Scholarlink

<p>This paper addresses uncertainty, multi-timescale control, and real-time computational challenges in AC/DC hybrid active distribution systems with high-penetration distributed renewables and heterogeneous controllable resources. A two-stage optimal scheduling framework is proposed by integrating deep reinforcement learning (DRL) with model predictive control (MPC). In the upper stage (15-min resolution), a deep deterministic policy gradient (DDPG)-based policy generates fast scheduling references for the energy storage system (ESS) and converter-interfaced resources, replacing repetitive large-scale online optimization and improving real-time decision efficiency. To reduce the risk that aggressive DRL actions are directly transmitted to the lower layer, a control barrier function (CBF)-inspired safety projection layer is further introduced to project the raw DRL actions into a safe set defined by SOC, voltage, and converter-capacity constraints. In the lower stage (1-min resolution), an MPC closed-loop feedback layer uses real-time renewable-generation and load measurements to correct renewable distributed generator (RDG) and voltage-source converter (VSC) setpoints. The offline training of the DRL policy can be parallelized over multiple source-load scenarios, while online deployment requires only a fast neural-network forward pass followed by short-horizon MPC correction, making the framework suitable for real-time and high-performance-computing-assisted applications. Simulations on a 51-node system (IEEE-33 AC feeder coupled with three DC subnetworks) demonstrate that the proposed method improves voltage regulation, maintains ESS SOC within limits, and enhances schedule tracking under renewable uncertainty.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Two-layer dispatching for hybrid AC/DC active distribution systems under renewable uncertainty: a DRL–MPC solution

  • 期刊论文

  • 发表日期：

    2026年6月25日

查看全文
  • Jujin Yu,
  • Dongxun Xie,
  • Yi Su

摘要

This paper addresses uncertainty, multi-timescale control, and real-time computational challenges in AC/DC hybrid active distribution systems with high-penetration distributed renewables and heterogeneous controllable resources. A two-stage optimal scheduling framework is proposed by integrating deep reinforcement learning (DRL) with model predictive control (MPC). In the upper stage (15-min resolution), a deep deterministic policy gradient (DDPG)-based policy generates fast scheduling references for the energy storage system (ESS) and converter-interfaced resources, replacing repetitive large-scale online optimization and improving real-time decision efficiency. To reduce the risk that aggressive DRL actions are directly transmitted to the lower layer, a control barrier function (CBF)-inspired safety projection layer is further introduced to project the raw DRL actions into a safe set defined by SOC, voltage, and converter-capacity constraints. In the lower stage (1-min resolution), an MPC closed-loop feedback layer uses real-time renewable-generation and load measurements to correct renewable distributed generator (RDG) and voltage-source converter (VSC) setpoints. The offline training of the DRL policy can be parallelized over multiple source-load scenarios, while online deployment requires only a fast neural-network forward pass followed by short-horizon MPC correction, making the framework suitable for real-time and high-performance-computing-assisted applications. Simulations on a 51-node system (IEEE-33 AC feeder coupled with three DC subnetworks) demonstrate that the proposed method improves voltage regulation, maintains ESS SOC within limits, and enhances schedule tracking under renewable uncertainty.