Development of a model to predict the skin friction coefficient with application to hydraulic turbines

Sammanfattning

This project aims to develop a model to predict the impact of surface roughness on frictional losses in hydraulic turbines using machine learning techniques. A neural network was developed and pre-trained using a synthetic dataset, where skin friction coefficients were estimated via three empirical correlations. These networks were then combined into an ensemble model. To enhance prediction accuracy, the ensemble model was fine-tuned using experimental data. The study evaluates different configurations for training, validation, and test splits to determine optimal performance. The final model demonstrates good agreement with experimental values, particularly when trained on a larger subset. The results show that transfer learning can effectively incorporate empirical knowledge and that the proposed methodology holds promise for efficient prediction of surface-induced flow losses in hydraulic systems.