Accelerating the learning process of deep reinforcement learning algorithms in distribution network reconfiguration...

Area of Science:

• Electrical Engineering
• Artificial Intelligence
• Optimization Techniques

Background:

• Distribution network reconfiguration (DNR) is crucial for minimizing power losses in electrical grids.
• Traditional DNR methods often rely on accurate network models, which can be complex and computationally intensive.
• Existing approaches may struggle with large action spaces and inter-dependencies within the network.

Purpose of the Study:

• To propose a model-free deep reinforcement learning (DRL) approach for distribution network reconfiguration (DNR).
• To develop an efficient method for managing the large action space inherent in DNR problems.
• To improve computational efficiency and performance in minimizing power loss and voltage deviation.

Main Methods:

• Implemented a loop-based strategy to manage the action space effectively.
• Utilized a modified Q-learning algorithm to address inter-loop coupling effects.
• Employed an innovative replay method to accelerate convergence speed.
• Tested the approach on standard IEEE 33-, 69-, and 119-bus distribution networks.

Main Results:

• The proposed DRL approach demonstrated significant superiority over traditional metaheuristic and mathematical techniques.
• Achieved substantial reductions in distribution network power loss.
• Improved voltage deviation across the tested networks.
• Showcased considerably faster computational times compared to existing methods.

Conclusions:

• The model-free DRL approach offers a powerful and efficient alternative for distribution network reconfiguration.
• The loop-based method and modified Q-learning effectively handle complex network dynamics.
• This approach presents a promising direction for optimizing power distribution systems.