In recent years, significant progress has been made in the field of time series forecasting, especially with the Transformer model becoming the mainstream model in time series research and being widely applied in various fields such as power, finance, and transportation. However, most previous models only capture sequence information at the same time scale, neglecting the hidden sequence information at different scales and spatial dimensions. In other words, they fail to effectively capture multi-scale and spatiotemporal information, leading to inaccurate forecasting results. Moreover, many models stabilize non-stationary sequences to improve sequence forecastability, which leads to the loss of non-stationary information and affects the distribution of the prediction results. To address these issues, this paper proposes a Multi-Scale Spatiotemporal Fusion model (MSFformer). The model segments the sequence into patches and employs a multi-branch structure to capture multi-scale information. Meanwhile, it adopts a spatiotemporal fusion attention mechanism to capture the spatiotemporal information of the sequence. In addition, we introduce a stationarity correction method, which restores and captures non-stationary information by constraining the differences in stationary distributions between the model’s input and output data. Experiments on a large number of real-world datasets demonstrate that this model has superior forecasting capabilities compared to most other models.