<p>In this paper, we propose DPatchformer, a novel deep learning model for multivariate time series forecasting. To address the issue of semantic loss caused by fixed patch operations, we introduce a deformable patch mechanism that dynamically adjusts the subsequence partitions based on the characteristics of the data, enhancing the model’s ability to capture multi-scale features. Additionally, we propose two Intra-patch enhancement–Intra-patch Attention and Intra-patch Linearization–to better model the dependencies between adjacent time steps within each patch. A simple yet effective Feature-Mixing MLP is employed to capture the inter-variable interactions in multivariate datasets. Extensive experiments on benchmark datasets such as ETT and Weather demonstrate that DPatchformer outperforms most state-of-the-art models, particularly in capturing complex temporal patterns. The ablation study systematically validates the effectiveness of three core components: the deformable patch mechanism, intra-patch enhancement, and Feature-Mixing strategy. Compared to the transformer baseline, our architecture establishes new effectiveness benchmarks for multivariate time series forecasting.</p>