A Method of Ship Trajectory Prediction Based on a C-Informer Model

The navigation of ships in complex environments is influenced by various uncertain factors, such as wind, waves, water depth, and ship performance, etc. It is challenging to precisely define and reflect the dynamic patterns of ship trajectories simply using mathematical models. To address this issue, a multi-step prediction method for ship trajectories based on feature engineering and neural networks is studied. The task of trajectory prediction is divided into two parts: data processing and model prediction. The data processing module preprocesses AIS trajecto-ry data using feature engineering methods. It starts by cleaning the raw AIS data, then uses the maximal information coefficient to select features highly correlated with the position prediction task. Additionally, a variable time interval information is introduced to address the problem of existing models only being able to select fixed time interval data for training and prediction. This module ultimately reconstructs high-quality ship trajectory sequences. The model prediction module constructs a ship trajectory prediction model based on C-Informer. It utilizes the multi-head Prob-Sparse self-attention mechanism of the Informer model to reduce the time complexity of the network model. Simul-taneously, it enhances prediction speed by generative decoding. By introducing a causal convolution module, the sensitivity of the model to neighboring time trajectory features is increased to compensate for the deficiencies of the Informer model in extracting local information. This adaption makes the model more suitable for ship trajectory prediction tasks. The experimental results based on Automatic Identification System (AIS) data near Nanjing port show that the C-Informer model for trajectory prediction has an overall mean square error (MSE) of 1.72×10－7 and a mean absolute error (MAE) of 2.43×10－4. Compared to the original Informer model, this represents a reduction of 28.6% and 31.9%, respectively. When training the C-Informer model with the selected feature combinations, the MSE and MAE are decreased by 57.7% and 42.1%, respectively, compared to using only latitude and longitude fea-tures. In predicting trajectories at different time steps, the C-Informer model reduces prediction time by up to 69.6% compared to the long short-term memory network model, with a maximum loss reduction of 75.8%.