LSTM Based Prediction and Time-Temperature Varying Rate Fusion for Hydropower Plant Anomaly Detection: A Case Study

Abstract

Data-driven based predictive maintenance is vital in hydropower plant management, since early detections on the emerging problem can save invaluable time and cost. The overheating of bearings of turbines and generators is one of the major problems for the continuous operations of hydropower plants. A reliable forecast of bearing temperature helps designers in preparing future bearings and setting up the operating range of bearing temperatures. In this study, the fusion algorithm between Long Short Term Memory (LSTM) neural networks based effective slide-window regression model with time-temperature varying rate based anomaly detection framework is developed for detecting component and temporal anomalies of 56 MW Francis Pumped Storage Hydropower (PSH) plant in predictable and noisy domains. Data sets of all sensors were collected for a period of ten year ranging from 2007 to 2017 used for the train and test dataset. The predicted upper guide bearing temperature values were compared with the actual bearing temperature values in order to verify the performance of the model. The data analysis results shows anomaly is validated on the PSH plant.