Multimodal machine learning based equivalent modeling method for active distribution networks

With high integration of distributed generators (DGs) and diverse topology reconfigurations, the dynamic states of active distribution networks (ADNs) become more complex, which poses great challenges to the dynamic equivalence of ADNs. In this paper, motivated by the similarities between Newton Raphson power flow calculation (NRPFC) and computation of convolution network (CNN), and based on the fact that recurrent neural network (RNN) can represent the differential algebraic equations (DAEs) of loads, we propose a multimodal machine learning based equivalent modeling method in order to track the dynamic behaviors of ADNs. First, the multimodal machine learning is divided into two branches, one of which is gated recurrent unit (GRU) +fully connected (FC) module to represent the DAEs for load modeling, and the other is CNN+ attention module to simulate the NRPFC to extract spatial feature changes of power system caused by disturbances or topology reconfigurations. Then, the robustness and generalization abilities of the proposed method are evaluated by different test systems, where the IEEE 14-bus transmission network is connected to distribution networks with different scales (i.e., IEEE 33-bus distribution network and IEEE 57-bus distribution network). The simulation results reveal that the proposed method can accurately capture the dynamic behaviors of ADNs under different operation conditions. In addition, the precision and generalization of the proposed method is tested with comparisons with other deep learning (DL)-based equivalent methods.