Enhancing Traffic Flow Forecasting with Spatio-Temporal Information: A Case Study on Highway Networks in Belgium

As the number of vehicles continues to increase every year, it is crucial to monitor traffic conditions effectively. Traffic forecasting serves as a valuable method in traffic management to support decision making. This study investigates the impact of utilizing spatio-temporal data for traffic flow forecasting, as opposed to forecasting each location individually. We compare spatio-temporal Long Short-Term Memory (LSTM), a machine-learning technique that utilizes spatially correlated sensor data in a network, and compare this to conventional LSTM and Seasonal ARIMA (SARIMA). Experiments have been conducted on counting loop data of a highway network in Belgium. We have investigated the performances of the forecasting model for different prediction horizons. The results show that spatial information improves the performance of the forecasting model and reduces the training time required. Further analysis show that the spatio-temporal LSTM improves majorities of detectors while the minorities detectors are degraded which imply the importances of network selection. The computational load of the spatio-temporal LSTM model is minimally impacted by the number of sensors in the network thus making it scalable. The results show that longer prediction horizons degrade the forecasting performance but the impact up to a 12 hour-horizon is not significant.