Accurate prediction of highway traffic flow is crucial for traffic management and network planning. Traditional forecasting methods often struggle with data limitations and complex traffic conditions. The widespread use of Electronic Toll Collection (ETC) technology offers a new opportunity to obtain large-scale, high-precision traffic data for improved forecasting. This study presents a short-term highway traffic flow prediction method using a self-supervised spatiotemporal Transformer based on ETC data. This powerful neural network model captures spatiotemporal dependencies in traffic data. It comprises a spatiotemporal Transformer for prediction and a masked autoencoder to handle missing data, enhancing generalization. We collected and processed extensive ETC data, applying the model to generate accurate traffic flow predictions. Comparisons with baseline models demonstrate that our method significantly improves prediction accuracy and reliability in short-term traffic forecasting.