Phase 1: GRNN Training

  • Trained on 90% of DoE simulation data
  • Normalised parameters (0 to 1)
  • Chose spread value for balance of error vs smoothness
  • Trained Generalized Regression Neural Network
  • Output: pressure increase

Phase 2: Genetic Algorithm

  • Fitness function defined as difference from target pressure
  • Implemented modified NSGA-II in MATLAB
  • Searched for blade parameters that matched required head
  • Final solution verified via full CFD Setup simulation