A multi-scale temporal-frequency fusion network based on MLP for long-term time series forecasting

Long-term time series forecasting has been widely used in extensive applications, such as weather forecasting and electricity consumption management. However, the complex patterns of time series in real-world contexts pose challenges to prediction accuracy, particularly as prediction horizons lengthen. Recently Transformer-based models have achieved excellent performance in long-term time series forecasting, but the self-attention mechanism for computation will cause high complexity when computing global correlations. In contrast, MLP-based models present robustness of prediction performance with the increase of input lengths but perform less well in capturing fine-grained features. To solve the problems above, we propose a novel multi-scale temporal-frequency fusion network based on MLPs, coined as MTF-Net, which uses patching methods to enhance the capacity of MLPs in capturing local features with different scales, and introduces frequency-domain MLPs to strengthen the capability of capturing the global seasonal patterns. By fusing predictions, the framework can complete the modeling of multi-scale temporal patterns in long-term time series. Extensive experiments results on six real-world datasets demonstrate significant performance improvements across multiple benchmarks compared with other state-of-the-art methods.