Deep reinforcement learning for real-time energy dispatch in smart grids with high renewable penetration
Аннотация
The increasing penetration of Renewable Energy (RE) in modern Smart Grids (SG) introduces substantial variability and uncertainty, posing critical challenges to real-time energy dispatch. Traditional optimization and rule-based methods, while effective under deterministic conditions, exhibit limited adaptability to stochastic RE generation and fluctuating demand. This study develops a Deep Reinforcement Learning (DRL) model for real-time dispatch in renewable-dominated SG, formulating the problem as a constrained Markov Decision Process (MDP). Actor-critic networks—Deep Deterministic Policy Gradient (DDPG), Proximal Policy Optimization (PPO), and Soft Actor-Critic (SAC)—learn adaptive policies that jointly minimize operational costs, enhance renewable integration, and maintain grid reliability. A modified IEEE 33-bus distribution system with high RE diffusion is simulated using historical solar and wind profiles, storage dynamics, and realistic demand patterns. A comparative analysis of rule-based heuristics, deterministic Mixed-Integer Linear Programming (MILP), and two-stage stochastic optimization proves that DRL achieves superior performance across multiple dimensions. SAC delivers the best results, reducing operational costs by 20%, achieving 92.8% renewable application, and minimizing loss-of-load probability to 0.5%, while maintaining real-time computational feasibility (0.41 s per dispatch interval). Constraint satisfaction validation confirms 99.8% voltage compliance and 100% thermal limit adherence. Scalability analysis of the IEEE 123-bus network reveals sub-quadratic training-time scaling and effective model transferability under parameter variations. Sensitivity analyses confirm robustness under varying prediction errors, dispatch granularities, and storage configurations. These results establish DRL as a scalable, reliable, and cost-efficient model for next-generation SG dispatch under RE uncertainty.