A Spatiotemporal Feature-Based Model for Traffic Flow Prediction

The accuracy and timeliness of real-time highway traffic flow prediction are critical for intelligent, proactive management. Currently, research on traffic flow prediction based on ETC gantry data, particularly for continuous multi-segment flow prediction, remains limited. To address this gap, this paper presents a spatiotemporal traffic flow prediction model that integrates Long Short-Term Memory (LSTM) networks with an attention mechanism. First, a training dataset is constructed, and the model framework is designed. LSTM is employed to capture temporal features, while the multi-head attention mechanism is used to uncover the inherent relationships among traffic flow-related features and capture global spatiotemporal correlations within the road network. Additionally, the model is optimized via a feedback neural network, and a multi-gantry fusion training strategy is implemented to enhance the accuracy of multi-segment predictions. Experimental results demonstrate that, at both 5-minute and 30-minute intervals, the proposed model outperforms the comparison models, achieving an average absolute percentage error of 8.173% for multi-gantry fusion prediction. Validated through six months of engineering practice, the proposed model shows high prediction accuracy and reliability for real-time highway traffic flow forecasting. The results suggest that, compared to traditional traffic flow prediction methods based on video or radar data, the proposed model offers a significant advantage in terms of cost-effectiveness and reliability. Furthermore, the wide coverage of ETC gantries makes this model highly scalable and applicable to diverse real-world scenarios.