QRCNN-BiGRU-Multi-attention for Prediction of Ship Traffic Flow in Compound Channels

In order to address the complex challenges in waterway traffic management in compound channels, a new model called Quantile Regression Convolutional Neural Network with Bidirectional Gated Recurrent Unit and Multi-attention (QRCNN-BiGRU-Multi-attention) is proposed for the prediction of ship traffic flow in compound channels. Quantile regression is utilized to estimate specific values at various quantiles and offer predictions across the complete range of quantiles. This method does not require assuming the parameter form of the distribution function in advance, thereby enabling interval prediction capabilities. Compared with other CNN-BiGRU models, this model has a simpler structure by serializing CNN and BiGRU, eliminating unnecessary network structures. Through a comprehensive analysis of the influencing factors in ship traffic flow for the compound channel, the main factors with significant impacts have been identified, such as the tidal height, wave height and hour time. Fast rejection test and crossing test are performed to judge whether a trajectory collected from automatic identification system (AIS) crosses a gate line of compound channels. The multivariate regression interval prediction is tested on a dataset of multivariate regression with 3 features. A 95% confidence interval is used. Results show that the proposed algorithm achieved the higher prediction accuracy than the traditional neural networks, such as BiGRU-attention, BiGRU, QRCNN-BiGRU, QRGRU and QRBiGRU..