Multi-Objective Optimization for Efficiency Improvement and Aerodynamic Excitation Reduction of a Diagonal Flow Fan Based on Machine Learning

Abstract

Abstract. The aerodynamic excitation within the diagonal flow fan is a key source of its vibration and noise problems. To enhance efficiency and mitigate aerodynamic excitation of diagonal flow fan, this study proposes a multi-objective optimization method based on machine learning. Using Latin hypercube sampling (LHS) and validated numerical simulations, a dataset is constructed by parameterizing the angle distribution curves along hub and shroud surfaces of impeller blade and stator vane as variables. Backpropagation neural network (BPNN), whose hyperparameters are optimized via particle swarm optimization (PSO) is trained as surrogate models, and nondominated sorting genetic algorithm II (NSGA-II) is subsequently used for multi-objective optimization. Results demonstrate that the established surrogate models and optimization algorithm possess high prediction accuracy and reliability. Under the design conditions, the optimized model, compared to the original model, achieves a 3.18% improvement in efficiency and significantly attenuates the pressure pulsation amplitudes at low-frequency stage and blade passing frequency (fBPF) which are mainly attributed to the improved internal flow characteristics, as reflected in detailed flow field analysis.