Abstract: Spatiotemporal forecasting of road traffic flow is a key task in intelligent transportation systems. Existing methods often rely on deep spatiotemporal attention mechanisms or graph convolutional networks to capture inter-road dependencies, but these designs typically incur high model complexity and computational cost. From empirical observations, we find that traffic flows within the same city exhibit highly repetitive patterns. Accordingly, we hypothesize that large-scale traffic sequences can be characterized by a finite set of prototype patterns. To validate this hypothesis, we propose PASTNet, a prototype-pattern-driven lightweight model for traffic flow forecasting. The core innovations of PASTNet are: (1) extracting prototype patterns from historical data and embedding them as priors to effectively guide the prediction process; and (2) employing a “star” operation to construct a compact network that still yields high-dimensional representations, thereby substantially reducing computation while maintaining strong performance. We conduct extensive empirical evaluations on large-scale real-world traffic datasets. Experimental results demonstrate that, compared with representative state-of-the-art traffic forecasting models, PASTNet achieves superior predictive accuracy and computational efficiency.