Abstract
High renewable energy penetration intensifies net-load ramping and increases the demand for operational flexibility in distribution networks. However, existing hybrid AC/DC distribution network planning studies rarely consider both renewable ramp stress and the network deliverability of available flexibility. Furthermore, directly generating discrete investment decisions through reinforcement learning (RL) leads to a large combinatorial action space, making planning feasibility difficult to preserve. To address these issues, this paper proposes a hybrid AC/DC distribution network planning framework under renewable ramp stress, considering network-constrained flexibility via RL with feasible projection. A normalized net-load-ramp feature is incorporated into the planning state to represent stage-dependent renewable ramp stress. Instead of evaluating flexibility solely from resource regulation headroom, the proposed network-constrained flexibility evaluation further considers voltage and line-flow constraints, and the resulting flexibility shortages are incorporated into the planning reward. A mixed-integer feasible-projection model maps the continuous planning priorities learned by proximal policy optimization (PPO) into discrete stage-wise planning schemes. Case studies on modified IEEE 33-bus and IEEE 123-bus systems demonstrate the effectiveness of the proposed framework. In the IEEE 33-bus high-renewable case, the proposed method reduces the discounted total cost by 32.18% relative to rolling mixed-integer linear programming (MILP) while maintaining a zero investment-infeasible-stage rate. For the IEEE 123-bus system, the selected PPO solution reduces the stage-summed load shedding from 57.01 MWh for the AC-only plan and 1.94 MWh for the best of 50 feasible random-priority plans to 0.554 MWh.