Adaptive hierarchical learning for uncertainty-aware distributed energy resource planning.

Area of Science:

• Electrical Engineering
• Power Systems
• Artificial Intelligence in Energy

Background:

• Increasing distributed energy resources (DERs) introduce complex, multi-source uncertainties into power distribution networks.
• Traditional planning methods struggle with implicit uncertainties from partial observability of security constraints by third-party operators.
• Existing approaches often rely on model simplification and predefined scenarios, limiting adaptability.

Purpose of the Study:

• To develop a data-driven framework for co-optimizing DER location, capacity, and operational strategies.
• To enable autonomous learning of implicit constraints without requiring full model knowledge.
• To address the challenges posed by multi-source uncertainties in distribution network planning.

Main Methods:

• An adaptive hierarchical learning architecture is proposed.
• A bi-level Stackelberg structure integrates Monte Carlo Tree Search (upper level) for planning scheme generation.
• Multi-agent reinforcement learning (lower level) learns operational policies from real-time data under partial observability.

Main Results:

• The framework demonstrates lower investment costs compared to traditional methods.
• Faster solution times were achieved while maintaining voltage stability.
• Superior scalability and adaptiveness to implicit uncertainties were validated on benchmark and large-scale systems.

Conclusions:

• The proposed adaptive hierarchical learning framework effectively manages uncertainties in distribution network planning.
• Data-driven optimization and learning from real-time data offer significant advantages over scenario-based methods.
• The approach provides a robust and efficient solution for integrating DERs.