Multivariate time series forecasting has broad applications in fields such as traffic planning and weather forecasting. In recent years, Transformer-based models have become a research hotspot, particularly those that segment time series into patches and treat them as tokens for learning, effectively capturing complex temporal patterns and enhancing Transformer efficiency. However, existing patch-based models face limitations in the fusion of multi-scale features. Most current models fail to flexibly adjust the importance of features across different scales, which prevents the model from fully exploring the potential relationships between different time scales and limits its performance in modeling complex periodic patterns. To address this issue, we propose the Multi-Scale Patch Fusion Transformer (MSPFT), a model designed to efficiently Multi-Scale Pyramidal Patch Embedding module. Unlike traditional multi-scale models that independently extract features at each layer, our proposed pyramid-style multi-scale encoding module employs a cross-scale connection mechanism driven by learnable weights, where the temporal representation from the previous scale is injected as supplementary information into the next scale. This allows fine-grained scale layers to incorporate long-range dependency information, enhancing the model’s ability to capture long-term trends. Additionally, we introduce a multi-scale fusion mechanism that enables the model to dynamically adjust the weights based on each scale’s contribution to the prediction, thereby improving the recognition of both short-term and long-term temporal patterns. Experiments on eight real-world datasets demonstrate that MSPFT outperforms most of the current state-of-the-art methods.