关键词: 配电网重构, 数据驱动, 分布式电源, 不确定性分析, 动态重构

Abstract:

The widespread application of distributed generation（DG） in distribution networks has brought significant uncertainty challenges to the safe and economical operation of systems due to the randomness of DG output and load fluctuations. Traditional reconfiguration methods generally suffer from bottlenecks such as low computational efficiency and difficulty meeting real-time control requirements under complex time-varying operating conditions. Therefore， a data-driven dynamic reconfiguration method for distribution networks considering DG uncertainty was proposed to achieve millisecond-level intelligent control of active distribution networks. The method adopted a decoupled framework of "offline complex optimization-online rapid decision-making". In the offline stage， Latin hypercube sampling was used to generate operating scenarios covering multiple uncertainties. An improved genetic algorithm was used to optimize a large number of scenarios， with the objective of minimizing the total daily network loss of the system. The globally optimal combinations of reconfiguration switches satisfying multiple physical constraints were obtained， and a high-quality "state-optimal topology" mapping dataset was constructed. In the online stage， a deep neural network decision model based on a "wide-narrow" architecture was constructed to accurately learn the nonlinear mapping between scenario features and optimal switching strategies. Simulation results on the IEEE 33-bus standard distribution system showed that the proposed model required only 0.012 s for a single online decision， significantly improving performance compared with traditional algorithms and meeting real-time response requirements. Meanwhile， compared with the fixed network topology， the proposed method reduced average daily energy losses by 1 163.9 kW·h， achieving a loss reduction rate of 36.9%， and increased the minimum node voltage per unit throughout the entire day from 0.899 to 0.947， effectively preventing local voltage violations. Under extreme unseen operating conditions， such as abrupt changes in photovoltaic output， the deviation between the loss obtained from the model decision and the theoretical optimal solution did not exceed 0.25%， demonstrating excellent robustness. The proposed method bridges the gap between computational complexity and real-time decision-making through deep learning， providing a promising real-time intelligent control solution with great engineering application prospects for active distribution networks containing highly uncertain DG.

Key words: distribution network reconfiguration, data-driven, distributed generation, uncertainty analysis, dynamic reconfiguration