which predicted switch states to fix, thereby reducing the number of binary variables. Case studies in a modified IEEE 33-node system demonstrate the effectiveness of the framework, achieving a 99% acceleration in model solution speed without compromising feasibility or optimality. This methodology combines data-driven prediction and optimization to offer a scalable solution for real-time grid reconfiguration in dynamic environments.