LSTM networks for vessel traffic flow prediction in inland waterway

Vessel traffic flow reflects the congestion and security of traffic in waterway, accurate prediction can help ensure the waterway safe and smooth. A novel deep learning model based on long short-term memory networks (LSTMs) was proposed for wide predictions from aspects of short-term, long-term and influence of water level factor. To improve the performance, traffic flow sequences were pretreated to remove trend and seasonality as stationary ones, and a time-window method was applied to provide multi-step lag observations as features to increase temporal correlation. Results show that vessel traffic flow is applicable for wide prediction terms with high precision, and demonstrate that the proposed method has strong memorizing ability to describe traffic flow features. Furthermore, LSTMs achieved better performance compared with some well-known prediction models, such as GRU, DBNs, LR, RBF-SVR and ARIMA, which illustrates LSTMs has excellent generalization.