Speakers
Description
Multi-objective optimization for a hydraulic turbine blade is a significant challenge due to the high computational cost of performing the number of computational fluid dynamics (CFD) simulations. It becomes more challenging while dealing with a high number of parameters. This research addresses this issue by proposing an innovative approach that leverages Autoencoder techniques to reduce the dimensionality of the problem, specifically the latent space representation of the original parameter space. This latent space is then used as the domain of optimization, potentially offering a more efficient search landscape. The study compares the performance of optimization runs conducted in the latent space against traditional optimization in the original parameter space. For both scenarios, the genetic algorithm ”Non-dominated Sorting Algorithm II (NSGA-II)” with an island model is utilized, allowing parallel exploration of the search space and promoting diversity in the population. Experiments encompass a range of multi-objective optimization benchmarks such as evaluations in terms of computational time. Furthermore, the study analyzes the trade-offs between time invested in training an efficient Autoencoder and subsequent gains in optimization speed. This research contributes to the field by offering a methodology for accelerating multi-objective optimization in the field of hydraulic turbine design.
