Bad data detection can provide reliable data dependence for microgrid operation decision-making. Due to the frequency of operation mode switching and difficulties of microgrid modeling analysis, traditional bad data detection method based on state evaluation has not been applied to microgrid. This paper utilizes extreme learning machine (ELM) to learn the historical data of microgrid for purpose of extracting the data feature; and detects bad data by DBSCAN clustering algorithm analyzing the feature. A bad data detection method based on ELM and DBSCAN is proposed. Taking advantage of the historical operation data of a four-terminal DC microgrid prototype, the simulation scenario is designed and result verifies the efficiency of this method. In addition, this paper contrasts it with several data mining algorithm, and it is indicated that ELM+DBSCAN are of high superiority on both algorithm performance and detection effects.