Hypergraph Kolmogorov–Arnold Networks for station level meteorological forecasting

Abstract

Weather forecasting plays a vital role in various domains, including disaster prevention, resource management, and energy optimization. Meteorological observation data are crucial for accurate weather forecasting, as they provide the spatial and temporal information needed to capture complex weather patterns. Graph Neural Networks (GNNs), due to their ability to handle non-Euclidean spatial relationships, can model the spatial dependencies among weather stations and have achieved good results in various weather prediction tasks. However, traditional graph neural networks, while effective at modeling spatial relationships between weather stations, often fall short in capturing the complex, multi-dimensional dependencies across different scales. This paper introduces a novel multi-information spatio-temporal hypergraph learning framework to overcome these limitations. By integrating neighborhood and semantic hypergraph convolutional networks, the framework effectively aggregates information from both spatially adjacent and semantically similar weather data, enabling it to capture intricate spatial and temporal features. Additionally, the Kolmogorov–Arnold Network (KAN) is introduced to enhance the model’s ability to learn dynamic, high-dimensional feature representations through the use of learnable univariate functions instead of fixed linear weights. Experiments on benchmark weather datasets show that the proposed method surpasses traditional spatio-temporal graph neural networks, providing higher accuracy and robustness in predicting complex meteorological phenomena.