Deep Reinforcement Learning Approach for Dynamic Distribution Network Reconfiguration Based on Sequential Masking

Area of Science:

• Electrical Engineering
• Artificial Intelligence
• Power Systems

Background:

• Dynamic distribution network reconfiguration (DDNR) is crucial for secure and economic power distribution network (PDN) operation, especially with high renewable energy source (RES) penetration.
• Data-driven solutions, particularly deep reinforcement learning (DRL), are gaining traction for DDNR due to enhanced data availability in PDNs.
• Existing DRL methods struggle with the vast and sparse action space of DDNR, often due to the radiality constraint, limiting scalability and optimality.

Purpose of the Study:

• To address the challenges of scalability and optimality in DRL-based DDNR.
• To propose a novel sequential masking strategy to decompose the complex action space of DDNR.
• To develop a data-efficient, safety-guaranteed, and scalable DRL solution for DDNR.

Main Methods:

• A sequential masking strategy is proposed to decompose the DDNR problem's complex action space into manageable sub-action spaces.
• A gated recurrent unit (GRU)-based agent is designed to process sequential data.
• An adapted soft actor critic (SAC) algorithm is employed to handle the decomposed action spaces.

Main Results:

• The proposed method demonstrates superior algorithmic performance and scalability compared to existing data-driven approaches.
• Case studies confirm the effectiveness of the sequential masking strategy in handling the radiality constraint.
• The GRU-based agent and adapted SAC algorithm provide a data-efficient and safety-guaranteed DRL solution.

Conclusions:

• The developed DRL approach, utilizing a sequential masking strategy, offers a significant advancement in solving the DDNR problem.
• This method overcomes the limitations of existing DRL techniques in terms of scalability and optimality for large-scale power distribution networks.
• The findings highlight the potential of this approach for enhancing the secure and economic operation of modern power grids with high RES integration.