STF-DKANMixer: Tri-component decomposition with KAN-MLP hybrid architecture for time series forecasting

by Junxiang Wei, Rongzuo Guo, Yuning Wang

Long-term time series forecasting is critical for domains such as traffic and energy systems, yet contemporary models often fail to capture complex multiscale patterns and nonlinear dynamics, resulting in significant inaccuracies during periods of abrupt change. To overcome these limitations, we introduce STF-DKANMixer, a novel hybrid architecture combining a Multi-Layer Perceptron (MLP) with the expressive power of the Kolmogorov–Arnold Network (KAN). Our framework begins with a DFT-based decomposition strategy: long-term trends and seasonal components are extracted directly via Discrete Fourier Transform (DFT), while the remaining residual is further decomposed into high-frequency details using a Haar wavelet transform with error compensation. In the Past-Information-Mixing (PIM) stage, each component is processed by a GELU-activated KAN module for superior nonlinear feature mapping before being fused by a novel deformable feature attention (DFA) block, which adaptively learns sampling offsets and weights to capture complex dependencies. Subsequently, the Future-Information-Mixing (FIM) stage leverages an adaptive weighted ensemble of multiple lightweight predictors, enhanced by residual connections, to generate the final forecast. Extensive experiments on benchmark datasets validate the superiority of our approach. STF-DKANMixer significantly outperforms state-of-the-art models, reducing Mean Squared Error (MSE) by up to 36.1% (12.3% on average) and Mean Absolute Error (MAE) by up to 28.8% (8.8% on average). Impressively, these results are achieved while using less than half the computational resources of comparable methods. Our findings establish STF-DKANMixer as a robust, efficient, and highly accurate solution, setting a new performance standard for complex, long-horizon forecasting tasks.