Abstract: Existing forecasting methods often struggle to adaptively capture and fuse the intricate dynamic patterns within multivariate time-series data, especially when faced with multi-scale variations, non-stationarity, and high noise levels. To tackle this limitation, we propose a wavelet decomposition and multi-scale patch mixing (WDMPM) model for multivariate time-series forecasting. The method first explicitly decomposes each input sequence into multi-resolution components via the discrete wavelet transform, separating low-frequency trends from high-frequency details. A parallel multi-branch architecture is then adopted to process each component independently. Each branch successively adopts a pre-processing convolution module for local feature enhancement and noise suppression, an attention-weighted multi-scale patch embedding module to adaptively fuse information across different temporal scales, and a dual-layer mixer module to efficiently integrate cross-time and cross-variable dependencies. Additionally, a patch-wise structure-aware loss is introduced to preserve the statistical and correlational consistency of the predicted series, improving the model’s robustness and structural fidelity. Finally, extensive experiments on six public time series datasets demonstrate that the proposed WDMPM achieves the most advanced accuracy on three datasets, obtains the second best performance on one dataset, and exhibits substantially higher computational efficiency than the mainstream baseline methods. These results validate the effectiveness of the proposed explicit multi-scale modeling framework and its ability to balance forecasting accuracy with operational efficiency in complex real-world forecasting scenarios.