A multivariate time series prediction model based on the KAN network.

Time series forecasting is crucial in various fields such as financial markets and weather prediction. Although mainstream deep learning models like RNNs and CNNs have made some progress in capturing short-term patterns, they still fall short in handling long-range dependencies and complex interactions. The transformer model enhances modeling capabilities through self-attention mechanisms, but its complexity and sensitivity to noise limit its applications. To address these challenges, we propose the KANMTS model, which integrates the advantages of Kolmogorov-Arnold networks (KAN) and multi-layer perceptrons (MLP). KANMTS leverages the non-linear mapping capabilities of KAN to improve the capture of complex patterns and dependencies in multivariate time series while maintaining the computational simplicity of MLP. Experimental results show that KANMTS outperforms existing methods in terms of forecasting performance and resource utilization efficiency, with a more pronounced accuracy improvement on large-scale datasets. We also explore the interpretability of KANMTS using symbolic regression and visualization methods. Furthermore, KANMTS demonstrates good generalization and integration capabilities, making it suitable for various application scenarios. This study provides a simple, efficient, and more interpretable solution for multivariate time series forecasting, advancing the field of time series analysis.