Search

This paper presents some metamodeling techniques to analyze the variability of the performances of an inductive power transfer (IPT) system, considering the sources of uncertainty (misalignment between the coils, the variation in air gap, and the rotation on the receiver). For IPT systems, one of the key issues is transmission efficiency, which is greatly influenced by many sources of uncertainty. So, it is meaningful to find a metamodeling technique to quickly evaluate the system's performances. According to the comparison of Support Vector Regression, Multigene Genetic Programming Algorithm, and sparse Polynomial Chaos Expansions (PCE), sparse PCE is recommended for the analysis due to the tradeoff between the computational time and the accuracy of themetamodel. [ABSTRACT FROM AUTHOR]

• ANNMLP provides the highest accuracy in predicting heat transfer coefficient. • SVR has the highest generalization capability to predict heat transfer coefficient. • XGBoost shows the highest accuracy and generalization capability for pressure drop. • Significant features for heat transfer coefficient: mass velocity and vapor quality. • Significant features for pressure drop: inclination angle and mass velocity. Machine learning techniques have great potential to predict two-phase flow characteristics instead of classic empirical correlations. In the present study, four different machine learning models, including multi-layer perceptron artificial neural network (ANNMLP), support vector regression (SVR), K nearest neighbors (KNN), and extreme gradient boosting (XGBoost), are employed to predict the heat transfer coefficient (HTC) and the frictional pressure drop (FPD) of R134a and R1234yf condensation flow in horizontal and inclined tubes. The dataset includes 348 points from previous works and the current research. To extend the data, an experimental study is also performed on the condensation of R134a in a horizontal tube for different mass velocities and vapor qualities. The results show that, in the best model, HTC can be estimated by ANNMLP with the mean absolute percentage error (MAPE) of 7.01%. The best prediction of FPD is achieved using XGBoost machine with MAPE of 10.87% on test data. Also, the feature importance procedure is implemented to recognize the most useful features. Based on the results, the mass velocity and the inclination angle are identified as the most influencing parameters on the prediction of HTC and FPD, respectively. [ABSTRACT FROM AUTHOR]