Generalizable deep sequence models for 4D trajectory prediction of tower crane loads

Predicting the trajectory of suspended loads is essential for proactive collision avoidance and improving situational awareness during tower crane operations. Existing approaches suffer from limited generalizability, simplified motion assumptions, and lack of real-time deployment. This paper presents a data-driven framework for 4D (3D + time) trajectory prediction using deep sequence models trained on realistic crane-operation data. A Unity-based simulation environment is developed to emulate rotary and linear encoders, and 29 participants operate the crane across randomized pick-and-place tasks, producing diverse motion trajectories. Six architectures — including LSTM-based Seq2Seq models with different attention mechanisms, ConvLSTM networks, and Temporal Convolutional Networks — under varying prediction horizons, temporal context, sampling rates, and sensor noise are evaluated. The Seq2Seq model with Temporal Attention achieves the best performance, with a mean 3D displacement error of 0.45 m on unseen logistic scenarios. A high-performing model is integrated into a real-time digital twin to provide feedback for operator training.