Time series forecasting is extensively applied across diverse real-world domains, such as finance, power systems, and meteorology, yet its accuracy is often constrained by the inherent complexity of temporal data. While exogenous variables can enhance predictive performance, their inclusion introduces additional challenges by increasing the intricacy of the time series structure. Moreover, traditional Multi-Layer Perceptrons (MLPs) frequently fail to effectively capture hierarchical feature representations across multiple scales, further complicating time series modeling. To address these limitations, we propose TKANMiXer, a novel Kolmogorov-Arnold Network(KAN)-based framework designed for variable fusion and multi-scale decomposition. Our framework comprises two key components: (1) incorporating a seasonal-trend decomposition module on the basis of variable fusion, (2) the multi-layer KAN architecture hierarchically captures. Specifically, TKANMiXer addresses the complexity arising from exogenous variables by disentangling temporal patterns, multi-layer KAN architecture hierarchically captures multi-scale temporal dependencies, significantly improving forecasting accuracy. Extensive experiments demonstrate that TKANMiXer achieves state-of-the-art performance in long-term forecasting tasks across multiple benchmark datasets.
Kong et al. (Tue,) studied this question.