Probabilistic power flow with topology changes based on deep neural network

Abstract

The uncertainty of power systems is rapidly increasing with the continuing development of renewable energy. Probabilistic power flow (PPF) is an effective tool for addressing these uncertainties. However, the high computational burden is a major bottleneck for the practical application of PPF. This paper proposes an efficient method for solving the PPF based on deep neural network (DNN). Stacked denoising auto-encoders (SDAE) is selected to extract the nonlinear features of the power flow model with discrete topology status. The following two aspects are investigated to improve the DNN performance: (1) construction of the feature vector that effectively characterizes the renewable energy, load, and topology and (2) knowledge transfer of DNN parameters to improve the training efficiency of the DNN for evolutionary scenarios. After training, the power flow solutions of all samples generated by Monte-Carlo simulation (MCS) can be directly projected through the DNN with high accuracy, rapid speed and low computational burden. Finally, the effectiveness of the proposed method is verified on the modified IEEE 39-bus and 118-bus systems.