Projectile Trajectory Prediction Based on TCN-BiLSTM-Attention Model

Accurate projectile trajectory prediction is crucial for modern battlefield operations, yet existing methods face challenges in precision and noise robustness. To address these issues, this paper proposes a hybrid TBAtt (TCN-BiLSTM-Attention) model. The model first employs Temporal Convolutional Network (TCN) with causal and dilated convolutions to extract multi-scale temporal features, then integrates Bidirectional Long Short-Term Memory (BiLSTM) with symmetric features to capture bidirectional dependencies, and finally introduces an Attention mechanism to dynamically weight critical time steps. This combination enhances global feature extraction and noise resilience. In noise-free conditions, the model demonstrates significant performance advantages with trajectory point errors of 49.674 m, 54.619 m, and 0.419 m in the x, y, and z directions for 20-second predictions, with mean absolute errors reduced by 43.869 meters, 22.164 meters, and 2.757 meters compared to TCN-BiLSTM. Under noise conditions, the model maintains robustness, achieving mean absolute errors of 5.366 m, 76.117 m, and 4.009 m for 10-second predictions along the x-, y-, and z-axes respectively, with corresponding prediction accuracy improvements of 99.3%, 97.4%, and 95.6% in trajectory point estimation. The proposed approach significantly improves prediction accuracy and stability, offering a reliable solution for real-time battlefield applications.