AGC Command Allocation Strategy for Virtual Power Plant Based on Meta-Reinforcement Learning

The inherent strong stochasticity and heterogeneity of distributed energy resources impede the ability of Virtual Power Plants (VPPs) to dynamically respond to Automatic Generation Control (AGC) signals. To address this challenge, this study proposes a meta-reinforcement learning (Meta-RL) based strategy for AGC signal allocation within VPPs. This strategy integrates a meta-learning framework with the Proximal Policy Optimization (PPO) algorithm, which can overcome the poor environmental adaptability of traditional strategies and the limited generalization capability of basic reinforcement learning. This study firstly develops a multi-objective optimization model balancing economic benefits and operational performance. Subsequently, it formulates a Markov Decision Process (MDP) defining state variables, action spaces, and reward functions, and designs an AGC signal allocation optimization strategy based on this framework. Finally, it conducts strategy simulations and comparative experiments across diverse task scenarios. The results demonstrate that under multiple operating conditions, the proposed strategy exhibits superior performance over alternative strategies in key aspects such as generalization capability and response speed. Specifically, it achieves a $10.4 \%$ to $59.5 \%$ improvement in regulation performance metrics and increases the net daily revenue by $11.85 \%$ to $60.20 \%$, which effectively achieves a balance between profitability and strategic stability objectives.