Accurate long-term multivariate time series forecasting is crucial in finance, energy, and traffic management. Despite advances in Transformer-based models, capturing both local structures and long-range dependencies remains challenging. Enhancing multi-scale dependency representation improves forecasting accuracy and stability. This paper introduces MSCAN, which integrates Multi-Scale Dynamic Attention (MSDA) for adaptive dependency modeling and Context-Aware Convolutional Attention (CACA) for refined local feature extraction. By combining attention and convolution, MSCAN effectively models temporal dependencies across multiple resolutions. Experiments on seven benchmark datasets show that MSCAN consistently outperforms existing methods across diverse forecasting tasks.