Post-Disaster Generation Dispatching for Enhanced Resilience: A Multi-Agent Deep Deterministic Policy Gradient Learning Approach

This paper proposes a reinforcement learning-based approach for dispatching distributed generators (DGs) to enhance operational resilience of electric distribution systems after a severe outage event. The increased computational complexities and sophisticated modeling procedure of resilience-based e...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:North American Power Symposium (Online) S. 1 - 6
Hauptverfasser: Abdelmalak, Michael, Hosseinpour, Hadis, Hotchkiss, Eliza, Ben-Idris, Mohammed
Format: Tagungsbericht
Sprache:Englisch
Veröffentlicht: IEEE 09.10.2022
Schlagworte:
Computational modeling
Dispatching
Distribution system
extreme weather events
Generators
Power supplies
Real-time systems
Reinforcement learning
resilience
Training
ISSN:2833-003X
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:This paper proposes a reinforcement learning-based approach for dispatching distributed generators (DGs) to enhance operational resilience of electric distribution systems after a severe outage event. The increased computational complexities and sophisticated modeling procedure of resilience-based enhancement strategies have pushed toward adopting intelligent-based algorithms, specifically for real-time control applications. In this paper, a multi-agent deep deterministic policy gradient learning algorithm is developed to dispatch distributed generators after an extreme event. The proposed approach aims to provide a fast-acting control algorithm for improved resilient operation of islanded distribution power systems. The problem is formulated as an iterative Markov decision process that consists of a system state, action space, and reward function. Each agent is responsible for dispatching a single DG and is trained to increase its cumulative reward value. A system state represents the system topology and characteristics whereas an action refers to DG power supply. A reward is computed based on the power balance mismatch value for each agent. Different failure scenarios are generated and used to train the proposed model. The proposed method is demonstrated on the IEEE 33-node distribution feeder system in the islanded mode. The results show the capability of the proposed algorithm to dispatch DGs for resilience enhancement.
ISSN:2833-003X
DOI:10.1109/NAPS56150.2022.10012133