Ship Traffic Flow Simulation Based On Generative Adversarial Network

With the development of intelligent shipping,ship traffic flow simulation,which dynamically reappears the navigation process of many ships at sea,has become an important research field in Marine traffic engineering.It plays an important role in the feasibility demonstration of Marine engineering,such as port planning and construction and waterway reconstruction,and can also provide auxiliary decision-making of traffic flow management for maritime and waterway departments.The key technology of ship traffic flow simulation is to model the dynamic trajectories of multiple ships.However,the calculation time of most current simulation methods increases greatly with the increase of the number of ships,so it is difficult to simulate all ship trajectories at the same time.In this thesis,the generative adversarial network,which has developed rapidly in the field of generative deep learning in recent years,is used to study the ship traffic flow data in the third alert area of traffic separation in the Qiongzhou Strait,hoping to build a more realistic model of large-scale ship traffic flow that is in line with the actual navigation state.The main contents of this study are as follows:(1)Construct a synchronous multi-ship trajectory simulation model Vessel-GAN based on Social-GAN.Aiming at the problems such as obvious decrease of model operation efficiency,single loss function and lack of trajectory guidance information when the number of simulation steps increases,The Long Short-Term Memory(LSTM)Network of the trajectory coding module is replaced by the Temporal Convolutional Network(TCN),so as to reduce the number of parameters in the long time ship traffic flow simulation and improve the computing speed of the model.According to the characteristics of maritime navigation,the collision avoidance loss function is designed based on the idea of binary classification cross entropy,so that the generated trajectory is more in line with navigation norms.According to the conditional generation idea of C-GAN,the end point guidance information is input into the network to make the simulated collision avoidance behavior more real.The experimental results show that the improved Vessel-Gan model,compared with other models such as original Social-GAN and Social-LSTM,improves the calculation speed by 36%,reduces the average displacement error by 28% and the end displacement error by 41% in terms of accuracy.(2)Construct ship interaction data set in Qiongzhou Strait.Analyze and clean the original AIS data collected by the maritime department in the process of management,including segmenting the data according to different voyage,removing the vacancies and duplicate values,using the isolated forest algorithm to detect outliers,and adopting cubic spline interpolation to obtain evenly spaced AIS data.Finally,according to the ETH and UCY data set structure,the sliding window method is used to screen the preprocessed data with interactive behavior in the space,and the ship interactive data set is constructed in Qiongzhou Strait.The random rotation method is used to expand the data set to weaken the influence of the fixed course on the ship trajectory simulation,and the data set is visualized to visually display the scenes in different time periods.(3)Design simulation experiments based on the real AIS data of Qiongzhou Strait,simulate and visualize ship traffic flow in a long period of time with the model,compare the simulated ship traffic flow diagram with the real ship traffic flow diagram,and analyze the difference of simulation quality and timeliness of different models.