Recently, patch-based modeling has gained popularity for efficiently capturing local structures in multivariate time series forecasting, though it can suffer from boundary information loss, locality bias, and limited modeling of global dependencies across temporal and spatial. In this paper, we propose SA-HPAFNet, a Structure-Aware Hierarchical Patch Attention Forecasting Network, which integrates multi-scale patch encoding, structure-aware retrieval, and hierarchical attention mechanisms for robust forecasting. First, each patch is encoded using Inception-based convolutions to extract multi-scale structural features and preserve boundary continuity. Based on the encoded most recent patch and a preliminary forecast, a dynamic query is constructed to retrieve the top-k most similar historical patterns. These patterns are then fused through gated attention to augment the current patch representation. The enhanced sequence is passed through a hierarchical module combining temporal convolutions and inter-patch attention to model long-term dependencies and cross-variable spatial dependencies. Finally, a conditional attention mechanism dynamically aligns relevant historical information with the forecast horizon to refine forecasting results. Experiments on eight real-world datasets demonstrate that SA-HPAFNet consistently outperforms state-of-the-art baselines in both accuracy and robustness, making it a general and effective solution for multivariate time series forecasting.