Sequential Control of Individual Switches for Real-Time Distribution Network Reconfiguration Using Deep Reinforcement Learning

This paper presents a study on Dynamic Sequential Distribution Network Reconfiguration (DSDNR), which considers the sequential control of individual switches in real-time distribution network operation. Traditional Distribution Network Reconfiguration (DNR) is formulated as an optimization problem that determines optimal switch states for network conditions. With the increasing variability introduced by distributed energy resources (DERs), real-time DNR research has evolved toward continuously identifying optimal switch states over time. However, in practical operations, the simultaneous change of multiple switches can compromise system stability, making it essential to optimize the sequential order of individual switch operations. This study introduces a new DNR formulation that integrates sequential switching optimization into real-time applications, enabling fast and stable operation under dynamic conditions. To achieve this, we propose a deep reinforcement learning (DRL)-based approach for sequential optimization and develop an agent learning environment that incorporates constraints on conditional loop networks and individual switching conditions, newly introduced in the DSDNR problem. The proposed algorithm is validated through case studies, including unbalanced systems, demonstrating that optimized sequential switching prevents partial outages, voltage instability, and line overloading. Furthermore, the results confirm that the method ensures stable system operation even in high-DER variability environments.