A novel deep learning based probabilistic power flow method for Multi-Microgrids distribution system with incomplete network information

Abstract

With the massive deployment of microgrids (MGs) and energy communities, various stakeholders have been involved in distribution networks. Due to the underdeveloped information infrastructure, especially in rural distribution networks, there is an increasing number of “blind areas” in the operation of distribution networks. The calculation of probabilistic power flow (PPF) with incomplete parameters have become an urgent issue to be solved for ensuring safe operation. Based on the deep learning and mechanism models, a novel PPF method is proposed for multi-microgrids distribution systems considering incomplete network information. Firstly, accessible power exchange data as well as public and independent information are utilized to realize equivalent modeling for microgrids area with incomplete parameters, based on a novel Kriging surrogate enhanced Gate Recurrent Unit-Temporal Convolutional Network (GRU-TCN). Then, the PPF calculation is effectively conducted by the distribution system operator (DSO) through the point estimation method (PEM), in which the equivalent GRU-TCN models and model-based power flow are integrated. Therefore, the complicated interactive iteration of the power flow equation is avoided, and the PPF calculation efficiency is effectively improved. In addition, user privacy is protected because only the trained GRU-TCN deep learning models will be used by DSO for the PPF calculation. The proposed method is validated in a modified IEEE 33-node distribution network, a modified American PG&E 69-node distribution network as well as the modified three-phase unbalanced IEEE 123-node distribution network including several MGs with unknown internal network parameters. The results show that the proposed method can improve the PPF calculation efficiency greatly while ensuring high-precision calculation results. The required evaluation time can be reduced by 64.01% and 99.31% compared with the DNN-based Monte Carlo sampling method and the traditional mechanism model-based Monte Carlo sampling method with complete information, respectively.