Forecasting plays a crucial role in time series analysis and is broadly applied across various domains. It is worth noting that the existing deep learning-based methods for multivariate time series forecasting often suffer from limitations in modeling strategies, leading to insufficient mining of intra-series features within individual variables or inter-series correlations among different variables. To address the above limitations, a Multi-scale Dual-path Transformer Network (MDTNet) is proposed to facilitate multivariate time series forecasting. For each univariate time series, the multi-scale dilated time attention module is designed, which captures the multi-scale characteristics and intra-series features individually by using dilated time attention mechanism and sliding window technology. For the whole multivariate time series, the multivariate attention module is provided, which explores the potential relations among different variables and seizes the inter-series correlations. Intuitively, the designed dual-path structure of MDTNet effectively captures both intra-series and inter-series features by using different attention mechanism. Experimental results of eight datasets from five benchmark application domains demonstrate the proposed method outperforms the existing approaches.