The rapid growth of renewable energy and increasing electricity demand pose challenges to the reliability and flexibility of traditional distribution networks. To address these issues, the construction of AC/DC hybrid distribution networks (AC/DC-HDNs) based on existing AC grids has become a promising solution. However, planning the expansion of such networks faces challenges like complex device and line topologies, dynamic fluctuations in distributed generation (DG) and load, and high power electronics costs. This paper proposes a time- and space-integrated expansion planning method for AC/DC-HDNs. The approach builds a distribution grid model based on graph theory, integrating the spatial layouts of AC distribution lines, DGs, main grids, and loads, while capturing dynamic load and renewable energy generation characteristics through time-series analysis. A modified graph attention network (MGAT)-based deep reinforcement learning (DRL) algorithm is used for optimization, balancing economic and reliability objectives. The simulation results show that the modified algorithm outperforms traditional algorithm in terms of both training efficiency and stability, with a faster convergence and lower fluctuation in cumulative rewards. Additionally, the proposed algorithm consistently achieves higher cumulative rewards, demonstrating its effectiveness in optimizing the expansion planning of AC/DC-HDNs.
Keywords: AC/DC hybrid distribution network; deep reinforcement learning; expansion planning; graph attention network.