| dc.description.abstract | Most hydroelectric power plants experience few faults during their lifetime. However, with the expected increase in volatile energy sources in the energy mix and the use of hydroelectric power plants as balancing mechanisms, the wear and tear on system components may rise. This is where fault detection and diagnosis can significantly improve maintenance regimes. A mechanism within the plant that signals when components are under distress can be both cost-efficient and increase the plant's reliability.
The hydraulic system that regulates the flow of water through the turbine is one of those systems that can be affected if the operation of the plants deviates from established norms. This thesis explores the feasibility of creating a model based on features from the Supervisory Control and Data Acquisition (SCADA) system.
The algorithm selected for this task is neural networks, specifically, a recurrent neural network (RNN) and long short-term memory (LSTM). The two models are used to predict the guide vane position with the use of input features such as oil level, accumulator temperature, and oil pressure. The RNN model proved to be the most accurate of the two, exhibiting low error rates and high R^2 score. The LSTM model struggled with accurate prediction, showing poor model fit metrics, even with the introduction of L2-regularization.
Additionally, an investigation was undertaken into the RNN model's performance on synthetic data with anomalous values, and it revealed a significant decrease in accuracy. | |
