Deep reinforcement learning-enabled methods for large-scale active distribution network planning with forbidden zones

The global shift toward sustainable energy has accelerated the integration of photovoltaic (PV) systems and battery energy storage systems (BESS) into low-voltage (LV) distribution networks. This transition introduces bidirectional power flows and complex geospatial constraints, such as forbidden zones, which conventional planning methods struggle to address. While deterministic mixed-integer programming (MIP) ensures optimal performance, it is computationally prohibitive for large-scale networks, often requiring hours of calculation. Conversely, traditional heuristics often lack stability and converge to suboptimal solutions under high distributed energy resource (DER) penetration. To bridge this gap, this paper proposes a Proximal Policy Optimization (PPO) reinforcement learning framework for LV distribution network planning. Our method explicitly encodes geospatial forbidden zones into the RL action space to ensure geographic feasibility. The proposed PPO framework was compared against five heuristic algorithms and MIP across three representative networks. Quantitative results demonstrate that PPO achieves near-optimal performance, with total investment costs and network losses remaining within 0.8%–1.0% of the theoretical optimum achieved by MIP. Furthermore, PPO exhibits superior computational efficiency; in a large-scale 69-node network, PPO generates optimal topologies in approximately 200 s, representing a speed-up factor of over 37.5 times compared to MIP, which exceeds 7,500 s. These findings establish PPO as a scalable and robust alternative for real-world grid modernization.