Bad data must be detected in the microgrid because they mislead the decision making of energy management systems (EMSs). The authors propose a robust detection approach that combines an improved robust extreme learning machine (R-ELM) and density-based spatial clustering algorithm with noise (DBSCAN). To resist the impact of outliers in training data, R-ELM applies robust estimation and orthogonal transformation to the ELM training process. After training, R-ELM is used to construct an error-filtering map to extract the characteristics of microgrid measurements. These characteristics are analyzed by DBSCAN to identify bad data. The detection performance of this proposed approach is verified by historical data from a four-terminal ring-shaped DC microgrid prototype. Compared with the back-propagation neural network and ELM, R-ELM is validated to have good robustness. DBSCAN is also verified to outperform traditional K-means clustering. Overall, the approach described here maintains its robustness against outliers and achieves fast and effective detection of bad data in the microgrid.