Since the high penetration of distributed energy sources complicates the dynamics of electrical power systems

accurate dynamic models are indispensable for study on the transient behavior of the microgrid (MG). In some practices

the lack of full detailed information results in failure of differential equation based dynamic modeling

which leads to a demand for a black-box MG modeling method. It is a critical challenge to maintain the effectiveness of the black-box model under a wide operating range and various fault conditions. In this paper

inspired by the mathematical equivalence between the recurrent neural network (RNN) and differential-algebraic equations (DAEs)

a dynamic equivalent modeling method

using long short-term memory (LSTM)

is presented to tackle this challenge. At first

the modeling equivalence and advantages of our basic idea are explained. Then

modeling procedures

including data preparation and design guidelines

are presented. Finally

the proposed method is applied to a multi-microgrid testing system for performance evaluation. The results

under various scenarios

reveal that the proposed modeling method has an adequate capability for representing the dynamic behaviors of a black-box MG under grid fault and operating point changing conditions.