<p>Multivariate Time Series (MTS) forecasting is pivotal across various industries, particularly in intelligent transportation and energy management. Existing state-of-the-art methods typically rely on heterogeneous architectures that combine Graph Neural Networks (GNNs) for spatial dependencies with separate temporal networks (e.g., RNNs, TCNs, or Transformers). However, harmonizing these disparate components often leads to structural incompatibility and necessitates complex fusion layers, limiting model efficiency and interpretability. To address these challenges, we propose DGNet(Dynamic Graph Networks). The core innovation of this model lies in creatively extending graph convolution operations from the spatial domain to the temporal domain, constructing a unified spatiotemporal modeling paradigm. This approach enables the parallel and consistent capture of spatiotemporal features without introducing additional fusion layers. To support this, we design a Graph Learning Layer (GLL) that dynamically generates task-specific spatial and temporal graph structures without predefined priors. These structures are processed by Dynamic Graph Convolution (DGC) layers to achieve spatiotemporal consistency. Extensive experiments on four real-world datasets demonstrate that DGNet significantly outperforms existing baselines in prediction accuracy while maintaining computational efficiency.</p>