<p>Long-term time series forecasting is of critical importance in diverse applications, including energy scheduling, traffic planning, and weather modeling. However, real-world time series often exhibit pronounced multi-scale dynamics, non-stationarity, and overlapping periodic structures, which pose substantial challenges for existing methods in jointly modeling long-range periodic dependencies, short-term local variations, and frequency-domain characteristics. To address these challenges, we propose WPFormer, a unified forecasting framework that integrates representations from the time, frequency, and period domains. At its core, WPFormer introduces a wavelet-periodic fusion module, which performs multi-scale decomposition via discrete wavelet transform while explicitly modeling intra- and inter-period dependencies through periodic segmentation. Furthermore, a multi-band filtering encoder is developed to distill critical spectral features by reweighting frequency sub-bands, coupled with a dual-path attention mechanism that simultaneously captures structured periodic interactions and global cross-variable dependencies. Extensive experiments on multiple long-term forecasting benchmarks demonstrate that WPFormer consistently outperforms state-of-the-art models, achieving superior accuracy and robustness across diverse scenarios.</p>