TimesKAN: A Generic Backbone Network for Time Series Prediction Based on Kolmogorov-Arnold Networks

The field of time series analysis has significant applications in various domains, including physiological signals, weather prediction, and power systems. However, the intricate data patterns inherent in time series often lead to suboptimal performance of conventional models. To address this challenge, this paper proposes a novel multivariate time series prediction model called TimesKAN based on the Kolmogorov-Arnold network, drawing inspiration from the Kolmogorov-Arnold representation theorem. This model consists of two key components: (i) KAN layers constructed using B-splines as learnable activation function feature extractors instead of commonly used convolution and transformer structures in traditional time series models. B-splines are smooth piecewise polynomial functions capable of approximating complex shapes and patterns effectively, enabling the network to capture subtle changes in the data without being affected by abrupt changes or discontinuities; (ii) We introduce embedding multivariate channel independence for KAN layers where each channel contains a single univariate time series that utilizes independent KAN weights for each KAN input while sharing independent channel weights across all embeddings. This extension allows KAN layers to effectively handle complex data models. Experimental results demonstrate that TimesKAN outperforms many state-of-the-art frameworks for long-term predictions on 85% of datasets and short-term predictions on all datasets. The feature extraction method based on the Kolmogorov-Arnold representation theorem significantly enhances both long-term and short-term prediction accuracy. Furthermore, TimesKAN exhibits excellent performance when dealing with periodicity and trend within time series data while also adapting well to unstable changes within such sequences. The Kolmogorov-Arnold representation theorem provides a promising alternative, presenting opportunities for further enhancements to contemporary deep learning models that heavily rely on convolutional and transformer architectures.