Highway traffic state prediction method based on spatial-temporal correlation and physical information constraints

Highway traffic state prediction is needed for intelligent traffic control and decision-making. However, current approaches face challenges in modeling spatiotemporal features, addressing data sparsity, and maintaining adherence to physical laws. This study introduces a novel traffic state prediction model that combines physical information with a hybrid Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) network approach (PI-CNN-LSTM) to achieve precise road traffic state forecasting. Initially, a hybrid CNN-LSTM deep-learning architecture was developed to capture spatial-temporal features. The CNN was used to extract the spatial topology correlation features of the road network, and the LSTM was combined to model the time series evolution law of traffic parameters, and the temporal and spatial dependence of traffic states was captured synchronously. Subsequently, to ensure the predictions align with traffic flow conservation laws, a physical regularization loss function was formulated using the Greenberg fundamental diagram and Lighthill Whitham Richards (LWR) conservation equation. The fundamental diagram parameters, including free-flow velocity and blockage density, were determined by the neural network within the loss function. Finally, the model's generalization capability under sparse data was enhanced by training on subsets of the Canadian Whitemud Drive dataset at sampling rates of 10%, 20%, and 30%. The experimental results showed that on working days, when the data sampling rate was 10%, the Mean Square Error (MSE) of PI-CNN-LSTM was 13.24, the Mean Absolute Error (MAE) was 7.21, and the Mean Absolute Percentage Error (MAPE) was 6.98%. When the sampling rate was increased to 30%, the MSE was 12.05, which decreased by only 1.19 compared to the sampling rate of 10%. This indicates that the physical constraints significantly enhanced the model's ability to adapt to the data sparsity. In the highly volatile traffic scenario on non-working days, PI-CNN-LSTM maintained a stable performance. The MSE was 10.75 at a 10% sampling rate and dropped to 9.01 at a 30% sampling rate. Among all the evaluation indicators, compared with the CNN-LSTM and PIDL models, the error was significantly reduced, which proves the validity and superiority of the proposed model.