XLinear: Long-Range Time Series Forecasting with MLP

Time series forecasting models are essential tools in various sectors. Among these, MLP (Multi-Layer Perceptron)-based models have demonstrated greater resistance to noise compared to Transformer-based models. However, traditional MLPs struggle to capture complex features, limiting their effectiveness in analyzing long-term dependencies.

To address this challenge, XLinear has been proposed, an MLP-based forecasting model specifically designed for long-range forecasting. XLinear decomposes the time series into trend and seasonal components. For the trend component, which contains long-range characteristics, an Enhanced Frequency Attention (EFA) mechanism is used to capture long-term dependencies through operations in the frequency domain.

Furthermore, for the seasonal component, a CrossFilter block is proposed to preserve the model's robustness to noise, avoiding the low robustness problems often associated with attention mechanisms. Experimental results demonstrate that XLinear achieves state-of-the-art performance on test datasets, while maintaining the lightweight architecture and high robustness typical of MLP-based models.