Research on multi-stage optimal dispatch strategy of microgrids with multi-resource clusters based on hybrid data-driven method

In recent years, microgrids have become increasingly essential to advancing new type power systems. As compact power systems that coordinate the operation of various distributed resources, microgrids can monitor the status and manage the operation of contained resources effectively. However, the large-scale integration of distributed flexible resources has substantially increased the communication and computational burdens on regional dispatch center. Moreover, traditional model-based centralized algorithms struggle to balance the diverse interests of multiple stakeholders while maintaining computational efficiency. Additionally, directly applying raw prediction data to original data-driven algorithm can hinder both training efficiency and accuracy. To address these challenges, this paper proposes an optimized scheduling strategy for microgrids based on hybrid, multi-type data-driven methods. First, a multi-stage model is developed to minimize the total operating cost for microgrids with diverse resource clusters. Especially considering that electric vehicle (EV) possesses energy storage capabilities, which serve as typical distributed flexible resources, capable of making rapid power adjustments. A combined Monte Carlo simulation and physical mechanism approach is employed to determine the dispatchable capacity boundaries of EV cluster. Second, a Conditional Generative Adversarial Network (CGAN) is designed to generate high-quality scenario sets, mitigating the challenges of low accuracy and slow convergence when agents are trained directly on predictive data, thus reducing the risks inherent to purely RL-based methods. Third, a multi-agent deep reinforcement learning (MADRL) algorithm is introduced, where the coordinated scheduling task is modeled as a Markov game. An improved entropy-constrained algorithm allows agents to make real-time optimal decisions based on local observations after training. Finally, a case study on a modified system is conducted to validate the effectiveness of the proposed approach.