Asynchronous Source-Load-Storage Interactive Control in Power Grids Based on Edge Sensing Networks

Hanjun Ling, Aiman Zhang, Xiong Yin, Sheng Wang, Jiadong Guo, Shengyuan Qin, Xiandong Liang, Yiming Lu, Zipeng Yu, Yiwei Huang

Abstract

The intermittency and volatility of renewable energy often result in local wind and photovoltaic curtailment, posing risks to grid stability. Meanwhile, the largescale integration of demand-side resources is shifting power systems toward bidirectional interaction, increasing uncertainty on both the supply and demand sides. To address these challenges, this study develops an asynchronous source-load-storage interactive control framework based on edge sensing networks. The method introduces a graph reinforcement learning-based coordinated scheduling mechanism, in which the network topology and electrical states collected at the edge are encoded as graph data. Spatial features are extracted through graph convolution, and an actor-critic structure is used for dynamic policy optimization, improving adaptability to topology changes. A heterogeneous photovoltaic market-trading strategy is further proposed, forming a coordinated trading framework driven by electricity price signals and integrating revenues from photovoltaic generation, load aggregation, and storage interaction. By combining edge-side processing with multi-algorithm fusion, a distributed regulation architecture is established to support real-time response under asynchronous grid conditions. Results on the IEEE 33-bus system show that the proposed graph reinforcement learning model achieves a total installed capacity of 6398.1 kW, a 13.55% improvement over DDPG. The PV-storage, wind-storage, and integrated source-loadstorage capacities reach 2351.7 kW, 2368.4 kW, and 1956.2 kW, respectively, indicating a balanced allocation. Node voltage deviation is reduced to 0.011, representing a 21.43% decrease relative to DDPG. The cumulative discounted reward converges within approximately 160 episodes, with a runtime of 61.2 s, improving convergence speed by about 30% compared with MADRL and 50% compared with DDPG.

Source: semanticscholar

Read the AI summary and key takeaways for traders on WOBR Quant Research.