Abnormal Detection Using Two-Stage Method in Power Plants

This research article presents a novel two-stage methodology for anomaly detection in complex systems, specifically targeting combined power plants. The first stage emphasizes feature engineering, which includes the use of Mahalanobis distance to eliminate redundant sensor effects, Kmeans clustering to group similar data points, and Principal Component Analysis (PCA) for dimensionality reduction. By transforming the data into a more suitable format, the methodology prepares it for the subsequent detection phase. The second stage employs an LSTM-Autoencoder, a specialized neural network architecture, to identify abnormalities based on the reconstruction error from normal operational data. The proposed approach was validated using data from an operational combined power plant, demonstrating superior performance in accuracy and computational efficiency compared to existing techniques. The findings indicate that this methodology not only enhances anomaly detection in power plants but also has potential applications in other complex systems, addressing the challenges posed by imbalanced data and the scarcity of labeled abnormal instances.