ADRL-VPP: An Adaptive Deep Reinforcement Learning Framework for Load Regulation in Virtual Power Plants

Abstract

Virtual Power Plants (VPPs) combine renewable energy sources with storage units to provide an integrated power generation system that improves grid stability and energy efficiency. Traditional load-balancing solutions struggle to respond to fluctuations in generation, electricity costs, and demand, frequently resulting in power outages and grid instability. To solve these issues, this study proposes an adaptive deep reinforcement learning system, ADRL-VPP (Adaptive Deep Reinforcement Learning for Virtual Power Plant Load Management), which enhances load management while ensuring stable, efficient VPP operation. Experiments were carried out using the VPP_LoadReg_Dataset, which included 10,000 samples of solar, wind generation, storage, load, time, weather, and electricity price variations. ADRL-VPP was evaluated against conventional backstepping and fuzzy adaptive controllers as baselines. The model uses a deep Q-network (DQN) with ε-greedy exploration and adaptive learning rate optimisation to discover optimal actions, including increasing, decreasing, saving, or doing nothing. The experimental results demonstrate that ADRL-VPP outperformed the baseline controllers, achieving 92% accuracy, 90% precision, 91% recall, 90.5% F1-score, and a Grid Stability Improvement Index (GSII) of 0.87. Overall, ADRL-VPP provides a strong, intelligent solution for dynamic load regulation in VPPs, demonstrating the promise of deep reinforcement learning for sustainable, adaptable power management.