To mitigate and eliminate the common spatiotemporal illusions in short-term traffic flow prediction, a novel pre-training long-short spatiotemporal interleaved Transformer model was proposed, based on the Transformer network and a self-supervised pre-training to fully supervised training framework. Long-term spatiotemporal heterogeneity was acquired by the self-supervised pre-training, and a spatiotemporal interleaving module was designed to interact and obtain the long-term spatiotemporal heterogeneous interactivity. A short spatiotemporal recurrent Transformer was designed to compress and extract the short-term spatiotemporal sequences onto a spatial slice, which represented the unique spatiotemporal features of the entire short-term sequence. Guided by the long-term spatiotemporal interleaved heterogeneous interactivity, similar features were matched on the future timeline to reconstruct the future short-term spatiotemporal sequence. Different traffic flow prediction models were compared in terms of accuracy and multi-step predictions in four traffic flow benchmark datasets and two traffic speed datasets. Experimental results show that the proposed model improves the accuracy of traffic data prediction compared to current state-of-the-art models.