Deep reinforcement learning framework for adaptive power control in grid-forming inverters: A multi-objective optimization approach

Abstract

A novel deep reinforcement learning system is introduced, revolutionizing grid-forming inverter control through an attention-based neural architecture with adaptive policy optimization. The system uniquely integrates real-time stability constraints with multi-objective learning, addressing the fundamental challenges of power system control under uncertain conditions. The approach employs a comprehensive state-space representation incorporating grid dynamics and historical information, complemented by an advanced attention mechanism that enables selective feature prioritization across varying operational conditions. The learning architecture combines a hierarchical policy network structure with a prioritized experience replay mechanism, achieving rapid adaptation and stable control performance. The result validation demonstrates improvements over conventional methods, including a 43.75% reduction in harmonic distortion (from 3.2% to 1.8%), a 46.7% faster dynamic response (8 vs 15 ms), and a 50% extension in stable operation range under weak grid conditions (operational down to short circuit ratio, SCR=1.5). The system maintains 96% inference accuracy while executing within 50 μ s, meeting real-time control requirements. Additionally, the system demonstrates superior power decoupling performance, reducing coupling effects by 80% compared to traditional approaches while maintaining stable operation across diverse grid conditions. Learning-based control systems in power electronics demonstrate strong generalization across various operating conditions while ensuring stability. Integrating deep learning with power system constraints opens up new applications for complex real-time control problems that require adaptability and reliability.

Topics & Concepts