Fault reconfiguration control strategy of islanded marine ranching power supply system based on deep reinforcement learning

•A fault reconstruction method is proposed for island-mode marine ranching power systems.•The method uses an improved Noisy-Dueling DQN for fast, autonomous power topology adjustment.•It accounts for both clean energy and complex marine environmental influences.•Noise networks improve exploration an...

Full description

Saved in:
Bibliographic Details
Published in:International journal of electrical power & energy systems Vol. 169; p. 110796
Main Authors: Wang, Yichun, Zhang, Bo, Wang, Rongjie, Desong, Jiang, Cui, Yabo, Sun, Zhouyang, Liu, Hao
Format: Journal Article
Language:English
Published: Elsevier Ltd 01.08.2025
Elsevier
Subjects:
Deep reinforcement learning
Fault recovery and reconstruction
Island AC power supply system
Markov modeling
Markov modeling
Island AC power supply system
Deep reinforcement learning
Fault recovery and reconstruction
ISSN:0142-0615
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:•A fault reconstruction method is proposed for island-mode marine ranching power systems.•The method uses an improved Noisy-Dueling DQN for fast, autonomous power topology adjustment.•It accounts for both clean energy and complex marine environmental influences.•Noise networks improve exploration and generalization in real-time reconstruction control.•Transfer learning accelerates model training by reusing knowledge from different operation states. When the islanded marine ranching power system fails, a rapid and accurate control strategy is crucial to maintain continuous power supply to critical loads and maximize the restoration of lost load power. To ensure the safe operation of the marine ranching power system, a fault reconstruction control method based on an improved Noisy-Dueling Deep Q-Network (Noisy-Dueling DQN) algorithm is proposed. Firstly, a reinforcement learning environment for reconstructing the marine ranching power system, including distributed generation (DG), is constructed. This enables the agent to continuously update model parameters through interaction with the environment in order to learn the optimal reconstruction strategy. To enhance convergence efficiency and generalization capability of the algorithm, noise networks are used to improve exploration ability of dueling DQN algorithm by implementing linear layers with learnable noise using an adaptive noise model. Finally, through analysis of fault reconstruction cases in different operating conditions of the marine ranching power system, it is demonstrated that this proposed algorithm can provide optimal reconstruction strategies within as short as 70 ms while achieving objectives such as maximum load recovery, minimum switching times, and minimum network losses. To illustrate, in the event of a fault occurring on lines b7, b18, and b25 at 8:00, resulting in the withdrawal of 58 kW of loads from operation, the proposed algorithm is configured to restore all the lost loads and minimise the network loss to 0.0112 kW.
ISSN:0142-0615
DOI:10.1016/j.ijepes.2025.110796