Optimization of one-dimensional characteristic prediction algorithm based on neural network model

Based on the large cascade data set of CFD, a fully connected neural network deviation angle prediction model suitable for a variety of blade profiles was constructed. The original empirical model in HARIKA algorithm was replaced. A large data set covering 58300 data sets of NACA65, double arc and multi-arc main blade profiles was constructed by Latin hypercube sampling. Benchmarked against eight machine learning regression models including fully connected networks and support vector machines in learning deviation angle prediction based on the NACA65 airfoil dataset, the fully connected model achieves a mean deviation angle prediction error of 0.06°, outperforming all other regression models. In the field of one-dimensional property prediction, the model combination trained with different blade profiles had higher prediction accuracy than one model trained with multiple blade profiles. Under transonic conditions, compared with experimental data of the AI222-25 engine’s two-stage fan, the optimized HARIKA algorithm demonstrated an average reduction of 9.06% in relative error for pressure ratio prediction, with a maximum reduction of 20.43%. This confirmed the effectiveness of the optimization method in enhancing the performance prediction capability of the HARIKA algorithm.