A causal prior-guided spatiotemporal traffic flow prediction model was proposed in order to address the challenges of identifying the driving factors behind traffic flow variations and improving prediction performance. The spatiotemporal feature fusion module was used to discretize the spatial dimension and integrate it with periodic embeddings in order to enhance the representation of complex nonlinear feature. The latent causal discovery module was utilized to introduce a lag-aware causal convolution and a causal filter in order to capture potential causal relationship among time series. The causal validation module was used to identify true causal relationship under a hypothesis testing framework in order to calculate time delay between nodes with causal effect. The causally validated structure was embedded as prior knowledge into a dynamic graph convolution module in order to guide information propagation and feature aggregation. The experimental results demonstrated that the proposed model achieved reduction in MAE by 4.9197%, 3.9518% and 2.1968% on PeMSD4, PeMSD8 and Chengdu-DiDi datasets, respectively, compared with baseline models. The effectiveness of the causal mechanism in traffic flow prediction was verified.