Considering that multivariate time series can reveal component information at different frequencies in the frequency domain, thereby capturing temporal and variable dependencies, we propose a Phase-Aware Attention-based Multiscale Multivariate Time Series Forecasting Model (PAAM). The model primarily consists of a temporal dependency learning module and a variable dependency learning module. In the temporal dependency learning module, we employ learnable frequency bands to partition the frequency domain representation of the time series into multiple scales and achieve information fusion through a multiscale aggregation module. In the variable dependency learning module, we model lead-lag relationships between variables by integrating a phase-aware attention mechanism and combine it with a global frequency domain relationship learning module to capture global variable dependencies. Finally, we achieve global feature fusion by integrating temporal and variable information. Experimental results demonstrate that PAAM outperforms the current state-of-the-art model, PatchTST, on most of the six datasets, with the mean squared error on the ETTh1 dataset decreases by up to 15.17% and an average reduction of 12.42%.