Traffic flow forecasting plays a vital role in intelligent transportation systems (ITS), supporting efficient route planning and traffic management. However, many existing methods rely predominantly on short-term historical data, which limits their ability to capture long-range temporal dependencies and overlooks crucial long-term periodic patterns. They also often fall short in effectively decoupling spatiotemporal features and modeling global spatial correlations. To address these challenges, we propose a Spatio-Temporal Masked Hourglass Transformer Network (STMHTNet), an encoder–decoder framework designed for accurate long-term traffic flow forecasting. STMHTNet integrates three complementary trend extractors: a long-term trend extractor with an hourglass transformer preceded by a masking operation, a periodic trend extractor, and a short-term trend extractor. This design enables hierarchical spatiotemporal representation learning and enhances forecasting robustness. Extensive experiments on four real-world traffic datasets demonstrate that STMHTNet consistently surpasses several state-of-the-art baselines across multiple evaluation metrics.