Resumo:
There is significant hydrokinetic potential in various flowing water courses, such as rivers, canals, and plant outflows. Turbine design can involve either designing based on empirically established parameters or determining these parameters through optimization techniques. This study focuses on Bayesian optimization, specifically efficient global optimization. The main objective is to develop a more efficient global optimization methodology based on multifidelity surrogate modeling, applied to turbine design. A new multifidelity surrogate modeling method based on Hierarchical Kriging improved by Radial Basis Functions is developed. The proposed method involves improving the surrogate model based on high-fidelity data and correcting the surrogate model based on low-fidelity data. The method is applied to analytical test functions and to the preliminary design of a shrouded hydrokinetic turbine, aiming to maximize hydraulic power by optimizing the turbine's geometric parameters. The geometric parameters of the nose cone, nacelle, and diffuser angle of attack are defined as design variables. The turbine is modeled in CFD with two mesh refinement levels, fine and coarse, used as sources of high and low fidelity, respectively. The rotor is approximated by an actuator disk with a fixed pressure drop. The optimization process compares the proposed method with two methods from the literature. The results from the analytical test functions show that the proposed method performed better for the test case with the highest number of variables and the most multimodality. For the real engineering case, the developed method resulted in an average hydraulic power improvement greater than the other two methods for the first quarter of the total computational budget and showed lower standard deviation during most of the optimization time. The optimized turbine geometry resulted in an increase in the downstream channel of the rotor and an increase in the nose cone diameter, culminating in a 21% increase in the power coefficient.
