SwinKAN: A Dual-Polarization Radar Extrapolation Model Based on Swin Transformer and Convolutional Kolmogorov–Arnold Networks

This study presents SwinKAN, a multivariate prediction model based on the Swin Transformer and the convolutional Kolmogorov–Arnold networks (KANs). The model aims to enhance the prediction of severe convective weather using dual-polarization radar data. SwinKAN employs a multiinput, multioutput architecture. It simultaneously predicts multiple variables, including radar reflectivity Zh, differential reflectivity Zdr, and specific differential phase Kdp, fully leveraging the advantages of dual-polarization radar. To enhance interaction between radar variables, SwinKAN integrates them within the channel dimension. This integration improves the model’s ability to capture complex spatiotemporal features. This study also introduces KAN to the extrapolation of dual-polarization radar data, a novel contribution. Additionally, the Swin–Kolmogorov feature extraction module (SKFEM) is designed to efficiently extract and integrate both local and global features. This significantly improves the model’s ability to predict precipitation intensity and evolution. To further optimize performance, the cross-feature fusion module (CFFM) is designed. This module enhances prediction accuracy, especially in regions with heavy rainfall. Experimental results demonstrate that SwinKAN outperforms existing methods across multiple evaluation metrics. These results validate the model’s superiority and effectiveness in severe convective prediction.