Flow characteristics prediction of the tri-eccentric butterfly valve based on the Multi-Scale Fusion Attention method and Kolmogorov-Arnold network

Abstract

The tri-eccentric butterfly valve is widely used in the petrochemical, energy, and other industries. The flow coefficient and hydrodynamic torque are key parameters of the tri-eccentric butterfly valve. Rapid and accurate prediction of these parameters can improve the control accuracy of fluid flow and ensure the efficient and stable operation of the valve. Deep learning techniques offer a promising approach for predicting flow characteristics. This paper proposed a novel network structure based on the Multi-Scale Fusion Attention module (MSFA) and Kolmogorov-Arnold networks (KAN) to predict the flow coefficient and hydrodynamic torque. The proposed MSFA module enhances the multi-scale perception ability and integrates both high-level and low-level features to improve information representation. The KAN network replaces the linear weights and activation functions of Multilayer Perceptron (MLP) with learnable B-spline basis functions, which enhances regression prediction performance. The results indicate that the proposed MSFA module effectively improves both the prediction accuracy and convergence of base models such as MLP and KAN. The MSFA-KAN model achieves a mean absolute percentage error (MAPE) of 2.61 %.