Cost-Effective automatic generation control for Renewable-dominated grids: Multi-Agent deep reinforcement learning approach

As power systems continue to increase in complexity, real-time frequency control has become progressively more critical for balancing power supply and demand by modulating generating unit outputs. This paper proposes a data-driven Automatic Generation Control (AGC) strategy that dynamically adjusts both the participation factors and governor droops of generators. Using Multi-Agent Deep Reinforcement Learning (MADRL) within a centralized training framework, agents are trained to determine optimal participation factors and governor droop settings in the AGC loop. Unlike traditional AGC methods that rely on fixed values, our approach directs each unit to minimize the overall system’s operational cost using local information. Training is performed a day in advance using forecasted data for load and renewable energy generation. The effectiveness of this cost-efficient AGC approach via the MADRL framework is evaluated on the modified New-England 39-bus system in a Software-in-the-Loop (SIL) environment, accounting for real-time load and photovoltaic (PV) generation fluctuations. Results demonstrate the proposed method’s efficacy in reducing operational costs and its suitability for real-time implementation.