To address controlled flight into terrain (CFIT) accidents, the Automatic Ground Collision Avoidance System (Auto-GCAS) was developed and implemented. This paper focuses on the trajectory prediction module within the Auto-GCAS. Aiming to overcome the limitations of traditional trajectory prediction algorithms, such as insufficient prediction accuracy and inadequate capability to capture features of nonlinear complex motion trajectories, a trajectory prediction method based on CPO-optimized CNN-LSTM-Attention is proposed. The Crested Porcupine Optimization (CPO) algorithm is employed to optimize the parameters of the CNN-LSTM-Attention network, primarily to determine the optimal learning rate and the optimal number of hidden nodes. Both qualitative and quantitative experimental approaches were adopted, using function simulations to model the complex motion trajectories of fighter aircraft. Comparative experiments were conducted to evaluate the prediction performance of four models: LSTM, CNN-LSTM, CNN-LSTM-Attention, and CPO-CNN-LSTM-Attention. The experimental results demonstrate that the CPO-CLA model exhibits significant advantages in model convergence speed and trajectory prediction accuracy, with notable improvements over the traditional LSTM model in terms of RMSE, MAE, and MAPE metrics.