A spatiotemporal network with multi-task learning for highway traffic flow forecasting

Multi-step highway traffic flow forecasting remains challenging due to complex spatiotemporal dependencies and heterogeneous node characteristics in highway monitoring networks. This study develops an MT-STNet-based forecasting framework for unified traffic flow prediction at entrance toll stations, exit toll stations, and gantries. Specifically, a directed graph is constructed to represent the highway network, where different node types are embedded into a unified framework to characterize traffic propagation patterns. On this basis, MT-STNet jointly incorporates historical traffic flow, temporal information, node identity, and road network structural attributes to capture temporal evolution, dynamic spatial correlations, and static topological dependencies. The model performs direct multi-step prediction for heterogeneous monitoring nodes within a single learning framework. Experiments are conducted on real-world highway data collected from Yunnan Province, China, from July 1 to August 31, 2024. Comparative results against the selected benchmark models, including SARIMA, SVR, and LSTM_BILSTM, show that MT-STNet consistently achieves the best performance among these baselines across all node types and the overall evaluation. In the overall pooled evaluation, MT-STNet yields an MAE of 2.839, an RMSE of 4.270, and a MAPE of 38.60%, outperforming the strongest baseline by 15.10%, 14.53%, and 18.55%, respectively. The proposed model provides an effective solution for short-term traffic flow forecasting in heterogeneous highway monitoring networks.