A deep learning-based general robust method for network reconfiguration in three-phase unbalanced active distribution networks

Abstract

By connecting deep learning and robust optimization, this paper proposes a general robust method (GRM) for distribution network reconfiguration (DNR) while hedging against the risk brought by uncertainties in active distribution networks (ADNs). The method is general in the way that it is applicable for both loss minimization and load balancing in three-phase unbalanced distribution systems with few a priori assumptions on the distributions of the uncertain distributed generator (DG) output and loads. An uncertainty set construction network based on deep neural networks is first proposed to adaptively construct the uncertainty set from historical data for DGs and loads. Then the robust DNR for three-phase unbalanced networks is formulated as a two-stage mixed-integer quadratic programming (MIQP) problem considering the worst-case scenario within this uncertainty set. Finally, based on the column-and-constraint generation (C-CG) method and duality theory, an iterative algorithm is devised to solve the GRM. Numerical tests on two unbalanced IEEE benchmarks have validated the effectiveness of the proposed method.