<p>Long-term time series forecasting plays a vital role in applications like financial market prediction, energy usage forecasting, and traffic flow analysis. Deep learning models, though widely used in this task, often struggle to capture long-term dependencies and periodic patterns due to the overlapping of variables with different periods. To this end, we propose a multi-scale temporal fusion model named MSTF, which extracts sequential features and periodic patterns across multiple scales using a Time Reverse and Transform block and a Dynamic Combination Reconstruction block. Unlike traditional Transformer-based models, MSTF emphasizes overall temporal continuity rather than individual time point values. Experiments on seven datasets against ten advanced models demonstrate MSTF’s superior performance, faster training time, and reduced model complexity, particularly excelling on high-dimensional datasets.</p>