CNN-Integrated MA-DRL Framework for Cost Effective Dynamic Frequency Support Through Intelligent Voltage Regulation in Grid-Forming Inverters

The domination of inverter-based resources has reduced effective inertia in autonomous microgrids (AMGs), leading to steeper rates of change of frequency and deeper frequency nadirs, which causes synchronization and stability issues. AI-assisted controllers offer a promising solution to synthesize inertia and fast frequency response, overcoming limitations of traditional methods. This article proposes an intelligent dynamic voltage regulation (IDVR) scheme tailored for grid-forming inverters (GFMIs) in an AMG with meshed network topologies (AMGMTs). By leveraging conservation voltage reduction (CVR) principles, IDVR dynamically regulates voltage to emulate inertia and provide dynamic frequency support. Therefore, IDVR offers a cost-effective solution that does not require energy storage, and resolves key limitations of CVR, including failure in reactive power control (e.g., inaccurate Q-sharing). To construct a fully decentralized control structure in a complex AMGMT architecture, a multiagent deep reinforcement learning (MA-DRL) framework is employed, which obviates the limitations on GFMIs’ synchronization arising from spatial variations in nodal voltage. Further, a novel 2D convolutional neural network architecture is proposed to address the computational complexity and cost efficiency of training and operating MA-DRL system in dynamic time scales. Simulation results in MATLAB/Simulink demonstrate the effectiveness of the proposed IDVR.