Graph oriented topology identification and state monitoring in distribution system

Abstract

In the past decades, the power system is integrated with increasing renewable energy resources (RES) to combat climate change and mitigate the energy crisis. The uncertainty and intermittency of RESs cause significant impacts on the distribution system operation. This change in the distribution system composition spurs the requirement of timely monitoring of the distribution system states. Different from the transmission system with redundant monitoring, the distribution system lacks sufficient measurements. Another distinguished difference is the frequent changing topology. Thus, there remains a gap in the distribution system topology identification and state estimation. In addition, the gap also exists in probabilistic power flow for the system planning. Therefore, this thesis focuses on addressing developing progressive state monitoring approaches based on graph-oriented artificial intelligence technologies. To timely identify the distribution grid topology, a power distribution grid topological generative adversarial network (Gridtopo-GAN) model is proposed for the topology identification considering the challenging situations of limited measurements and meshed structure. Specifically, an innovative topology preserved node embedding architecture is introduced to represent and compress the numinous topologies such that the topology identification of large-scale systems with varying topologies can be handled. The bad measurement data and missing data inspire the application of the GAN with the generative capability to render the robustness to the proposed topology identification model. Numerical simulations represent the effectiveness and time saving of the proposed model. To timely track the states of distribution systems with high penetration of RES, an unrolled spatiotemporal graph convolutional network (USGCN) is developed for distribution system state estimation and forecasting that is exposed to complex correlations among the renewable power outputs. Specifically, the proposed unrolled spatiotemporal graph model can capture the spatiotemporal correlations simultaneously to obtain ameliorated forecasting accuracy. Then, the node-embedding is proposed to represent the hidden spatiotemporal correlations so that automatically learning the correlations and distribution system parameters can be achieved. Furthermore, the multiple stacking spatiotemporal convolutional layers can achieve the ahead-of-time system states. The simulation results verify the accuracy and efficiency of the proposed model. To represent the uncertain distribution system states quantificationally, a graph-aware deep learning network (GADLN) is proposed. To fully capture the mapping from the fluctuated power injections and the uncertain system states, the convolutional operation is introduced to aggregate the correlations among renewable power outputs to facilitate the PPF. In this way, improved effectiveness and speed-up calculation can be achieved in the proposed model. Moreover, the numerical results show the superior of the GADLN over the state-of-art with accurate and effective manners. To calculate the probabilistic power flow (PPF) with complex correlations, a graph attention enabled convolutional network (GAECN) is proposed to approximate PPF. Specifically, the graph attention enabled convolutional layer is proposed to aggregate the correlations of the power injections during the training process. Within this layer, the full self-adaptive graph convolutional operation is proposed to capture and learn any implicit correlation automatically so that significantly enhanced accuracy can be achieved. The improved accuracy and efficiency achieved by the proposed model are indicated by the simulation results.