Multivariate Time Series Forecasting (MTSF) predicts multiple variables by analyzing their interdependencies, with applications in weather and energy forecasting. Existing models often overlook local patterns and variable-specific temporal correlations. To address this, we propose PAMCNet, a hybrid network combining patch-dependent Attention and Mamba channel modeling. The main structure of the model includes the Patch Clustering Attention Block (PCA-Block) and the Mamba Channel Prediction Block (MCP-Block). Specifically, PCA-Block mines local correlation patterns by partitioning patches and clustering methods, and uses the patch Attention layer to captures the correlation between local variables under each patch cluster. MCP-Block designs a bidirectional Mamba structure to capture the global variables correlation, in which two Mamba blocks are connected by denoting loss. The module also uses distribution loss to align the predicted variable distribution with the future data distribution. Experiments on benchmark datasets show PAMCNet improves prediction accuracy while maintaining efficiency. Visualization confirms its effectiveness in capturing local variable correlations.