Thrust Coefficient Prediction In Hydrokinetic Turbines Using Supervised and Unsupervised Machine Learning
Abstract
The aerospace industry seeks to enhance turbine efficiency, but traditional simulations are costly, resource-intensive, and require specialists. By creating an online, AI/ML model that can predict the performance of turbines, the traditional, high-cost, and resource-demanding methods of turbine design can be surpassed. To address this, hydrokinetic turbine data from the University of Manitoba was analyzed using supervised and unsupervised AI/ML models to predict the thrust coefficient across various parameters. Supervised learning, implemented through IBM Watson, utilized the Random Forest Regressor, Decision Tree Regressor, and Snap Decision Tree Regressor, achieving the lowest RMSE of 0.027. Key features like "Type"and "TSR"were identified as most influential when optimizing turbine thrust. In contrast, unsupervised learning, using K-Means clustering, yielded a low Silhouette Score of 0.2101, reflecting poor separability in the data. This comparison demonstrates the superior predictive power of supervised methods for turbine performance optimization, as unsupervised learning struggled to capture the nuanced relationships between features and outputs. This knowledge can be applied to create a model that utilizes supervised learning to predict the thrust coefficient when given the parameters. This will aid in optimizing the design of turbines at lower costs than are currently present.
