TATBformer: A Divide-and-Conquer Approach to Ship Trajectory Prediction Modeling

Ship trajectory prediction, a cornerstone in time series forecasting, demands increasing precision amidst the complexities of modern maritime navigation. Traditional prediction methodologies have been constrained by the non-uniformity of AIS data and often overlook the intrinsic ship information. Addressing this oversight, our work integrates ship-specific data, proposing a divide-and-conquer model architecture. For trajectory data, we employ the Temporal Convolutional Network to extract implicit temporal insights from the vessel, which are subsequently channeled through a transformer for feature extraction. Simultaneously, ship-specific data is processed through a three-layer Multi-Layer Perceptron, recognizing its pivotal role as a reference for individual vessel predictions. This approach not only introduces a richer data set but also counters the potential computational surge induced by transformers. Additionally, we introduce an innovative composite loss function, infusing norm-based regularization to reduce dependencies on non-critical parameters. Empirical evaluations confirm that our approach surpasses classical time series forecasting techniques, delivering superior prediction accuracy with minimized errors.