A multi-model predictive framework for unsupervised anomaly detection in univariate time series data from hydraulic turbine units

The maintenance of stable operation in hydraulic turbines is essential for optimizing energy generation and minimizing unforeseen outages in hydropower facilities. Early detection of deviations enables prompt alerts and facilitates predictive maintenance, thereby improving system reliability. This study introduces a multi-model methodology employing unsupervised learning on univariate time-series data for anomaly detection, obviating the need for labeled datasets, which are often scarce in industrial settings. The findings demonstrate a compromise between detection accuracy and computational efficiency. The Isolation Forest ( i Forest) model yielded the best overall performance, with 99% accuracy, 97% precision, 100% recall, and a 98% F1-score, while requiring minimal training time (0.16 s) and prediction time (0.12 s), rendering it suitable for real-time surveillance. Neural network techniques such as the Long Short-Term Memory Autoencoder (LSTM-AE) and Autoencoder achieved comparable accuracy (90%–91%) but necessitated significantly greater computational resources, thus limiting their practical deployment. K-means exhibited perfect precision but low recall, whereas One-Class Support Vector Machine (OC-SVM) provided high recall but entailed very lengthy training and inference durations. Overall, the i Forest model is recommended as the most balanced and efficient solution for real-time anomaly detection in hydropower systems. • Assessed five unsupervised anomaly detectors using real operational hydropower data. • Isolation Forest reached 98% F1 and excelled in speed over the other tested models. • LSTM-AE and autoencoder models improved anomaly detection but used more compute. • K-means and OC-SVM showed precision–recall trade-offs impacted by sensitivity. • Future work uses multivariate data, improved interpretability, and edge tools.