Predictive Maintenance for Hydraulic Turbine Unit: A Comparative Deep Learning Approach Using Internet of Things Data in Real-Time

Accurate prediction of the remaining useful life (RUL) of hydraulic turbines is crucial for optimizing maintenance, reducing downtime, and enhancing operational efficiency. This study evaluates advanced neural network approaches, including convolutional neural networks (CNNs), long short-term memory neural networks (LSTM NNs), and a hybrid CNNs-LSTM NNs framework to estimate RUL using a high-dimensional time-series dataset. Rigorous preprocessing techniques, such as square root and Yeo-Johnson transformations, were applied to normalize the data and improve model performance. The models were assessed with metrics like mean absolute error (MAE), mean squared error (MSE), root mean squared error (RMSE), and the coefficient of determination (R2 score). Experimental results highlight that the hybrid CNNs-LSTM NNs model outperforms standalone CNNs and LSTM NNs by effectively combining spatial and temporal data patterns and achieving an R2 score exceeding 70% despite dataset imbalances. A detailed benchmarking analysis offers a deeper understanding of the strengths and limitations of each architecture, guiding the selection of models tailored to specific operational needs. This research advances predictive maintenance frameworks in the renewable energy sector, bringing attention to the potential of hybrid deep learning approaches for precise RUL estimation in industrial applications.