• Deep learning
• Fault detection
• Forecasting
• Hydropower plants
• Machine learning
• Signal processing

In the transition toward sustainable and decarbonized energy systems, hydropower plants (HPPs) stand out as reliable and flexible energy sources. Their ability to provide dispatchable power and balance the intermittency of other renewables like solar and wind makes them central to future energy strategies. However, modern hydropower systems are becoming increasingly complex due to the integration of advanced monitoring technologies, industrial IoT (IIoT), and the demands of Industry 4.0. These complexities introduce new challenges in terms of monitoring, fault detection, and predictive maintenance.

This thesis focuses on data-driven fault forecasting and anomaly detection in hydropower plants, using real operational datasets collected from two stations—one in Italy and another in Spain—over a multi-year period. The datasets consist of high-frequency readings (one-minute intervals) from over 1000 heterogeneous sensors, including temperature, pressure, and vibration data. The main objective is to explore how signal behavior evolves over time and how anomalies can be predicted using both classical and advanced machine learning methods.

Initial data preprocessing involved signal cleaning, handling frozen values, and filtering out outliers caused by sensor failures or maintenance cycles. Then, statistical techniques and seasonal decomposition were used to establish baseline models for identifying deviations in sensor patterns. However, given the strong nonlinear dependencies and the presence of complex correlations between signals, deep learning models were later employed to improve prediction accuracy and model robustness.

LSTM-based architectures were trained for multivariate time-series forecasting, while autoencoders were used for unsupervised anomaly detection. Additionally, 1D CNNs were explored for learning temporal dependencies and identifying local irregularities. A hybrid model combining these deep learning strategies was also proposed and benchmarked against classical models in terms of performance and computational efficiency.

The hybrid approach proposed in this work achieved the best overall performance, showing strong generalization across different fault types and seasonal variations. Special attention was given to the explainability and interpretability of the model results, addressing one of the key challenges in the deployment of AI-driven solutions in critical infrastructure.
The real-world nature of the datasets used in this study introduced several practical challenges, such as sensor drift, missing data, noise from operational fluctuations, and varying maintenance schedules. The thesis addressed these challenges by developing a robust preprocessing pipeline and by integrating domain knowledge from plant operators into the design of the monitoring framework.

Furthermore, the work emphasized the importance of dimensionality reduction for dealing with high-dimensional sensor spaces. Techniques like Principal Component Analysis (PCA) and t-SNE were applied to better visualize the data structure and to improve the computational feasibility of downstream models. This enabled not only better performance but also improved understanding of sensor relationships.

The methods were evaluated in terms of several metrics, including accuracy, recall, F1-score, computational efficiency, and early detection time. The results showed that hybrid models incorporating both deep learning and domain-informed preprocessing significantly outperformed traditional thresholding or rule-based techniques. Moreover, the proposed framework was validated on both training and test datasets to ensure generalization and real-world feasibility.

Finally, the thesis discusses potential directions for future work, including the integration of digital twin technologies for hydropower systems. By creating a synchronized virtual model of physical assets, digital twins can enable continuous monitoring, simulation of failure scenarios, and more efficient maintenance planning. The research paves the way for the implementation of scalable, interpretable, and intelligent monitoring systems in modern hydropower plants.