Abstract
To address the limitations of medium- and long-term four-dimensional (4D)trajectory prediction models, this paper proposes a hybridCNN-LSTM-attention-adaboost neural network model incorporating a multi-strategyimproved snake-herd optimization (SO) algorithm. The model applies the Adaboostalgorithm to divide multiple weak learners, and each submodel utilizes CNN toextract spatial features, LSTM to capture temporal features, and attentionmechanism to capture global features comprehensively. The strong learner model,combined with multiple sub-models, then optimizes the hyperparameters of theprediction model through the natural selection behavior pattern simulated bySO. In this study, based on the real ADS-B data from Xi'an to Tianjin, thecomparison experiments and ablation studies of multiple optimizers are carriedout, and a comprehensive test and evaluation analysis is carried out. Theresults show that SO-CLA-adaboost outperforms traditional optimizers such asparticle swarm, whale, and gray wolf in handling large-scale high-dimensionaltrajectory data. In addition, introducing the full-strategy collaborativeimprovement SO algorithm improves the model's prediction accuracy by 39.89%.