Forecasting time series with multiple variables is a significant technique that utilizes past data from multiple variables to anticipate future results, serving as a vital tool in diverse fields including traffic flow estimation, financial market projection, and load demand prediction. In contrast to forecasting of univariate time series, the process of multivariate time series forecasting requires consideration of interdependencies among the variables. Despite the notable progress achieved by transformer-based models in this domain, there persist challenges in accurately representing the relationships between variables and capturing a wide range of features across various scales. This paper proposes a new model named PatchMGT (Patch-based Multiscale Graph-enhanced Transformer) to tackle these issues. PatchMGT segments the time series with multiple variables into patches and Graph Neural Networks (GNN) is employed to capture variable dependencies within each patch. Moreover, the adjacency matrix learning module is given to eliminate the need for prior knowledge. A transformer architecture utilizing grouped multihead self-attention is utilized to efficiently capture features across multiple scales. Additionally, a temporal channel attention mechanism is devised to assign suitable weights to different feature scales, thereby augmenting the model’s ability to represent information. Drawing upon comprehensive experiments carried out on empirical data, the efficacy of PatchMGT is validated, demonstrating superior performance when compared with existing state-of-the-art model.