부산대학교 도서관

Edge intelligence (EI) is an emerging interdiscipline to advance the coordination of artificial intelligence and edge computing. EI sinks the computation and decision-making process from centralized clouds to the edge node in proximity to terminal devices, which is robust to the unacceptable communication delay or disconnection. In this paper, we propose an EI-based framework for online scheduling of soft open points with energy storage (ES-SOPs), a novel power electronic device, to enhance both spatial and temporal flexibility in power distribution networks. The proposed framework empowers the edge computing via hybrid deep reinforcement learning (HDRL), which seamlessly combines advantages of both data-driven deep neural networks and physics-based ES-SOPs model. Inside the edge computing node, a deep neural network first learns a set of parameters from the historical data and ES-SOPs local status. Then, the outputs of the deep neural network are fed into a physics-based ES-SOPs model to construct its objective function, where rigorous operation constraints are included. Finally, this model is solved to obtain near-optimal ES-SOPs online scheduling. Case studies on a modified IEEE 33-node system demonstrate the effectiveness of the proposed framework under different levels of uncertainties and its superiority over safe DRL and model predictive control-based methods.