Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Event-Driven Deep RL Dispatcher for Post-Storm Distribution System Restoration

Farshad Amani, Faezeh Ardali, Amin Kargarian

TL;DR

The paper addresses the need for fast, reliable crew-dispatch decisions during post-storm restoration when information arrives progressively and solvers are impractical in real time. It introduces an event-driven DRL framework with an actor–critic policy and a feasibility mask, trained via PPO on a dynamic, partially observable MDP using lightweight hazard surrogates. Key findings show that the DRL policy reduces energy-not-supplied and critical-load restoration time while achieving runtimes in the millisecond range, closely approaching rolling MILP performance but with far greater replanning flexibility. The approach is practical for utilities, scalable to larger feeders, and modular enough to incorporate improved hazard models and network checks, offering a robust, real-time tool for resilient outage management.

Abstract

Natural hazards such as hurricanes and floods damage power grid equipment, forcing operators to replan restoration repeatedly as new information becomes available. This paper develops a deep reinforcement learning (DRL) dispatcher that serves as a real-time decision engine for crew-to-repair assignments. We model restoration as a sequential, information-revealing process and learn an actor-critic policy over compact features such as component status, travel/repair times, crew availability, and marginal restoration value. A feasibility mask blocks unsafe or inoperable actions, such as power flow limits, switching rules, and crew-time constraints, before they are applied. To provide realistic runtime inputs without relying on heavy solvers, we use lightweight surrogates for wind and flood intensities, fragility-based failure, spatial clustering of damage, access impairments, and progressive ticket arrivals. In simulated hurricane and flood events, the learned policy updates crew decisions in real time as new field reports arrive. Because the runtime logic is lightweight, it improves online performance (energy-not-supplied, critical-load restoration time, and travel distance) compared with mixed-integer programs and standard heuristics. The proposed approach is tested on the IEEE 13- and 123-bus feeders with mixed hurricane/flood scenarios.

Event-Driven Deep RL Dispatcher for Post-Storm Distribution System Restoration

TL;DR

The paper addresses the need for fast, reliable crew-dispatch decisions during post-storm restoration when information arrives progressively and solvers are impractical in real time. It introduces an event-driven DRL framework with an actor–critic policy and a feasibility mask, trained via PPO on a dynamic, partially observable MDP using lightweight hazard surrogates. Key findings show that the DRL policy reduces energy-not-supplied and critical-load restoration time while achieving runtimes in the millisecond range, closely approaching rolling MILP performance but with far greater replanning flexibility. The approach is practical for utilities, scalable to larger feeders, and modular enough to incorporate improved hazard models and network checks, offering a robust, real-time tool for resilient outage management.

Abstract

Natural hazards such as hurricanes and floods damage power grid equipment, forcing operators to replan restoration repeatedly as new information becomes available. This paper develops a deep reinforcement learning (DRL) dispatcher that serves as a real-time decision engine for crew-to-repair assignments. We model restoration as a sequential, information-revealing process and learn an actor-critic policy over compact features such as component status, travel/repair times, crew availability, and marginal restoration value. A feasibility mask blocks unsafe or inoperable actions, such as power flow limits, switching rules, and crew-time constraints, before they are applied. To provide realistic runtime inputs without relying on heavy solvers, we use lightweight surrogates for wind and flood intensities, fragility-based failure, spatial clustering of damage, access impairments, and progressive ticket arrivals. In simulated hurricane and flood events, the learned policy updates crew decisions in real time as new field reports arrive. Because the runtime logic is lightweight, it improves online performance (energy-not-supplied, critical-load restoration time, and travel distance) compared with mixed-integer programs and standard heuristics. The proposed approach is tested on the IEEE 13- and 123-bus feeders with mixed hurricane/flood scenarios.
Paper Structure (21 sections, 15 equations, 2 figures, 4 tables, 2 algorithms)

This paper contains 21 sections, 15 equations, 2 figures, 4 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Event Context and Modeling Assumptions
  3. Hurricane Wind Field and Wind–Induced Damage
  4. Flood Depth Field and Flood–Induced Damage
  5. Correlated Multi–Hazard Sampling and Access Impairment
  6. Progressive Outage Discovery and Repair Times
  7. Stochastic Scenario Sampling for Dispatch
  8. DRL for Crew Dispatch After Events
  9. Dynamic MDP and Reward
  10. Policy, Value, and Memory
  11. Event-Driven Interaction
  12. Learning: PPO with Masked Actions
  13. Deployment
  14. Case Study
  15. Experimental Setup
  16. ...and 6 more sections

Figures (2)

  • Figure 1: Deployment dataflow. Surrogate wind and flood models generate runtime streams, and an event-driven loop builds the state $s_t$, masks infeasible actions, selects crew and switching assignments, and applies them with a low decision runtime.
  • Figure 2: (a) RL training losses over epochs and (b) average episodic reward over epochs.